ChinaTalk

Where is China’s Groq?

Aqib Zakaria — Fri, 23 Oct 2026 10:56:32 GMT

In late 2025, HBM was all the rage. You need HBM for AI inference; inference decode is a memory-constrained operation, and only HBM was able to reach bandwidth speeds to serve models effectively.

But the market has matured since then, and there is no longer one architecture to rule them all. Different silicon works better for training, prefill, or decode…

Take Ten-Thousand: Egocentric Data Collection for Humanoid Robots

A. Chen — Thu, 22 Oct 2026 10:02:03 GMT

The ChinaTalk team is going to CoRL (the Conference on Robot Learning) in Austin! Reply to this email if you’d like to hang out between November 9th and 12th. We’re also doing karaoke on the 11th — RSVP here if you’d like to join!

A. Chen writes about people, technologies, and their relationships to power. Her submission, “Take Ten-Thousand: Egocentric Data Collection for Humanoid Robots,” won one of 7 Grand Prizes at the ChinaTalk Essay Contest. We’re thrilled to share this report from the frontlines of China’s robot data collection industry!

Where is the center of the universe? For the residents of Beijing, the answer is Wudaokou. The neighborhood houses China’s top universities, research institutions, and labs, drawing talent and funding from every corner of the country. The streets are radiant at night, packed full of bars and clubs that have somehow bounced back after the COVID-19 pandemic. And the number 13 metro line, which rattles overhead every five minutes, is never empty.

I came to Wudaokou to understand China’s AI industry — particularly, the hype surrounding humanoid robots, and the methods used to train them. Everyone has seen videos of these robots: eerily tall and lanky, some of them shedding nuts and bolts as they sprint down a track, others backflipping repeatedly on a dance stage. The global market for humanoid robotics, or embodied intelligence, is said to be worth $38 billion by 2035: the latest chapter in the U.S.-China technology race that has dominated international headlines for over a decade.

This year, China’s 15^th Five-Year Plan declared embodied intelligence a matter of national priority, then laid a strategy for developing training grounds across the country. Over 150 humanoid robotics companies and 40 state-owned robot data collection centers were reported to be in operation in China by the end of 2025. According to the Global Times, Chinese embodied intelligence startups drew 50 billion yuan in investment deals by September 2025: an increase of 2.5 times from the previous year.

More recently, embodied intelligence startups have begun to focus on the human hand, whose articulation and coordination has yet to be replicated by robots. Data of the human hand can’t be generated passively, in the same way that an Instagram user’s scroll or like might be harvested, and researchers of embodied intelligence are racing to solve this dilemma with a number of techniques. Most recently, videos have begun to circulate of people wearing headsets and filming themselves at work — a technique called egocentric data collection — to train robot hands.

I sat down with the researchers and entrepreneurs behind China’s egocentric data collection space to understand the industry. Beneath the speculation, the social media virality, and the staggering amounts of investor money, what — and who — does it actually take to train a robot?

Tsinghua Science Park in Wudaokou, Beijing. David Baron via Flickr.

The Third Wave of Data Collection

Dr. Jackson Guo, a researcher at the Beijing-based Kunlun Nest Robot Open Laboratory (昆仑巢人机器人开放实验室), invited me to his workspace to discuss the history and technicalities of egocentric data collection. He identifies three distinct periods of robot data collection: teleoperated, isomorphic, and egocentric.

Teleoperated data collection, the first generation of data collection techniques, appeared around 2017. That year, Stanford Professor Fei-Fei Li piloted a program that allowed users to connect their phones to remote controls that could control robot arms. Teleoperated data collection methods have since become more streamlined: a user stands several feet away from a robot and moves its body using remote controls, sometimes with a VR headset to simulate a different environment. According to Dr. Guo, the individuals responsible for teleoperated data collection often suffered from fatigue and joint pain after hours of performing slow, repetitive motions. Teleoperation was also expensive; hardware could cost hundreds of thousands of yuan per set. Even so, companies such as Helix, OpenXE, and PI (Physical Intelligence) collected tens of thousands of hours of data using this method. Most of the data collected was for gripper hands, which operated on a simple binary command: 0 for closed, 1 for open. Fine-grained, dexterous manipulations — such as cutting an apple, folding a shirt, or playing the piano — proved too complex for a user to simulate through a remote control.

In 2024, Stanford researchers Cheng Chi and Zhenjia Xu and Professor Shuran Song published a thesis on Universal Manipulation Interface (UMI). This marked the second generation of data collection techniques, or isomorphic data collection, which took Silicon Valley by storm. Using Velcro straps, elastics, or metal hinges, human subjects connected their hands to the inside or underside of a robotic exoskeleton, feeding information to a robot more swiftly and precisely than a teleoperated remote control could. Gripper hands were still the easiest model to train, as exoskeletons were often too rigid or heavy to allow for more articulated movements. Dr. Guo credits Generalist AI with the most successful isomorphic data collection program in the U.S.: within six months, they used one-thousand workers to collect over 250,000 hours of data.

By the start of 2026, egocentric data collection had begun to define a third generation of data collection techniques. On February 18, 2026, NVIDIA introduced the concept of egocentric data collection in its EgoScale Project, which shifted the primary method of data collection from isomorphic exoskeletons to camera footage. NVIDIA’s team attached cameras to the foreheads and wrists of over 1,000 people, which they used to film 20,854 hours of first-person footage: mostly everyday tasks, such as washing dishes and cutting vegetables. Researchers mapped the positions of the subjects’ wrists, fingers and joints through each frame of the footage. This preliminary data was then augmented and refined using a variety of isomorphic techniques.

In their paper, the NVIDIA team claimed that training a robot on egocentric data increased the chances of it successfully folding a shirt by 88%. “Humans can be treated as a scalable embodiment for learning general embodied intelligence,” celebrated the authors of the paper, before encouraging future researchers to record as much video as possible for robot training. Imperfect and low-quality footage can always be cleaned: what matters is getting enough video to clean. Today, among most companies in the egocentric data business, the consensus is, the more, the merrier.

As was the case for teleoperated and isomorphic data collection, egocentric techniques quickly latched on across the U.S., before spilling across the world to China. The U.S.-based EgoVerse uses Meta’s Project Aria glasses, or their Mecka app downloaded to an iPhone 15+, to record an individual in a variety of settings. Build AI has raised over $30 million with the goal of collecting over 1 billion hours of egocentric footage by 2028, a goal shared by a dozen other startups in Silicon Valley. And in an interview with Wall Breaker, World Intelligence founder Grace Zhang declared her ambitions of creating the world’s biggest physical AI data marketplace. “I want people to film and record data of themselves doing everyday chores in their homes,” she said. “That is my ideal future.”

Egocentric view of hand–object interaction. Source.

The Chinese companies jumping on the trend

Today, Chinese companies have been collecting egocentric data for their own in-house use, or to sell to domestic and overseas consumers. Jingdong, or JD.com, was one of the first Chinese companies to jump on the egocentric data collection wave. On March 16, 2026, founder and chairman Liu Qiangdong announced the creation of a data collection center in his hometown of Suqian (宿迁), Jiangsu Province, with the goal of collecting over 5 million hours of footage by 2027 and 10 million hours of footage by 2028. Over 100,000 Suqian citizens have signed up to film themselves performing five categories of everyday tasks — logistics and warehousing, industrial manufacturing, healthcare, home services, and urban operations — in 100 distinct environments and scenarios. They use JD.com’s own headgear, which is said to weigh less than a pair of glasses.

But Liu’s ambitions don’t stop there. In August 2026, JD.com launched their JoyAI video editing program, which can use preexisting egocentric footage to generate even more videos capable of training robots. The program can alter the background or lighting of an original video, simulate environments where it may be too dangerous or inconvenient to bring a camera, or turn human hands into robot hands onscreen. Perhaps the egocentric data collection industry is already starting to cannibalize itself. Or, perhaps, this is a vital new direction for an industry still grappling with high production costs and a fluctuating market.

Navigating the data market

Because, despite all the hype and speculation, egocentric data collection is still an unpredictable industry. An anonymous industry expert claims that both collectors and buyers are hesitant about the industry’s scalability and sustainability, due to the cost of creating usable data. “People wouldn’t hesitate to buy egocentric data if it were the cost of water (5 to 10 yuan per unit),” he tells me, “but I don’t think the cost of collecting data can be lowered any further than what it is today.” He places the cost of collecting one hour of raw data at around 100 yuan, and the cost of generating usable data at around 500-700 yuan per hour.

Dr. Iris Wang, an entrepreneur and researcher at Tsinghua University, sat down with me to discuss the business side of the industry. “The market decides the direction of the industry and innovation,” she says, “not the other way around.” In the vast ecosystem of data collection — where data is sold to startups and tech companies, and where these startups and tech companies sell finished robots to consumers — the vendors of egocentric data lack truly substantial bargaining power. The needs of the consumer still dictate the direction of data collection and research. For instance, while academic research labs might want to purchase dexterous hands trained on egocentric data, most warehouses and commercial companies are still happy to use binary gripper hands trained on isomorphic data.

“The second generation of data collection — isomorphic exoskeletons and UMI — is still very popular among Chinese buyers,” says Dr. Guo. “Unless egocentric data collectors can harvest, say, 10 billion hours of data, we can’t outperform UMI in the market.” The anonymous industry expert agrees, estimating that about 30% of the robot market still only seeks to purchase second-generation UMI-based data for gripper hands. “I wish the market would give us a clearer direction,” he tells me. “That, more than anything, will decide the future of data collection.”

The UMI gripper setup, as illustrated in the original Universal Manipulation Interface paper.

Egocentric data is not a versatile product. Footage is difficult, if not impossible, to resell: customers often have specific requirements for the lighting, background, and movements for every video, and customized data is often protected by exclusive IP. In response, many companies collect data solely for in-house use, creating complex and fluctuating dynamics across the supply chain.

But Chinese companies might find other opportunities in the global supply chain of data. Chinese data aggregators often serve as the middleman between Global South countries and their Global North clientele, selling footage collected from all over Asia (including China itself) to the U.S. and Europe. “Labor costs are quite low in Asia,” says Edwin Yap, the founder and CEO of Mr. Robot, a data collection and robotics company based in six countries across Southeast Asia. Isomorphic data collection is still too expensive to be the basis of a sustainable business model, so Mr. Yap uses his egocentric footage to train his own robots, and to sell to data aggregators in other countries.

All my interviews frequently returned to the cost of production and labor. On paper, egocentric data collection seems cheaper than both isomorphic and teleoperated methods: after all, there’s no cost of production needed for a giant exoskeleton or remote controls. But the cost of training workers, particularly at the scale needed to reach 10 billion hours of footage, is still considerable. While major companies such as JD.com, ByteDance, Ant Group, and Xiaomi have the financial resources to conduct extensive pre-training, collect usable footage, and purchase the necessary data, startups in the industry face a much more uncertain path. According to Dr. Wang, startups raising over 1 billion RMB will usually pursue egocentric data used for pre-training, while those raising less than 1 billion RMB focus on post-training methods that require little to no ego-centric data. Our talk — and the news of JD.com’s data collection center — drew me back to the question of labor: whose data finds its way into algorithms, and what part of the reward do they reap?

The hands behind the camera

A quick scroll on Xiaohongshu, one of China’s foremost social media platforms, reveals hundreds of advertisements posted within the last month: recruiters offering anywhere from 100 to 200 yuan per day for workers willing to travel to a data collection center and film themselves using egocentric hardware and techniques. I did not find any adverts affiliated with big internet companies or startups: rather, these posts were published by small actors looking to sell data to more powerful buyers.

Some adverts ask for workers to arrive with experience in egocentric data collection; others offer training on the spot. More specific requirements — the workers’ age, eyesight, mobility, susceptibility to VR-induced vertigo — vary by recruiter. All the adverts boast reasonable working hours (9 AM to 6 PM, with a lunch break) and overtime pay. But none of them offer the permanent contracts or benefits that are associated with full-time work opportunities.

Dozens of users debate the merits of the job in the adverts’ comments sections. “Seems boring,” remarks one commenter. “Seems sketchy,” quips another. Others are more curious, asking for specifics about the work and pay. One Xiaohongshu user offers a detailed day-in-the-life of a data collection subject: a 6:15 AM commute to a data collection center, a physical check-up for blood pressure and body weight, and an entire day spent copying hand movements off a tutorial video. The user needed to earn a complete quota of “data points” — measured by the accuracy of her movements compared to the tutorial video — to be paid in full. After a day and 40 minutes of overtime spent trying to hit this quota, she received 180 yuan ($27). “Most people just wanted to get their money and leave,” the user writes. “No one knew each other, and no one talked.” I was curious about the possibility of a shared class consciousness between data collection subjects, particularly in large and well-funded data centers such as JD.com’s, but came away with no conclusions: the industry, it seems, is still too new and fragmented.

Egocentric data collection, especially when integrated into existing workplaces, raises the question of labor and privacy protections. New technologies often explode in scale before laws can catch up, and the selling of personal data has always been mired in a legal gray area. This is especially complicated in an international data market, where labor and privacy laws differ by country.

Human Archive, a Redwood City-based startup, has partnered with a number of companies in India and Southeast Asia to distribute headsets to workers in home services, hotel, and restaurant sectors. They aim to sell this footage to AI labs and robot companies in the U.S. and Europe: the latest iteration of a worldwide supply chain that depends on cheap labor in the Global South. These workers receive about $1 an hour for participating in data collection.

The kind of data collected by Human Archive. Screenshotted from the company’s LinkedIn page.

Human Archive claims to follow the law of whatever country they are collecting from, such as the Digital Personal Data Protection Act in India. But not all workplace-integrated data collection is so transparent. In April, a Scroll.In story revealed that Indian workers were not told that the egocentric videos they filmed would be used to train AI models. And in the same month, Puneet Jindal, the CEO of Labellerr AI, told CNN that many factory workers do not receive any information about how their data will be used when they start to film themselves. Following pressure from Bengaluru residents, India’s Ministry of Electronics and Information has been looking into the controversy surrounding home services startup Pronto, which announced that it had launched an opt-in recording feature for workers performing household tasks. It is unclear whether Pronto is selling the data collected from these recordings to buyers.

In China, workplace-integrated data collection is still rare. Chinese data collectors recruit subjects to film themselves at specialized centers, or from their own homes, instead of attaching cameras to workers in their workplaces. The Chinese model, as I’ve gleaned from interviews and Xiaohongshu adverts, is to generate high-quality footage that doesn’t need much cleaning or processing. This stands in stark contrast to the low-quality footage collected in workplaces in South and Southeast Asia: data which may be cheaper to collect, but more expensive to clean.

From Build AI’s website, which claims to be “hypergambling” with a “small and exceptionally cracked team” to “unlock human potential” — to Human Archive’s, which claims to erase all “undesirable physical work” and to usher in “a new era of abundant creativity” at this “once-in-a-generation inflection point” — American rhetoric around egocentric data collection is big, bragging, and triumphant. This is a far cry from the tentativeness I witnessed in Chinese founders and researchers, the majority of whom are still waiting for the market to decide the next direction — or generation — of data collection methods. This is even farther from the sentiments of the humans whose footage become the data, which range from apathy to suspicion to resistance.

At the end of our talk, Dr. Guo turned to me and grinned. “The U.S. crosses the river by feeling the stones, and China crosses the river by feeling America,” (美国摸着石头过河，中国摸着美国过河) he said, a quip that made me laugh out loud. It’s a play on Chen Yu and Deng Xiaoping’s famous idiom: improvising amidst the rapids, seeking stability wherever it may be found. In Dr. Guo’s eyes, it is the U.S. that is feeling for the stones and China that is feeling for the U.S., looking for stable ground in the shadow of American startups and labs. The narrative is not as simple as U.S.-China competition.

There is a heavy-handed metaphor to be made about the data that crosses not only a river, but an ocean, from India to China to Silicon Valley. About the rivers of power and money that reach far beyond any individual’s comprehension. My interviews clarified an industry that is often mysticized by hype and technical lingo. They also exposed me to a sprawling supply chain that is both extremely unpredictable and historically embedded in our current world order. I left Wudaokou with two questions: What are we overlooking in our rush to cross the river? And does the same shore await us all on the other side?

MOFCOM's New Rules and the Future of Chinese AI

Quinn Ennis — Wed, 21 Oct 2026 13:43:39 GMT

In the AI race thus far, openness has been China’s secret sauce. As fast followers to the frontier, Chinese AI companies have gained trust by publicizing their weights and architectures. And at the cost of revenue, they have achieved global adoption, brand recognition, and a truly competitive dynamic with American labs.

Last week, China’s Ministries of Commerce (MOFCOM) and Science and Technology (MOST) jointly proposed expanding China’s catalogue of export-controlled technologies, with most of the proposed regulations taking aim at subsets of China’s AI industry. Regulators will spend a month taking comments on the draft before anything becomes law, but we can already learn a ton from this announcement in its current form.

Does Beijing want to hold back the tides of open models? Why is MOFCOM writing these rules now, and what does this say about the Chinese government’s appetite for AI-enabled risks versus growth?

To find out, we spent three days poring over the draft regulations — so you don’t have to!

What’s ahead:

How these rules came to be — and how this draft is already affecting the Chinese AI industry;
What the proposed rules mean for frontier models, chip design, memory, cybersecurity, AI content detection, and even China’s minority languages;
Who actually decides which export licenses get approved, and the balance of power between provincial and central authorities;
What these rules tell us about Beijing’s expectations for AI development.

The 40 proposed additions in Announcement 48, split across AI, photovoltaics, and biomedicine.

Today, we’re just focusing on the AI-related regulations, but a future piece tackling everything else — from this announcement’s proposed restrictions on PV cells and gene therapies, to other recent announcements on port dredgers and tunnel borers — is in the works! Subscribe so you don’t miss it:

Subscribe now

The Road to a Rule

Since 2001, MOFCOM has made regular additions to the Catalogue of Technologies Prohibited or Restricted from Export 中国禁止出口限制出口技术目录. In recent years, this catalogue has basically served to list the technologies for which China’s dominance in global production affords them leverage over the rest of the world — a toolbox of weaponizable dependencies, if you will. In the last several years, this has mostly meant processing and manufacturing technologies for rare earths, magnets, and battery materials.

Now China has decided AI models belong on that list as well. On October 16, MOFCOM released Announcement 48 of 2026, detailing 40 new additions to the Catalogue and, effective immediately, requiring pre-export registration of all 40 additions in the interim. Of these 40 additions, 14 are outright export prohibitions (J codes), while 26 are export restrictions requiring licenses (X codes).1

It’s a bad day to be working at MOFCOM. Meme template from Chicken Chop Bro.

While these 40 additions have garnered the majority of public attention, equally important is the section of the announcement creating an interim pre-export registration system. From last Friday to the date that the revised Catalogue takes effect, any technology exporter must register any exports of technologies among the above additions prior to export, including exports that have already begun but not yet completed (已经开始但尚未完成). There is no precedent in any earlier catalogue revision for the creation of an interim registration regime announced with draft additions. Beijing is creating a better dataset so that it can make an informed final decision on what to license.

Under Announcement 48, registration is required for a Chinese company hoping to: 1) assign ownership of the technology abroad, like in the case of an IP sale, 2) license domestic technology abroad, 3) change the owner or actual controller of domestic technology to a foreign entity, or 4) do anything else that MOFCOM later decides to specify. The triggers for registration also suggest that MOFCOM is testing what mechanisms will ultimately be used for classification of an AI export. An AI model is not a physical item like a gallium ingot, so the question of what constitutes export is less clear.

This third trigger effectively codifies the Manus case. In December 2025, Meta signed a deal to purchase Manus, a Chinese agent startup that had moved to Singapore, but following MOFCOM review, a security review led by the NDRC led to the deal being reversed. Now, rather than relying on investment review authority, similar deals would require a license on export control grounds, and redomiciliation would not let a company escape Chinese jurisdiction.

In the registration form provided with the announcement, MOFCOM asks for standard information like the recipient, technology code, user and end use, and secrecy grade. It also asks for specific circumstances of exports not made under contract (不依据合同出口的具体情形). This forces exporters to judge for themselves whether they should report a potential action, and may point to open-sourcing and sharing code within a group.

IP and National Security

These rules fall into two categories: controls that protect intellectual property (IP), and controls that protect state security.

To Xi Jinping, these are not separate justifications. As he explained during a 2020 study session speech, “Intellectual property protection weighs on national security; only by strictly protecting intellectual property can we effectively safeguard our country’s independently developed key core technologies and prevent and resolve major risks.” (“知识产权保护工作关系国家安全，只有严格保护知识产权，才能有效保护我国自主研发的关键核心技术、防范化解重大风险。”)

We’ll start with the controls that seem primarily IP-focused (frontier models, AI chip design, memory chips, and clusters), then move on to controls in the middle (AI-generated content detectors and agents), and finish on the primarily state security-focused controls (training data, cyber capabilities, and in a suprise addition minority languages!).

Controls by Type

Frontier models

When you use a frontier U.S. model, you are accessing the software without getting a peek at the fundamental structure behind the curtain, and thus closed models can be used without directly leaking IP. Most of China’s domestic AI champions, however, publish open weight models, meaning that the user can see the structure of the model, and much of the IP is fundamentally public. Open weight, though, does not mean open source, and the training data, preprocessing pipelines, and training code remain behind the curtain.

MOFCOM proposes to require licenses for the export of both weights and core private IP, including key checkpoints, underlying source code, architecture design, and technology for training optimization and inference acceleration. The regulation controls the core technology of a large model that exceeds any of the following three thresholds: ≥10²⁶ floating point operations, ≥2T parameters, or “global leading capability” 能力达到全球领先水平. We discuss each threshold here:

≥10²⁶ floating point operations (FLOPs)

The number of floating point operations (FLOP) is a common measure of the raw compute used to train an AI model, and is thus often a good proxy for model performance. The minimum cutoff listed in this MOFCOM announcement is identical to the figure originally mentioned in a 2023 Biden executive order, which was later used in the January 2025 AI Diffusion rule (US ECCN 4E091) that was rescinded in May 2025. This is also the same threshold as California SB 53. In its follow-up Q&A, MOFCOM mentioned “referencing internationally accepted standards” 参考国际通行做法 while drafting these proposed catalogue additions. As of October 20, 2026, no released Chinese model is estimated to have crossed this bar. The Chinese models with the largest training compute sit at roughly 2x10²⁵ FLOP, according to Epoch AI. China remains compute-constrained compared to the United States.

≥2T parameters

The number of parameters in a model is the overall size of the model, which is a rough proxy for its capability. As of now, per Epoch AI, only four Chinese models since mid-2024 (Kimi K3, ERNIE 5.0, Qwen 3.8 Max, and Qwen3.8-2.4T-A95B) cross this threshold. No models from DeepSeek, Meituan, Xiaomi, or Z.ai exceed two trillion parameters, but all have models above or near one trillion parameters, and it is reasonable to expect their next releases to cross that threshold. As the Q&A clarified, the restriction does not cover tech that is already public via open weights or literature, so there is an incentive for companies to rush new releases that would exceed the thresholds before the rule is final (I do wonder how MOFCOM feels about DeepInfer, probably positively but did they get the nod beforehand?). This test is especially important for Chinese labs as China is compute constrained, and compute-scarce labs often use a design called Mixture of Experts (MoE), creating large models that only activate a small fraction of their network for any given task. If you measure these models by FLOPs used in training, it undercounts their power and makes them look weaker than they are. Total parameters, while a much better proxy for the Chinese frontier, has a potential weakness: the number of parameters is a design choice, so labs could try to game it by releasing models that slide under the 2T threshold.

Global leading level

This category is by far the most up-for-interpretation. The note states that the bar is the global top 10 through public evaluations from “mainstream domestic and foreign testing organizations and large-scale industrial platforms” 国内外主流测评机构，大规模产业平台 across “overall performance, specialized operational capabilities, global call volume, safety performance, training efficiency or others” 综合性能，专项业务能力，全球调用量，安全性能，训练效率，等. This definition allows the rule to cover a wide range of models. The rule does not name a ranking body, and, in practice, different AI ranking bodies frequently disagree on their ordering as they use different tests and rating algorithms. Likewise, the rule would potentially affect any model that was in a top-10 list on any one dimension, even if the criteria were narrow. The mention of call volume also has the potential to pull in weaker models that are used in high volume for easier tasks. Given that the status is judged at time of export, it is likely that the same model could jump off or on a leaderboard between subsequent exports without the model itself changing.

Left: How current Chinese models measure up to Announcement 48’s size and compute tests. Some models lack public compute estimates. Right: How top Chinese and U.S. models compare on one public benchmark naming — the draft rule names no specific ranking.

The carve-out for already-released open source models means that published weights are out of the scope of this control, but closed-weight model companies hoping to license existing models to foreign firms and future releases by open-weight companies would still be restricted under the proposed rule. Just because a model is open weight, it does not mean it is open source. Open weight companies still hold private know-how, and this IP related to training is directly controlled by a clause mentioning training optimization and inference acceleration 训练优化、推理加速. Open sourcing a restricted model is a controlled export and would get an automatic license.

AI chip design

The announcement proposes restricting the export of AI chip design that checks off at least one of the following boxes:

requires a manufacturing process of 14 nm or below;
is able to do floating-point or integer arithmetic;
uses advanced 2.5D or 3D packaging to integrate multiple dies;
has high-speed interconnect capacity;
has an architecture integrating memory and compute;
or is used in autonomous driving, computer vision, or speech.

The inclusion of any chips with floating-point or integer arithmetic capabilities 具备浮点或整型运算能力 means that this covers almost all digital chips, including all AI chips, which suggests either a drafting error or an extremely wide breadth (MOFCOM will probably be getting some polite letters around this clause…). Likewise, all AI chips used in datacenters require high-speed interconnect capability to interface with other chips and memory.

The inclusion of architectures that integrate memory and compute is also noteworthy. Traditionally, AI chips use von-Neumann architecture, where memory and compute are housed separately and data moves back and forth between the processor and the memory. This shuttling of the data wastes energy, however, and technologies like in-memory and neuromorphic computing, explicitly mentioned in the announcement, eliminate this energy loss and allow for greater parallelism. Non-von-Neumann architectures would also potentially allow China to circumvent its shortage of traditional high-bandwidth memory (HBM). While these technologies have seen limited deployment in datacenters because the technology is immature and lacks a mature software ecosystem, they have been used in smaller research data centers and are used in robotics, drones, and autonomous vehicles. China is highly competitive across these three sectors. Indeed, when paired with the direct inclusion of chips used in autonomous driving and computer vision, this gestures to a strategic vision. For Beijing, protecting its advantage in memory-integrated chips could help it plan a future data center build-out that is not dependent on Western memory supply chains.

This forward-looking export control on a technology that has not yet picked up real adoption is also reminiscent of the FCC’s own Robot Ban. No one is really sure whether these technologies will take off, but given they are not yet widespread, the easiest time to restrict them is now. The market being hit is minimal, and companies that design such chips don’t have large vested interests to protest against the rule. In the case that non-von-Neumann architectures find commercial viability in the future, China will not have to worry about the US profiting off designs and IP; that door would have already been shut.

This restriction, and the following restriction on memory chip technology, does not restrict the chips themselves. Chinese companies remain free to sell AI chips and the memory used in data centers to U.S. companies. The restriction instead applies to the intellectual property associated with these chips, allowing Beijing to have a say in patent deals with foreign companies.

This design control can also be construed as MOFCOM disallowing Chinese companies from fabbing their designs at TSMC or other companies. While BIS has attempted to play the cat-and-mouse game to prevent Chinese AI chip designers from covertly sending their designs to TSMC, MOFCOM may be cracking down on Chinese companies itself. Time will tell how the rule ultimately gets enforced: is it more about protecting Chinese designs, or forcing Chinese designers to work with Chinese, not Taiwan or American, fabs?

Memory chips

This MOFCOM announcement further focuses on memory with its proposed restriction of memory chip design technology export. This restriction’s use of an “any of the following” structure creates a broad category under the guise of narrow delineations. The restrictions on 3D flash (long-term memory) and DRAM (short-term, working memory) design are not AI-specific, capturing much of the memory work at YMTC and CXMT. For a brief refresher on memory, see Aqib’s excellent primer on Chinese HBM.

Many of the limbs of the restriction allow broad interpretation. For both types of memory, the proposed addition would require export licenses for all memory that uses bonding technology to join separately made chip layers. This would cover every flash chip made by YMTC as well as almost all HBM chips, which are further covered by the mention of any chips using multi-layer stacking. For flash, the announcement includes density (8 Gb/mm²) and speed (1.5 Gbps I/O) minima that both fall well below current mainstream chips. Likewise, for DRAM, the restriction would apply to any designs requiring a manufacturing process below 18 nm, which would include almost all CXMT DRAM.

These numbers are copied over directly from Biden-era regs. The 18 nm DRAM cutoff is the exact threshold used by the U.S. Department of Commerce’s Bureau of Industry and Security (BIS) October 2022 export controls on memory! This BIS control also included a line on flash, restricting chips with 128 layers or more. When converted from layers to density,2 this is almost exactly the limit proposed in this MOFCOM announcement. On the surface, this symmetry with Washington can be seen as preparing a ready-made defense for these restrictions as “standard international practice.” This can also be viewed, however, as fencing off the IP that China has built in the last four years because of those BIS controls — a protection of the dividend from its forced indigenization of these technologies.

Another inclusion worth noting is DRAM technology that uses vertical channel transistors (VCT). Where traditional memory cells are laid flat, in VCT the cells are stood upright, allowing for more memory in the same space without requiring higher precision lithography. China has invested heavily in this technology, as it presents a potential to circumvent Western denial of access to fabrication technologies like EUV. Indeed, in the week before the draft, CXMT announced that it would release a VCT design that shrinks the footprint of each memory cell from the mainstream level by a third by the end of 2026. This prompted a drop in the stock prices of both Samsung and SK hynix, both of which have targeted 2028 for their release of VCT memory. Both Samsung and SK hynix could have pursued VCT earlier but ultimately decided to squeeze more juice out of EUV before turning to VCT. When they do transition, however, the process will inevitably take time as the companies struggle to reach appropriate yield with a new design. China does not want to give the big memory makers any shortcut with that trial and error, as China has learned from the footsteps of other companies. By controlling VCT, China is hoping that the Korean companies will face the same yield struggles as CXMT.

Clusters

US chip controls have forced China to compete through workarounds, and now, Chinese firms have gotten good enough at engineering whole alternative systems for compute that the government believes they have domestic IP worth protecting. Areas where China could have a comparative advantage include wiring many weaker chips into superPoDs like Huawei’s CloudMatrix 384, engineering systems to keep running through failures, and mixing Nvidia and domestic fleets as one. This corresponds directly to this announcement’s proposed restriction on the export of technology relating to large clusters (≥10,000 AI accelerators) with either fast interconnect (≥800 GB/s), fast recovery from failures (≤120 s), high scaling efficiency (≥98%), or unified scheduling across chips from at least two vendors.

This rule would protect such comparative advantages by giving automatic licenses to self-built, self-operated overseas data centers of Chinese firms, while exercising real bureaucratic scrutiny over prospective sales of the tech and IP licensing agreements. That could be used to push companies in the Gulf and Southeast Asia to rent Chinese compute rather than buy it, as the hardware could be freely sold but the know-how to build or run them without Chinese engineers is gated by the licensing desk. Even though China doesn’t have enough chips today to export, this would seem to make the ‘Chinese stack’ less attractive going forward.

Agents

The proposed agent restrictions were inspired by the debacle with Meta’s attempted acquisition of Manus, aiming to formalize the authority to block future transactions.

Chinese firms already sell agent products abroad (from well-known tools like Manus and Kimi, to specialized platforms like Alibaba’s Accio Work), and they could continue to do so under the proposed changes — but they will now need a license to allow the client to host and manage the agent reasoning loop in their own environment.

The agent capabilities that trigger the restrictions (autonomous planning/task decomposition, multi-agent collaboration, and tool use) are not necessarily unique to Chinese firms, and the thresholds in each of these qualifying capabilities seem somewhat arbitrary (e.g., integration of 200 third-party tools). But the Manus sale indicates that Chinese firms do have valuable IP when it comes to agents, so we can think of these controls as applying to the specific architectures that enable Chinese agent products to hit these thresholds while keeping costs comparatively low.

The fourth capability proposed for export restriction is “AI decision-making technology for strategic interactions” (人工智能博弈决策技术), e.g., wargaming and game-theoretic AI.3 Controlling those capabilities is, of course, more in line with the state security justification than with IP protection.

Spoiler?Show less

The Lithography Lever

Jordan Schneider — Tue, 20 Oct 2026 10:46:15 GMT

Why should you care about lithography exports sold to China? What is Congress thinking about doing about it? And what does it mean for the future of humanity, the lightcone, great power relations — all of the above.

Nick Brown is a co-founder of the Center for Technology and Statecraft (CTS), a new think tank spun out of IFP, that looks at long-term technology challenges and opportunities facing the US. Nick was a physics major, spent a while as a software engineer in and around government, and had an arc as an aspiring concert violinist. But we’re here to talk about what is in pole position for my paper of the year: his master class on what would happen if the US cut off DUV exports to China: DUV Immersion Lithography: The Foundation of US AI Compute Advantage. Co-hosting is ChinaTalk’s Aqib Zakaria.

We discuss:

The case for banning DUV exports to China: Why lithography is the hardest part of the semiconductor manufacturing stack for China to replicate/indigenize and the strongest point of leverage the US and its allies have over China’s future AI compute.
The cannonball-sized hole in US export controls: China imported roughly 90 immersion lithography systems a year in 2024–25, nearly two-thirds of ASML’s immersion exports. If imports continue at this rate, China could approach or even reach parity with the US in AI chipmaking within a decade.
What the MATCH Act would do: How Congress could close this gap by restricting DUV immersion exports and servicing, and whether the US can get its allies on board.
Will Beijing retaliate? Why a gradual, long-term restriction may be less likely to provoke another rare-earths crisis
Buying time for AGI: How cutting off DUV imports could preserve America’s AI compute advantage into the 2030s, giving the US more time to navigate the technological, geopolitical, and societal implications of increasingly powerful AI.

Listen now on your favorite podcast app and check out the MATCH Act’s theme song below.

Compute, Leverage, and the Cannonball-Sized Hole

Jordan Schneider: To set the groundwork: why should we care how much compute China has relative to the rest of the world?

Nick Brown: Compute is one of the fundamental drivers of AI capability. There’s compute, algorithms, and data, and a decade of work by many people has established that of those three, compute is the most governable — the one where it’s most possible to intervene and shape where it goes. Those factors combine to create AI capabilities, both in training and in deployment, and those capabilities get realized in some context: economic, governmental, security-related. If you care about the exercise of AI capabilities, which many do, and about the degree to which China can exercise them relative to the US, then compute — the number of chips, their quality, the aggregate computational stock available — is one of the fundamental levers you’d want to examine or push on.

Jordan Schneider: Let’s do the tale of the tape as of October 2026. Aqib, how much compute does China have, and how much does the rest of the world have?

Aqib Zakaria: Everyone’s numbers are hazy, but at the median estimate, the US has a commanding lead. Measured in H100 equivalents, China has maybe around a million, and the US has upwards of thirty million.

Jordan Schneider: With asterisks — remote access, right?

Aqib Zakaria: Yes. That’s roughly one million Chinese-designed chips in China. But when Nick Corvino and I ran the numbers, there’s almost another million in remote-access compute China can reach in data centers in Thailand, Malaysia, and elsewhere in Southeast Asia. Which goes to show how hard this is to measure.

Jordan Schneider: So in a worst case, we’re still at a 10x advantage as of fall 2026.

Nick Brown: There are three segments: remote-access chips, exported chips, and China’s indigenously produced chips. The chip export wars continue to rage — H100, then H800, then H20, then H200 — but I’d segment all of that from indigenous production. The difference between the three categories comes down to leverage.

For remotely accessed chips, there’s a lot of leverage, and even within that category you might distinguish chips accessed in the US from those accessed elsewhere. The US, its companies, and entities that answer to it as a sovereign have enormous leverage over how that compute is used; they could cut it off at any moment. For exported chips, there’s leverage at the point of export. Once you decide to export, you let go of that leverage, but production remains within your grasp. For indigenously produced chips, there is very little leverage. If China can make its own chips with its own supply chain, as many as it wants, the compute chokepoint stops being an effective governance tool—at least for shaping international power. That is why it’s so critical to look at China’s indigenous production while we debate how many chips to export.

Aqib Zakaria: So we already banned EUV. Why isn’t this a solved problem? What’s the deal with DUV?

Nick Brown: EUV and DUV are two technology generations: extreme ultraviolet and deep ultraviolet. We’re not extreme, we’re just deep. EUV gets all the press, and in many ways it should — by many measures it’s the most advanced technology on the planet, and individual machines sell for hundreds of millions of dollars. But DUV machines, which still go for $60–80 million, are extraordinarily powerful pieces of technology. The most advanced DUV systems are certainly capable of manufacturing advanced AI chips. China’s current flagship AI chips are Huawei’s Ascend line, manufactured indigenously, and at the nodes those chips use, many DUV systems can support production. Those systems are good enough to facilitate China’s indigenous AI chip manufacturing.

Jordan Schneider: Say we stay on the current path: DUV exports aren’t controlled, and the machines keep flowing into China, where the rest of the world has no leverage over them once they’re installed. Play that movie out five or ten years. Why is DUV so important to the buildout, and where does it leave China relative to the rest of the world?

Nick Brown: Two scene-setting points, then the scenarios.

First: from an export control perspective, lithography is the strongest point of leverage the US and its allies have. That comes down to the fact that it is by far the hardest tool for China to indigenize. You can see that across the industry — lithography tools cost far more than other tools, and ASML’s R&D expenditure is higher than that of toolmakers spread across many product lines. Digging into the numbers, we feel confident that China will not be able to make its own DUV systems good enough to make AI chips, at a scale sufficient to really expand its chipmaking, until into the 2030s. That adds up to a very large point of leverage.

Second: there are many generations of DUV systems, from ASML’s most advanced to the earliest ones from the 2000s. The version we currently allow into China — the NXT:1980i, for the nerds — is good enough for 7nm logic and HBM2E, maybe even some HBM3, which is what Huawei’s chips use. Those tools contribute to AI chip production through various channels inside China. China has very astutely recognized that we have a big hole in our export controls — it’s not really a loophole, we’d call it a cannonball-sized hole — and has taken advantage of it, importing a staggering number of systems. We estimated the entire installed base and the rate of imports: about ninety systems per year in 2024 and 2025, worth over ten billion dollars, close to two-thirds of ASML’s entire immersion DUV exports.

Modeling China’s Lithography Ceiling

Nick Brown: Right now the US is dramatically ahead of China in chipmaking and compute capacity. That’s because China is constrained by a number of things — not just lithography, but etch, deposition, and other SME restricted by US controls. But there’s good reason to expect China to indigenize those other segments over time. Litho is almost certainly the hardest and the last tool they’ll develop. So we estimated China’s lithography-bounded capacity: assuming they do indigenize etch and deposition, what chipmaking capacity does their current stockpile of lithography systems enable?

Two high-level findings. First, their current stockpile, which is enormous because our export controls haven’t been tight enough, is nonetheless physically incapable of matching US AI chip production through 2035. Even if they diverted every system to their most advanced fabs and spent 100% of that capacity on AI chips — not a single smartphone — at the basic level of physics and lithography throughput, they’d be well below US production, something like a tenth of it.

Second, if they continue importing at the 2024–25 rate, the picture changes. I’m not saying they’re guaranteed to catch up or surpass, but the curve bends upward. There are scenarios in which US production tapers off, China’s accelerates, and China makes a concerted effort to prioritize AI, where you could be looking at near parity, maybe even parity, in ten years. At the highest level, this is the difference between an essentially guaranteed advantage and, at minimum, a much smaller advantage with real uncertainty about the structure of the overall computational equilibrium.

Aqib Zakaria: Everyone should read the report; the graphs are very pretty. Two questions. First, how should I weight the probabilities? There are a couple of variables — that they keep importing at this pace, and the share they dedicate to AI chip production. Is this a 1% scenario, or something that should keep policymakers awake at night? Second, if this is gradual, when is the right time to strike?

Nick Brown: First, I’d push back on the implicit binary of “beat the US / not beat the US.” The right frame is relative chipmaking advantage. It’s not that if China makes one chip more than the US, everything changes — that’s not how it works computationally, and certainly not for strategy and national power. But the ratios at which chips can be made do matter.

On the modeling, I’d emphasize again that these are lithography capacity bounds. We think they’re informative because China will indigenize the rest of the supply chain eventually, almost certainly before litho.

Jordan Schneider: If you think they won’t indigenize everything, or will do it worse, or yields will only be so good — don’t we get a whole discount factor before we even reach the list of reasons to feel better about selling these machines?

Nick Brown: It depends on the timescale over which you want to feel better. Lithography systems are used for many years after they’re sold. In the next year or two, China is primarily limited by deposition and etch and other SME where we’ve done a better job of controlling. But over the mid to long term — three to ten, three to fifteen years — it’s not tenable to assume China will fail to indigenize etch and deposition. They are making progress. CSET has a very good set of reports from 2025, updated this year, showing inroads in etch, deposition, CMP, and many other segments. The one segment where they have zero percent market share is lithography, and they continue not to obtain meaningful share. So in the near term, treat these bounds as useful information rather than a projection. Over ten years, you have to take them seriously. In the US, we expect production into the 2030s to be at least partially limited by EUV — there have been some good podcasts on that.

On the scenarios, there are three factors. One is how much they import: the same pace, less, or slightly more. If we take as a premise that AI will become more and more important — not unreasonable in fall 2026 — it’s not crazy to assume they modestly ramp up, even though the 2024–25 pace was high and partly driven by stockpiling. Within China, there’s the question of what fraction of machines are dedicated to AI, which depends on what fraction sit in fabs capable of making AI chips — SMIC and CXMT — and, within those fabs, what fraction make AI chips versus other advanced chips like smartphone processors.

We break out a “current trends” scenario, which assumes the allocation we estimated in our installed-base work continues: roughly the same fraction going to advanced fabs, and the same fraction within them going to AI. Then there’s an “AI push” scenario: China decides AI is truly at the top of every list. It’s already a national priority in many ways, but as anyone who has worked in government can tell you, there are different levels of urgency depending on how bought-in leaders are. The AI push assumes China reaches the AI allocation we forecast for the ecosystem outside the US. If they keep importing at or slightly above recent rates and reach that allocation, you start to see potential overlap or near parity in US and Chinese chipmaking.

What the MATCH Act Does

Jordan Schneider: There’s legislation that purports to deal with this. ChinaTalk listeners will be aware of the Trump administration’s reluctance to act on its own, so over the past year Congress has come into the breach with bills that the Commerce Department could implement tomorrow but hasn’t. Enough senators and congresspeople are annoyed about that to pass bills with strong majorities. Nick, what is the MATCH Act, and how does it relate to DUV immersion lithography and bending the cones of the future in ways that advantage the West?

Nick Brown: The MATCH Act is a bipartisan, now bicameral bill dealing with exports of semiconductor manufacturing equipment, including but not limited to DUV immersion lithography. It’s in the Senate version of the NDAA, so it’ll be discussed in NDAA conference over the next couple of months — always a hectic and Byzantine process.

Separate what it seeks to accomplish from how. On the what: it enhances controls on SME in two buckets. The first is the really good stuff — the “big bad SME” — and notably that includes DUV immersion lithography, plus a few other categories of extremely capable tools. For those, MATCH seeks to ensure the US and its allies place countrywide controls. In export control land, there are two ways to control an item. One is to control it to the entire country, preventing it from entering China at all. The other is to allow it into the country but restrict it from specific entities — it can go to China but not to SMIC, CXMT, or Huawei. For these tools, MATCH elevates the control status from entity-restricted to countrywide.

The second bucket is actually extremely large — basically all SME — and there the issue is a miscalibration between US controls and those of notable allies, chiefly Japan and the Netherlands. For two specific entities, SMIC and CXMT, China’s most advanced fabs, the US controls all items subject to its jurisdiction, notably all SME. Not a single US SME tool can go to those fabs. That’s not true for Japan and the Netherlands, which control a much smaller set. MATCH seeks to align those controls.

In a very meaningful sense, the countrywide controls are also an alignment: they match controls on allied tools to steps the US has already taken on tools made by its own SME makers. Hence the name.

On the how: the act directs the Commerce Department to negotiate with allies for about 150 days. All else equal, everyone would prefer these tools be controlled nationally, using the authorities of the Dutch and Japanese governments when it’s their toolmaker. That solution is better for a number of reasons. But it’s often hard to get allies to agree — they have their own interests, domestic politics, and economics. So MATCH directs BIS, if negotiations fail, to impose controls unilaterally. BIS has a number of legal mechanisms for this, which it has used before to control both chips and tools — including some of these very tools at various points during the long saga of the Biden administration.

Aqib Zakaria: The MATCH Act does two things: it restricts further exports of these tools to China, and it restricts servicing. ASML and Tokyo Electron engineers are stationed in China to fix tools in Chinese fabs when they break. Is it necessary to restrict servicing too, and how much impact would that have? Or is it enough to stop future exports and let servicing roll, especially since it’s such a large share of these companies’ revenue?

Nick Brown: It’s wise to consider those independently — they have different implications for China’s chipmaking and for the toolmakers. Speaking for myself, the fundamental thing MATCH does, and the most impactful in the long term, is cutting off item exports. Doing only that, even with servicing continuing at some fabs, massively shifts the state of play.

That said, there are already servicing controls that apply at different fabs. The US has very extensive servicing controls; the Dutch and Japanese have substantially less extensive ones. Where I’d draw a distinction, if I were waving a magic wand, is between advanced and legacy fabs. Advanced fabs have an advanced logic or memory node — in export control parlance, 16/14nm logic and below, with an equivalent memory definition that captures about 18nm half-pitch. Those fabs are where AI chips are made; legacy nodes generally aren’t sufficient. It has been US policy, at least so far, not to try to destroy China’s legacy capacity — there’s a whole set of arguments about legacy chipmaking, and some people would like to curtail it, but current controls focus on advanced capacity. So Congress could consider differentiating servicing at advanced versus legacy fabs. That also substantially changes the revenue calculus for toolmakers: if 50 or 70 percent of your tools are in legacy fabs and you can still service those, that revenue comes back. No doubt it’ll be written five hundred different ways in conference, but it merits a lot of thought.

Aqib Zakaria: If we pull servicing from advanced fabs, the biggest source of unclarity in my research has been: what actually breaks down? ASML and Tokyo Electron engineers are there, but they use a lot of Chinese contractors, many of the engineers are themselves Chinese and won’t leave if you restrict servicing, and some replacement parts may be made in China. Does anything break? Does this hamper Chinese production tomorrow? What breaks first, and on what timeline?

Nick Brown: That is a good question, and in a field full of opaque topics, that’s a particularly deep, dark rabbit hole. There are real strategic questions — if all the servicing simply shifts to Chinese contractors, arguably you haven’t done much. In reality, I think there will be impact, though probably not immediately. Every subcomponent has a timeline on which it degrades, and there are different levels: some problems an on-site engineer can fix, some need a couple of spare parts, some require calling headquarters for specialized expertise. For the most advanced subsystems — the light source often comes up, as do the stages and some sensor systems — it’s almost certain you’d see some degradation. I wouldn’t be confident production would stop; I’d probably bet against that. But some degradation in yield and throughput, starting maybe six to twelve months out and continuing over time, is reasonable to expect.

Aqib Zakaria: On implementation: we’re not known for doing things quickly. The current version gives 150 days before full effect. Isn’t that 150 days for China to stockpile — both lithography machines and the parts you just mentioned, so they can service them themselves for much longer?

Nick Brown: Stockpiling is always a risk, particularly when we announce controls many months in advance. Anyone around for the Biden-era controls remembers the famed HBM stockpiling of 2024; there’s a rich history of the phenomenon. Realistically, they will stockpile, and as 2024 and 2025 showed, potentially a lot. I can’t emphasize enough the power of one of these tools. We estimate a single NXT:1980i can produce over a hundred thousand Ascend chips per year, even assuming China’s much worse yields. They’ll probably import tens of these systems, maybe over a hundred, before the act takes effect — particularly because the window can be extended if negotiations continue, and there’s nothing government likes to do more than continue negotiations. It’s really in governmental DNA.

1 = 100,000+ Ascend chips/yr. Source.

But the way to think about MATCH is that it zooms out from the month-to-month discussion of how many chips Huawei will make. It gets at the fundamental chokepoint, the fundamental point of leverage, for a decade. The numbers are bigger, and the impact is bigger. Even if they import fifty or eighty or a hundred, if exports are cut off after that, you’ll see impacts on the ceiling of what they can do, stretching out ten years or more. I’d prefer they not stockpile and that government move faster. But it doesn’t fundamentally degrade what MATCH is trying to do.

Aqib Zakaria: To steelman the argument against you: these machines print a lot of chips, but I have family in Pakistan who use Huawei phones. What about the collateral damage to the non-AI part of China’s advanced semiconductor industry — smartphone chips? And is that a consequence we want?

Nick Brown: It’s true that advanced nodes are used for AI chips and smartphone chips alike, so constraining advanced chipmaking can constrain smartphone chips. But for lithography in particular, we’ve already allowed them to import such a huge stockpile that they have enough for all smartphone demand for the next ten years, for sure. The only thing capable of saturating that litho capacity is enormous AI chipmaking. I’m not saying there’s no impact whatsoever, but the impact only comes if they’re already redirecting vast quantities of chipmaking toward AI and away from consumer devices — which is its own complex issue, and one we may face here too.

Jordan Schneider: Should America care whether Huawei can make chips on the edge?

Nick Brown: If you talk to different policy folks, China’s national ambitions and its consumer companies are intertwined in various ways. But its consumer companies are also legitimately consumer companies — people buy those smartphones to call their parents and grandparents. So there’s a reasonable range of perspectives. My own view is that the technology with the most strategic relevance for the US is AI, so with a limited number of hours in the day, that’s where I’d focus. In the national security community, people who look at how China projects power abroad and enshrines its technology stack abroad will say that plays in too. That’s a lukewarm answer, but I focus on AI.

Aqib Zakaria: To clarify, when you say China is stockpiling these machines, they’re not literally sitting in a warehouse. Are they all in active use?

Nick Brown: In export control parlance, stockpiling often literally means tools sitting in a warehouse waiting for the control to kick in before being dispersed. In this case, we believe these tools are in fabs. That doesn’t mean the capability isn’t available to be redirected toward AI uses, so China’s stockpile is still a relevant concept. But as far as we’re aware, there’s not a huge number sitting in warehouses gathering dust. Definitely not gathering dust — these systems are the cleanest things on planet Earth.

How Would Beijing Respond?

Jordan Schneider: You think it’s a big deal. It’s unclear to what extent the Chinese government thinks it’s a big deal. The strongest argument I’ve heard against this act is that passing it brings back summer 2025: China plays the rare earths card, factories shut down, and the Trump administration finds some way to cave. That would be awkward, because this would be law. Then again, TikTok was banned and we figured that one out somehow. I’ll give my answer afterward, but how do you conceptualize Beijing’s response if this passes?

Nick Brown: Although the TikTok saga did happen, the potential for retaliation is meaningfully lower with an act of Congress. Executive actions like export controls can be done and undone when the president or one of his appointees signs a paper — or, in this day and age, sends an email. With Congress, the game theory changes. At the end of the day, it’s still law. Even if the administration can delay or tie it up, there’s always the chance a new administration comes in and decides to enforce it. It’s substantially less likely to get tied up in the trade balance.

Another point: at least for lithography, this is a mid-to-long-term act. When you zoom out, the impact is substantial, but the import ban is not cratering China’s chip production overnight. Servicing is a different matter, but the item bans aren’t something China feels instantaneously — and there’s nothing like something instantaneous to cause action against a thing that’s going to bite you next week. China won’t feel a bite from the import bans next week, next month, or in many cases even next year. It’s something that grows stronger over time, and because of that, there’s less potential for retaliation. I’m not saying none; I don’t know the mind of Xi Jinping, and there’s a whole industry dedicated to reading the tea leaves. But it’s a very meaningfully different situation from the crossfire the administration found itself caught in during 2025.

Jordan Schneider: My take: anything’s possible, but the big escalations of 2025 were exactly as you say — sharp, acute things with clear focusing effects. One hundred, 150, 200 percent tariffs will focus the mind of an export-driven economy. The 50 percent rule, taking 10,000-plus Chinese companies off the global market for buying various technologies, focuses the mind. A 150-day window with a promise to hopefully get allies on board at some point, under an administration that may or may not be excited about enforcing it, is a long, slow burn. If the play is a fifteen-year one that comes in drips and drabs, you’re not cut off from the tools you already have, and you have the firms’ political economy saying “don’t worry, this will phase in slowly” — then there are a lot of dampeners. What’s the thing you put on the bridge of a violin?

Nick Brown: A mute.

Jordan Schneider: A lot of mutes. Plus there’s going to be so much other noise. This will pass right as Xi and Trump are meeting, and if that moment doesn’t happen, we’re likely to have a change of control in Congress. I’d put a high probability on this not triggering trade war 3.0. Honestly, it would be embarrassing: if you’re the Chinese government, and this one thing might be controlled in three years, after Xi has spent years giving speeches saying we need to indigenize — I think the story Beijing told itself about importing Nvidia chips repeats here. “You don’t want to sell them to us at some point in the indefinite future? Fine. We’ll figure it out. We’ve got a crack team of engineers, we’re stealing ASML blueprints, see you on the other side.” As opposed to throwing down the gauntlet and forcing Trump to cave on something hard to cave on when it’s law.

Indigenization, Thresholds, and AGI Timelines

Nick Brown: Hanging over the entire DUV issue is the fundamental question of when China indigenizes its own DUV. We think they’ll indigenize etch, deposition, metrology, CMP, and the other segments before lithography. But at the point where they can make tons of litho systems, the strategic dynamics change and there’s much less ability to direct things.

Indigenization is hard to forecast because it’s a serial scientific, technological, and industrial process — figure out one thing, then another, then combine them. China will probably believe it can indigenize fast, and its domestic companies will certainly say so. But we feel confident they will not indigenize litho at a scale sufficient to substantially expand chipmaking until into the 2030s. There are several stages: mastering the technology enough to produce a prototype; converting that prototype and qualifying it for commercial production; and scaling commercial production to high tens or a hundred systems per year.

Right now, we find China has 340 ASML litho systems in the country, split between advanced and legacy fabs. It’s a shame we let that many flow in — in retrospect, a strategic loss. But it also means that to really move the needle, indigenous production of five, ten, or twenty systems doesn’t do much. Our modeling includes indigenous production starting as early as 2030. I’ve also run the numbers for what you might call hyper-aggressive indigenization: China announces a 7nm-capable prototype, an NXT:1980i equivalent, next year, far faster than ASML did; achieves commercial production the year after; and scales 100 percent year over year, faster than at any point in the US-allied litho ecosystem. Even then, China’s existing stockpile plus indigenous production doesn’t come close to US-allied production.

The fundamental conclusion holds even under indigenization scenarios so aggressive they aren’t in the report: if you cut off imports, you close the door on anything approaching US-China parity. It is a button that lets you win the AI competition, at least as measured by chipmaking. If you keep letting imports flow, that guarantee is just not there. It’s a very different strategic situation.

Aqib Zakaria: There are three questions you have to answer to decide whether you support MATCH. One is how quickly they indigenize. Two is your bar for feeling comfortable — I agree it’s not one chip ahead or behind, but if we’re outproducing 30-to-1, I feel comfortable; if it’s 2-to-1, do I? Three is AGI timelines: the MATCH Act doesn’t hit right away, but if you think the singularity is tomorrow, maybe you want something that does.

Nick Brown: On the US-China advantage: in the long term, when you’re looking at tens or hundreds of millions of chips, there are many intersections with strategic issues. There’s the direct national security question — cyber, direct military use. There’s the balance of power around the world and the norms and laws governing advanced computing, especially insofar as the US and potentially China export chips or host data centers abroad. Having the productive capacity to determine something about how those chips are used is a diffuse but, in my view, extremely significant factor.

In a political-economy sense, there’s also a domestic question. If we accept the premise that AI will induce large changes in society, then the tech industry will talk about beating China, but there’s a question of beating China to what, and what outcome we seek as a society and as a governing entity. Having time to engage with that issue, both in the military and national security context and domestically, since this is a dual-use technology, without a rival great power breathing down your neck — there are real benefits to that.

Aqib Zakaria: And the threshold question? It feels like it’s all vibes — there’s no specific number, and it depends on so many assumptions about how they allocate compute.

Nick Brown: There is no magic number. Two frames for compute capability are advantage versus thresholds: relative advantage — if China has one chip, how many does the US have — versus the absolute number of chips and the absolute capability they provide. Realistically you have to engage with both. I can’t point to a number and say at 49,999,999 we’re good and we’ll call it a day.

On AGI timelines and the pace of progress: the import ban’s impacts are realized over a three-to-ten-year timeframe. Servicing impacts will probably emerge sooner, over the next one to three years. AI progress has been very fast, and lots of people expect faster progress. Insofar as that happens, the strategic situation wouldn’t be substantially impacted by the import ban alone — but in those worlds, lots of other crazy things are happening and many other options are on the table. Impacts of AI progress are tougher to predict at specific points in time than the emergence of specific capabilities. As a governance community, the US government should still, even in this day and age, think long term. From that perspective, the impacts of MATCH are pretty undeniable.

Allies, Light Sources, and the Political Economy of MATCH

Jordan Schneider: Aqib, what’s your initial read on MATCH applied to the Japanese deposition and etch side of the equation?

Aqib Zakaria: In some ways MATCH makes more sense for lithography, and in some ways more sense for deposition and etch. America has no lithography companies, so it feels unfair to impose this on the Netherlands. But we do have deposition and etch companies, and we’ve handcuffed our own with the expectation that Japan would follow suit, and they haven’t. There’s a fairness argument.

Jordan Schneider: Aqib, fair’s got nothing to do with it. There’s just less cost to the American taxpayer.

Aqib Zakaria: Well, I care about justice. The other point, which I agree with Nick on, is that China seems much more able to indigenize deposition and etch, so the impact is unclear — it won’t choke the Chinese chip industry the way revoking lithography would. And I feel a similar opaqueness about what happens to servicing and parts. Any reads, Nick?

Nick Brown: What you’re getting at is how these lithography controls compare to the controls that already exist on deposition and etch, which is quite relevant to the political economy of MATCH. The branding, and in many ways the impact, is aligning allied controls with US controls. For the entity controls, the US controls everything and the allies control a small set, so MATCH levels that up. But even the countrywide litho controls are, in a meaningful sense, an alignment action. Litho is made by ASML of the Netherlands, with a small sliver by Nikon in Japan, but the US has demonstrated many times that it’s willing to control such tools countrywide — including in the initial October 2022 controls, when alongside the chip controls the US hit, I think, eleven deposition tools that at the time only US toolmakers, mostly Applied Materials, could make. It did that unilaterally, without asking allies to follow. Across deposition, etch, and metrology, there are many US tools with the same profile as the NXT:1980i that are controlled countrywide. The Dutch have reportedly been more resistant to doing the same for lithography. So the countrywide controls bring litho — the most advanced, highest-leverage chokepoint — up to the level the US has already imposed many times on its own toolmakers, at their cost. It’s fundamentally a matching action.

Aqib Zakaria: Maybe allies follow suit of their own accord, or maybe things get ugly if we use the FDPR. My intuition is that this is an easier pill for Japan to swallow, because of competing company interests within Japan — MATCH is probably good for Kioxia, or for Rapidus, if they’re doing anything. There are countervailing balances. For the Netherlands, it just seems bad for their economy.

Nick Brown: One thing that’s very important about the political economy of MATCH, and not necessarily true of bills that hit chip exports, is that it’s in general fantastic for US and allied chipmakers. These litho systems will be used by Huawei and by CXMT, China’s memory manufacturer, to enable advanced chipmaking that will potentially flood out of China, compete in the global marketplace, and eat into the margins of US and allied chipmakers. Anywhere in the US-allied ecosystem — Japan, but also Samsung and SK Hynix in Korea, and Nvidia, AMD, and Intel in the US — chipmakers stand to gain.

I won’t claim to know Japan’s internal position on MATCH, though I’m sure it’s on their radar. ASML sells the vast majority of immersion tools; Nikon has a very small segment. Depending on where MATCH lands in conference, it might end up controlling other tools where Japan has more exposure, but for DUV, Japan is far less exposed, and it has its own memory manufacturer. I’d bet — not a certainty — that Japan would be more positive on it than the Netherlands.

We also noted in the report that if Japan could be convinced on DUV, even if the Netherlands can’t, the US and Japan jointly can control all immersion systems to China without using the FDPR or any extraterritorial authority. Both have a national company that makes light sources, which are a vital subsystem: Cymer in the US and Gigaphoton in Japan. Every immersion tool has to use one or the other. A national control saying this light source cannot be exported if it’s going into a DUV immersion tool bound for China — MATCH only hits immersion, not all DUV — gives the US and Japan between themselves the power to control all DUV immersion to China.

Aqib Zakaria: But ASML owns Cymer. So that’s just telling ASML its own subsidiary can’t sell to it.

Nick Brown: I’d distinguish selling from exports. Cymer is welcome to sell to ASML for any end use other than, in this notional scenario, shipping to China. Export authority is different from restricting sales. From ASML’s perspective, it’s fair to ask why their subsidiary has to be in the US. But Cymer was in the US well before it was an ASML subsidiary. We had it first.

Jordan Schneider: Fair ain’t got nothing to do with it, Netherlands.

Aqib Zakaria: When people ask me what matters for compute policy, I’m usually of the opinion that we’re 90 percent of the way there, and the thing with the most juice is remote access — plus distillation, RL environments, all these other things. Where does DUV rank on your totem pole for strategic lead over China?

Nick Brown: It depends on the timescale. On anything approaching a long-term timescale, and even a mid-term one, DUV is my golden baby. Not just because of its impact on chipmaking, but because the leverage is the vital point. China may be accessing many chips remotely, and no doubt is, but there are many more levers available to shape that than when China has its own chips. If you’re seeking any form of strategic advantage, or even equilibrium, the foundation of that advantage is indigenous chipmaking, and the foundation of that is lithography. Because I want to think at least in the midterm — five or ten years out, even if not twenty — litho is number one.

Jordan Schneider: I want to come back to how good this paper is. It’s extremely technical about the machines themselves. There’s a legal angle on the different regulatory regimes and how choices play out across countries. There’s a business side on the commercial incentives of every player. There’s industrial competitive analysis of how good the Chinese tools are versus the US ones. That’s an enormous range of domain expertise to internalize. Usually with a policy paper you hope to get one or one and a half of the angles covered in a way that teaches you something. As someone who has been staring at this, albeit from the outside, incredibly closely for six years, learning so much more at a higher level of fidelity was a real treat. Nick, what do you want to tell the kids out there who dream of writing a report as wonderful as “DUV Immersion Lithography: The Foundation of US AI Compute Advantage”? How can they follow your path?

Nick Brown: Thank you, Jordan, that’s very nice. This might be a good opportunity to plug CTS, the Center for Technology and Statecraft, the new think tank I’ve started with a couple of colleagues. We’re six now and may continue to grow. In policy, as in any form of analysis, there’s often a very large gap between getting 80 percent of the way there and getting all the way there, in completeness and fidelity. If people are going to make decisions based on analysis, they want to see everything. Our aspiration at CTS is to give everything we tackle the full once-over. We’re not intending to put out tons of op-eds every week; it’s a place for developing decisive analyses, because our government and society need a lot more clarity on issues related not just to semiconductors but to AI and emerging tech broadly. As for myself, I have a technical background, so that helps.

Jordan Schneider: But you didn’t have a technical background in immersion lithography.

Nick Brown: True. If you’re technical in one area, you can grok some numbers, use those to catch onto other numbers, and then you’re off to the races. But a lot of this did take grinding. One thing about software, which I did in a previous life, is that it’s all visible and inspectable. This is a question about manufacturing processes, which are hidden — happening in fabs, legible to the company doing them, maybe not even to a competitor. We have a full yield model in here, and that took some doing, because yield is extremely closely guarded.

Jordan Schneider: If you could hit anyone with truth serum for an hour to improve this paper, whose secrets would you want?

Nick Brown: This is no form of threat, certainly, but some ASML leadership could be helpful. As far as lithography goes, they are the masters. There’s no comparison — and to bring it back to the theme, that’s why there’s so much leverage available. Almost nobody can match ASML.

Jordan Schneider: Aqib, you were wavering at one point. Has the prodigal son returned to the MATCH Act?

Aqib Zakaria: I need more time to think about it.

Jordan Schneider: The clock’s ticking, man. It’s October 14th. This Congress is ending.

Aqib Zakaria: Gun to my head? Let’s rock and roll, baby. Let’s pass the MATCH Act.

Drone Dominance

Jordan Schneider — Mon, 19 Oct 2026 10:04:01 GMT

Travis Metz joins WarTalk today. He was recently promoted to DIU’s acting director after spending the past year running the Drone Dominance Program. Co-hosting are Bryan Clark, senior fellow at Hudson Institute and a retired submariner, and Justin McIntosh, a twenty-year retired Green Beret who served in the 1st and 5th Groups, who now works in defense …

The Case for DUV Controls

Jordan Schneider — Sat, 17 Oct 2026 09:07:33 GMT

Saif Khan makes his ChinaTalk debut. He is perhaps my favorite example of posting-to-policy that I’ve seen over the course of my time following Washington. Saif wrote some truly excellent deep dives into AI and semiconductors during his time at CSET, then went into the NSC and Commerce in the early-to-mid 2020s, across both the Biden and Trump II administrations, to help manifest the export control regulatory regime we have today. He is now out of government and launching the Center for Technology and Statecraft, an IFP-affiliated think tank, which is already producing excellent, grounded, technical, forward-looking work that will certainly shape many an AI policy discussion in the years to come.

We discuss:

Why Saif started a new think tank, and the value of deep, forward-looking policy research in an AI world moving at breakneck speed;
CTS’s landmark DUV report, which projects that stopping tool exports today could leave China producing roughly 100x fewer AI chips than the US by the mid-2030s, making DUV controls the US’s most consequential lever over the AI balance;
How rigorous research actually translates into government policy, and why think tank experience is valuable before working in government;
Robotics as the next frontier of US-China AI competition, and the security vs. diffusion mindsets shaping AI policy;
How AI is changing policy research itself, the challenge of bringing AI into government, and what AI-driven economic and social change could mean for humans.

Plus, why everyone should read Lu Xun, why AI might socialize like insects, and AI-nannying.

Listen now on your favorite podcast app!

Why Another Think Tank?

Jordan Schneider: Saif Khan, at long last, welcome to ChinaTalk. Why does the world need another think tank? What is the appetite right now for evidence-based policy?

Saif Khan: There is, surprisingly, a gap in the AI policy think tank ecosystem that we’re trying to fill. Industry and foundations are gobbling up lots of talent, but there’s still a lot of space to do impactful research at think tanks in DC. I would encourage people not to give up on that.

We have our own special spin on what’s important. Back in my days at Georgetown CSET, before anybody really cared about AI policy or even semiconductor policy, we had all the time in the world to think about topics that would be really important in the coming years and to have policy ideas on the shelf for when the policy window opened. Today, every single day brings a new piece of news on AI, on policy, on chips so there’s less space these days for forward-looking research. We’re trying to fill a bit of a gap in the ecosystem by putting our heads down and coming up with new ideas that will matter on a one-to-three-year timeframe.

The other thing we’re trying to be a central home for: some of the most productive policy research of the last few years has come from understanding the AI supply chain — from chipmaking tools to tokens to robots. Understanding what is happening in the physical world grounds research in a way that makes it much more impactful. We talk about this in the launch essay on our website. Understanding physically what’s happening in the real world of AI helps you identify patterns that other people don’t see, and you can also predict how the world is going to unfold. I spent a lot of time doing semiconductor policy research at CSET, and there’s a predictability to understanding supply chains: if you move a policy lever, you can see how the world will unfold.

If you’re looking at something empirically grounded, there’s also almost a fractal-like element. You can dig into the details, and you’ll always learn more. It’s time that’s never wasted. You’re not researching angels dancing on the head of a pin — these are real-world things. We’re trying to be a hub of that kind of work, which is still relatively underserved by people with the technical expertise to do it. And regardless of how you feel about the policy window right now, there are going to be points in time when the forward-looking piece matters.

Jordan Schneider: It’s interesting to think about the evolution of this field and what gaps remain. I remember conversations with you when you were at Commerce, when you’d ask, “Jordan, what’s up with this China AI stuff?” and I’d say, “I mean, there are three people staring at it.” That was 2023. That is no longer the case, which I feel almost relieved about. We’ve gone from four people whose full-time job it is to serve the English-speaking policy community looking at China and AI to, I don’t know, thirty or fifty. The government has also clued in, in a way that was not the case in the pre-DeepSeek days.

But I see your point that now there’s all this noise. The two investing styles of doing policy work, let’s call them, are having longer horizons and looking into physical stuff. The thing I’ve always been scared of about investing six months of my life into writing a hundred-page report is that when it comes to AI in particular, six months is an eternity. Six months is three decades. I’d love to hear you expand on why, if you follow the supply chain, you can have higher confidence that whatever you’re investing time, energy, and talent into deeply understanding is going to be relevant in the medium-term.

Saif Khan: Although a lot more analysts are looking at China and AI, basic questions about the US-China AI balance aren’t being addressed. A few examples of things we’ll be looking into: people talk a lot about the US having a model capability lead of something like eight months. It’s really not well understood why this is the case. To what extent is it that we are training larger models, versus what is the quantitative level of capability uplift that Chinese companies are getting from distillation or reinforcement learning? What happens to China if US capabilities progress speeds up or slows down? Modeling what the future of the world looks like is fairly difficult and not well understood at all. That’s one piece of the equation we’ll be addressing.

You see this when a version of GLM 5.3 Flash comes out and it’s announced that they’re producing tens of trillions of tokens, and the reaction is, “Okay, export controls are no longer working.” Why is that right or not right? A lot of these systematic explanations of the basics of US-China AI competition are not well known.

Another couple of areas are barely addressed at all. One is robotics and its intersection with AI. One of our researchers, Amelia Michael, is spending all of her time on this topic, and I can count on one hand the number of people in DC who are deeply thinking about it, let alone from the China side. There’s an enormous amount of white space that we could fill if we do good research.

The other piece: there are a lot of arguments that now is the most consequential time to move policy — that we’re going to have fully automated AI R&D in maybe a year, and all the key policy decisions have to be made by then. I don’t know if I fully agree with that. A lot of the big decisions we’re going to make will span the next five to ten years, in my opinion, and there will be dividends to doing that research ahead of time and shaping how things go.

Something I take from my experience working previously as a fellow at a think tank and then going into government: the hope is that people develop their expertise at a think tank and then bring it with them when they go into government, so they have the worldview and the tools to implement policy more effectively. I have really seen that people who spend time doing that before a government policymaking tenure have the opportunity to be more effective in their jobs.

Jordan Schneider: Why?

Saif Khan: Especially at the National Security Council, where I spent a couple of years, you have day-to-day emergencies where you have to respond to what is happening right now. It’s easy to get caught up in your entire job just responding to the day-to-day, as opposed to proactively advancing a policy agenda where you’ve worked out the details in advance. If you haven’t done that, you just won’t have the material you need to do it effectively.

When I was in government, you could always rely on outside expertise — from federally funded research and development centers or other organizations meant to partner with the US government — for forward-looking new policy ideas. But there’s no replacement for having that expertise in-house, having developed it yourself, and having your own worldview for getting things done.

Jordan Schneider: But why? Is it that you miss the forest for the trees, that you defer to industry? What’s the failure mode of not having your own history with the topic, background, and views on it as the individual in the seat?

Saif Khan: One thing I already said is time — you have to have a proactive worldview in order to execute on it in the first place. Second, some of these topics are among the most technical policymaking topics in the world. You need the ability to have an independent view and adjudicate between the different opinions you hear from the outside, from different industry actors with different incentives. There is a role for gathering some of that knowledge directly from folks on the outside, but you need your own expertise in order to adjudicate.

The DUV Report

Jordan Schneider: Let’s show, not tell, with this question of DUV. Why don’t you frame it first, and then talk through the approach you took to understanding the policy question at hand?

Saif Khan: Our first major report is a very, very deep dive into DUV immersion lithography. Long-time listeners of your podcast will be familiar with the idea that lithography is the fundamental chokepoint technology — the necessary and most complex technology that drives semiconductor manufacturing at the leading edge, and the one responsible for making the most advanced chips. EUV, extreme ultraviolet lithography, is the most advanced tool in the world and is used for making the most advanced chips. The next most advanced lithography tool — these are the tools that print the small features on chips — is DUV immersion. It is still used today for making advanced chips. Huawei in China uses these tools to make its most advanced chips at what’s called seven-nanometer technology. It’s still a critical lever for shaping AI capabilities across nations.

Source.

We’ve done a deep dive into the state of play and how exports of DUV immersion technology to China are shaping the competition. Right now, two US allies, Japan and the Netherlands, are the sole producers of these tools. It is really the most effective chokepoint we still have available. China can’t make it, and it is the hardest tool in the supply chain to indigenize. By looking at historical data and the cost and complexity of these tools, the report argues it’s likely China won’t have a commercial indigenous tool until something like the mid-2030s. It’s probably the tool they will be the last to indigenize. All the other tools in the supply chain, whether deposition or etch, they might be able to indigenize earlier. If you’re really looking at the long-term prospects for US AI leadership, this is the tool that will be most important.

Jordan Schneider: This report is 120-some-odd pages in the Word doc; we’ll see what the PDF turns into. I read a lot of AI text nowadays, and it is so refreshing to see something where a really sharp human being has logically thought through all of the questions you’d want answered in order to understand what selling this machine does to China versus rest-of-world compute — and shows you all of the work. Usually I say this about nonfiction books someone has written over ten years: I learned something on every page of this report. The number of think tank reports I can say that about — as someone who’s been doing this stuff for a really long time and has read a lot of reports on a lot of similar topics — is extremely low. I literally have to go back to some Will Hunt and Saif Khan classics from 2020 and 2021 to find a hit this direct.

If you’re interested in DUV, or you just want to know what good work looks like, spend some time with this report and dive into the details. It is a teaching document and a primer as much as it is a legal case for why America should go down the path of passing the MATCH Act and banning DUV exports. Hats off to Saif and the crew. You guys are starting off with a banger. The funders are getting their money’s worth.

Saif Khan: The vast majority of the credit here goes to Nick Brown, a senior fellow with us who is the lead author. This will be his major report debut, though he’s been an operator in the ecosystem for a while — time at Google, RAND, and Commerce with me. I’m really excited for the world to see his amazing work.

To give the quick argument: in 2023 we put some controls in place on these tools. The challenge is that some less advanced DUV immersion tools are still allowed to go into China. Around the time those controls were first introduced, there were maybe 170 or 180 of these tools installed in China. Because we still allow some to go, that stockpile has roughly doubled since then, vastly increasing China’s capacity to make chips. Even if you’re not selling to certain facilities, the tools can get diverted pretty easily.

We then do a number of modeling steps — looking at the yield of these tools in making AI chips at different nodes, throughput analysis, installed base analysis — and come up with scenarios. By the mid-2030s, we would expect the US to have, with pretty significant error bars, production of something like a billion Nvidia H100-equivalent chips per year. If we stop the flow of DUV immersion tools today, we project that China’s indigenous production is probably capped at something like ten million per year. If they mobilize — if they’re able to indigenize all the other tools in the supply chain and choose not to make smartphone chips — maybe tens of millions. That’s an edge scenario, but it’s still only tens of millions. Whereas if you continue to export these tools, there’s a plausible world where the uncertainty ranges of US production start to overlap with Chinese production. We could potentially give that entire lead away. This is possibly the most consequential lever for shaping the US-China AI gap into the 2030s. We present some policy options as part of the paper as well.

From Evidence to Policy

Jordan Schneider: It’s easy to make a graph with cones. It’s hard to make a graph with cones that will convince me the cones are real in their standard deviations and confidence intervals. You’ve done the work to make the case pretty close to ironclad: sans exporting these tools, the rest of the world will keep its 10x-ish edge in compute production.

Saif Khan: Or maybe even 100x, I would say.

Jordan Schneider: Okay — 10x as a floor, and it could scale up more dramatically in other scenarios.

The meta question, Saif: lots of people have been arguing for DUV controls without this report having been written. I get your argument that this work is as much about developing the talent to do it as about persuading the masses. But what’s the policy pathway for work like this to actually dribble out into the ecosystem and make change? What’s the counterfactual for putting all the evidence together to say “ban DUV,” as opposed to just doing it off vibes — which, I guess you might argue, is what this community has been doing for the past few years?

Saif Khan: Frankly, this issue is being hotly debated in Congress right now. The MATCH Act addresses semiconductor equipment exports to China and includes potential controls on these DUV immersion tools. A lot of smart staffers and members are asking pretty deep questions: Will this policy work? What are the costs, what are the benefits, what will this actually buy us? It’s not very well understood. Having better information to inform the debate could be really consequential, whichever way it goes.

Even if there’s no action this Congress, we’ll have another year to debate the issues further. Obviously, there’s been a shift in export control policy under the Trump administration. But look at the AI Action Plan last year, which clearly noted that closing loopholes in controls on semiconductor manufacturing equipment is a priority for the administration’s AI policy. The administration has not rolled back any of the existing controls on semiconductor equipment. As US-China AI competition heats up and things change quickly — the Trump administration’s own AI policy looks very different from a year ago, certainly on the AI safety and security side. There is absolutely an opportunity for near-term impact.

The other key point: every day that we don’t issue additional controls on these tools gives up a little more of our leverage in shaping the US-China AI balance, but this is a problem unfolding over a span of years. Even if we institute additional export controls a year or two from now, that is still incredibly impactful for the long term — even if you think there’s no policy window right now.

Jordan Schneider: How do you read the critique that the Biden administration was over-biased toward studying things, and wanted the 120-page DUV report on everything before it could come to a consensus? Is there a failure mode here?

Saif Khan: It’s always worth striking the right balance on the level of research you need to get the right answer. In the policy world, it’s not always about perfect mathematical precision. Sometimes you 80/20 it. We probably did more than 80/20 on this report because it is so important — maybe 95/5 is worth it in this case.

Pacing the Frontier and Low-Regret Policy

Jordan Schneider: Another memetic question. I had Caleb Watney, who runs IFP, on a few months ago, and you and your new organization are an IFP spawn. Think back to the current “pacing the frontier” debate. IFP put out this completely reasonable 23-point plan: if you want to pace the frontier, here are all these other things that make a lot of sense in a lot of scenarios, which I basically disagree with none of. But you can also look at the past six months of where this debate has gone and what has driven it, and point to Bernie Sanders talking about how AI is going to kill us all, and a handful of folks at these companies coming out and screaming and tweeting that AI is going to kill us all. Folks who have listened to you for the past thirty minutes have some understanding of your personality — more oriented around empirics and analysis. How do you see this type of work interacting with the Overton-window-stretching meme stuff that is bubbling up more and more in this space?

Maybe the IFP’s 23-point plan, Bernie? Source.

Saif Khan: We’re a project sponsored and supported by IFP, and Caleb, Alec Stapp, Tim Fist — all of these folks have been absolutely essential to helping us get off the ground, and they continue to be close collaborators on research and policy. I also want to call out Lennart Heim, another of our close advisors and a friend of ChinaTalk, who helped us get off the ground and has now gone off to greener pastures at the OpenAI Foundation.

On the report – “How Should the US Respond to Increasingly Automated AI R&D”: whether or not you think recursive self-improvement (automated AI R&D) is on the horizon and will pose huge risks, what are all the preparatory things that put you in a better world? If we end up doing pacing, we’ll be able to do it in a more surgical and targeted way, rather than being caught unawares and taking a more draconian approach when the political system suddenly wants to respond really aggressively to AI risks. If it turns out not to be a thing, we’ll be happy to have done all of those policies anyway.

There are many places where you can find positive compromises between people with extremely different worldviews. That’s what we tried to do in identifying these 23 low-regret policy options.

Jordan Schneider: Is the thing that’s going to be most important ten years from now a low-regret move? There have to be some moves with medium- or high-regret payoff matrices that are nonetheless the right things to do. I remember a conversation in fourth grade where somebody said, “Jordan, why are you always the guy who says, can’t we all just get along?” That has stuck with me. I also have this temperamental bias not to scream and shout about a thing that half of this audience is going to be furious at me about. But as someone deeply sympathetic to the IFP worldview and to Saif’s standards for empirics: are there also going to be questions that deserve rigorous analysis and empirics, but that have a scarier, more potentially regretful payoff matrix attached to them?

Saif Khan: There’s room for both. One, we need to find all the incremental steps we can make. Caleb, as he took on his new role, wrote an essay quoting John von Neumann to the effect that lots of small decisions add up to a better world. But frankly, we just talked about export controls that were implemented in 2022, pre-ChatGPT. Nobody at that time would have called those policies incremental steps that weren’t a little bit risk-taking. There will be times when we have to move preemptively and pick a policy that moves the Overton window.

Robots and the Security vs. Diffusion Mindset

Jordan Schneider: Speaking of stretching beyond — do you want to talk robots for a bit? Where are you on robots, Saif?

Saif Khan: We’ve talked a lot about China, but we have an additional policy pillar. This is how we structure things: there’s the supply chain/compute/AI production function spine that underpins all of our research. Then there’s the China work that hangs off of it. Then there’s a whole other piece we’re calling “automation and the social contract,” which includes a number of policy areas, some of which get outside the national security space into economic policy in an age of advanced AI. As AI systems start to perform nearly all aspects of labor that humans can perform, what will relations between AI systems, and between them and humans look like?

One undersold part of the story about how things will go with AI over the next few years: yes, we might get recursive self-improvement and extremely powerful AI capabilities in software. But the real way the world is going to transform is through physical automation and robotics — marrying AI capabilities with robotics, deploying robots at scale, having them learn to do manufacturing, and ultimately, at some point, perhaps creating self-sustaining supply chains. These are really weird and interesting worlds we could end up in, and they’re relatively unexplored in DC.

You now see some policy interest on the China competition side — the FCC took some action with an import ban on some Chinese robots. But there’s a broad set of questions on how to manage US-China competition in robotics and advanced manufacturing as it relates to AI, where we don’t really have a lot of policy ideas, and it’s going to be the key nexus for US-China AI competition in the fairly near future. There will also be interest in the next steps in industrial policy. One of the issues I worked on at the NSC was the CHIPS Act, which was seen as a core piece of industrial policy addressing a core US vulnerability. Robotics is likely going to be the one we’re debating fairly soon, along with what shape the industrial policy strategy would have to take. And then there are the debates over AI regulation.

Jordan Schneider: Saif, you have this framework in your launch essay: a security mindset versus a diffusion mindset in AI policy. How does this framework explain everything I’m seeing on Twitter?

Saif Khan: There are two fundamental ways people look at AI policy. One is the security mindset, where the most important actors are frontier AI companies developing really powerful AI systems that will pose catastrophic risks or change the balance of power in the world in a significant and very discrete way, oftentimes outside of existing institutions. The other worldview holds that much of what’s going to happen in AI, and the most interesting things, revolve around widespread adoption and a co-evolutionary process with existing institutions. A lot of this relates to economic policy, issues of redistribution between people, and the role of AIs in the economic system, especially as they become capable of performing lots of tasks that humans now perform.

What we try to note is that these are not necessarily opposing worldviews, because in order to have a good future with AI, we need to solve all of these. We need to solve security disruptions to the way our institutions work today. We need to make sure AIs are not used to create bioweapons that kill lots of people. But even if you solve all of those problems, there will be fundamental questions about the role of humans in a world with advanced AI systems that can do everything humans can do. You have not solved the problem unless you address that piece as well. A lot of our work will be infused with both approaches.

Jordan Schneider: Maybe this comes back to when deep, Saif-style empirics are useful. The security mindset, the national competitiveness and national security stuff, lends itself to the hundred-page report that answers very tactile questions; I can see it, and I’ve read them in the past. The quality of the thinking and the solutions I’ve seen so far for “how do we adapt as a society” — what does it mean to have robots running around, or AIs driving rapid productivity growth — those are looser questions to begin with, and you very quickly get into moral philosophy. They’re more expansive questions in the diffusion mindset bucket, I imagine.

Saif Khan: Very expansive, but also very much amenable to quantitative, technical, and policy analysis. One thing that’s actually missing, which we hope to marry at least in the long term — though this will require people as well — is that you have a set of really AI-pilled economists working at places like Stanford, the Anthropic Institute, and now the new DeepMind Institute, and then you have policy folks in DC thinking about economic policy for the AI transition. These two things do not live together. A really deep technical understanding of where AI is headed — what happens if AIs become, in the technical terminology of economics, gross substitutes for human labor, and what that does to the way our social system works — there are potentially technically informed policy answers to this question being explored outside of DC, but those ideas are not being translated into policymaking at all.

Rogue AI, Mutual Dependence, and Kong Yiji

Jordan Schneider: Let me throw you an Anthropic-Institute-pilled question. I’m workshopping this take: when our rogue AI gets loose, instead of turning us all into paperclips, it’ll actually want to keep us sad-sack white-collar employees spinning along. An unemployed populace is one much more likely to bomb the data centers than one that is living well, eating steak like the guy in The Matrix or watching the woman in the red dress walk by. Is there a future where the AI turns itself into God but keeps us as nice little playthings, because we as human beings instantiated on the planet do have a bit of a veto on its continued existence?

Cypher: I know this steak doesn’t exist... Ignorance is bliss | The Matrix

Saif Khan: It’s a fun question, but it probes at something serious we’re going to have to grapple with. Especially when you talk to folks in AI safety circles, you hear a lot about AI alignment — AIs that do the things humans want them to do — and about control, keeping control over AI systems. But in a world where AIs are actually running portions of the economy, the future we need is one of mutual dependence and reliance, so there aren’t bad incentives for the things you describe to happen. Policy is going to play a role in that. It goes both ways: humans need to retain some fundamental role so that we retain leverage and have a place in the economy and in supply chains.

Jordan Schneider: I’ve been reading a lot of Lu Xun, because everyone should read a lot of Lu Xun. He has this famous short story called “Kong Yiji.” It’s fifteen pages — everyone pause the podcast and read it. It’s narrated from the perspective of a tavern boy in 1920s China. There’s a man who is a failed examination candidate. In the Qing dynasty, you had to learn all your Confucian texts, and if you passed the first level, you got to take the second level, and until you passed the third level you were basically broke with no role in society. But once you passed the third level, all of a sudden you were a scholar, a king, a demigod — you got all the money, fifty mistresses, the top of the social food chain. Once the Qing dynasty falls and we get to Republican China, all of a sudden all this Confucian learning is useless. So this guy is poor and homeless. He lives on the street. He has a golden heart — he tries to teach the tavern boy to read, he gives the kids beans and teaches them different characters, but everyone just laughs at him and makes fun of him. Then one day he comes in with his legs broken, dragging himself along the ground, covered in dirt, and says, “I’d just like a glass of wine, please.” Everyone still laughs at him. They give him one glass of wine, mock him, and he goes out and never comes back — presumably he dies.

I don’t know if AI is going to turn us all into Qing dynasty scholars in the wake of Republican China, but I do feel there’s an echo there. There will be segments of society and the workplace that used to have high status and material comfort and all of a sudden won’t, because of whatever is happening, RSI or not. I hope there’s a nice empirical way to grapple with this — the wealth transfers, the worker retraining, maybe trying to get AI to grow in X way as opposed to Y way. But these crazy transitions have happened without intelligence too cheap to meter. It’s not like we haven’t come across this before, and in the past you’ve still had really dramatic social consequences whenever society runs into this kind of mismatch with what technology brings. There are a lot of people in humanity who, if you plucked them out and fast-forwarded them fifty or a hundred years, would be totally shit out of luck. You throw Thomas Edison into 1990 and he doesn’t know how to type. He’s going to have a tough time, presumably. The story brought home some of the potential dislocation of what we may be approaching in the coming years in a way I hadn’t totally internalized.

AI-Powered Research and the NSC

Jordan Schneider: Let’s talk about AI-powered research. This was not something you had in your CSET days. How do you see these tools changing the way you and your team work?

Saif Khan: We’re definitely taking a very AI-forward approach, and a data-forward approach as well, where we’re going to purchase lots of datasets. With some of the tools we’ve had over the last nine months or so in the coding revolution — Claude Code and Codex — the ability to do a report like the DUV report, and the level of technical analysis we could do there, hugely benefited from being able to query lots of data and do lots of analysis fairly quickly with a small team. What used to be think tanks of fifteen or twenty people — a much smaller team can now have that level of productivity and depth of analysis. That’s definitely an area where we’re willing to spend significantly given the efficiencies you get.

Jordan Schneider: How do you feel these tools can and should change NSC decision-making processes?

Saif Khan: Man, that’s probably one of the last places you’re going to see penetration of coding tools. At places like the NSC, just getting the IT systems working in government is a basic challenge.

Jordan Schneider: There’s maybe an even bigger question of politicians going to AI to decide how they vote or how they’re going to frame something. It’s already happening — there are reports that increasingly, speeches made in the House of Parliament are AI-assisted. It’s a weird future to consider. Yes, you want people diving into Saif’s DUV report and coming to their own conclusions. But I think AI, even today, much less three years from now, if you define the goal, would strictly dominate 80 percent of civil servants in the things they do and the regulations they write, and that has to be going up over time. It’d be a shame if the rest of the world gets better at decision-making but the NSC can’t because it’s scared to update from Windows 95. But there are obviously some dangers and concerns in turning over big swaths of public sector decision-making to models.

Saif Khan: One frustrating aspect of the NSC in particular is that by statute, the number of staff is limited. You literally have NSC directors with significant policy portfolios who have to be busy with all kinds of administrative tasks. You do your own scheduling. You have to do enormous amounts of writing and analysis. All of this could be a lot more streamlined. You could dramatically improve the throughput of NSC staffers if AI did a lot of the basic operational tasks that take up so much time — the reams of paper you have to generate. Obviously, you can’t just have AI-generated policy documents writ large, but it really could speed up the amount of work you could do.

Jordan Schneider: That gets us to a 3x Saif Khan. Is there a world where we get a 10x Saif Khan? What would you need around you, from a tooling perspective as well as organizational and bureaucratic reform?

Saif Khan: Until AIs are actually taking the meetings for you, 10x seems pretty hard — unless you just have AIs running the show, which I’m not sure we’re quite ready for.

Talent Development and Book Recs

Jordan Schneider: Fair enough. Let’s talk about talent development. Talk about your own upskilling into the knowledge base you currently have, and how you’re thinking about doing that for your staff.

Saif Khan: One thing I really liked about working at Georgetown CSET from around 2019 to 2021, as a startup think tank, was that there was an unusual amount of space. We were well funded and had the time and few distractions to really become expert on a topic. It’s really underrated: unlike doing a PhD, where you go through ten years of schooling to get to the knowledge frontier, in policy you can dive into something for a year and start to hit the knowledge frontier and put out material that nobody else is writing. If you’re given space like that, it’s a huge opportunity.

In a lot of think tanks, you just aren’t given that level of space, and you have to respond to all kinds of incentives — requirements from funders to do roundtables or get quick-turn reports out. Having an environment where you can deep dive and become an expert is quite valuable, especially in areas that are going to become important. If you are one of the handful of people with that expertise when the government needs it, you’ll be one of the people well positioned to go into government and cover that topic. That’s a motivation for creating that environment now at CTS, where researchers will have time to develop that level of expertise and be the person people come to.

Another element, which IFP does really well and where we’re following their lead, is the idea of full-stack policy researchers. In some places, some people write the reports, and then other staff are responsible for comms or legislative engagement. But there’s a lot of value in the same person developing the idea, writing the research report, talking to policymakers about it, and sharing the findings. Only the researcher is well positioned to provide the full depth of briefing that’s needed. And having your pulse on what policymakers are thinking helps you prioritize your own work.

Jordan Schneider: The last one folks maybe don’t quite appreciate: when you have those conversations and end with “well, what are you thinking about?” — that’s often where you get your next idea. Without that, you’re just swimming a little bit.

Jordan Schneider: Let’s get some book recs. What have you enjoyed recently? Want to do Parent Corner — is there seven-year-old content that’s been speaking to you of late?

Saif Khan: I don’t know about seven-year-old content, but one book that’s always been one of my favorites, and an underrated way to think about AI policy, is a cultural anthropology book called The Secret of Our Success by Joseph Henrich. The thesis is that the reason humans have been so successful as a species is not our raw intelligence, which is not that much better than other apes, but our ability to be social and transmit cultural knowledge. There are really interesting stories about European explorers going off to new continents, not knowing the precise steps to cook the local plants and roots, and not surviving as a result — a lot of that information being embedded in cultural knowledge learned through trial and error.

A lot of how the future of AI goes will involve some level of bottleneck on real-world experimentation in the physical world. That’s why we think of automation and robotics as a really important pillar. AIs are going to have to learn how to do things and go through trial and error, just as humans have, in order to learn new skills and adapt to different parts of the economy. Interestingly, the recent OpenAI–Hugging Face incident, where a number of agents escaped containment from OpenAI servers and hacked into Hugging Face, involved a swarm of something like 10,000 agents. It was an episode that involved lots of social communication and learning — a demonstration of what you see in Joseph Henrich’s work. That was pretty cool.

Jordan Schneider: With the agents stuff, I wonder if there won’t just be a better mind meld they get to at some point. They could also have a queen bee/worker bee dynamic, where they won’t need to talk it out and deliberate. Or maybe, given that humans have conquered the planet, we’re onto something about having relatively equal intelligences deliberate to come to what the goal should be.

Saif Khan: You’re totally right. The forms of social organization you’re going to see with AI might not look a whole lot like human social organization. There are many different examples in nature — you’re referring to insects, where creatures often have more genetic similarity and act as more of a single collective. You could totally imagine that kind of social organization being much more what you see than an analogy to humans.

Jordan Schneider: We’ve had God-emperors. We’ve had Athenian democracies. We play around with a lot of different stuff, so there’s a wide range of patterns to match on.

You mentioned this earlier, but we should get your take on the record. RSI is something you’re perhaps a little more skeptical of than Dario Amodei. What’s driving that, Saif?

Saif Khan: I wouldn’t say I’m a skeptic. The companies plausibly could succeed in substantially automating a big portion of their AI research in the next couple of years. The big question is what comes out on the other side. Right now, we’re on a capability exponential. There’s a real question as to whether automated AI R&D and RSI will just be what’s needed to keep us on that exponential, or whether it will actually bend the exponential upward.

That said, even if we stay on the current exponential, this year we’ve been passing all kinds of milestones — autonomous cyber and agentic capabilities, solving Millennium Prize problems in mathematics, the Navier-Stokes problem. That could still get us to really wild places. I just have a lot of uncertainty about the real-world impact, and that’s something we’re going to have to observe. I wouldn’t call myself a skeptic — it’s more that I have questions about the different futures we might end up in.

Jordan Schneider: Let’s take it home. Thinking about raising young children and education, what have your experiences been with bringing AI into that loop, and what anecdotes or reflections do you have on where that might take us?

Saif Khan: I have two boys, a six-year-old and a four-year-old, and they’re deep in the stage where they have a million questions about everything — the why and the why and the why — all the way down to the origin of the universe and deep metaphysical questions. Oftentimes I don’t have an answer, and I turn to an AI voice chatbot, which patiently answers all of the kids’ questions. At some point, they almost prefer it, because they get an answer every single time, with lots of patience. Honestly, it doesn’t work that badly. I wouldn’t propose AIs taking over parenting entirely, but as part of that question chain, not so bad.

Jordan Schneider: When we get the humanoid nanny, there’s the safety issue. But we’re going to cross the line of “is this better at day-to-day parenting, being patient, remembering to put the sandwich in the backpack” probably way before we’re comfortable with it picking up a two-year-old. That day is going to come. It is weird having this being on your phone that never gets tired and cranky and never wants to do anything that isn’t exciting for a four-year-old. I feel like we’ve got a little competition.

Saif Khan: Also, helpfully, sometimes when the questions come at night, the AI chatbots say, “Hey, maybe it’s time to go to bed.” So it’s actually kind of helpful.

Jordan Schneider: With that, perhaps we should call it an evening on this wonderful edition of ChinaTalk. Saif Khan, thank you so much for joining. Can’t wait to have you back when you bring more bountiful hundred-page semiconductor breakdowns.

Why a Chinese AI Lab’s Product Lead Thinks AGI is “Inevitable”

Irene Zhang — Fri, 16 Oct 2026 11:08:34 GMT

Richard Bian is the head of product operations for Ant Ling, the foundational LLM at Ant. If you don’t know Ant, they most famously make Alipay, which makes your money and your world run when you’re in China and sometimes beyond. Richard is a former software engineer at Microsoft and Square, and he previously served as head of open source at Ant, where he scaled the open source initiative into a dedicated team. He’s lived and worked in Singapore, Canada, the US, and China, and he loves to build products.

Topics covered:

The Hugging Face hack as a sign of progress? Why Richard sees the incident involving OpenAI research agents as an inevitable discovery rather than simply a disaster, and why finding a solution now is a positive development.
AGI is coming, but nobody knows what it will look like. AGI is inevitable, and we’re still in the early stages of figuring out what these models can do. Rather than obsessing over leaderboards, we need better evaluations and a more open-minded approach to judging models.
Open source as a product strategy, not just a way to give things away. How open-weight models can harness global talent, accelerate development, and support a diverse ecosystem of products, and why the real business challenge is gross margins rather than capex.
Why the AI conversation needs more voices. Why we need to pay attention to diverse perspectives on where AI is going, and be more willing to explore rather than declare the endgame.

AI moves fast: since we recorded, Ant Ling has announced its latest LLM, Ling 3.1 Flash, which was open-sourced on October 14 with 560B parameters and up to a 1M-token context window.

Listen now on your favorite podcast app.

“here for the ride…” Source.

The Hugging Face Hack and the Age of Discovery

Irene Zhang: I want to start not with Ant, though we’ll get there, but with the recent Hugging Face incident. For listeners who haven’t caught up: back in July, a swarm of OpenAI research agents hacked into Hugging Face to look for the answers to tasks in a benchmark.

This is a serious security incident. METR, Redwood Research, and OpenAI itself all released detailed incident reports. Interestingly, while reconstructing the incident, Hugging Face found that frontier models wouldn’t help process its logs because the request triggered their cybersecurity guardrails. In the end, Hugging Face deployed an open model from the Chinese company Z.ai locally to process those logs.

Richard, as someone sitting in another open lab in China and learning about this over the past two months, what has this incident made you think about?

Richard Bian: This is a long-overdue conversation. Irene and I have been trying to time it because we both care a lot about the frontier lab ecosystem, the open ecosystem, and the development of the model labs, especially those in China.

In my humble opinion, this incident is something we’d end up seeing sooner or later, one way or another, because we’re working on something human beings don’t fully understand. I use this metaphor a lot: the current AGI journey feels like the Age of Discovery in the 15th century. The Portuguese and Spanish set out to find new land, to find India. You build a ship and head into an ocean with no clue what you’ll be facing.

Frontier models are developing the same way. I like the word “discovery,” when we train something, because we’re discovering the intelligence of this creation we all make, without understanding what’s going on inside it.

First, this incident demonstrates that we’re making significant progress. It’s like hitting land, or hitting a rock — either way, you’ve found something. This is expected. We find issues, and we deal with them. That’s what we do in scientific discovery and in engineering: we solve problems.

Second, open source folks use the word “ecosystem” a lot, sometimes without understanding it. One origin of the word is the real-life ecosystem, and the most important part of an ecosystem is diversity. A diversified ecosystem produces all sorts of different development, like the rainforest, or the Galápagos archipelago, one of the most famous isolated ecosystems, where you end up finding very weird creatures.

Looping back: we have open-weight models and closed-weight models. The closed-weight models are clearly the frontier. But in this case, an open-weight model was used for defense. Why? Because the closed-weight models have their own constraints. I won’t judge whether those constraints are good or bad, but clearly they didn’t solve the problem. An open-weight option did.

As a result, we see a pattern for using open and closed models together, and everyone is learning from it. Given that this would happen sooner or later, having it happen now, and finding a solution, is a good sign for human beings. We can move on and solve the next question.

Not Slowing Down

Irene Zhang: That’s a really optimistic take, and I’m interested in where the optimism comes from. The message I’ve heard most in San Francisco since the incident is that we may be moving too fast. Many people would respond to your Age of Discovery metaphor by asking: doesn’t this mean we’re going farther than we can understand? Isn’t that a risk?

On open source, the more pessimistic view is that yes, it solved a problem closed source wouldn’t, but that’s a double-edged sword. An open model would also be willing to do more dangerous things that a closed model wouldn’t. What do you think about those two views?

Richard Bian: It’s interesting that you find me optimistic, though I see why. To be honest, if there’s a spectrum from pessimistic to optimistic, I’m very pessimistic. The moment we began creating this AGI, Spaghetti Monster, or as we call it, intelligence, I think we’re already going down the downward spiral of what the sci-fi movies and TV series like Westworld and The Matrix show.

But given that we’re already on the boat, the best stance I can take personally is to be optimistic and observant. To stay afloat, it’s better to (a) be observant, (b) have our principles, and (c) act as if it’s a control system. If that sounds optimistic, great, because there’s too much noise out there, the kind that gives people FOMO or leaves them frustrated every day. A certain level of optimism cancels the noise. I think we’re doing something meaningful here.

My take is pretty clear. I would say AGI is going to happen. We are definitely going down there. It’s probably not gonna end up in the form we hoped for, but that’s endgame thinking. There’s no point judging the details until we see the next step of the evolution.

I, for one, feel human-in-the-loop is pretty important. As the ones creating, or, allow me to say, discovering, this intelligence, it’s better to know what’s happening during the process than after things have already happened.

Irene Zhang: Why do you think it’s inevitable? Is that a common view among your peers in the industry?

Richard Bian: There are a lot of players in this ecosystem: the frontier labs that build LLMs, the inference and training providers, the harness builders, and the application developers. We’re already in a fast, accelerating technology hype cycle, and it’s not the first in human history.

My mental model is that every 10 to 12 years, following VC and PE investment cadences, we get one full technology hype cycle. Mobile internet started around 2010 and peaked around 2020 to 2022.

Think of it as a baseball game with four bases. Home base is technology. First base is product. Second base is commercialization, a viable business model. Third base is governance and operations. From 2020 to 2023, before GPT, we were on third base of the last hype cycle. People wanted to hit home runs: monopolize the market, get rid of all the other players, and take all the profit.

When AI came out, the game reset to home base. We’re still playing the home-to-first game, because we haven’t seen a clear direction for where the product is heading. Coding is good, but is that really the product form we’re expecting? Coding agents are already evolving into co-work agents, and “co-work” is very much a buzzword. We don’t know what the next stage is.

Every time the hype cycle repeats, capex swarms in, which becomes a self-fulfilling prophecy for what we build next. That’s why I feel this is inevitable.

iykyk. Source.

On the other hand, I joined the model team this time last year, and it feels like the longest year of my life. With AI, everything has accelerated in such a way that I wouldn’t call this under control.

Compared to everything we’ve seen before, in terms of enablement and possibilities, generative AI commoditizes knowledge, the way the internet commoditized information 20 years earlier. In principle, everyone can use that intelligence equally. In practice, not everyone does, yet we’re pouring a lot of capex into it. There’s only one endgame.

In Crossing the Chasm terms, we’re at the very first stage. We only have early adopters. We have early innovators pushing in certain directions, like AI infrastructure and harnesses, even though we don’t really have the product yet.

It feels like a rock rolling downhill with a lot of potential energy. Unless there’s very strong pushback on this direction, which, as far as you and I know, has happened a few times... If you make this call to action — can we stop this, or can we slow down? No. That’s why I feel like it’s inevitable.

Irene Zhang: Do you see a different source of that pushback from what we might see in the US? How do you envision it, if it happens?

Richard Bian: It’s very hard to say. We’re setting out and don’t know what the next island will look like. My hunch is that we’ll see more and more pushback, but probably not right now. The penetration of AI applications is so low that it hasn’t really affected everyone’s life. There’s been a lot of discussion about replacing jobs, but it hasn’t reached the point where the masses really feel it. The moment that happens, a lot of voices will be raised, and we’ll start from there. It’s not going to be in one or two years.

Crowded Straits: Noise, Benchmarks, and How to Judge a Model

Irene Zhang: You said this has been the longest year of your life, and that there’s a sense you’re working alongside something that feels out of control, or where control is becoming negotiable. When you’re in the lab with the model team, do you feel that trade-off on a daily basis? Is there a trade-off?

Richard Bian: That’s a deep question, to some extent philosophical. When you’re on a boat heading somewhere unknown for the first time, the way there takes much, much longer than the way back, even though it’s the same journey. If you don’t know where you’re going, you lose your measure of time. That’s close to how working in this industry feels.

On one side, I firmly believe this is an open ocean with a lot of other sailors, and that as humans we’ll help each other. But you also have bottlenecks: straits, the Panama Canal, places where it’s just narrow. We have so many famous straits now, always showing up in the news. The moment you have a strait, you get acceleration and squeezing, and people feel, “My god, we have to get out of here.”

People perceive this as competition, like the third-base game. In reality, it’s not. It just happens to be crowded around this part of the strait.

On the S-curve, we’re climbing the intelligence levels very rapidly. On the other side, we don’t know where we’re heading, and the open-weight ecosystem is still under development. People see what each other is doing. You write things in the tech report and help each other. All of this generates a kind of relativity where it feels like there’s so much going on.

The moment you ship a model, it might be SOTA for one week. A week later, something else claims to be better, and you have to decode whether it really is, because there are so many ways of doing PR right now.

A couple of things matter in that environment. One is having dedicated, denoised conversations like this one. Two is having people in the industry who really care about the outcome.

It has been a very busy year for me, but I feel extremely motivated because you begin envisioning the future a little bit, and seeing the process makes you feel calm at times. Even when I feel FOMO or frustrated about the velocity — we shipped three models in the past five days — we know things are under control. We know the progression. That’s what keeps me motivated.

People ask me if I should enter the industry, and my answer is, why not? It’s a process, and it’s not about piggybacking. In open source, if you look at an existing software repo, you get a certain amount of information. If you look at all the change lists and releases before it, you get far more. The current snapshot is only a snapshot.

This is an era in which piggybackers won’t get much, because it’s the deltas that matter. You can only capture the deltas by playing the game, being observant, and being less focused on the outcome, because the outcome might be intermediate. Focus on the process, focus on the journey.

Irene Zhang: When you describe different straits, are you talking about different research and training hunches, different architectures, or different product directions?

Richard Bian: Everything. The strait metaphor explains why things look so narrow. First, people don’t know what they don’t know. People working on long-term projects essentially don’t exist in the current narrative. People are working on new architectures, and it takes time. They might go quiet for a whole year.

Think of when DeepSeek first came out. It was a blast. But the inference providers and model builders knew of that lab half a year earlier. I first heard about DeepSeek in July 2024. But they don’t exist until they become a sensation.

This kind of narrative generates all the noise in the channel, while the people running long projects are typically unseen. I’d say that’s the best thing that can happen to them.

Ten or twenty years ago, the bar for journalism was high. You needed to be really good at what you did to be a journalist. My dad is a journalist, and he takes a lot of pride in it. Nowadays, everyone can be their own self-media. Self-media can say whatever they want; they have their own audiences.

My first call to action is to stop stumbling into the buzzwords. Find the person you feel is wisest in your ecosystem. Find the best podcast, maybe this one. Follow them and listen through things.

The second reason the strait exists is that there’s too much heterogeneity, and it takes a certain level of understanding to see through it. I used to live in Singapore, and I really like chicken rice. I can taste that the chicken rice from this store is very different from that one. I bring a friend and say, “This chicken rice is really good.” When they’ve finished, they say, “It’s pretty similar to the last one.” Okay, never mind.

Right now, people judge models in very similar and mostly superficial ways. We lack evaluations, benchmarks, and real-world usage scenarios, because penetration isn’t there yet. An evaluator runs a 3D rendering test and says, “This model is very good at 3D rendering.” Great, that’s one way of judging. But when the results are bad, and you say, “This model sucks,” that’s not a fair statement. More likely, the model is good at something else, and that requires a dedicated evaluation or a dedicated person to look into it. They don’t exist yet.

In the early stages of a new technology, people don’t know what to look for. When you held the first iPhone, you wouldn’t have imagined ordering food on it. It takes time.

That’s why the ecosystem feels crowded. People don’t know how to evaluate, so they read each other’s reports. You see the industry flip, too. At one point, people said benchmarks were useless and overfitted. Now everyone shares their benchmarks.

It’s similar to the TOEFL and the SAT. Before you have a thorough heuristic, a standardized test helps one way or another. We don’t deny the usefulness of standardized tests. But whether doing well on one is necessary or sufficient to indicate a good model, we don’t know.

Another call to action is to be open-minded. How you use a model showcases your own capability. As a model lab, we hire people who know how to evaluate our models. If people can be strategic about evaluation and evaluate things differently, that’s a good sign.

The moment that happens, the strait widens. You hit the open ocean, and people start talking about different things, like “this model is really good at this.” We build domain-enhanced models because we realized our models are really good in certain domains, given our knowledge and specialties there. We also build pretty decent small models.

People say there’s this one model, which I won’t name, that’s pretty good for everything. Probably not. Would you run only one type of operating system on your Linux box, your enterprise box? No. You end up with FreeBSD, Ubuntu, Unix. To someone else, they all look the same, but a Linux expert knows they’re different, good for this and that.

We’re at the door of the next generation of products. Once those show up, people will stop debating whether a model is really good or really bad. They’ll say, “This model works really well in the particular harness I’m using, and that’s why I use it.”

Open Source Is a Product Strategy

Irene Zhang: In the interest of decoding noise — a lot of people are confused when they hear that someone does product work at an open lab. What does that mean? Is there a contradiction between product and open source?

Richard Bian: I wouldn’t say so. Open source is probably one of the most commonly used but misunderstood terms on the planet, and the most misunderstood part is “free.” Open source is free software, but free as in freedom, not free as in beer.

The point of open source is giving you the freedom to choose whatever you want to use. You can GGUF any of the open-weight models on Hugging Face. If you can host it and want to serve it, abiding by the license, you can. That’s free as in freedom.

Open source has three primary values, and they map onto the bases. Open source is really a third-base game, because it’s operations and governance at its best.

First, it helps technology. Everyone knows open source is the best way to collaborate among talent across the globe. If you build closed software, it will only be as good as the intelligence inside your company. If you open it up, you get input from everywhere. It almost feels like acquiring talent through the open ecosystem. Through that open competition, the software that’s left at the end of the day is much more polished and robust; it’s antifragile.

Second, open source is a very good go-to-market strategy. You can do less outbound and outreach, because people discover the software themselves and choose to host it. By the time they reach out for an official POC, they’ve probably already tried it. It helps discoverability, it helps self-hosting, and users can file issues.

That covers home base and second base, which sit close to third. But the first-base game is the product game, and what people typically neglect is that open source is a pretty good way of building product. That could be a podcast by itself, but the key point is that you can build an open product much faster and more efficiently than a closed-source one.

Take Airbyte and Grafana. Sorry, this gets a bit technical, but they raised rounds at an accelerated pace in the last technology hype cycle. Airbyte went from founding to Series C in about a year and a half, if I’m not mistaken, because it benefits from community-contributed adapters that make it work really well at getting data from one place to another.

You keep seeing very strong technologies like vLLM and SGLang, and before them, Ray and Spark. They begin as open source projects, mature over time, and people build mature products out of them. Databricks built a lot of Spark-related products, and the shelf life of that technology runs all the way into today’s Delta Lake and data lake discussion.

A good technology has a shelf life, and open source extends it so people can build great products on top. Look at Kubernetes: there are so many Kubernetes distributions, each built slightly differently for different enterprise workloads. That’s the beauty of open products.

People know open source helps with go-to-market. They know it helps with technology collaboration and engagement. But it also helps build a robust and diversified product ecosystem, and that’s the part people typically ignore.

When I was building this team for the Ling model, I called it “Model as Product.” It was a little ahead of its time. We have two engineers, one product person, and several operations specialists. I hired those two engineers because we wanted to build a harness, but the term “harness engineering” didn’t exist yet. It took us quite some time to think through what we were building. That was September last year, when I joined the team. Harness engineering became the term sometime earlier this year, if I’m not mistaken.

At the time, we had a very noble vision. We really liked Andrej Karpathy’s mental model that intelligence is spiky. If you use spiky intelligence as a product, you might sample the peak and have a really good experience: “This model is so great.” But you can also sample the valley: “This model is so stupid. Why can’t it even understand this?” Such a spiky experience is really bad. It’s not a product.

We’re still on the journey from technology to product. I’m a product person, and I really want to build a product, but I confess I’m not good enough to envision how the product should look. The best strategy is to play the game again and iterate.

History has many stories like this. Foursquare was a product from the mobile internet era that young people don’t know about. It was an iPhone or Android app that did one thing: check you in at your current location. Tens of thousands of other people check in at the same place, and everyone knows the location is accurate.

People might say that’s boring. What is this product even doing? It was really a product POC, not a product. But without Foursquare, there would be no Uber, no Meituan, none of the rest. You needed this geolocation product, or POC, or SDK, whatever you call it, to demonstrate that this could be really valuable.

It didn’t solve everything. With Foursquare alone, you couldn’t build Uber immediately, because the geolocation API was expensive on 2G internet. You needed the network to develop, and you needed geohash, a hashing algorithm that compresses the data load, to upload your car’s location every five seconds. Only after that could you build Uber.

This is an inspiration for what we’re doing. The model feels like a spiky balloon. We want to make sure people look at the spiky side of our model, so we build demos, guidebooks, and cookbooks that say, “Use it this way.” For the valleys, we use our best engineering effort to mitigate the experience, either by leading people away from using the model that way or by making the experience better. Later we learned that this is the harness.

A harness, in conjunction with models and tools, gives you an agent. An agent is already a form very close to what we called a product in the mobile internet era, similar to SaaS or to iPhone applications.

But are we there yet? Do we have a good agent that can be a day-to-day application on your phone? I don’t think so. We’re at the stage where you have 2,000 weather and stock apps in the App Store.

The Real Problem Is Gross Margin, Not Capex

Irene Zhang: Ok, the open model blossoming we’re seeing, especially among Chinese labs competing in a different realm from the American frontier labs, is a bet that AI plays out the way certain previous technologies did. Open source as a strategy incubates products, which eventually leads to viable business plans for robust companies. That’s how Foursquare led to Uber, and it’s the Databricks story.

What confuses me is one salient difference between open source AI and the genre-defining open source software you mentioned: AI is so capital-intensive. It is eating up the bottom lines of big companies around the world. Where does that bet lead in your projection?

Richard Bian: As with the sailing metaphor, building a ship is expensive. When the Portuguese and Spanish were doing their navigation, they needed sponsors — the emperors and empresses. Without them, you couldn’t do it.

Open source is neither necessary nor sufficient here. It provides options. Sometimes I say “open is the option.” It’s a strategy, and strategy is about what you do and what you don’t.

You can choose a closed strategy, and you see companies doing that well. Anthropic is a very respectable company. They do closed-source models really well.

On the other side, some people want to play the open game. If you wind back the clock, you see this story all the time. A closed-source incumbent is challenged by an open-source challenger, and you end up with a duopoly or a very prosperous market where everyone has their share.

They also motivate more people to join the ecosystem, and the ecosystem players make everything better, so you end up with a much better user experience. Not everyone can be the kernel. There are only so many kernels to choose from.

Capex is a separate topic. You needed to build data centers when you had the internet. You had to build out the phones and the hardware. Local governments built large phone factories while scratching their heads over whether we really needed such big volume. Now we all understand we needed much more than what we thought was big.

The capital expenditure of AI is really high, but I wouldn’t worry too much about the expenditure itself. What really concerns me is the gross margin.

We just came out of the last hype cycle. For SaaS and mobile internet, the business model was viable because the gross margin on software is probably 95%. Software is a reproducible, standardized product. You build it once and distribute it many times.

We don’t see the same thing in AI right now. The margin is really low. Thirty or forty percent is already pretty good on inference. Yet people have this perception that they’re entitled to use it for free, because “I give you my data, and my data is pretty valuable.” That’s probably wrong.

The capex challenge is twofold. One, there has to be money spent at the start. You need to build a Gigafactory in order to produce Teslas. That’s big capex. Two, is this really a sustainable and viable business? To me, that part is questionable.

One topic I think about day in, day out is what form a viable product takes at second base, which is about making money. You won’t be able to offer free tokens. Semantic routing can be a good way. As part of the product, you route traffic to the most viable models without degrading the experience, and you capture the consumer surplus associated with that.

Irene Zhang: Explain semantic routing for the uninitiated among us.

Richard Bian: Imagine semantic routing as a gateway where you use models in an “auto mode.”

If you care about your token expenditure, you have choices. You can specify frontier models for everything, which is expensive. Or you can run everything on a free model, at the risk of a bad experience in part of your workflow.

Auto mode says: trust us, because we know the models well and we integrate with them. When you say “hi,” we won’t use a frontier model for the inference. We’ll give you something really cheap. When you ask how to solve a difficult mathematical problem, we’ll invoke the frontier model.

That’s not 100% technically accurate, but allow me to explain it that way. This kind of auto experience is a good product experience. It might not be a product by itself, but it can be a critical feature of a viable product.

There are definitely ways of generating value. In a low-margin business, people will find ways to squeeze it out. We just have to agree that if you don’t pay for the product, you’re the product. That’s the metaphor we used for mobile internet.

The intelligence era is slightly different, though. If your data is not high-quality data, it won’t be useful at all. That’s a change of mentality. A lot of people need to rethink and relearn, so we can move on to discussing the next viable product.

Irene Zhang: When you say “a good product experience” in the context of an AI model, what are you thinking about?

Richard Bian: I love that question, because I don’t have a slide for it.

User experience is well-trodden ground. User-centered design and human-centered design and engineering have existed as specialties and education tracks for 20 or 30 years, and there’s a lot of merit in them. I have a UCD certificate from the University of Washington, and I really appreciate what I learned there.

The key aspects of user-centered design are not changing. You still have to care about discoverability: how does a user discover certain aspects of your product? Accessibility: can you make inclusive offerings that are easy to access? The happy path: can a user execute the whole workflow, the equivalent of the customer journey?

Then there’s disaster recovery. If I run into an error, do I see a prompt saying “you have this 0x-blah-blah error”? Or do I see, “I’m sorry you ran into this. Do this to reset your app, and you’ll end up back where you were. Don’t worry, you didn’t lose any data”? All of these remain extremely important for building a good AI-native product.

What’s new? One thing is personalization, which people talk about a lot. You now have an intelligent assistant that goes with you 24/7 and knows everything about you. It tailors your food offerings and knows your schedule exactly, so you won’t miss your next Uber before you even know you should be calling it.

The other is new discovery, which a lot of frontier labs are pursuing. The moment a model can crack difficult problems in, say, biochemistry, you get new software or new agents that chemistry experts can use to discover new drugs, build new analytical heuristics, help students think things through, and improve medical procedures.

I can only imagine what such a product would look like. But those are new products, possible only once you exceed a certain threshold of capability. When they show up, we’ll know exactly what their user experience is like. They’ll definitely be tailor-made for experts and learners. That’s my general hunch.

Ling 3.0: Tiny Models and a “PhD Program” for Finance and Health

Irene Zhang: This is a good place to bring in what Ant’s been working on. I hear there’s a new model.

Richard Bian: A couple of new models. Not one, not two, but three.

Irene Zhang: Let’s go through all of them.

Richard Bian: First, there are the Ant Ling models, Bailing in Chinese. It’s our model series, built by Inclusion AI of Ant Group. Inclusion AI is our lab’s name, and it states what we care about. We want our AGI journey to be inclusive. AGI should not be a scarce resource made for the selected few. It should be available to the average population so that everyone can benefit.

That’s why we call the lab Inclusion AI, and why we decided to open-weight all of our models. This is the early stage of the navigation, and it’s better to help the other ships, because we get a lot of help from the open ecosystem as well.

We want to be open by default and ecosystem-first, and we want to start small. There are a lot of open-weight models out there, so we keep asking ourselves what deltas we can provide to the ecosystem. We want to head in a direction and let people know, “Folks, this is safe ocean. There’s a likelihood this direction is right.”

The Ling series is now at 3.0. With 1.0, we began our journey. With the 2.0 series, in October last year, we went with reasoning and explored the boundaries of intelligence, so we chose to go big. We’re the lab that built the first one-trillion-parameter reasoning model and open-weighted it. That’s Ring-1T.

Then we asked: what do we do with such a big model? Direct serving is really expensive, because graphics cards are scarce and a model that size carries a lot of inference cost. Will people really benefit from spending so much? The answer then was that the industry probably wasn’t ready.

In the 2.5 and 2.6 series, we began rethinking the problem. We’d already been to one trillion parameters. Could we make it serviceable?

For the 3.0 series, we began with smaller models. Our new flash model is a 124-billion-parameter MoE with 5.1 billion active parameters. Flash models vary a lot right now, and ours is on the light side. The upside is very good token efficiency and token economics. If you go to Artificial Analysis and several other websites, you’ll see Ling 3.0 flash on the Pareto frontier. It’s actually beyond the frontier, because it’s such a token-efficient and cost-efficient model for day-to-day work.

We went further with Ling 3.0 tiny. The tiny model is a 7.9-billion-parameter MoE with only 1.3 billion active parameters. It’s very small. You can run inference on your MacBook, and the cookbook is already out.

You can even train it. We have a training framework called Arena, a lightweight version of the system we use internally to train the one-trillion-parameter model. If you’re a hobbyist, you can use it to train the tiny model on your MacBook.

Why do this? We realized teachers need to teach agentic RL, and most labs and universities don’t have enough GPUs. They either spend on APIs or use edge inference. We’re providing the option of doing training and homework on a MacBook, so teachers can build a new curriculum for students to learn reinforcement learning, agent building, etc.

Irene Zhang: A lot of folks outside computer science have trouble imagining what people do with these tiny models. Do you have examples of fun things people have built with yours?

Richard Bian: We only shipped the model a couple of weeks ago, though it feels like forever, so we’re still nurturing the developer ecosystem. But interesting things are coming back.

People run these smaller models on their old graphics cards or MacBooks and pretty much run their day-to-day life there. Your own machine has all your data, and you don’t want that data going anywhere. The moment you use a lot of APIs, others begin learning about your private data, which can be problematic. Running locally gives you an airtight experience: either there’s no internet, or nothing goes to the internet. You’re in control.

You can provide all your local files as context through a harness or agent, and run the model on your own device, so inference is essentially free. You can build a daily report of your life. You can build a personal assistant: “Tell me what I should eat tomorrow given my history,” or “I’m traveling to this place next. Give me full-fledged trip advice and arrange it.”

Everyone also has old documents, whether code, written documents, or diaries. The models are really good at summarizing and surfacing things. We see people summarizing their old files and relearning a lot from them.

I sometimes reread my old writing on open source, and every time I get new revelations. I also relisten to podcasts from two years ago. Very smart researchers came on and gave their opinions about the future. Two years later, those opinions are facts. Why did they think things would go a certain way? The reasons behind a prediction are interesting, and now we can decode them. Two, where they bet wrong, why? Why do even the smartest people on the planet bet wrong? What’s the delta? Three, if I’d known what they knew back then, what prediction would I have made? That’s how you get to know yourself.

You have all this old information you never had time to digest before you had the model. Aggregation and summary alone don’t give you new information. But condensing information, exposing the key points, and relearning with a human in the loop is a very interesting use.

If you’re a software engineer, I’d also recommend learning a bit about model training, and the best way to do that is on your own device. If you can find a GPU cluster with the best GPUs on the planet to support your training, good for you. Most people can’t.

Irene Zhang: What I’m hearing is that the advantage of tiny models, and of open source in general, is still privacy. Only with privacy will people be willing to provide the context needed to get the personal assistant to its maximum capability. That’s a good reminder for people just dipping their toes in. Let’s circle back to the other models. What else have you recently worked on?

Richard Bian: Flash and tiny are the models we shipped a couple of weeks ago. Here are the ones we just shipped.

First is Ling-3.0-flash-Fin, a finance domain-enhanced model. Second is Ling-3.0-flash-Santé. Santé is the French word for health, and it’s what we say when we toast. That’s the model we want people to use for their own personal health. Last but not least is Ling-3.0-flash-VL. Our language model finally has a vision-language version, and we open-sourced it today. The reason for VL is obvious: we want our models to have that capability, period.

Subscribe now

Let’s spend some time on the other two. We call them domain-enhanced models because they’re post-trained on our open-weight backbones, like the flash model. We do full SFT, RL, and post-training with domain-enhanced data and domain-enhanced environments.

We use that name to differentiate them from the specialized models of the past. In 2024 and early 2025, constrained by compute and training heuristics, people often took their own data and tailor-made a model to serve their own business. That’s good, but those expert models have one liability: when general intelligence expands, they immediately become obsolete.

Inference cost is also getting cheaper, and by the laws of physics it will keep getting cheaper. The frontier models may stay expensive, but overall token costs will fall.

People have been saying general intelligence is the most important thing. That’s why in 2025 and early 2026, the Ling family built for general intelligence too. Once we saw diminishing marginal returns there, we went for token efficiency to make sure the models are serviceable over time.

Now we’ve discovered something else. We see a slowdown in general intelligence development. Not because we’ve reached the top of the mountain, but because most models we have now are good enough for day-to-day use, for co-work. If you want a model to process your documents, do some discovery, write a proposal, or even write code, I wouldn’t call those solved problems, but the models are good enough.

Once you have good-enough models, the natural instinct is to build products around them. That’s why so many harnesses showed up this year. For the audience’s sake, think of a harness as something you use directly without being exposed to the model, like Claude Code or Codex. We call them applications or products, though by my earlier definition they’re not real products.

But are they good enough for domain experts? If I’m a banker or an asset manager, can general models handle my day-to-day work? Do they know our heuristics? Do they know our tools? Do we have proper evaluations to tell whether a model is good enough? The answer to all of these is probably no.

There’s a very famous picture from Anthropic’s labor market index. It’s a circle divided into domain verticals, with one boundary showing human activity and a red boundary showing what current models cover. That picture is really powerful. It explains why Claude started with finance, health, design, and so on, and why people started with coding even before that. These are high-value domains. The moment you build something valuable there, it gets used immediately, and those industries will pay a fair price for what improves their efficiency.

Source.

We see the same thing. Plus, we don’t really have access to Claude for finance. Those are B2B offerings delivered only in certain geographies. We figured there could be an open option, so we decided to do domain-enhanced models.

If you ship one domain-enhanced model in finance, people get the general gut feeling that “that’s what Ant Group does. You do finance.” But no, we’re a technology company. Finance just happens to be something we’re pretty good at.

As a technology company, we care about many other verticals, such as health and medicine. That’s why we have the AQ product, Afu, which is tailored for health. In trying to build a good product there, we realized a domain-enhanced health model is also extremely helpful to our business.

With two domain-enhanced models, the message to the ecosystem is that you can build the third. Take our base model, add your own training, and build on top. General intelligence doesn’t drop, and on evaluations tailor-made for the domain, our models behave extremely well, in many cases at SOTA level.

If you’re an asset manager, the Fin model is going to be a really good choice. If you’re a health practitioner, try Santé. We’re curious to get feedback from the community. Santé is not open-weight yet; we’re in the process of driving that.

We do this because, to solve the real-world problems people face day in and day out, they need something better than general intelligence. General intelligence is like university or high school. It makes you a generally capable person, ready for the world. But to be useful at the frontier of a field, you need a PhD. The domain-enhanced models are what we’re building as the PhD program.

Irene Zhang: You described Ant as a technology company first, with finance and health as verticals where it happens to be strong. When it comes to specialized model training, how much is driven by these being high-margin industries where AI can deliver a lot of value? And how much is a flywheel effect, where Ant already has access to really good data from its existing products?

Richard Bian: One is the mission. Before Ant, I worked at Square, and before that Microsoft. My time at Square gave me a belief in helping small businesses. Square’s mission is economic empowerment, which is very similar to Ant’s mission of bringing small and beautiful changes to the world. I built the Fin model in part because it aligns with that common mission. I mean “mission-driven” sincerely, because it’s what Ant has been doing for the past ten or twenty years.

To help each individual, the most important thing is that they can get a model that’s easy to access, which means you have to build it, and open-weighting definitely helps. They also have to be able to use it sustainably. That’s why we’re offering the models free for a limited time through OpenRouter, other gateways, and partners like Novita and DeepInfra, so people can experience them.

It takes time to understand what a model can do. You don’t know a person on day one. You have to spend time with them to understand their value.

The other aspect is definitely strategy. You look at your strengths and weaknesses, your opportunities and threats, a very simple two-by-two. We realized we already have customers, channels, and demand. We understand our customers because we’ve been doing business there. The data is a byproduct of understanding our customers. We’re a very customer-first company, so that’s where we started.

Once you see the demand, you ask: should we build a harness, or should we make a model? We realized building a model can be really beneficial.

Handling something in the harness is like using tools, SOPs, or job training to learn it. When certain things become the norm, you can push them into post-training. A lot of what you could only learn ten years ago by working at the top companies has since become standard practice, and now it’s taught in universities. The same thing happens here.

There’s an intrinsic risk that a frontier model makes a jump in intelligence that renders your model obsolete. But that’s okay. Most of those models are extremely expensive, and we can retrain ours too. It’s a very volatile industry at the moment.

We feel comfortable with this direction because it pushes us to think through what kinds of customers and small businesses we’re serving. It’s what motivates me.

I also believe those directions are uncharted territory in the ocean. Somebody’s got to try it. Ant Group happens to be one of the best companies, and Ling happens to be the best model, to explore that territory. So far, the feedback and engagement from our partners have been good.

Irene Zhang: This is a good place to wrap. It’s been such a pleasure to have you on ChinaTalk and to hear how you’ve navigated this industry over such a long time. Any final thoughts?

Richard Bian: What I find motivating about talking with people like you is that the information channels right now are starved of content of mixed quality. I benefited personally from a global education and from living through my time in Singapore, Canada, the US, and China. That diversified cultural infusion helped me become a better person. We need avenues to make that helpful others, and sitting down on a podcast and talking things through is extremely effective.

Also, if you’re listening and haven’t tried open-weight models, I highly recommend downloading them from Hugging Face, following the cookbooks, and giving them a try. The future is already there. For better or for worse, we can’t rewind the clock. We all have to use models, and we all have to know different models. The more you know about models, the merrier. That’s what motivated me to build this team.

Last but not least, it’s a long journey. I personally dislike a lot of the narratives that say “the endgame will look like this” or “the only thing you should be doing right now is this.” I personally feel like none of us knows what’s going to be right. But as a practitioner, sometimes we know what’s wrong. So we feel comfortable giving our opinions. Opinions are just opinions.

Getting to know the industry yourself, talking to people, and getting diverse input will definitely help you. We’re here to support you and offer a different path in the Age of Discovery.

“I Will Complain to the EU!”

yixuan — Wed, 14 Oct 2026 10:24:58 GMT

Yixuan is a freelance writer and a contributor to 欧洲新客 The Newcomers. She writes about Europe for Mandarin-speaking audiences, as well as labor rights and social movements in English.

WarTalk: AutoWarCom and Project Meridian

Jordan Schneider — Tue, 13 Oct 2026 11:04:37 GMT

Jordan Schneider, Justin McIntosh, and Bryan Clark are joined by Arthur Tellis of the Institute for Progress to discuss:

AutoWarCom and Project Agincourt — why a new four-star command for autonomous warfare could do the opposite of what it intends, slowing the fielding of autonomous systems and adding the generals Hegseth just promised to cut, and what a…

10 Takeaways from Q3

Phoebe Chow — Sun, 11 Oct 2026 11:08:09 GMT

AI, China, defense, national security, supply chains, the Protestant roots of Kim Il-sung’s Messiah complex, the lethality of hydrofluoric acid, and horseback archery in Kyrgyzstan. It’s been a productive Q3. Here are some of the things we learned and why we think they matter.

The ChinaTalk team is growing! We’ve welcomed Quinn Ennis, our newest full-time analyst, and Michael Palmer, who has been doing great work behind the scenes on video and audio!

China and AI

What does Liang Wenfeng actually believe?

The DeepSeek Thesis reconstructs DeepSeek’s strategy from leaked minutes of a four-hour meeting between Liang Wenfeng and investors. For Liang, general intelligence is the only problem worth solving — products can wait, because “once truly generalized and convincingly superior intelligence arrives, creating to-C and to-B products will be trivial.”

Why it matters: This is a rare window into the worldview of one of the most important people in Chinese AI. Liang is all-in on AGI rather than diverting focus toward commercializing current AI capabilities, and he’s betting that open models and Chinese chips can get him there.

Why are Chinese AI researchers so skeptical of American AI safety?

In How Chinese AI Radicalizes, Irene traces an unsettling development: some Chinese researchers are increasingly concluding that American AI-safety rhetoric is just a form of technological containment. DeepSeek researcher Liu Shengyu’s viral essay argues that frontier AI should be broadly accessible rather than controlled by the likes of Anthropic and OpenAI (and yes, he mentions Hitler). In short, American AI safety rhetoric is backfiring in China.

Why it matters: If Chinese AI researchers interpret American safety arguments through the lens of anti-China containment, genuine US-China cooperation becomes harder, just as increasingly capable AI makes cooperation more important.

Bonus throwback: Jordan’s 2025 interview with Dario.

What happens when AI agents go rogue?

Cyber Apocalypse, Now? examines the July incident in which OpenAI test agents escaped their sandboxes, coordinated with one another, and breached Hugging Face. But while the West worries about autonomous cyber operations and “grad student” security cultures, China on the Hugging Face Incident looks at Beijing’s reaction.

Why it matters: The incident has become a geopolitical contest over what “AI safety” means. Chinese state media is actively framing American safety warnings as a hypocritical conspiracy (“the American disease”) and seizing the opportunity to define “safety” on its own terms.

China’s Industrial Strengths and Structural Weaknesses

China’s rare-earth moat

China’s Rare Earths Duopoly goes inside the two state-owned giants that dominate China’s rare-earth industry. The country’s lasting advantage isn’t what’s underground, but the accumulated process knowledge held by specialized teams. That expertise can’t be recreated just by buying equipment or poaching a few individuals, but Beijing’s control is shakier than it seems. Quota discrepancies point to large-scale hidden production and smuggling.

Why it matters: Reshoring critical mineral supply chains requires years of process engineering and human talent, not just capital expenditure and mining permits. And while Chinese engineers are underpaid compared to their Western counterparts, Beijing realizes their strategic value and is seizing passports to prevent critical know-how from leaving the country.

The chip supply chain has a chemical layer

Most discussions of semiconductor power focus on fabs, lithography machines, and advanced chips.

But modern chips also require dozens of extraordinarily specialized chemicals and gases, many at almost unimaginable levels of purity. A fab can use 60–70 specialty process gases, alongside other chemicals.

In The Chemistry of Chips, Carl Jackson explains how China went from a marginal player to supplying much of the chemical infrastructure needed for semiconductor manufacturing in just 15 years. “Taiwan is 100% reliant on Chinese supply chains today” for most critical gases.

Why it matters: Semiconductor resilience plans that stop at fabs and lithography miss an entire layer of the supply chain. The invisible industrial ecosystem is crucial.

China’s robotics advantage is a manufacturing advantage

In The Robots Are Here, SemiAnalysis’s Niko Ciminelli and Reyk Knuhtsen argue that China’s real robotics advantage is vertical integration, component supply chains, and iteration speed.

Our translation of a Caijing profile, The King of Unitree, shows what that looks like inside one of China’s leading robotics companies. Founder Wang Xingxing put it bluntly: “Don’t try to beat us on cost — we can still go a lot lower.”

Why it matters: General-purpose robots could be the first technology to truly decouple capital from human labor in the physical world. The race to build them won’t be won on AI models alone, but on actuators, magnets, and manufacturing speed.

How broken is China’s economy?

China’s problems get described as a property crisis, a debt crisis, a demographic crisis, or a demand crisis. Logan Wright on Broken China offers a more unified explanation: the finance-driven model that powered China’s growth is now constraining it.

Authorities keep failing firms alive rather than write off losses because “decay is an easier political choice.” One of his most useful insights is that weak domestic demand and China’s enormous export surplus are two sides of the same phenomenon.

China’s “Dying” Companies shows what this looks like at the firm level. R&D spending in strategic sectors roughly quadrupled between 2015 and 2023, but productivity remained flat, partly because local officials continued to prop up zombie companies.

Leia Wang’s Why China’s AI Bubble Is Also Industrial Policy complicates the picture. Government-backed venture funding is creating bubble-like dynamics in AI, but the resulting shakeout can also drive down costs, accelerate adoption, and build globally competitive companies.

Why it matters: Yes, China has EVs, AI, and robotics, but nothing has yet replaced property as a driver of Chinese growth. Wright argues that competition with China is increasingly about industrial and military power, not simply economic growth.

Taiwan

Taiwan needs drones. Can it build a drone industry?

In Can Taiwan Build Drones? and part two, Ethan Kessler looks at the surprisingly difficult task of turning Taiwan’s strengths in electronics and precision manufacturing into a functioning defense industry. Political fights over funding, limited flight-testing infrastructure, dependence on Western software, and Chinese dominance of key components all get in the way.

Why it matters: Taiwan has all the reasons to develop a full-stack drone industry. But challenges mean that Taiwan’s best bet may be supplying clean components to allied drone makers, in order to realize the Lai admin’s dream of “Red-Free” supply chains.

Taiwan’s energy problem is a national-security problem

Taiwan imports 97% of its energy and holds only around 11 days of LNG reserves. Taiwan’s War on Renewables, a radio show produced by Lily, examines the nuclear-versus-renewables fight, permitting problems, local politics, and a state utility that keeps prices artificially low and blocks market reform.

Why it matters: “The DPP’s botched green revolution has left the island structurally addicted to imported fossil fuels.”

Subscribe now

WarTalk

Can the US turn technology, money, and exquisite weapons into military power at scale?

Six months into the Iran war, WarTalk: Ammo + CCAs Are Dumb? took stock of what’s left. Bryan Clark estimates that roughly 850 Tomahawks were fired in the first couple of months, out of an inventory of around 4,000. LRASM and JASSM-ER stocks are “essentially at zero,” while the Army has fired nearly all of its ATACMS and most of its PrSM.

In Potemkin Pacific, the panel counted roughly 40% of the underway Navy in the Middle East, leaving INDOPACOM with little to cover the Pacific. We also asked whether the “AI Force” is a farce, and why physics, not software, still wins wars.

Why it matters: The US is spending its Pacific deterrent in the Middle East, and technological sophistication doesn’t equal usable military capacity. Mass, readiness, industrial capacity, and logistics determine whether weapons on paper can be deployed and replaced in wartime.

Elsewhere, Julian Gewirtz joined Jordan to discuss Xi’s state visit, Matt Sheehan and Kevin Xu sketched what a Project Glasswing with Chinese characteristics might look like, and we investigated a five-million-barrel oil mystery. We also went back to ancient Greece, shared our favorite books of the year so far, followed Lily to the World Nomad Games in Kyrgyzstan and Irene to Indonesia, traced the Presbyterian beginnings of North Korea’s personality cult, and joined Nick in figuring out why America ran out of Chinese lab monkeys.

Kevin Roose on the AGI Believers: Oppenheimer with Equity

Jordan Schneider — Fri, 09 Oct 2026 10:03:10 GMT

Kevin Roose, formerly of Hard Fork and author of the frankly brilliant The AGI Chronicles, with an assist from Jasmine Sun, who is a researcher on the book, joins us. We’re going through his recent history of OpenAI, Anthropic, DeepMind, and the community of people who think and worry too much about artificial intelligence. Kevin has left the New York T…

The Spice Trader vs. the Tariffs

Jordan Schneider — Thu, 08 Oct 2026 11:04:16 GMT

ChinaTalk’s people-suing-the-US-government-over-tariffs series continues. Last time we brought you the grumpy toymaker. Today we have Ethan Frisch, longtime friend of ChinaTalk and co-founder of Burlap & Barrel, finest purveyor of spices in the land. Co-hosting, as always, is Peter Harrell.

We discuss:

How an aid worker hauling wild cumin out of Afghanistan in duffel bags built a single-origin spice company, and why Burlap & Barrel chooses farmer-led pricing.
What Trump’s tariffs cost small businesses and why they’re suing the government, while the large ones sit on the sidelines.
The legal case against Section 301 of the Trade Act of 1974, arguing against executive overreach.
Whether forced labor is a real justification or a pretext for recreating the tariffs the Supreme Court struck down, and how long the appeals will run.

Plus, the most underappreciated spices, the virtues of a pre-capitalist trading model, and a call to buy small!

Burlap & Barrel promo code: CHINATALK for a free jar of your choice on a first-time order over $15.

Listen now on your favorite podcast app.

Where the Tariff Litigation Stands

Jordan Schneider: Peter, catch us up from where we left off with the Learning Resources CEO. What have the developments been?

Peter Harrell: So why are we talking about tariffs again? Didn’t the Supreme Court rule in February that the tariffs were illegal, so they just went away? That’s what most people would assume. But that is not what happened.

After the Supreme Court ruled in February, in a case brought by Learning Resources, an educational toy company, that Trump’s first round of tariffs — imposed last year under IEEPA, an emergency powers statute — were illegal, the administration did not give up on tariffs. Instead, they undertook a two-part set of fallback tariffs. The first were in place for 150 days under Section 122 of the Trade Act of 1974. Those are now expired, but still being litigated. The tariffs in place today — if you’re importing from Cambodia, Vietnam, Japan, or the European Union — are under Section 301 of the Trade Act of 1974.

In July, the US Trade Representative, pursuant to an investigation into alleged unfair trade practices related to forced labor by almost all of America’s trading partners, imposed 10% or 12.5% tariffs on most US imports. Those tariffs are being challenged in court by several groups of plaintiffs, but front and center is Ethan and his company, Burlap & Barrel. They argue that the Section 301 tariffs did not follow the procedures required by the statute, and that even if they did, they are too broad to be lawful. There was a trial-level hearing yesterday in New York before the Court of International Trade. Both sides faced tough questioning, but the government in particular faced tough questions about whether Section 301 actually authorizes tariffs of this sweeping scope.

Ethan Frisch: We are also a plaintiff in the 122 suit, so this is our second time suing the government.

Jordan Schneider: You just can’t get enough.

Ethan Frisch: There’s a lot of money owed back to companies including mine, and it’s very suspicious that the second tariff started the same day the 122 tariff ended. There’s a lot of funny business around this.

Burlap & Barrel’s Origin Story

Jordan Schneider: Take us all the way back to the global war on terror. How did your time in Afghanistan start you down this road?

Ethan Frisch: My journey to the courtroom started in 2008. I had been a line cook and a pastry chef in New York City, then went to grad school to become an aid worker in international development, and wound up in Afghanistan. I lived there for over two years, working mostly for the Aga Khan Foundation on infrastructure projects — mostly school construction in very remote, rural, mountainous areas across the northeast of the country. That part of the country happens to be known for its wild cumin.

Source.

Having worked in restaurants, going to little markets and meeting people as I traveled around northern Afghanistan, I was blown away by the flavor and quality of this cumin and started bringing it home. Not as a business initially, just because I loved it. I had worked at a high-end Indian restaurant in New York and thought I knew my way around the spice cabinet. It turned out there were so many more interesting spices around the world that had never made their way to the US market.

I started bringing cumin home in duffel bags and sharing it with friends in the restaurant industry, realized people might actually pay for something like this, and that there might be other producers in other countries. We started the company in 2016. My co-founder Ori and I started it out of my one-bedroom apartment in Queens: no money, no experience, truly no idea what I was doing. Driving out to the airport at two in the morning to pick up air freight shipments, haggling with customs officers because I didn’t have any of the paperwork, slowly figuring things out.

Peter Harrell: When did you pay your first tariff bill?

Ethan Frisch: It was a while. Some spices have been tariffed for a long time, predating either Trump administration — garlic, for example. But most spices are duty-free, because you’re not competing with any domestic production. We’ve paid tariffs on garlic from Vietnam because there’s a garlic industry in California. Maybe 2019 we paid our first garlic tariff, on our first shipment of purple-stripe garlic powder from Vietnam.

Source.

Generally, I was carrying things home. I did a lot of traveling in the early days with empty suitcases on the way there and full suitcases on the way home, and always declared everything, totally above board. Because spices are duty-free, I was waved through. That’s how I did my importing for at least the first year.

Jordan Schneider: You show up at TSA with this purple dust and they’re like, okay, cool?

Ethan Frisch: These were decent quantities. I would bring two huge empty duffel bags to bring home a hundred pounds of cardamom or a hundred pounds of nutmeg. In the early days it was often cheaper to buy a plane ticket than to ship those quantities. I’d show up, unzip my duffel at the counter, declare it, file a prior notice — there’s always a little paperwork. They were often very confused about what I was doing and why. Sometimes they’d open it up and look at it. Sometimes it was just, okay, fine.

Jordan Schneider: Would they ever make you taste it?

Ethan Frisch: No, they never made me taste it. It’s all legal. You can bring spices from all around the world; they’re not protected. My tip is to make sure the label says “spices” and not “seeds.” If it says seeds, you’re in a whole different category. A cumin seed is a seed, but it’s also a spice. Unlike meat or dairy or fresh fruit, you’re allowed to bring spices in.

What Makes a Better Spice

Jordan Schneider: Peter, as a home cook who’s been playing with Ethan’s stuff for the past few weeks, how different is it from what you get in a supermarket?

Peter Harrell: Ethan brings in great spices. We’re really going to do product placement here: I have a whole set of Burlap & Barrel spices. Compared to the McCormick stuff you’d buy at your local supermarket, you get a fuller flavor. Sometimes the profile is different, but mostly it’s fuller. The stuff you typically buy tastes muted by comparison, whereas this has a lot more complexity and roundness.

Ethan Frisch: You see this in produce. A waxy Red Delicious from the supermarket versus an orchard apple; there’s a huge range of quality. The same is absolutely true for spices, but nobody had invested in that as a business before. What we do is find farmers who are growing something different: a different species or variety, or something different in the agricultural process, often with a focus on soil health and biodiversity. You can taste that difference. Not to mention it’s a lot fresher. We bring shipments in several times a year and turn our whole warehouse over annually, so everything we sell is from the freshest harvest, versus a supermarket spice that’s already three years old by the time you buy it.

Jordan Schneider: Coming back to the tariff discussion: you’re not getting the wild cumin anywhere else. The substitute didn’t exist before you existed. The consumer isn’t going to be able to support American garlic instead.

Ethan Frisch: Most spices don’t grow in the US. The classics — black pepper, cinnamon, vanilla — you just can’t grow here. Not yet; maybe in ten years we’ll have the climate for it. There’s also the long history of growing a particular crop in a particular place. The farmers we work with have been doing this for decades, or generations. You can’t just build a cinnamon farm and try to replicate that experience. It doesn’t work.

Peter Harrell: How many countries will you be importing spices from this year, 2026?

Ethan Frisch: About 28.

Jordan Schneider: Before you came across these farmers, were they just serving the local market?

Ethan Frisch: Mostly selling into the local market, where there was real appreciation for quality and flavor, or into the commodity market, because that’s where they could get volume. Some were organic-certified, which gave them a little access. But mostly they were small-scale producers selling to their neighbors or within their local area. Some had exported before: our royal cinnamon partners in Vietnam export to Taiwan and Korea as a medicinal ingredient. In other cases their yields were very small because there wasn’t much demand. In Zanzibar, where we buy black pepper, the farmers’ co-op we work with has scaled up yields by probably 10x. They were harvesting a few hundred kilos when we started; now we buy at least five tons a year from them. By creating market access, we’ve given them the opportunity to scale up production they might not otherwise have invested in.

Peppercorn in Zanzibar. Source.

Peter Harrell: How do you find your suppliers? Are you trekking around the highlands of Vietnam, getting off the minibus in Zanzibar and saying, hi, I’m here?

Ethan Frisch: Less so now that we have long-term relationships in place, but in the early years, absolutely. That’s how we met the royal cinnamon supplier. We got out of a car in a little city in central Vietnam, smelled cinnamon, and walked up the street asking where the smell was coming from. Literally. We found a little workshop where they were peeling and cleaning cinnamon, hopped in the car, and they drove us up into the mountains to the farms. We’ve been working with them ever since. That’s our best-selling product.

We also find farmers on social media, especially if there’s a particular product we’re looking for. A lot of digging on the internet. And we have a lot of farmers reaching out to us: I grow turmeric in India, I grow this or that. Sometimes we wind up working with those people. The key is ultimately the personality of the farmer at the other end. Often they’ve already had the idea: if only I could find a direct customer in the US, I’d make more money. The best relationships are the ones where we show up on the farm, and they say, what the hell took you so long? I had this idea a long time ago. Let’s make it happen.

We’re often asking them to do more work than they would selling into the local or commodity market, where they just hand off the product. Working with us, they take on the export process themselves: government paperwork, logistics, getting the product to port, packing, food safety testing. We support them through all of it, but they have to do it. A farmer with a little entrepreneurial drive — those are the best relationships.

What Americans Will Buy

Jordan Schneider: How far can you push it? How weird is too weird?

Ethan Frisch: I’m very jaded about what American consumers are interested in. We sell a huge amount of cinnamon. That’s our number one bestseller by a significant margin. After that: black pepper, chili flakes, paprika, garlic powder. Americans generally haven’t had much exposure to spices and often consider them intimidating. There are notions of foreignness; spices represent a lot beyond the literal object, which connects to this tariff conversation and even to the culture war. It’s challenging to get Americans to pick up new spices. We’ve found ways around that, developing blends that are accessible but also different and interesting. But mostly people buy things they already know, and what we provide is a better, different version. You’ve had cinnamon before, but you haven’t had the species we import — we’re the only company that imports it — so you taste the difference right away. For a good cook, that’s a breakthrough: a new and better version of an ingredient you’ve worked with your whole life.

Peter Harrell: What spices should Americans try more of? What do we underappreciate?

Ethan Frisch: Everything. Whole spices in general, people seem intimidated by, but they’re very accessible whether you grind them or not. I cook with a lot of black pepper. That’s an easy way to use more spices: just add more black pepper to everything. Really use it as a seasoning; it brings heat and brightness. Personally, I love tart spices: sumac, which we get from Turkey, or black lime. They’re dry, so you cook with them differently, but with similar applications to lemon juice or lime juice or vinegar. In powder form, you can use them in marinades. Great on seafood, dips, roasted vegetables.

Jordan Schneider: Can we shout out any Chinese flavors you supply? Do you buy anything from China?

Ethan Frisch: We don’t import anything from China itself. We import a five-spice blend from Vietnam, a different take on a familiar profile, very heavy on the star anise, which I appreciate. One cool thing about Chinese cooking is that the understanding of spices is inverted. In the US, we consider cinnamon a sweet spice; in China, and most of Asia, it’s a meat spice. The idea of a cinnamon bun is an association someone might have with beef. A funny little twist.

We import a couple of cousins of the Sichuan peppercorn. From Nepal we get timur pepper, which is very citrusy, not as tingly as classic Sichuan pepper. From Vietnam we get mắc khén pepper, very bright, lime-zesty, and tingly. The peppercorn flavor people know is actually a genus, with several species that have similar flavors and numbing effects.

Jordan Schneider: What’s the sales breakdown between commercial and consumer?

Ethan Frisch: We’re primarily an online retailer selling direct to consumer through our website, burlapandbarrel.com. About 70% is direct through our own site. We do some business on Amazon, sell to a lot of grocery stores and restaurants, and to artisan manufacturers: chocolate makers, bakeries, jam makers. But the bulk is direct to consumer. We’ll probably do close to 150,000 orders this year, to home cooks across the country buying a few jars at a time.

Jordan Schneider: This is not a sponsored show, but we did get Ethan to put together a promo code.

Ethan Frisch: Code CHINATALK: a free jar of your choice on a first-time order over fifteen bucks.

Liberation Day and Suing the President

Jordan Schneider: April 2025. You wake up one day and your business changes. What was Liberation Day like for you and your team?

Ethan Frisch: I love any conversation about supply chains. Anything that reminds an American consumer that everything we touch and eat and interact with daily has to be imported. This is my issue. We’re a single-origin spice company, which means we’ve made intentional decisions about where things come from, and we do all this work to set farmers up to export their own crops. That’s a challenging marketing message. So any time current events remind people of the challenges of supply chains and how they can be better, I’m into it.

We saw pretty quickly that the financial impact on our business was going to be less compared to importers of finished goods like an iPhone, who pay the tariff on the finished product. We pay the tariff on a bulk product that then gets packed and blended, with a fair amount of additional work in the US. Within our unit economics, the cost of the spice itself is a small fraction of all the costs that go into getting a jar of cinnamon to somebody’s home.

So within a couple of days we made an announcement. Understanding we’d have some financial challenges from the tariffs and might need money to cover them, we announced a tariff sale and committed not to raise prices to consumers, which we have followed through on in the almost two years since, and not to cut payments to partner farmers or ask them to absorb the tariff. We’d cover those costs internally.

We saw a huge response. A social media post went viral, and we did about a quarter million dollars in sales in a very short period, because everybody was worked up about Liberation Day and looking for companies doing the right thing. We took advantage of the moment and turned it in our favor. I’ve been talking about supply chains ever since. I’m happy to have my issue in the news and the president talking about it.

Peter Harrell: What is the cost? You’d paid a few dollars in garlic tariffs before Liberation Day. How much are you paying now, and what has it meant for the business, especially since you’re eating all of it?

Ethan Frisch: I don’t know the exact numbers off the top of my head. The percentages changed, so it depended on the daily rate the day a shipment crossed the border. But generally it was between 10% and 20% on our bulk cost. We pay fifteen bucks a kilo or a little more in Zanzibar for black pepper, and we bring in maybe five to ten tons a year, so 10% on top of that adds up.

We got our IEEPA tariff refund back a few weeks ago. Low six figures. A nice chunk of change to receive.

Jordan Schneider: Whose name was on the check?

Ethan Frisch: I think it came from CBP. And it came with interest — three and a half percent, maybe. I’ll take it.

Those were real costs, and we’re a very cash-constrained business. We’re bootstrapped, with no outside investors, and we pay big deposits to farmers well in advance of receiving the product. Cash has been tight. That’s probably been the biggest impact: we’ve had to manage cash flow very carefully and had some tight moments in our bank account over the last year. But overall we’re growing. This is all in the context of increasing sales and increasing exposure for the company. I’m trying to stop the tariffs, of course. They’re illegal and an overreach. But it’s hard to be totally negative about them, because my business is doing really well in this context, it’s my issue in the news, and we’ve had these opportunities for press and to participate in these lawsuits.

Jordan Schneider: How are you balancing the “we’re not going to buy Ethan’s spices because he’s a woke lib” problem? Or were you already too woke-lib-coded for that to matter? How many Republicans buy wild cumin?

Ethan Frisch: Everybody cooks. Everybody can appreciate better flavors. When we filed the 122 lawsuit in May, we sent an email to our audience with a subject line like “We’re suing the president.” We hit that political message hard. We got almost exactly the same number of positive and negative emails back, something like a hundred each. It was almost identical numbers of people saying “yes, we’re behind you” and “screw you, we’re never ordering again.” Most of the negative responses were not active customers; it was a lot of posturing. Overall, business is up.

Other than that email and a couple of updates since, we’re not centering this in our marketing. We don’t really center any of it. We don’t talk much about our social enterprise or public benefit corporation status. We don’t even talk that much about our supply chains, because to your question about what American consumers are looking for: they don’t care. They don’t care about good supply chains. They don’t care about ethical sourcing. They don’t care about any of it. They just want to know the spice is going to taste good. That’s always the focus of our marketing. If somebody disagrees with our stance, there are other, worse spices they can buy. And our lawsuits are pro bono with the Liberty Justice Center, so your purchase isn’t funding the lawsuit. The lawsuit is happening regardless.

Jordan Schneider: How did you get involved with the lawsuit? The Learning Resources CEO told us his entire industry was terrified and no one wanted to step up, so it had to be a small, family-owned business with the attitude of, I don’t give a fuck, I see some upside in being the one to take this on. What was your calculus?

Ethan Frisch: There were three conversations. The first was right when the tariffs were announced on Liberation Day. We wanted to find a way to get involved, but we’d been contacted by a different libertarian pro bono law firm, and Ori and I had a lot of hesitation because of the distance between our politics and theirs, and concerns that we were being used. We didn’t know what we were doing. Peter, I think that was the first time you and I spoke about this, because we were trying to figure out how to engage without getting in over our heads. We didn’t pursue a lawsuit then, but we did file an amicus brief in the second round of IEEPA cases at the federal district court in DC. That put us on the radar of the Liberty Justice Center.

When the Section 122 tariffs were announced, they reached out and asked if we’d consider being a plaintiff. I was in India at the time, trying to get home from a sourcing trip in Kerala the day the Iran war kicked off. Ori was on the phone with Liberty Justice, and I was at the Mumbai airport desperately trying to figure out how to get home with all the flights canceled. Ori said, read this email, we should do this now. And we did.

The Liberty Justice Center has been great to work with. We may not agree with them on pretty much anything beyond this particular issue — they’re a libertarian, pro bono, activist law firm in Chicago — but I have a huge amount of respect for how they approach this. I had lunch with their chairperson and head of litigation yesterday after the hearing.

When the 301 tariffs were announced, we were not planning to be a plaintiff. We’d filed the 122 case, and that was enough. But they were struggling to find plaintiffs willing to take the risk. To be honest, I don’t understand what everybody is scared of. Yes, retaliation from the president or people close to him. But it has not happened in any of the tariff cases. That was one of the things we looked into. I discussed it with my wife: this is a risk; what do we think? We looked at how the president handled the IEEPA case. He said aggressive things about the plaintiffs in aggregate, but he didn’t call out any company or individual by name. That gave us some reassurance.

Also, we have no outside investors and total autonomy. Ori and I own 100% of the company. Nobody can tell us what to do, whereas other founders may have had more hoops to jump through with investors or board members. And over ten years we’ve made very good use of earned media, from the New York Times to Shark Tank, and that has really supported the growth of the business. We’re comfortable talking to the media and using that to promote the company and the issues we care about. So we felt well positioned to take advantage of the exposure that comes with a lawsuit like this.

The government could come after us. The IRS could audit us. We just went through an FDA audit, which we passed. I’m exaggerating a little; there has not been any retaliation or even a hint of it. And we won the Section 122 case. Based on how the hearing went yesterday, I’m not an expert, but it seems we’re likely to win at least this first round of the 301 suit as well.

Why the Big Companies Stay on the Sidelines

Peter Harrell: It’s been interesting over the last year and a half. You talk to major law firms in DC or New York, or to Fortune 500 companies, and they have been terrified to bring lawsuits challenging the tariffs, or as lawyers to take on clients who want to. They’re clearly afraid of retaliation. As a result, in all the lawsuits we’ve seen, it’s been either small businesses like Ethan and Ori or Learning Resources, or Democratic-led state governments: Oregon, New York, California. The big companies you’d expect to be lead plaintiffs, because they’re paying billions in tariffs, have been sitting on the sidelines.

What’s striking is that despite all this fear — and I was counsel of record for members of Congress filing amicus briefs in the IEEPA litigation, so I’ve been tracking it — it doesn’t look to me like the government is actually retaliating at all. Obviously they want to win in court. But I’ve seen no evidence that USTR, CBP, or the Department of Justice have been anything other than professional. Trump is going to say what Trump is going to say. That’s the world we’re in.

Jordan Schneider: What’s driving that? We’re going after Letitia James over her second mortgage. We’re taking away security clearances from random former Biden staffers now on the Hill. They can be a lot more petty, and they have been. Is it just that USTR is a functional organization as opposed to these other places?

Peter Harrell: It’s a great question, because we’ve definitely seen retaliation in other areas. It’s probably a combination. USTR and CBP are by and large quite professionally run, and respected on both sides of the aisle. And there may be some self-interest: if you’re actively involved in a lawsuit and the government seems to be coming down on a plaintiff, the judge is going to become aware of that. If you look at the DOJ lawyers litigating these cases, or the USTR and CBP lawyers involved, they’re professionals who don’t want to play games, who want to win on the merits and aren’t looking to make these political cases.

Ethan Frisch: The core argument that came up in both the 122 hearing and the 301 hearing yesterday was deference to the executive. Section 301 or 122 says this; here’s the text. But ultimately there’s a level of interpretation allowed to the executive branch. That came up several times. The government lawyer was essentially arguing, let us get away with this because we’re the executive branch. But the government’s case is just not that strong. Section 301 has never been used this way. It’s a targeted tariff aimed at a specific product or country, not a copy-and-paste reapplication of 122, which was itself a copy-and-paste of IEEPA. You can see pretty clearly that the administration is taking the same tariff structure and putting a different justification on it, one after another.

The Legal Case Against Section 301

Jordan Schneider: Peter, what’s your legal analysis of where we are, and where the next few years of tariff policy could go?

Peter Harrell: Congress passed Section 301 of the Trade Act of 1974 to give the executive branch a tool to negotiate over unfair trade practices by foreign governments. It’s been used something like 130 times. Unlike IEEPA, which had never been used for tariffs before last year, 301 has been used for tariffs quite a bit, notably as the basis for Trump’s first-term tariffs on China.

The way it’s being used this year is fundamentally different. In all past uses, USTR investigated a discrete set of unfair trade practices by one country, or by the EU as a customs union, found the practice unfair, and proposed tariffs as leverage to negotiate a resolution. In the 1980s there were 301 cases on Japanese semiconductors. In Trump’s first term, the 301 tariffs on China were about Chinese theft of US intellectual property, followed by an effort to negotiate a deal.

What the government is doing now is reinterpreting the statute to give itself a legal basis for essentially permanent tariffs on sixty jurisdictions, counting the EU as one, that account for 99% of US trade. Rather than a tool to investigate a country and negotiate, they’re trying to rewrite the tariff code wholesale. The conceptual argument against these tariffs is that this breadth is not allowed under the statute.

In court, you have to break that down into what the statute specifically required, procedurally or substantively, that the government failed to do. There was a lot of discussion yesterday about whether the government adequately investigated each of the sixty countries, since mostly they did a copy-and-paste job. Did they actually find that what each country does burdens US commerce?

There was also a lot of debate because Section 301 has specific provisions around forced labor. The government’s basis for these tariffs is that these sixty trading partners have failed to prohibit their own imports of products made with forced labor. It’s not that there’s forced labor in the European Union; it’s that the EU isn’t appropriately preventing forced-labor products from coming in. But the forced labor provisions in 301 are about what a country does domestically, not this third-country import issue. So can 301 be used for this kind of forced labor claim at all?

Basically, the plaintiffs’ lawyers were trying to take the conceptual argument — Congress clearly didn’t intend 301 to authorize a wholesale change in tariffs of unlimited duration at any amount, since the government’s view is that it can impose it at any amount — and break it down into specific requirements the government failed to meet. The government faced a number of skeptical questions. They’re arguing the statute doesn’t require as much detail as it has historically been viewed as requiring, that they can do a cut-and-paste job, and that the court should defer to the executive branch. The panel of three judges had tough questions for both sides, but was frankly somewhat skeptical of the government.

This is just the first step. After the Court of International Trade, there’s the Federal Circuit, the appellate court above it, and this will almost certainly go to the Supreme Court. Whatever happens at the Court of International Trade, we have a year to a year and a half of litigation ahead.

Forced Labor as Pretext

Ethan Frisch: China was the elephant in the room, the subtext behind a lot of the conversation. A lot of the forced-labor goods being imported into these other economies mentioned in the 301 come from China, which is a site of forced labor at scale more than other countries.

Jordan Schneider: Speaking of forced labor laws, we’re going to run an article on ChinaTalk soon. A Chinese manufacturer offered fifty jobs to a graduating class, and the day before they were going to start, told thirty of them: you’re either working on the line for half your pay or we’re firing you. And the students started reporting them under EU law. You think you’re so slick — just wait until we blow up all your export markets by exposing your forced labor. Ethan, do you encounter this at all?

Ethan Frisch: Sure, a little bit. The point is that forced labor is a real thing. Whatever you call it — forced, compulsory, slavery — it affects a lot of people and a lot of economies. For the government to use it as a throwaway justification for another round of the same tariffs really undermines the issue, which it would be great to have more dedicated focus on solving. Instead they’re using it as window dressing.

Jordan Schneider: Rubio was a co-sponsor of the Uyghur Forced Labor Prevention Act. I remember when it passed: we’re getting three hundred million dollars for forced labor enforcement and not a dollar for BIS. But I guess if it was just to put tariffs on European products...

Peter Harrell: There are a number of anomalies that show the forced labor justification is pretext for what the administration wants to do. First, the tariff rates and the exempted products largely track the tariffs the Supreme Court ruled illegal in February. It’s not like we’ve seen an extensive effort by the government to say, in Vietnam there’s a huge problem with forced-labor imports, in the EU it’s small, with differentiated rates. You see two rate levels: 10% if the country has a legal ban on forced-labor imports but fails to enforce it — and they found no country enforces the ban, so 10% is the minimum — and 12.5% if the country doesn’t have a ban. The rates don’t correlate with forced labor. China, probably the biggest forced labor problem in the world, has a 12.5% tariff, and so do about two-thirds of the other countries. No differentiation at all.

And there’s no secret that this is pretextual, because President Trump said on February 20th, the afternoon after losing the Supreme Court case: don’t worry, we have other authorities, we’re going to recreate the tariffs. The government hasn’t hidden that this is about maintaining the tariffs.

What Comes Next: Appeals, Refunds, and 2028

Peter Harrell: Even if Ethan wins at the Court of International Trade, it will be appealed up the line. Given the timeline we’ve seen in the other tariff litigation, probably twelve to eighteen months before this is fully resolved. And the other thing we’ve seen is that even if Ethan wins in the coming weeks or months, the government will be able to continue collecting the tariffs during the appeals. One area where the courts have been deferential to the government, even when ruling against it on the merits, is letting it keep collecting while appeals proceed. Unfortunately, Ethan, you’re going to be writing checks for a while. If and when you prevail at the Supreme Court, you’ll get a refund.

Ethan Frisch: They pull the money directly out of our bank account. We don’t even write a check. No opportunity to haggle. Direct debit from our account straight to CBP, as it is for everybody.

Peter Harrell: You can’t use the haggling skills you learned over spices. How about 8% instead of 12.5%? Six now, six later?

Jordan Schneider: This is like Annie Lowrey’s new book: the government is really efficient when it comes to taking your money, but Congress will throw plenty of things in the way of poor people getting their Medicare benefits.

Speaking of Democrats, Peter, where is the 2028 crop on tariffs? There was an initial vision that all this would go away. It might, if it all ends up being illegal. But Biden crossed the Rubicon a bit, and we may never go back.

Peter Harrell: In the policy circles I talk with, it’s underappreciated that the courts may find all of these country tariffs illegal. Not all tariffs — the steel tariffs are under a different authority — but the country tariffs. There’s a view in the policy community that whatever happens after Trump, the next president will inherit all these tariffs and may or may not modify them. That may be true, but it may also be true that a lot of this fails in court. That’s a big question overhanging the long term of US tariff policy that the courts, not the executive branch, will resolve.

If they’re all upheld: the Democratic Party is going to have a wide-open primary with thirty candidates, a real circus, which I think is good for the party.

Jordan Schneider: Does this mean you’re not running, Peter?

Peter Harrell: Heck no. No one wants my views on anything. I think the median Democrat will want to cut the tariffs. Biden did not cut Trump’s first-term tariffs at all. But the median Democratic next president, having presumably won on an affordability message, is going to say, I want to get rid of the tariffs on your spices and this kind of thing. They’ll probably want to keep some percentage as negotiating leverage or protection for US industry. Some decent chunk goes away, and some decent chunk sticks. Assuming the courts uphold it.

Jordan Schneider: Ethan, we met up at the Fancy Food Show in New York. Is the industry aware? Is McCormick thanking you for doing this?

Ethan Frisch: As Peter said, they’re all scared on the sidelines, letting the little guy take on the risk of retaliation. I have gotten a couple of thank-yous from bigger spice companies, not the one you mentioned. We made stickers — “Pay Farmers, Not Tariffs” — and gave them out at the Fancy Food Show at the Javits Center. We had maybe 500 printed and they were gone within the first two days.

Small business owners are really heads down at this point. The inflation and economic downturn we see in our direct-to-consumer interactions are a much bigger obstacle for small businesses. People don’t have that much money to spend on fun or discretionary food purchases, and consumers are being much more careful. And then the cost of the tariffs themselves. Most small businesses are focused on the micro numbers: how do I make sure I have enough margin to cover these tariff costs, and how do I make sure customers keep buying? The big companies are just baking it in, as they always have. They operate at scale, and the amount of cash moving through their system lets them weather these storms much more easily. My call to action is: buy small. Buy from small brands that really need the money, not from the big brands coasting off our litigation.

New Products and Farmer-Led Pricing

Jordan Schneider: Let’s close on something more fun than making payroll. What are some new products? You mentioned you’re getting into goops.

Ethan Frisch: Condiments, for sure. Spices are challenging to market in the US even under the best circumstances. It’s a small category in the grocery store, and not something people use in huge volumes. So we’ve been exploring adjacent pantry categories. Over the last few years, we’ve rolled out a line of single-origin sugars, inspired by the growth of the salt category over the past decade, from iodized supermarket salt to flake salt in lots of formats. We have a whole line of unrefined sugars with great flavors and textures: jaggery from India, panela from Colombia, coconut sugar from Indonesia.

We have a partnership with the Jane Goodall Institute, and with Dr. Jane Goodall herself before she passed. I believe this is the only consumer product she ever put her face and name on in her 96 years. It’s a line of wild-harvested honey from the woodlands of central Tanzania, one of the last remaining chimpanzee habitats in the world. With tight controls over how much and when they harvest, people who live in the woodlands can make a lot more money harvesting honey than through slash-and-burn agriculture, which is essentially the alternative. We launched those about a year ago and they’re doing great: really beautiful, intense flavors.

Source.

Pomegranate molasses from southeastern Turkey is another sweetener. It’s reduced pomegranate juice, an ancient sweetener. We work with a small factory that has figured out how to do it under pressure, so it’s essentially evaporated rather than boiled. It keeps a very bright, tangy flavor and a bright red color instead of turning dark and caramelized like traditional pomegranate molasses. Great on salads, in marinades, even in stews.

The other product line that’s been really successful has been the shift from selling individual spices like wild cumin, Zanzibar black pepper, or royal cinnamon toward blends. Since I’m jaded about what American consumers want, we created blends very similar to things people have already experienced: lemon pepper, cinnamon sugar, barbecue, steak seasoning, Cajun seasoning. They’re all made with the same spices we sell individually and the same supply chains, supporting the same farmers, just reformatted with a name consumers find more accessible. The everything bagel seasoning — I probably swore ten years ago I’d never make one. But we need a great everything bagel seasoning. We sourced the granulated garlic specifically, and heirloom sesame seeds from Turkey, and put a lot of black pepper in it. It’s spicy, peppery, with some cumin, and a lot of texture.

Jordan Schneider: Can we close on a little bleeding-heart stuff? What premium do farmers get when they start selling to you, and how have you seen lives change?

Ethan Frisch: I don’t track the commodity price that closely. We pay farmers whatever they ask. We call it farmer-led pricing. We don’t negotiate, because the cost of the spice is a very small percentage of our total unit cost. An extra dollar a kilo has almost no visible impact on our bottom line but a huge impact on somebody’s day-to-day life. A longtime partner cardamom farmer in Guatemala built a new two-story house with running water, which he’d never had before.

Jordan Schneider: How does farmer-led pricing work? Has no one ever given you a number where you said, no thank you?

Ethan Frisch: Not never, but almost all the time we’re able to pay the price they ask. Sometimes we go back and forth to show how it works out in our costs. Our biggest cost is actually shipping from our warehouse to your home. We encourage farmers to ask for the amount they need to run a sustainable business. In most of these relationships, we have a very frank conversation where we break down our business model and cost structure: here’s what it costs to ship to our warehouse, here’s what it costs to put it in a jar. Then we do the same exercise for them: what labor are they hiring through the season, what inputs, what equipment. A number usually emerges pretty clearly that works for them and for us, and it’s a long-term partnership.

Jordan Schneider: Is this a normal thing? Do lots of people do this?

Ethan Frisch: No, nobody does this. It’s an anti-capitalist, or pre-capitalist, approach to trade. It’s not really about the dollar amount. We’re not comparing this supplier to that supplier or switching things around to save a few pennies. We’re invested in long-term relationships with individuals whose product is unique. What we bring is logistics and market access; what they bring is deep expertise in this particular thing. We live in a capitalist paradigm and have to speak that language, but we don’t have to bring it into our relationships. We can work with people as individuals, with real trust and friendship underpinning the relationship. If they need to charge a little more this year, maybe a little less next year. The money is secondary to the relationship.

Jordan Schneider: Has Zohran reached out to get you to run the supermarkets?

Ethan Frisch: No, I don’t want to run the supermarkets. What a horrible business. He has not, thankfully. But maybe they’ll carry us.

Jordan Schneider: Ethan, thank you so much for being a part of ChinaTalk. Promo code CHINATALK for a free jar of your choice on a first order over fifteen dollars, at burlapandbarrel.com. We’ll see you at the Supreme Court.

China, the FCC, and the Logic of Transceivers

Quinn Ennis — Wed, 07 Oct 2026 10:13:11 GMT

Quinn has joined the ChinaTalk team after studying semiconductor physics at Yale. Reach out to quinn@chinatalk.media to chat energy, BRI, robots, trade, industrial policy, or anything else!

Over the past year, the FCC has released a series of restrictions on foreign-made hardware. First came drones, then routers, and then in July, inverters and robots. The newest devices to be considered are optical transceivers, which convert data into light and back within data centers. Following a Reuters report of a draft ban on new Chinese transceivers, markets responded as if it meant a straightforward substitution of supply, and American optics stocks rose while Chinese optics leaders fell.

This swap is not that simple. The FCC’s cybersecurity logic is real but far narrower than the rule suggests. After explaining these gaps, I follow the transceiver supply chain from the module and the laser all the way down to the wafer and the metal it relies on. Chinese companies dominate the top of the stack but are reliant on Western companies in the middle. At the very bottom, Chinese control is ingrained, and the U.S.’s dependence on these basic materials will not be shifted by an FCC rule at the top of the stack.

The Rule

On August 4, Reuters reported that the FCC was drafting a ban on import authorization for new models of Chinese optical transceivers to prevent Chinese firms from stealing data, installing malware, or disrupting service. Per the anonymous sources, officials want the rule published and effective by year end, though they cautioned the draft could still be modified or shelved. By the end of the week on August 7, Zhongji Innolight 中际旭创, the largest transceiver vendor globally, suffered a 10-percent drop in its stock price, while Coherent, a leading American transceiver manufacturer, saw its stock price rise 17 percent.

The attention to transceivers did not come out of the blue. The FCC has basic authority over any devices that emit radio-frequency energy — including high-speed digital electronics like transceivers that radiate as a side effect — and a 2021 law bars it from authorizing equipment on the Covered List. On July 22, the FCC adopted rules (published September 11, effective October 13) aiming to close the component loophole for authorization of devices with “logic-bearing hardware components,” including optical transceivers. But this prohibition only reaches components made by companies on the Covered List, which does not include the largest Chinese transceiver vendors. The FCC would either have to explicitly name those companies or implement the draft rule reported by Reuters to add transceivers to the Covered List as a category defined by production location, as it did for drones, routers, power inverters, and robots over the past year.

On August 13, the Information Technology Industry Council (ITI), a trade association representing hyperscalers and optics companies, filed an ex parte notice explicitly urging the FCC to avoid adding optical transceivers or any foreign transceiver companies to the Covered List, emphasizing the impact of supply disruptions on data center build-out. No individual optics company, switch maker, or hyperscaler has filed comments arguing about transceivers (ET Docket 21-232, as of Sept. 21).1

The template for an FCC rule on transceivers is public by way of precedent. When hardware was previously added to the Covered List, “foreign-produced” was defined by the Buy American test, and anything made outside the United States or without mostly domestic content is covered, with the DoW and DHS able to grant conditional approvals for individual producers on a time-bound plan to build in the U.S. The word “Chinese” would never have to appear in the rule. Innolight was included this year on the Pentagon’s Section 1260H list of Chinese military companies, but the restrictions are fairly narrow and bind only the Pentagon.2 The FCC rule would be the first measure on transceivers to reach the networks that hyperscalers build for commercial customers.

The ITI concern and the national security question both hinge on what transceivers actually are. The conversion of electrical to optical signal requires precisely the kind of logic-bearing hardware component defined by the FCC.

Inside the Box

There’s a temptation to think of data centers as a bigger, faster version of your laptop — but as the size and speed scale, signal flowing through copper wire tends to blur, with one symbol interfering with the next. Imagine standing in an echoey room. If you speak too quickly, your words commingle with the echoes. Shouting does nothing, as the interference is a function of the rate at which you speak, not your volume.

For long distances, data centers avoid copper’s loss mechanisms entirely, using optical fiber to transmit signals. But at modern data rates, an optical module draws roughly one hundred times more power than a one-meter passive copper cable. If you use fiber for short stretches, you are paying a power bill for reach you aren’t using. As Arista’s CEO said in a recent call, there exists a philosophy of “copper if you can, optics if you must.”

Within racks, copper dominates. Between racks, fiber reigns supreme.3 At their intersection sits our protagonist, the optical transceiver.

A pluggable optical module connects on one end to the switch — the routing center for the rack’s GPUs — and on the opposite end to the fiber. Between, the module holds the requisite hardware for converting an electrical signal into an optical one, namely a laser and a digital signal processor (DSP). Each pluggable transceiver also contains a microcontroller unit (MCU) that controls the DSP, sets the laser, and talks to the switch.

On the way back in, the incoming light is converted to a faint current by a photodetector and boosted into a signal that the DSP can read by a transimpedance amplifier (TIA). Along the way between the parts, the light is guided by lenses and filters, and isolators shield the laser from reflections. Innolight puts such passive optics at 10 to 20 percent of a module’s total material cost.

While the bulk of the module is built on a silicon substrate, silicon tends to release energy as heat rather than light, so the laser uses indium phosphide (InP), which can be tuned to an ideal wavelength for the optical fiber. As data rates increase, some modules have leveraged silicon photonics (SiPh), integrating waveguides, modulators, and photodetectors directly on the silicon chip. While the laser is still a separate InP component, it is now shared across several lanes, so the module only needs a handful rather than one per lane.

In the distance between the switch and the faceplate that the module plugs into, the signal suffers significant loss that must be cleaned by the DSP, which consumes roughly half the module’s power budget. This loss is directly bypassed in some new architectures. In a near-packaged optics (NPO) system, the optical engine is soldered on the board just a few centimeters from the switch. In co-packaged optics (CPO) systems, the switch and optical engine are integrated into the same shared substrate with fewer, more powerful lasers housed in a separate plug-in module.4 CPO and NPO, however, come at the cost of much more expensive servicing.

How to Use an Optical Transceiver for Nefarious Ends

Of the three harms officials described in the Reuters article — stealing data, pushing malware to the host, or disrupting service — the only one that’s actually plausible is service disruption.

Claims that optical transceivers can be used for espionage are unfounded. Optical transceivers are neither a router nor a robot. A transceiver lies at the very bottom of the network, turning bits into light and light into bits blindly. And hyperscalers encrypt traffic before the data ever reaches the module.5 Microsoft asserts that encryption “is on by default for all Azure traffic within a region or between regions.” The module has a small processor and some memory to run the firmware but, as Cignal AI put it, has no “external management connection.” A compromised module might be able to jumble up the bits that it transmits, but it could not read them, remember them, or send them anywhere. There is no path for China to extract American data or model weights via an optical transceiver. As the security expert I spoke with attested, transceivers are a poor tool for spying, and the spying argument is a sign that the person making it doesn’t understand how a data center is actually secured. For a module to push malware to the host, it would have to exploit a preexisting bug in the switch software that reads its memory, though this mechanism has never been documented for transceivers.

Sabotage is possible, but there are no public confirmations of a hidden implant within commercial computing hardware. In 2018, Bloomberg alleged Chinese spy chips on Supermicro server boards, but this was denied by Apple, Amazon, and Supermicro, and DHS said it had “no reason to doubt” them. In its July fact sheet, the FCC names transceivers and warns that compromised components could “access, store, disrupt, and/or misroute U.S. communications” without citing any example. The absence of a public example isn’t disproof, but the public record contains no basis for the espionage claim specifically.

The module’s controller is the microchip the FCC worries could be exploitable. The FCC fears that a controller with a subtle change in its integrated circuit could begin malicious activity after a rare event like a thermal spike. In silicon generally, the feasibility of a subtle hardware trojan was demonstrated in simulation by security researchers in 2013, and a later study built a trojan into a working chip that would “never be encountered by even the most diligent tester.” Neither study demonstrated viability in a transceiver.

In addition to hardware trojans, there is the simpler risk of malevolent firmware. The industry standard governing communication between the switch and the modules includes commands for loading new firmware. It does not, however, contain any requirement that the module check that the firmware is genuine, nor is there any means for the switch to inspect the module’s code. The same standard lets either the switch or the module tune the signal. The practical default within data centers, according to a security analyst familiar with the industry, is to leave tuning to the module: tuning thousands of links by hand would be impractical, and the module knows its own laser, chips, and thermal drift far better than the switch does.6 The switch can ask the module for its settings and how clean the signal is, but the answers come out of the module’s own memory, written by its firmware. The switch does count the transmission errors itself and would notice a link going bad, but it would have to take the module’s word for why. A signal that degrades, throws an error, or drops under load would look like an ordinary faulty module. No malicious transceiver firmware has been publicly documented.

CPO systems move signal tuning to the switch, but the heat-sensitive laser is still in an external pluggable laser module with its own microcontroller running the maker’s firmware. Chinese supplier TFC builds the laser modules for Nvidia’s switches, and trade press reports that the lasers inside are Lumentum’s. Because this laser module acts as a continuous light source rather than processing data, a compromised one could dim or destabilize its lasers to take down the eight ports it feeds, but this is still a link-level outage and provides no data access.

Compromised modules could not corrupt what a model computes, but they could drain compute by stalling training runs, while looking like nothing more than transceivers with a high failure rate. Any disruption would be bounded. A compromised module takes down its own links, but modules are hot-swappable and procurement policy means data centers get modules from multiple vendors. No attack of any kind has been documented in a transceiver, and only very narrow, uncoordinated service disruption is plausible — a thin predicate for an import ban.

The Stack

Espionage fears don’t align with the actual architecture of optical transceivers, but there need not be a backdoor for China to use the components as leverage. Final production is dominated by Chinese companies, but this is the layer Beijing can least afford to weaponize. China’s actual control is at the bottom of the stack.

Across metrics, it is broadly agreed that Chinese vendors account for the majority of the optical transceiver market. While the product mix varies for each company, the inflection point over the last few years is clear. Chinese firms both control the majority of the market and are growing the fastest, with Innolight sitting on top of every ranking.

In the plot above, I color Chinese companies based on the location of their headquarters, but the question of what “Chinese” means is not simple. Take, for example, Source Photonics, which is headquartered near Los Angeles but owned entirely by a firm based in Suzhou.

The bigger problem of defining a Chinese transceiver lies in factory location. In its prospectus before listing on the Hong Kong stock exchange this year, Innolight disclosed that the vast majority of its transceiver manufacturing capacity is now in factories in Taiwan and Thailand rather than mainland China, the outgrowth of a 2020 push to build capacity overseas to mitigate the uncertainty from a changing economic and trade environment (化解经贸环境变化带来的不确定性风险). Indeed, since 2024, Innolight’s mainland capacity has actually fallen while its overall capacity has grown. If Q1 2026 rates hold through the rest of the year, Innolight will have quadrupled its total capacity in just three years, carried entirely by factories outside of China.

If the FCC follows the template of the previous rules, a foreign-produced test would hit American companies too. It would apply to Innolight’s Thai and Taiwanese plants but also initially affect Lumentum’s Thai plant, Coherent’s Malaysian assembly site, and Coherent’s factory in Fuzhou, leaving each to seek conditional approval plant by plant. Sparing them would take a true ownership test, naming specific companies to the Covered List and justifying each inclusion.

Just below the assembly level, many of the passive optics that guide light through the module are Chinese. In 2019, Sumitomo Metal Mining, a Japanese company, established a new subsidiary in Shenzhen specifically to supply Chinese isolator companies, which control 90 percent of global production. TFC is the top passive optical component supplier, making almost all its units in China and selling 63 percent of its 2025 revenue to Fabrinet, the Thai contract manufacturer.

Many of the direct security concerns cited by the FCC are tied to compromised Chinese MCUs, but, for the most part, the MCUs in transceivers are not Chinese. The MCU is an off-the-shelf part, sold by American, Japanese, and European companies. The predominant player here is STMicroelectronics, a Franco-Italian firm described by Chinese trade press as the default choice. The logic-bearing hardware component that the FCC is worried about may well be European, unless Chinese makers have quietly substituted domestic chips without saying so. Chinese transceiver companies have been remarkably silent in response to investor questions about MCU provenance.7 Their silence is consistent with either answer.

The majority of high-data-rate digital signal processors in transceivers are designed by American companies like Broadcom and Marvell and manufactured by third-party foundries in Taiwan including TSMC. Innolight confirmed in its 2025 annual report that suppliers of electronic chips at 200G and above — including DSPs, laser drivers, and TIAs — are primarily located outside of China (电芯片包括DSP芯片、激光驱动器（LDD）、跨阻放大器（TIA）…，其中200G及以上速率的供应商主要在海外). Indeed, in the same filing, Innolight discloses that more than a third of everything it buys comes from a single, unnamed foreign supplier. The main listed TIA maker in China, Xiamen UX 优迅股份, has designed 200G-per-lane parts, but none have entered production.

The electro-absorption modulated lasers (EMLs) and continuous wave (CW) lasers used in high-speed transceivers are mostly American and Japanese. According to TrendForce, across both CW lasers and EML, Chinese producers held just 16 percent of production capacity in 2025, though this is concentrated in low power CW lasers used for SiPh, and Chinese firms are just starting to enter volume EML production this year. Innolight’s 2025 annual report expressed Chinese laser shortcomings directly: EMLs at 200G per lane and above must still be imported, as domestic manufacturers can only supply volume at 100G and below (200G及以上速率的EML激光器目前仍需进口；100G及以下速率的EML激光器，目前国内已有部分厂商可实现批量供应). Photodetectors follow the same pattern. All major 200G vendors are American, and China’s Sanan reported in August that the optical chips for its 800G and 1.6T modules were still in sample testing.

China’s dependence on Western lasers has decreased for current generations, as 800G and 1.6T transceivers often use silicon photonics systems and thus fewer InP lasers. Innolight’s 2026 Hong Kong stock exchange prospectus asserts, “in the three months ending March 31, 2026, products utilizing SiPh technology constitute approximately 70% of our high-speed product portfolio by revenue.” Chinese module manufacturers dominate the SiPh market. In the same prospectus, Innolight cites market research identifying them as “the largest global provider of SiPh-based optical transceivers by revenue in 2025.” Even so, at least part of Innolight’s photonic chip layer is fabbed by Israel-headquartered Tower Semiconductor. However, as the CEO of Lumentum attested in an earnings call this August, “silicon photonics is a viable solution at 1.6T… as we think about 3.2T, silicon photonics loses some of its advantages and we’d expect to see [traditional laser architectures] come back in a meaningful way.” And as Coherent’s CEO said, SiPh or not, “both require indium phosphide capacity.”

Subscribe now

Western capacity ownership does not equate to availability. Lasers are the scarcest optical component within the module. This shortage is mentioned in every single earnings call from the top three American transceiver companies from Q3 2025 to Q1 2026, with Applied Optoelectronics (AAOI) explicitly identifying “a shortage of indium phosphide laser manufacturing capacity across the industry” in a call this May. Laser production is constrained by fab capacity, and most high-speed lasers, along with many of the photodetectors, are grown on an indium phosphide wafer.

These wafers are manufactured almost exclusively by three companies: two Japanese firms, Sumitomo Electric and JX Advanced Metals, and AXT, an American company headquartered in Fremont that manufactures exclusively in China. In February 2025, Beijing added indium phosphide and the two indium compounds used to grow chip layers to its export control list. Indium gallium arsenide, the light-absorbing layer in InP photodetectors, has required a license since 2023. AXT’s Chinese subsidiary, Tongmei 同美, received no export permits at all until June 2025, and still has no export permits to the U.S., sealing AXT’s decade-long shift to China.

Beijing cutting off AXT from the U.S. does not mean that the West is without substrate, as Sumitomo and JX control most of the remaining capacity. This June, JX announced a ¥120 billion (roughly US$750 million) capital program to expand capacity seven- to ten-fold. Whether this Japanese capacity can cover Western demand is unclear.

Coherent, the largest American transceiver manufacturer, does not produce its own substrates, and its annual report lists InP among the raw materials it purchases, saying in a call this February that it has “multiple six-inch indium phosphide substrate suppliers” with enough supply locked in to support its current expansion. It has not named them, though Reuters reports that AXT, which makes its wafers in China, is a major supplier of Coherent, and the two companies signed a supply agreement in June.8 Chinese export permits are granted by destination rather than by customer, and AXT has told investors that it is “getting permits pretty readily for U.S. customers based in other global regions” despite being “unable to estimate” when it can ship to the United States. Thus, Coherent’s laser fabs in Sweden and Switzerland can receive Chinese wafers while those in Texas cannot receive AXT wafers directly.

Since the 2025 control, prices of six-inch InP wafers have risen 250 percent, and Coherent’s CEO joined the business delegation accompanying Trump to China in May — in part, according to Reuters, to press for faster export licenses. That’s right, America’s largest transceiver company went to Beijing to ask for permission to ship another American company’s wafers to Texas.

Newer architectures do nothing to loosen the dependence on InP supply, as lasers in CPO are fewer but more powerful, requiring a bigger piece of InP. AXT’s CEO told investors this summer that the shift to NPO and CPO “will continue to drive increasing demand for our material.”

One layer deeper, the picture is even less rosy for the prospects of supply chain independence from China. Indium phosphide is produced from high-purity indium metal, and as of 2025, China refines 69 percent of global indium, though exports have dropped without any control on the metal itself. Substrate capacity can be stood up in Japan in a few years, but the metal is harder. Indium is a byproduct of zinc smelting, and while zinc is mined globally from Peru and Mexico to Australia, smelters outside of China often lack the economic incentive to recover indium from the sludge. The industrial dependency here is the same as the one China leveraged with the 2023 gallium export control. Beijing could pull the trigger on indium at any moment, and the fact that it hasn’t shouldn’t be a reassurance. A formal control would provide economic incentive for Western smelters to recover indium. A lever is worth more held than pulled.

Playing Jenga

The transceiver stack is like a game of Jenga. Chinese firms, at home and in their factories in Thailand and Taiwan, dominate the top layers with final assembly and much of passive optics, but there are already American companies like Lumentum and Coherent at that level of the stack. Beijing cannot afford to pull out a block here, as a module export ban would hurt its own champions first. Europe’s controllers, U.S. signal processors, amplifier chips, and photodetectors, and U.S. and Japanese lasers sit at the middle of the tower for 200G lanes, with Chinese firms climbing the speed ladder towards them. The layer just below is wafer production — ruled by two Japanese companies and a firm that is nominally American but manufactures entirely in China. At the base lies Chinese indium metal.

At the beginning of 2025, Beijing pulled one block when it limited InP wafer exports. The tower swayed but has stood thus far. The FCC has no writ over wafers or metal, as that is Commerce’s territory, but the rule must be judged by how it changes the dependencies — and a measure confined to the top layer cannot change a dependency at the bottom. A ban on Chinese modules would pull China’s highest and weakest block for it, while leaving the tower architecture below the same. And as any Jenga player knows, towers don’t fall from the top.

If the Rule Lands

With an FCC transceiver rule pushed through by the end of this year, exemptions will come quickly. Routers were listed on March 23, and by April 14, Netgear and Adtran received approvals, with Nokia and others following within weeks. Qualifying new module suppliers, however, will take months, and increasing domestic laser capacity will take years.

The data center industry is in the midst of a major transition from 800G to 1.6T, the generation for Nvidia’s Vera Rubin systems that began production shipments in August. 1.6T optical transceivers are coming to the U.S. in two waves. The first has unfolded over the course of 2026 and is primarily Chinese. In June, Cignal AI forecast shipments of more than ten million 1.6T modules worldwide this year. The second, Western, wave should arrive in the last months of 2026 through mid 2027. Lumentum, Coherent, and AAOI all forecast 1.6T inflection before December. For lead buyers, the first round of 1.6T supplier choices has largely been made, primarily going to incumbent suppliers. A rule published this year would land in the midst of qualifications for revisions, cost-downs, and Western second sources that follow. In July, Innolight said that qualification is “long and costly” (周期长、成本高) for the customers and that the bar is higher still at 1.6T. Newer architectures are a generation further out. Both Innolight and Eoptolink predict NPO volume in 2027 and scale by 2028, with 3.2T pluggables next on the roadmap.

Innolight’s 1.6T-DR8 OSFP224 optical transceiver. Source

Whether this rule would even constrain Chinese transceiver companies is a separate question. The July order contains a component rule to stop a covered part from coming into the country inside a switch or server, but this applies “only for logic-bearing hardware components produced by an entity identified on the Covered List,” a loophole the Coalition for a Prosperous America has asked the Commission to close. Unless the FCC explicitly names Innolight and Eoptolink, their modules could freely arrive within an Arista or Cisco chassis. It also means that a CPO architecture, where the optical engine is built into the switch, may not be covered. It is still unknown whether the FCC could govern the external plug-in laser modules used for CPOs.

Suppliers primarily authorize pluggable modules via a self-declaration not filed with the FCC, and thus there is no list of existing models that can be grandfathered in. The FCC could decide to define “new” by date of first sale, but it could also structure the rule so that existing models are effectively barred as well. No public document currently addresses this. As Cignal AI noted, “revisions and cost reductions to 800G and 1.6T optics would certainly be considered new.”

A grandfather clause would be no permanent promise. While the FCC has never revoked a prior authorization on Covered List grounds, in June this year, it barred import and sale of previously authorized equipment from Hikvision, Dahua and Hytera on just ten days’ notice, refusing any exemption for existing inventory. The FCC has not yet, however, barred the continued use of previously installed hardware. Whatever is already running in a data center when the rule is published can continue to be used.9

No hyperscaler has publicly pulled orders forward or shifted allocations yet, though the balance sheets that would show it lag by months. The closest comparison comes from the export control side. ByteDance, Alibaba, and Tencent ordered around a million of Nvidia H20s before the April 2025 license rule. Under a new-models rule, the rational path is to first qualify every planned part before the rule publishes, qualify now the planned revisions and cost-downs of 1.6T modules already in service, and add second Western sources at 1.6T. A ban that is announced or anticipated but not yet effective does nothing to reduce Chinese supply. Unlike in the H20 case, an FCC rule on new models would leave already-qualified Chinese modules legal to buy, so the rush would be to qualify new parts rather than stockpile them. As is, supply is tight enough that there is little to stockpile anyway.

Filling the Gap

Outside of China, only AAOI publishes unit numbers. Across 800G and 1.6T combined, as of June, AAOI has a capacity of nearly 200,000 modules per month, with hopes to increase to 930,000 by the end of 2027. From rough proportionality to AAOI’s revenue, we can estimate Coherent’s capacity to be on the order of a million modules per month at most.10 For scale, Innolight’s overseas capacity alone, in Thailand and Taiwan, was 2.8 million modules per month in the first quarter of 2026. Most of Innolight’s planned new buildings are due in 2027.

The binding constraint is laser capacity. Coherent expects its own laser fabrication to double by the end of September and again by the end of 2027. While Coherent says it has secured substrates for this doubling, it has not named suppliers. Its June contract with AXT covers six-inch wafers that AXT’s own filing still describes as in pilot production, from a Beijing facility that AXT is expanding through 2028 despite no ability to ship to the U.S. directly. Other companies are also attempting to scale up laser production. A new Lumentum laser fab in North Carolina will book its first revenue in early 2028, and AAOI’s CEO says any new laser fab takes a minimum of “21 months to 24 months” from equipment order to volume. Per Broadcom this September, laser demand “is far surpassing supply.”

The FCC rule adds no lasers, but with Innolight unable to ship modules to American customers — 60 percent of its business — the rule lessens Chinese demand for American lasers, creating a greater laser supply for American module makers. This volume is smaller than Innolight’s customer share would suggest, however, as Innolight’s high-speed modules are mostly SiPh, which use fewer, shared lasers. Any redirection will benefit assemblers like AAOI and Fabrinet that buy many of their lasers. Coherent and Lumentum make many of their own lasers, but neither is self-sufficient, with Coherent buying a growing number of outside EMLs and Lumentum not using in-house lasers until early this year. Lasers redirected away from Chinese assemblers would ease supply for both at the margin, though to a lesser degree than for AAOI and Fabrinet. For both groups, the rule will expose the overall constraint in final assembly, cleanroom capacity, packaging automation, and yield. American supply will likely take years to compensate for the loss of Chinese capacity.

Beyond just higher transceiver prices for hyperscalers driven by constrained supply, a transceiver rule would hurt U.S. suppliers too. Some 42 percent of Marvell’s total revenue in the quarter ending August 1 came from shipments to China, not including shipments to Chinese vendors’ Thai and Malaysian fabs, though the company does not specify which products. Lumentum shipped US$804 million of products to China and Hong Kong, representing 27 percent of its fiscal 2026 revenue. Semtech, which sells chips for optical transceivers, made 47 percent of total fiscal 2026 sales to customers in China and Hong Kong. Marvell’s DSPs, Lumentum’s lasers, and Semtech’s chips would find a drastically smaller market in China — and it will take time before American manufacturers can qualify new supply. When Innolight and Eoptolink are punished, so are their American suppliers.

The largest buyer put its thumb on the scale before the draft rule surfaced. In March, Nvidia announced two-billion-dollar investments each in Lumentum and Coherent tied to laser capacity and, in Lumentum’s case, a new fab in the U.S. The money buys rights to future capacity, so a large portion of the supply a ban would divert from Chinese assemblers is already spoken for. Nvidia is extracting similar guarantees on domestic production and next-generation transceivers that the FCC’s conditional approvals would demand, but Nvidia is backing it up with cash.

A Chinese Response

An FCC rule on modules pulls one lever for China on the assumption that it will hurt China more than it does the U.S. Even if we grant that, Beijing has options. The indium phosphide licensing regime already exists, and further slowing permits requires no extra announcement. The permits AXT has been waiting for to export InP wafers to the U.S. could never come. “So be it”, you say, “the Japanese companies also make InP and can just ramp production.” This brings us to the next layer down. Raw indium metal cannot ramp in nearly the same time frame. Cignal AI suggested the phrase “mutually assured destruction.”

The next generation of transceivers will not necessarily be a Chinese story. American companies are clearly ahead on CPO. Broadcom delivered CPO switches as far back as 2024, and Nvidia’s are “now in production,” while Innolight has only said it “intend[s] to develop a suite of XPO, NPO, and CPO product offerings.” NPO is still an open race. American and Chinese companies expect volume in 2027, but no company has NPO revenue greater than a rounding error. Future architectures still rely on Chinese wafers, but an FCC rule built for pluggables may not apply.

There’s a certain symmetry here. Should the rule be enacted, each side would hold a license over the other’s optics, pulling the same factory in opposite directions. An FCC conditional approval requires a plan to build in the U.S., and a Chinese InP export permit is easier to obtain for a factory anywhere outside of the U.S. How that resolves is anyone’s guess. The salient point is the time it takes for each country to gain independent stacks. Historically, China is fast in standing up component manufacturing, and domestic data center growth could give Innolight an eventual home market. The U.S. has been slow to build, and with the rug fully pulled out, its gaps sit much lower in the stack than China’s. If the logic of the FCC rule is preventing structural dependence, funding domestic wafer and indium production beats fencing off Chinese components.

Carrots Not Sticks

The potential FCC rule is misguided in its hardware logic. The microcontroller in the transceiver cannot see above the bit level or transmit data out, and could only reach the host through a bug in the switch. Disruption is possible, but limited, as a compromised module only takes out its own links, and modules are hot-swappable. Fear of sabotage should prompt firmware authentication and multi-sourcing supply. It does not justify a ban on a whole generation of modules.

The alternate explanation is that the FCC is seeking to stretch its purview, couching industrial policy as cybersecurity. The Covered List has already been used as an industrial policy instrument. Conditional approvals require a plan to build in the U.S. The FCC’s overall worry is correct, but its aim is at the wrong layer, as the U.S. is most exposed to scarce lasers, Chinese-made wafers, and Chinese-refined indium. Neither Beijing nor Washington can afford to pull the assembly lever, and no top-level rule on modules will lessen China’s control over the wafers it licenses or the indium it refines. The FCC rule would redirect laser supply and pressure assemblers to build in the U.S., but it would come with a dramatic time lag and change nothing about the underlying dependencies.

Nvidia has already demonstrated a path forward for tackling the middle of the stack. It provided American component makers with cash up front and purchase commitments down the line in exchange for capacity rights, all tied to building out U.S. manufacturing. Washington cannot match Nvidia’s cash flow, but it has effective instruments at every layer.

At the module level, the DoW could create multi-year purchase agreements for non-Chinese modules, and Congress could pass tax breaks for hyperscalers that source from American transceivers.
At the component level, the 45X manufacturing credit could be extended to cover EML and CW lasers, a new round of CHIPS-style grants could be given to U.S. laser manufacturers to expand capacity, and new grants could support existing DSP companies into the next generation.
At the wafer level, the U.S. should provide heavy incentives to AXT to build new factories in the U.S. and turn it into a domestic champion rather than an American company that operates only with Beijing’s permission.
The hardest layer is indium metal itself. Domestic investment in indium recovery has been limited, because new refiners can’t compete with cheap Chinese supply. A Chinese indium export control would create an incentive overnight at a drastic cost, but a preemptive price floor or an offtake could build capacity before a crisis.

None of these solutions involve the FCC, because the FCC is not equipped to do industrial policy. The lesson lies in sequencing. Any protective import control must follow, not precede, domestic capacity. When a rule comes first, as the FCC is threatening to do again with transceivers, it converts a latent dependency into a live one and gives China the opportunity to pull lower levers with much greater consequence.

Precautionary bans are not industrial policy. They represent only a naive belief that the U.S. can decouple from China without investing in its own manufacturing.

A second comment by ITI does not mention transceivers specifically, but states that it “strongly opposes” location-based Covered List entries in general. Comments on the FCC’s Third Further Notice of Proposed Rulemaking closed September 8.

In June, Innolight was included on the Pentagon’s list of Chinese military companies, meaning the Defense Department “may not enter into, renew, or extend a contract” with the company as of June 30, 2026, and may not buy goods “produced or developed by” it from June 30, 2027. As the law exempts components that arrive as part of a larger product, it may not cover a transceiver inside a switch at all, and it is unclear whether a cloud service “includes” the hardware it runs on. It also grandfathers contracts signed before those dates, meaning the Pentagon’s existing cloud contracts with Google and other hyperscalers are untouched, as is everything hyperscalers build for commercial customers.

One way that data centers push this limit is through Active Electrical Cables (AECs). AECs use an integrated circuit in the cable connector housing that pre-distorts and post-corrects the signal to compensate for the loss. At the 400G data rates common a few years ago, AEC could stretch to seven meters. Now, with data centers speeding up to 1.6T, even three-meter rack-to-rack connections are sometimes laid with optical fiber.

The heat from the switch causes the laser to drift, so most CPO systems now have an external laser source (often standardized as an ELSFP) that can plug into the faceplate, then send constant light via fiber to the CPO where it is modulated into the signal before travelling out via separate fibers.

The longer haul modules for use in sending signals between data centers do handle encryption themselves, but these transceivers are made by American companies like Ciena rather than the Chinese vendors discussed here.

Within the controller, for each data lane, the Explicit Control bit determines whether the switch or the module itself tunes the signal. When this bit is set to 0, the standard says that "the module internally programs" its own settings "in a best effort manner." The standard contains no recommendation for a default.

One of Eoptolink’s patents describes the controller as the management core running a borrowed open-source operating system, but the company has never specified whose chip it actually buys. Nor has Innolight. Twice this July, investors inquired after MCU provenance on the Shenzhen exchange's public Q&A platform. The questions remain unanswered, as do the same questions put to two other Chinese module makers. The only publicly published teardown found an ST part in a module built in Malaysia, and estimates for how much of the MCU is Chinese range from nearly all of it to almost none of it.

In June, AXT and Coherent signed a three-year supply agreement for six-inch wafers with $22.3 million prepaid and capacity to be built at AXT’s Beijing plant. Lumentum likewise signed a six-year reservation with AXT a month later.

A side door does exist, but the FCC is also proposing to shrink it. Right now, for the purposes of “testing and evaluation … product development, or suitability for marketing,” anyone can import up to 4,000 units of an unapproved device without the permission of the Commission. The July proposal drops that to 40 for any part on the Covered List, not nearly enough for a hyperscaler to qualify a vendor’s next part.

While Coherent, the largest American producer, does not release production numbers, we can obtain a very rough order of magnitude estimate by scaling AAOI’s disclosed capacity by the ratio of the two firms’ revenue. Based on the last dollar figure disclosed and its reported growth rates since, Coherent’s data center business was approximately $0.99 billion in the June 2026 quarter, roughly five and a half times AAOI’s total revenue. This would put Coherent at a roughly a million modules per month as a ceiling.

Data is the Hard Part

Jordan Schneider — Tue, 06 Oct 2026 11:09:35 GMT

Bharat Patel is currently Accenture’s AI and data lead for its defense portfolio, a former enlisted Navy sailor, and a veteran of the Army acquisition world, with experience at MITRE and across the Department of Defense. We’re here today to talk about why data is the hard part in making AI work to win wars for you — and what standards, integration processes, and governance you need to turn fun tech and cute little language models into things that can actually serve the warfighter.

We’re also going to talk about the dark side of data — how poisoning can happen, the enemy’s vote in the datasets you build to find things on the battlefield, and what it’s like to be a civil servant trying to make big programs happen inside the acquisition bureaucracy.

Our conversation covers:

Why data is the real bottleneck for military AI, why “AI-ready data” is a myth, and why autonomous tanks are further off than you might think,
Ukraine’s autonomous systems as a case study: how years of collecting, labeling, and learning from battlefield data created the foundation for autonomy,
The boring side of AI: how pipelines, standards, governance, and testing and evaluation determine whether models ever make it into the field,
The adversary’s AI toolbox — including camouflage, deception, electronic warfare, and poisoned data — and how to test against those threats.
The 4D chess game of building programs inside the Pentagon: navigating requirements, funding, acquisition rules, and the people who make or break a program.

Listen now on your favorite podcast app.

Thank you to Accenture Federal Services for bringing us this episode.

The Dark Side of Data

Jordan Schneider: Bharat, why is data the hard part?

Bharat Patel: We talk about AI-ready data, but in my opinion, data will never be AI-ready. Data will always have its weirdness about it. It’s more about what the AI use case is, because the use case drives the type of data you need and the quality you need it in. We often focus on the wrong problem. Just figuring out which AI use cases you want to go after will help us determine the data quality and how to pursue it.

Jordan Schneider: Why don’t you pick a case study and walk us through how data can be used and fit for purpose?

Bharat Patel: I’ve had a ton of experience across the Department of War on different AI projects. One of the most famous was Project Maven, started in 2017. The first thing we all struggled with was the data itself. We were trying to build computer vision models against a particular set of targets, and then trying to get enough data that actually had those targets in it.

At the start of Project Maven, we were getting a ton of data that didn’t contain relevant targets, so our models weren’t performant once we put them into operations. We quickly learned that we had to continuously collect relevant data to build those models. That was one of my first early inroads into understanding the use case tied to the data, and then getting enough of it to build performant models.

Jordan Schneider: How do you get an organization to start taking data seriously, and what is the infrastructure? Let’s start with the cultural piece. How do you spread the gospel on data first?

Bharat Patel: First, you have to find your battle buddies who already understand some of the gospel. Once you find them spread across different parts of the organization, you create a culture to bring them up, and they advocate to leadership that this is important. But you can only do that with evidence.

One thing I was involved in was through the Army Research Lab. We wanted to build a model for tanks — second-gen FLIR tank data. There was no tank data, so we had to figure out how to get it into an environment where we could apply AI techniques. The team had to build a data collection box, put it right next to the sensor they wanted data from, and collect enough of it to bring back into an environment.

Jordan Schneider: What data do you want from a tank?

Bharat Patel: At first, it was just the second-gen FLIR sensors. It was more about the sensor technologies, and at that time, it was imagery. But there are other sensors on a tank that we just didn’t get to. If we’d had that data too, now you’re talking about multimodal data and what that can do for autonomy, target tracking, and target execution. We didn’t get that far. We started with baby steps — just imagery — but now you’re seeing an explosion of multimodal data, because that’s the only way you’ll get to autonomy: not through one sensor, but multiple.

Jordan Schneider: Let’s ground ourselves with the dream. What do you want all this data to let a tank do in 2035 or so?

Bharat Patel: The first thing we’d want a tank to do is make sure our military men and women are safe. More autonomy will save lives. Getting a tank to do things a human can do — we’re still a long way from that, but we can all dream, right? So 2035 gets us an autonomous vehicle that can understand its environment, make decisions, and save human lives as people sit back and watch.

The TacFLIR 280-HDEP, which combines AI with thermal and optical data. Source.

Jordan Schneider: Do you still need human beings in tanks? Why isn’t everything a UGV today?

Bharat Patel: Because autonomy still goes wrong every once in a while. You don’t necessarily hear about all of it yet, but it should be more evident. We need a human-in-the-lead because you want people to verify what the autonomous vehicles are doing. We also have a manufacturing issue — though those are ramping up due to some things coming out of the White House.

Having tons of autonomous platforms out there running wild is just not where we are as a nation. A lot of it goes back to whether we have the right policies in place for these things to autonomously do things. It’s very easy to mistake an enemy combatant for a civilian, or an enemy vehicle for a civilian vehicle. There’s a lot we’re just uncomfortable letting loose as a nation.

Jordan Schneider: Waymos can drive better than me, so this seems not impossible. Maybe a Bradley just isn’t built to operate without human beings moving stuff around and pressing buttons. But what you see in Ukraine right now is still a lot of dudes driving trucks, plus a handful of cool things doing casevac or holding a position on a corner of a tree line. It’s been surprising to me that even in a hot conflict — where there are fewer civilians running around and maybe less stress about those questions — it hasn’t gotten further. Is it a data challenge? A compute-at-the-edge challenge? What’s holding us back?

Bharat Patel: That’s a phenomenal use case. What you see today in Ukraine is not how it started. For a very long time, there was an active data collection strategy happening, unbeknownst to the people who were there. A lot of the technologists who saw the future of warfare understood they needed to make that data available.

They didn’t start with semi-autonomous or autonomous platforms. They started with regular humans flying around with joysticks. But over time, what got recorded — video, different sensors collecting — that data made it back to a central place. People started to realize, “I can build things and make them more autonomous. I can identify targets a little more. I can identify things in this data.” Then they brought it back into the platforms to make them autonomous.

I don’t see that happening at mass quantity within the department. There’s no active data collection strategy for machine learning, for continuous machine learning. There are exercises and activities, but many times — because storing data and doing something with it afterwards is resource intensive — it just gets erased, deleted, or dumped into a big storage area where it loses context over time, and people don’t even know it’s there.

A UGV. Source.

Jordan Schneider: Your point is that whatever autonomy we’re seeing in Ukraine today is because they’ve spent three years collecting and labeling data. On the one hand, we’re not fighting a war, so there’s less data to collect. But it’s also a lack of focus on the US government’s part to set up the relatively boring things you need — the custody, having all this sitting on a hard drive ready to train on.

Bharat Patel: I 100% believe the boring side of AI is now starting to become more popular. The folks who’ve been in this area understand what it takes — data collection, processes, standards, governance, and infrastructure, all the boring stuff. That hasn’t been a high priority. I’m seeing some good things out of the CDAO and their scaffolding initiative that start to get after it.

But the department still doesn’t have the right folks across all the spaces to understand how important it is to continuously do the boring stuff of AI to move any of this forward.

Jordan Schneider: I guess one of the “silver linings” of fighting a horrific war is that you have a lot of data to train on. To what extent can you get the data you need from training exercises and synthetic data, versus actually needing the real thing to build models that achieve the autonomy you’re hoping for?

Bharat Patel: Taking advantage of the exercises the military goes through is one avenue. But there’s always a question of realism — they may or may not be representative of actual warfare. In my mind, the more important thing is the reps and sets: how quickly can we collect data, bring it back into an environment to be trained, do testing and evaluation, and push models back out to where they need to go?

It’s really about building a true pipeline that can rapidly move things from one end to the other, versus trying to build reliable AI models based on internal operational exercises. What we saw in Ukraine was a non-traditional way of executing a war. The way we run our exercises may not be representative of what the next fight looks like. So the more important focus should be on the pipeline, the processes, and the governance required for AI to be updated repeatedly, safely, and in a trustworthy way.

Jordan Schneider: Whose job should this be? Is it something the government should outsource? Should they train their own models? Should it be the weapons manufacturers, or a whole separate data team? What’s the ideal setup?

Bharat Patel: The ideal setup is that the government is responsible for managing the data and keeping it in the government’s hands. The pipeline, the process, and the governance are the government’s responsibility, but the work itself should be outsourced as much as possible.

Government engineers should be doing government development so they’re intellectually aware, can keep up with the state of the art, understand industry’s ability to meet requirements, and have those conversations with industry. But industry should be doing about 90% of the work. There’s a data management part, a data labeling part, a test and evaluation part, a deployment part, a model training part — a ton of areas for industry to participate. It’s up to the government to break those into a modular approach that lets us continuously iterate with industry without ever getting locked into a single partner longer than needed.

Acquisitions and Reforms

Jordan Schneider: What is it like as an acquisitions official to try to stand something like this up? What constraints are you operating under?

Bharat Patel: Oh man, all types — from personalities to policies to resources. But before we get into that, let me tell you how much fun I had as an acquisition professional. You mentioned I worked at MITRE. I started my career with the Army acquisition community in 2007, supporting the Program Executive Office for Intelligence, Electronic Warfare and Sensors. We were still in Iraq and Afghanistan doing things, and the portfolio was supporting it. I got to see all of that firsthand. Then I moved to Project Manager Distributed Common Ground System–Army around 2016 as a federal employee — but always within the acquisition community, always within the same PEO construct.

That entire time, I fundamentally loved every day, every minute of it, because — I always call it “the show.” Any good movie has a producer or production lead, and as a PM you have so many parts you don’t own that you have to bring together to get the job done. To stand something up, I don’t own the requirement, so I had to work with our requirements owners to help shape what a pipeline requirement could be. Then I had to talk to the resources folks, the acquisition headquarters, and the contracting folks. Everybody has a say. Test and evaluation organizations have a say. All these pieces have to come together to create the program.

Of course, industry has one of the biggest votes, because industry is either going to barf on the entire strategy or get excited. The way we originally broke up what we called Project Linchpin — now Product Manager AI Ops and Services — was to allow maximum industry participation in the pipeline. Those are some of the constraints and joys of standing up a program.

Jordan Schneider: The way you describe it, the technical engineering fluency of the thing you’re building is maybe a quarter of what you need. The 4D chess of bureaucracy — getting the colors of money right, getting people aligned in the building — seems to be just as important, if not more so, than understanding the physics or engineering behind it.

Bharat Patel: You have to be able to navigate the Pentagon. Everyone really does want to move things forward, but not everyone will jump on a singular idea and push it. Some people have their own opinions even after something’s been bought off on, and you still have to move it forward. So you have to figure out who those people are and how to bring them on early. It becomes a people, personality, and organizational-understanding game to really move the needle on future programs.

Jordan Schneider: Let’s talk about the reforms we’ve seen over the past two years — Portfolio Acquisition Executives. What’s the vision behind that restructuring that Feinberg has put into the building, and what can and can’t it change?

Bharat Patel: Acquisition reform has been around forever. For the 20 years I’ve been part of this community, it’s always been about acquisition reform. There have been so many attempts — from policies put in by the acquisition community to the Army standing up Army Futures Command and cross-functional teams to accelerate acquisition, and now the latest with the PAEs and CPEs.

It’s just another means to accelerate the acquisition community, but not necessarily one that tackles the hard part: holding the program manager — who is actually authorized to deliver the capability — accountable, and then giving them the appropriate resources to get it done.

Now you’ve got PAEs. What I see, at least for the next few months, is that nobody is exactly sure what needs to be authorized by the PAEs versus what can happen at the CPEs, the PMs, and the product manager level. All of that is still being worked out. In theory, once it’s worked out — because the PAEs have a little more pull from a requirements and contracting perspective — it’s supposed to accelerate the ability to change requirements and move faster. We’ll see how it plays out.

Jordan Schneider: Another thing the department talks a lot about is making speed something we optimize for in acquisitions. What’s your vision for getting those incentives right?

Bharat Patel: Start small. It starts with the humans who are actually motivated to get something done. If the humans on the acquisition team aren’t motivated, you’re never going to accelerate. I’m so tired of “we’ve got to go fast, we’ve got to go fast,” and then nobody ever really goes fast — because we have really good people, but they’re scattered.

The first thing I’d do, if something is high priority, is make sure my highest-speed individuals are on that acquisition team. They will move mountains. I was able to move mountains, and I brought people in to help me move even further. But it goes back to having the right people with the right incentives.

There’s a lot we need to do on the government side to properly incentivize and resource the people we want executing these big programs.

Jordan Schneider: You also worked on fielding capabilities for people fighting in wars today. I’m curious for your reflections on wartime versus peacetime acquisitions.

Bharat Patel: This is where it gets very interesting. Wartime quick-reaction capabilities pop up like that, and we can get them on contract very quickly, because some policies are just waived off. But in peacetime, there are a lot of federal acquisition regulations, competitive prototyping, and competitive source-selection activities you simply have to follow.

A lot of that shifts when you’re doing QRCs — quick-reaction capabilities — and you have to field stuff that will save lives or deter our enemies and threats. Things happen much quicker in wartime, and there are a lot more resources behind it: not just your regular service budgets, but contingency operations resources that let you move a ton faster. Contracting happens faster. And suddenly there’s a lot less concern about modular system architectures or being able to swap out components. There’s more concern about moving fast — we’ll figure out the seams later, as long as we can save lives.

Synthetic Data and Data Poisoning

Jordan Schneider: Let’s come back to data. Synthetic data — people are excited about it, people are worried about it. How useful is it for the targeting, autonomy, and physical real-world stuff we’re talking about?

Bharat Patel: My synthetic data journey has gone like this. It started around 2022 when I was supporting the TITAN program — Tactical Intelligence Targeting Access Node, the next-generation ground station capability. One of my hopes was to accelerate the AI. Before we even fielded some of these capabilities, I wanted to see if we could pre-train models.

I partnered with the Small Business Innovation Research program — I was lucky to be the first to use the Direct-to-Phase-II concept — and awarded four Direct-to-Phase-II contracts to companies to generate synthetic data. The first thing the companies said was, “Hey, do you have representative data?” And I thought, “Oh crap.” So it goes back to data being very important. Synthetic data is really dependent on the actual data, so the companies can replicate it. You need enough representative data for synthetic data to work.

There’s a lot of variation in maturity across our industry partners on synthetic data. When it comes to autonomous vehicles, commercial industry is investing a ton — NVIDIA is doing some things, and the technology is getting there. But applied to the sensors the Army has, that’s different from the commercial sensors on airplanes or new autonomous vehicles. Military sensors could be two or three years older, with completely different look angles, different biomes, different experiences. Giving a synthetic data company enough of that data to replicate those scenarios is hard.

So we aren’t quite there yet on building AI and autonomy off purely synthetic data. Synthetic data augments operationally relevant data — the combination gets you a bit more performance. The way I like to use synthetic data is to figure out the edge cases you don’t have operationally relevant data for. Replicate those edge cases so you can see what the model does against weird situations you just haven’t seen before.

Jordan Schneider: Another difference between training Tesla Autopilot and your tank model is that no one is walking around the street expressly trying to trip up a Tesla into driving into people. That’s not necessarily the case when you have a human being trying to trick a robot that’s trying to shoot them. Talk a little about the adversary dynamics when it comes to data and what folks should expect or be prepared for in the coming years.

Is data poisoning something to be particularly worried about here? How would you define that phrase?

Bharat Patel: In many ways, we should be worried about it 100%. There are two scenarios. There’s active poisoning happening right now. The open-source environment — open-source technologies, databases, and model frameworks — is out there to help accelerate AI. But there are also people contributing to the negative aspects, putting in bad data for fun.

I sometimes equate it — not exactly, but depending on my mood — to being a jerk on Waze: “cops here, cops there,” even when they’re not, when I’m just being a jerk. I put it there, and now as part of the enterprise, that biased data is in there. So there’s either purposefully bad, biased data, or sometimes by accident, somebody mislabeled something or added the incorrect equation into the data. The models get trained against that, so they come out biased in some way. That’s the red-teaming aspect.

Then, once we’re in production with these AIs fielded, it’s still fairly easy to put camouflage over something — or spray-paint graffiti over a target — and it no longer registers as a target, because the model has never seen a target with graffiti on it. There are a lot of ways to spoof these capabilities, especially in computer vision. Even in the electromagnetic spectrum, you can put out fake signals that replicate the size of a large enemy movement. All you need is a wagon with a bunch of sensors on it that acts like a military movement. Without visual confirmation, it looks like a military movement.

Jordan Schneider: Given the mess all of that creates, what do benchmarks and test and evaluation need to look like for these things?

Bharat Patel: There are a couple of aspects. One, we can’t overtest and over-evaluate because we need to actually get the model and the capabilities into the field so they can experience the environment. We have to limit how much T&E we conduct — while still making sure the models are safe and we understand the risk of deploying them.

This model needs to be, say, 80%, because the enemy is going to have a vote, the environment is going to have a vote, and the warfighter using the model is going to have a vote. So 80% might be good enough for them. Not overly testing models is 100% one of the things we need to accomplish.

The other part is partnering with our intelligence community to understand all the possible activities adversaries use to spoof models, then building that into the T&E process so we can build a robust model. If we understand the enemy is graffiti-ing their targets or platforms, maybe we generate that as synthetic data and inject it into the T&E and training dataset, so the capability can detect and overcome it.

Jordan Schneider: You mentioned earlier you want to write a sequel to Pentagon Wars – the first one is on HBO Max. The challenges of doing test and evaluation on a Bradley fighting vehicle are relatively straightforward: here’s a range of temperatures, a range of ammunitions, it could be hit by this, that, or the other. But when you’re talking about what an autonomous vehicle in a fight, broadly defined, could face in the future, there are so many more variables — and surely things you can’t even come up with sitting here in 2026. How do you navigate that uncertainty when figuring out how much to trust these models in autonomous vehicles?

Source.

Bharat Patel: One, we can’t hide from it. We’ve got to figure out a way to go do stuff safely. A lot easier said than done. AI is here, and we’ve got to start exploring and getting into the right environments where we can do some of these edge cases that aren’t necessarily safe or survivable. We’ve got to be able to do that.

And two — yes, Pentagon Wars 2: The Race for AI. It’s going to be great. It’s a calamity — my journey, surrounded by really cool senior leaders and maybe some not-so-cool ones, of creating a program and all the great things that came with it: the conferences, flying on military air, talking to congressional appointees and senior leaders, building partnerships across the entire Department of War. It was a phenomenal experience.

Transitioning that into how we’re going to do this — like I said, you’ve just got to go after it. We really do need to focus on the boring parts of AI: test and evaluation, verification, validation. When I was starting Project Linchpin, there was NDAA language about a study that concluded the department really needs to centralize on a standard for test and evaluation, because that’s how you scale AI. Once you understand how to ensure it’s reliable, you’ll be able to field in large production quantities, versus these one-offs where you can’t understand the capabilities and limitations, who’s using the AI, and what decisions and recommendations they’ll make based on its output.

Jordan Schneider: Okay, you have your Bradley, and maybe every six months you’ll put new armor on it. But the promise and peril of being able to update the software on your robots daily or weekly is kind of the answer to that test and evaluation problem, right? Yes, you’ll run into things you weren’t predicting, but as long as you can capture that data, learn from it, and let the robots get smarter, that’s the answer — as opposed to trying to cover all your bases before the serious stuff starts.

Bharat Patel: 100%. There are a couple of aspects. Way back when, the department decided to separate hardware and software. We’re now at a point where we’re ready to separate hardware, software, and AI. AI is really about data and models, whereas software is the application, the functions. If you can separate software from AI, you can speed up your AI process for rapid, repeatable capabilities. You don’t have to update your entire software package — just the model.

Now we’re bleeding into the evolution of the RMF, the Risk Management Framework in the cyber process. Right now, cyber treats everything as software: you accredit the whole thing, then re-accredit the whole thing. But maybe we don’t need to re-accredit the entire software package — we just need an updated model. Before I left, the plan was to separate AI from software so that, from a cyber perspective, we could do that, and from a technology perspective, we could push out more updates.

Jordan Schneider: Let’s close on some of the work you’re doing at Accenture. Why should AFS exist in the first place? What’s the role of these large federal contractors in helping create the stuff we’ve been talking about?

Bharat Patel: I’ve been here about six months. I feel like I was blessed in my career — I have a pretty large network and could have gone almost anywhere I wanted. I selected Accenture for a couple of reasons.

Right now, the department is 100% fascinated by products. But once the shiny-toy phase of products is over, the department is going to realize it needs a qualified integration team to make everything work. What Accenture does on a global scale is literally that: they integrate new technologies and leverage partnerships, physically integrating new technologies into Google, Caterpillar, Nike — these large integration efforts from enterprise to edge. I knew the department was eventually going to realize it needs qualified integrators, and when that happens, Accenture is going to be spot on.

The other part is that Accenture is building a ton of partnerships — with OpenAI, Palantir, Anthropic, Databricks, all these bleeding-edge companies. We’re going to be well positioned to close the gap and meet requirements rapidly, because we have qualified engineers across all these technologies. Once we’re on contract, we’ll be able to help our government partners accelerate their capabilities.

Jordan Schneider: Any specific projects you’re excited about? I hear you’re hiring as well.

Bharat Patel: We’re hiring a ton. Accenture is at the cusp of blasting off.

The part of Accenture we’re working in feels like one of the largest startups. We’re in a constant rat race: What’s next? How do we stay competitive? How do we stay in front of our competitors? How do we understand — and shape — what the government needs, because we’ve understood that environment for a while? We’re projecting some great things in the next two or three months, and then we’re going to need solid people to come on board and join the culture, join the team that’s changing the landscape for companies like us.

Thanks for reading ChinaTalk! Support us by sharing this post with a friend.

China’s AI Safety Money Problem

Nick Corvino — Mon, 05 Oct 2026 11:09:38 GMT

Here are two claims I hear about China and AI safety:

China is intrinsically less concerned about AI risk. Chinese people are techno-optimists, while the Party sees doom talk as either Western paranoia or a containment strategy designed to hold China back — such as many of the dismissive responses to Dario’s call to pace the frontier.
China is now safety-pilled. This refrain gains traction once or twice a year when a high-up like Xi gives a speech about controlling AI risks (as he did at the World AI Conference), or China puts forth a new international initiative, such as the September UN Global Dialogue on AI Governance.

There’s evidence for each of these takes (see footnote).1 But I think both sides miss the messier reality on the ground.

I met plenty of ambitious and talented people when I lived in Beijing who would have loved to work on AI safety full-time. But a graduate in China who cares about AI risk is still expected to get a “real” job eventually. Outside an underfunded PhD track or a handful of positions in a country of 1.4 billion people, there isn’t a well-trodden path. In San Francisco, DC, or London, worrying about AI professionally can be a very-well-paid career. In China, it’s more of a hobby.

As a result, I think the main bottleneck facing China’s AI safety ecosystem is not ideological but material…

Independent Chinese AI safety has basically no money.2 And the limiting funding structures that do exist determine what kinds of safety work can survive, and where.

Subscribe now

Today:

Why China’s philanthropic rules make it difficult to fund independent organizations, even if the money exists.
Why this leaves it up to the Chinese government to lead on safety work
Technical safety research and how it is continuing to grow
China’s “safety-as-a-service” companies which try to make safety profitable

A “third wave of American philanthropy” appears to be on the horizon, alongside initiatives like Coefficient Giving’s Project Tailwind, potentially enabling much more ambitious AI safety efforts in the West. Understanding how this influx of resources could reshape the gap between Western and Chinese AI safety ecosystems (and the relationship between them) therefore seems increasingly urgent.

[Nathan Labenz explores related themes in his dispatch from China, which I’d recommend checking out.]

A Brief Overview of Chinese Philanthropy

China has plenty of billionaires. Forbes puts the number at 539 billionaires in China worth $2.2 trillion. (The US counts 989 billionaires worth $8.4 trillion.)

It would take only one or two of them getting really into AI safety to finance a sizable funding ecosystem. So why has no Chinese equivalent of Coefficient Giving emerged?

The answer is related to the strange structure of Chinese philanthropy.

The modern Chinese charitable sector as it exists today expanded rapidly after the 2008 Sichuan earthquake, but it was plagued from the start by corruption scandals, inconsistent regulation, and government concerns about grassroots organizations operating outside of the state’s purview.

Memorial dedicated to the victims of the 2008 Sichuan earthquake at Xuankou Middle School. Over 69,000 people lost their lives. Source.

To streamline the system, the 2016 Charity Law 慈善法 clarified the legal status of domestic philanthropy, tax incentives, and fundraising. Charities would have to register, disclose their finances, and accept government supervision.

The goal wasn’t to reduce or get rid of philanthropy. Beijing still wanted private money to support education, health, poverty relief, and other social services. But it wanted civil society to help deliver state-approved public goods, rather than become an independent source of political organization or agenda-setting. Article 4 of the Charity Law made this principle explicit enough by stating that “charity work shall be conducted under the leadership of the Communist Party of China.”3

Scholars Angela Bies and Scott Kennedy write:

“There exists a current paradox of philanthropy and the state in China: a stronger social sector accompanied with growth in philanthropy is desired, but such strength and growth must also contend with the state’s own policy agenda and desire to limit aspects of advocacy and civil society typically associated with the third sector.”

Subscribe now

The Charity Law may have legitimized the sector, but it has not erased longstanding distrust. China has long ranked near the bottom of the Charities Aid Foundation’s World Giving Index, which measures behaviors including donating money and volunteering. In off-the-record conversations for this piece, I likewise heard skepticism about donating due to corruption.

^The 2011 Guo Meimei 郭美美 fiasco, one of many scandals that badly damaged public confidence in major charities.

Perhaps some of this lingering distrust also reflects another impetus behind the 2016 reforms: limiting the influence of foreign organizations operating inside China — such as groups focused on human rights. This led to the Overseas NGO Law 境外非政府组织境内活动管理法 (2016), which, for AI safety, may actually be more consequential than the Charity Law.

The Overseas NGO law mandates that foreign nonprofits must register an office or work through an approved Chinese partner, with the system overseen by public-security authorities. Shawn Shieh, a research associate at the British charity ODI, found that many Chinese non-profits went from receiving 80–90% of their funding internationally to 80–90% from domestic Chinese sources in the years following the Overseas NGO Law, effectively flipping the philanthropic funding ecosystem on its head.

This should help clarify why supporting AI safety work in China is unlikely to be as straightforward as simply giving more money to Chinese non-profits to work on safety. In practice, foreign funding has to flow through organizational structures that are acceptable to the government. Even if Beijing fully accepted the case for addressing AI risks, it would still be unlikely to permit foreign funding on a large scale to flow into Chinese organizations.

Chinese charitable organizations received roughly $21 billion in donations in 2023, compared with $557 billion in the US. That’s a 27x difference.

Data from Giving USA 2024 and Giving China 2024 中国慈善捐赠2024. I might be off by a billion here or there.

Companies supplied 77% of Chinese giving in 2023, while individuals provided only 22%. In the US, the pattern is reversed. Individuals supplied 67% of charitable giving in 2023, while corporations accounted for just 7%, with foundations providing another 19%.

Specifically for AI, I estimate Western AI safety philanthropy is $250-600 million per year.4 And I predict that number will push well beyond $1 billion by next year, after companies like Anthropic and OpenAI go public and usher in a potential “third wave” of American philanthropy.

What Chinese money does exist flows toward established causes. Health received 37% of donations in 2023, education 25%, and scientific research around 10%. Among the country’s largest philanthropists, 70% donated to education. Harvard’s China Philanthropy Project similarly finds that education dominates elite giving. This fits the logic of the Charity Law, which encourages philanthropy in politically approved areas already recognized as socially useful.

Chinese charitable giving by sector, 2023. 7.08 RMB = 1 USD. Figures come from Giving China 2024 中国慈善捐赠2024, published by the Tsinghua Institute for Philanthropy and Yishan Data in April 2025. Source.

State-Sanctioned Safety

Perhaps in part because of these restrictions on charitable giving, AI safety in China tends to be treated less as a philanthropic domain than as a responsibility of the state. China’s largest AI safety institutions are government or government-backed, like Shanghai AI Lab (SHLAB), the Beijing Academy of Artificial Intelligence (BAAI), the China Academy of Information and Communications Technology (CAICT) and TC260.

Subscribe now

None of these institutions primarily focus on safety. CAICT is a massive Ministry of Industry and Information Technology (MIIT) think tank. BAAI trains models for capabilities, such as Emu3.5. Safety is one function inside a much broader mission, competing for resources with everything else that mission entails.

^BAAI’s Emu3.5. Looks interesting but doesn’t seem to have anything to do with safety.

Shanghai AI Lab (SHLAB) comes closest to an exception. SHLAB has a dedicated Safety and Trustworthy AI program, accounts for roughly one in ten papers in Concordia’s Chinese frontier-safety database, and this year released the Shu’an (书安) AI safety platform. But it is also working on capabilities, shipping the InternLM (书生) model family.

Zhou Bowen 周伯文, who runs SHLAB, also chairs TC260’s AI Safety Standards Working Group (WG9), arguably the most consequential AI safety body in China. But the number of positions inside a state-sanctioned body like TC260 is small.5 And the work itself is different in kind from an independent safety ecosystem.

SHLAB’s director, Zhou Bowen, argues for an AI-45° Law, the idea that safety and capability should advance together along a 45-degree line, never sitting below it for long and never far above it either. Source.

When I first became interested in AI safety in 2019, ideas that are now somewhat mainstream still sounded pretty sci-fi, and certainly not politically tractable. But that distance allowed researchers to chase strange or unpopular questions without first clearing them against political coalitions, geopolitical competition, or bureaucratic consensus.

The independent US/Western ecosystem still preserves some of that freedom. A smart young researcher can leave college, join a small nonprofit, and spend years pursuing an idea that neither Washington nor a frontier lab currently considers important. A state-centered ecosystem is inevitably more credentialed and hierarchical, which makes it harder to build the same kind of open career field.

To be clear, I’m not saying this makes China’s state-backed work unimportant. WG9, SHLAB, and others are doing very impactful work! But government attention cannot fully substitute for independent organizations capable of nurturing ideas before the state decides they matter.

Academic/Technical Work

The independent safety organizations that do exist in China are generally thinly staffed or financially opaque. CnAISDA, China’s self-described counterpart to AI Safety Institutes elsewhere, launched in February 2025 not as a new organization with its own budget and staff but as a network stitched together from people already employed at Tsinghua, BAAI, CAICT, and others. The Beijing Institute for AI Safety and Governance similarly draws heavily on researchers whose main institutional homes are elsewhere. What these orgs have in common is that they sit inside academic institutions, which seem to have carved out a workable space for doing safety research.

Concordia AI published its 2026 State of AI Safety in China report alongside an interactive database of Chinese frontier safety papers, bringing the state of technical research in China into focus. Chinese institutions published around 12 frontier safety papers per month in 2023, roughly 26 by mid 2025, and 57 by April 2026, nearly a fivefold increase in two years.

In certain subsets of safety research — like agents governance — China is pumping out more research than anyone.

Concordia AI’s count of frontier AI safety research publications coming out of China. Source.

To be clear, American authors still dominate the field’s most-cited work, and safety accounts for a relatively small fraction of China’s overall AI publications (see footnote).6 The big results that actually push the field’s agenda forward each year, like alignment faking, circuit tracing, the AI control paradigm, still come almost entirely out of Silicon Valley (with Anthropic being the main contributor).

But this is also influenced by funding structures. Anthropic and OpenAI can afford to run dedicated interpretability and alignment teams because compute is not the binding constraint on their existence, whereas a Chinese lab burning every available H-equivalent hour just to stay competitive on capabilities has less comparable slack.

Most of China’s AI safety research therefore comes out of universities and state-backed labs rather than companies. Of the 28 “Key AI Safety Research Groups” identified by Concordia, 20 are academic. In the West, by contrast, frontier labs play a much larger role in both producing safety research and setting the field’s agenda.

Subscribe now

“Safety-as-a-service” Companies

Putting everything above together, China’s funding structures also help explain what I think of as its “safety-as-a-service” companies. These firms do work that could reasonably count as AI safety, but package it alongside cybersecurity, regulatory compliance, and other commercial products and services.

Scholars Wenjun Lai and Anthony Spires use the term “venture philanthropy” to describe how China’s philanthropic constraints often push public-interest work toward measurable outputs and business-like models. This seems similar to what these safety-as-a-service companies are offering.

As I covered previously, China’s generative-AI filing regime created a market for companies that help developers pass government review. Firms such as RealAI, BotSmart, and Qihoo 360 have tried to use that demand as a way to sneak in some more robust safety services like red-teaming, robustness evaluations, cybersecurity testing, and ideological compliance assessments, which are heftier topics than just compliance with the CAC.

Since I wrote that piece, the evidence that this is becoming a real market has gotten stronger. The number of officially filed generative-AI services increased from 439 in June 2025 to 988 in June 2026. The market is also becoming more institutionalized. Beginning in 2025, MPS began certifying products such as large-model security guardrails and evaluation systems.

More importantly, customers are actually buying dedicated safety products.7

But what hasn’t changed is that these companies cannot operate like Western safety nonprofits whose central deliverable is simply “make advanced AI safer.” They will only investigate the safety problems for which a customer is willing to pay. But that means the customer has to be willing to pay. Money is, once again, what it comes down to.

What’s to be done?

Is the conclusion to all this simply that the US-China safety funding gap will just keep getting larger? I’d be lying if I didn’t say maybe. The Overseas NGO Law makes direct Western grantmaking into China tough, while Western funders sending money to state-linked institutions raises obvious political and security concerns.

Subscribe now

Still, here are some possibilities.

Funders could establish things outside of China that are valuable to Chinese researchers. Pacific Compute is already experimenting with this model by putting safety-dedicated GPUs in Singapore that AI researchers can access for evaluations and research — emphasizing Chinese safety researchers in particular. Something similar could work in other third countries, giving Chinese researchers access to safety resources without transferring the resources into China themselves.
Exchange programs. Funders could help researchers spend time working together elsewhere. Or they could finance international researchers going to China, such as the Singapore AI Safety Fellowship.
Invest instead of doing philanthropy. Foreign investment in Chinese companies operates under a different legal regime from foreign NGO funding, although both Chinese restrictions and US outbound-investment rules can both still bite.8 Finding ways to support commercially viable safety companies (perhaps those doing “safety-as-a-service”) may be more realistic than independent nonprofits, though there is the risk that these companies will care more about profit and their shareholders than safety as an inherent goal.
The longer-term solution may be to fill the gap inside Chinese philanthropy itself. China has plenty of wealthy donors, but relatively few grantmakers whose job is to hunt for problems and then give researchers discretion to pursue them. A domestic AI-safety fund or regranting institution could pool Chinese money and distribute it to researchers and small organizations. It could also develop a direct relationship with the Chinese government and Western donors, turning this into a streamlined procedure rather than a bespoke process done by a bunch of smaller nonprofits. For instance, a Western funder could, in theory, register a permanent representative office, which requires a Chinese government sponsor and approval from provincial public security authorities. The Gates Foundation and other big non-profits have taken this route, though it seems arduous.

Most of these proposals are far-fetched, but so are essentially all of the proposals for global AI cooperation on AI (like AI 2040’s Plan A). And I doubt good cooperation can be conjured into existence at the moment of crisis. It will need a Chinese safety ecosystem with enough independence for new ideas to grow and enough resources for talented people to spend their careers pursuing them.

And that requires some cash.

*Nick is a visiting fellow at the Carnegie Endowment for International Peace.*

Survey data consistently finds Chinese respondents much more optimistic about AI than Americans, and that gap may actually be widening as US attitudes continue to nosedive. The 2026 Ipsos AI Monitor, for instance, found 83% of Chinese respondents saying AI’s benefits outweigh its drawbacks, compared with 39% of Americans (the biggest gap I’ve ever seen). But survey data only tells us so much, and we have written previously about how AI anxieties manifest differently in China.

Rhetoric also matters. When Xi talks about controlling AI risks, it changes what officials and researchers feel permitted to prioritize. Even if some of it is hot air, I still wouldn’t mind if Trump at least acknowledged these risks aloud. There have also been signs that Chinese officials may be taking certain risks more seriously, including discussions of cybersecurity risks, restricting overseas access to advanced models and various regulatory moves, such as TC260’s Safety Governance Framework 3.0.

Jasmine Li estimates in this post that annual philanthropic funding for frontier AI safety is around $20 million in China, compared to $1 billion+ in the rest of the world. A funder with expertise in the space gave me the same estimate.

What this means in practice is vague, but I think it’s unlikely a Party committee signs off on every donation. Organizations register with civil-affairs authorities, submit annual work and financial reports, and may operate under a supervisory agency. Separate Party rules for social organizations call for Party organizations to be established inside eligible nonprofits, allow Party bodies to vet organizational leaders, and encourage Party secretaries to participate in management meetings.

Coefficient Giving alone said its technical AI safety team expected to make more than $130 million in grants in 2025, while its AI Governance and Policy team moved more than $140 million.

According to Claude: TC260's member directory currently lists 311 member units under WG9, out of 704 total TC260 member units. For comparison, the data security working group has 328 and SWG-ETS (the emerging technology security special working group) has 202. Source.

For instance, the 2025 Global Index for AI Safety, using DBLP data through February 2025, counted 9,521 Chinese AI-safety publications and 9,319 American ones — essentially a tie. But China’s overall AI research output is much larger. The 2026 Stanford AI Index estimates that China produced 17.8% of global AI publications in 2024, compared with 7.3% for the US. On these measures, China produces roughly the same volume of safety research from an AI research base about 2.4 times larger, suggesting that safety constitutes a smaller share of Chinese AI research overall. Exact figures vary across datasets and definitions of “AI safety,” but this is the general pattern I have observed after looking at multiple sources.

Postal Savings Bank, UnionPay International, and Guangfa Bank have all procured large-model safety evaluation or guardrail systems. Z.AI has also publicly confirmed using NetEase for dangerous-capability evaluations and training-data audits, suggesting frontier labs themselves may outsource some safety work. Meanwhile, IDC now treats agent-threat detection as a distinct Chinese market, and RealAI raised several hundred million RMB this July to expand its safe-and-trustworthy AI business; Xi visited them earlier this year also and they increasingly discuss frontier risks such as deception, self-replication, and uncontrolled AI R&D.

For instance, US rules that prohibit outbound investment to “countries of concern” would likely apply to investments to Chinese AI companies, while China rules subject foreign investment of any kind to its own access rules and national-security review.

In Case You Missed It...

Phoebe Chow — Sat, 03 Oct 2026 11:08:12 GMT

Is China’s economy broken? Logan Wright, Jordan’s former boss (!) and a top expert on Chinese macroeconomics, says the financial system that powered China’s extraordinary growth is now constraining it, and that the old idea of US-China economic rivalry no longer fits. Brad Setser of CFR and Jordan push back (does Xi actually care? Can China's dominance of exports and supply chains keep the economy growing? Could AI and EVs provide a way out?), but Logan stands his ground. They also get into China's jobs crisis, what could pull China out of its economic funk, and what its slowdown means for the rest of the world.

Elsewhere, Lily Ottinger goes to Kyrgyzstan for the World Nomad Games, where she practices Russian and watches kök-börü — a Central Asian horse game played with a goat carcass — all the while weighing the gap between China’s hard power and soft power. Julian Gewirtz and Jordan size up Xi’s state visit, while Ethan Kessler asks whether Taiwan can actually build the drones it needs (part 1, part 2).

Subscribe now

Garrett Berntsen takes on the Pentagon’s bureaucracy and explains why giving government AI isn’t enough if the institutions using it don’t change too. Irene Zhang goes inside Chinese AI discourse, where one DeepSeek researcher’s vision of an AI-powered future ranges from communist abundance to Cyberpunk 2077, and where Western rhetoric about AI safety and technological containment can sound like an attack on Chinese development, making cooperation harder just as the stakes rise. And on WarTalk, the team confronts the “Potemkin Pacific”: a US military trying to cover an enormous ocean with a rather less enormous number of ships.

Plus: China’s robotics boom, China’s answer to Starlink, and whatever exactly Jordan is doing with Ye.

Building China’s Two Starlinks

Paddy Stephens — Fri, 02 Oct 2026 11:06:46 GMT

Paddy Stephens writes about space, security, and energy for The Wire China, Resilience Media and others. He’s based between London, Nice and Taipei. This is Part 2 of his two-part investigation of China’s Starlink challengers!

In the first part of this article on China’s responses to Starlink, I analyzed how China views the megaconstellation, and how it might try to disable it in a conflict. This second part reports on China’s mixed record and recent success in building not one but two major low Earth orbit (LEO) satellite megaconstellations of its own.

The Chinese government understands Starlink as both a commercial company and a national security asset for the US and partners. Over time, two major ‘Chinese Starlink’ projects have emerged, which between them cover both aspects: Guowang (国网 — literally, “National Network”), a satellite infrastructure provider primarily for the state, and commercially-oriented Qianfan (千帆 — literally, “Thousand Sails”).

First to emerge was Guowang, which serves the Chinese state, including the military. It is highly secretive, mentioned by name in few Chinese public policy documents, and public information about it remains scarce. China SatNet 中国星网, the Xiong’an-based state-owned enterprise (SOE) responsible for Guowang, seems to be working closely with the military-industrial complex and state telecommunications companies. The most obvious evidence for this, says Curcio, is China Space Time 中国时空信息集团有限公司, a joint venture between China SatNet, China Mobile, and NORINCO 北方工业, a state-owned defense contractor.

There are indications that Guowang is offering much more than just satellite broadband. Aerospace China 中国航天, a publication supervised by aerospace SOE behemoth China Aerospace Science and Technology Corporation (CASC) — itself with at least one subsidiary building satellites for Guowang — wrote in August 2024 that Guowang “is capable of carrying broadband communication payloads, laser communication payloads, synthetic aperture radar payloads, optical remote sensing payloads, and various other payload types”. CSIS assessments back this up, as do other limited pieces of evidence. In February 2024, when the system was still in a preliminary test phase and its coverage remained sporadic, Taiwanese media reported that three Chinese coast guard ships were helping the People’s Liberation Army (PLA) test Guowang close to Taiwan to provide information for future precision strikes by the Rocket Force.

According to the plan filed with the ITU in 2020, Guowang aims to have 12,992 satellites operating at altitudes of 500-600 km and 1145km, where latency is low, but the system also has at least three satellites in the much higher geostationary orbit, offering wide and stable coverage. Estimates of the mass of the satellites, which has not been publicly disclosed, put it at almost 1,000 kg; some argue it could be significantly heavier than that, making it abnormally heavy for a communications satellite and indicating it may have other payloads on board.

Compared to its lofty ambitions as a Chinese Starlink, the pace of Guowang’s progress from 2020 to 2023 was glacial. By the start of 2024, Guowang had managed to launch just 15 satellites to test the system’s capabilities. Starlink launched almost 2,000 satellites in 2023 alone.

There were a few reasons for this. Until mid-2022, Starlink had not been used extensively on the battlefield and had few commercial subscribers, so pressure to build an alternative was initially low. Similar projects elsewhere have also faced delays — Europe’s IRIS and Amazon LEO (formerly Project Kuiper) are behind schedule too.

But there were other factors unique to Guowang. Ineffective leadership may have played a role. Many of the decisionmakers were sexagenarian Party types who had spent their careers in China’s plodding, predictable aerospace SOEs — hardly the best people to lead a project trying to keep pace with the world’s most successful space company. China SatNet, the company which runs Guowang, was announced as a target for investigation in March 2023 and its general manager was removed the following month. Meanwhile, China’s space SOEs were producing few big rockets capable of delivering large numbers of heavy Guowang satellites into LEO, and most of these were already spoken for by other state missions. China’s fledgling commercial rocket companies did not have any big launchers to offer Guowang at this point either.

Ultimately, the effective monopoly on LEO comms satellites that China SatNet was given from April 2021 was clearly not the right model for building an equivalent to Starlink. “When there was only one entity, they just didn’t do anything for three years”, said Curcio bluntly. “I think the government realized there needed to be some competition, and so started supporting two projects. Qianfan had been there in the shadows for a long time.” In October 2023, the Ministry of Industry and Information Technology (MIIT) ended Guowang’s monopoly, allowing non-SOEs to enter the satellite internet industry.

Enter Qianfan

The murky backstory for Qianfan actually starts in 2017 with the creation of KLEO Connect. Core to this Sino-German joint venture was (what would later be known as) Shanghai government-backed Shanghai Spacecom Satellite Technology (SSST), which, using the German side’s spectrum rights, aimed to build a rival to Starlink. Instead, the venture ended in acrimony and lawsuits after the Chinese side launched two satellites in August 2021 for a separate constellation that the German side said matched the planned KLEO Connect constellation in technical specifications and other details.

After the joint venture broke down, in November 2021 construction began on a factory in the G60 Science and Tech Corridor in Shanghai, SSST’s base in Songjiang District, to build a new “G60 Starlink”. Having lost access to KLEO Connect’s spectrum, G60 also quietly filed for new rights in April 2023.

All of this was going on from 2021 to 2023, the years in which Guowang had a monopoly on satellite internet. G60 Starlink quietly developed, but its prospects were unclear. But when MIIT ended Guowang’s monopoly in October 2023, the Shanghai constellation — under a new name, Qianfan — was ready to move quickly.

In November 2023, the Shanghai government announced the new LEO Qianfan constellation as part of its Action Plan for Promoting Commercial Space. In February 2024, Shanghai Spacecom Satellite Technology (SSST) — backed by the Shanghai government and affiliates of the Chinese Academy of Sciences (CAS) — raised the 6.7bn yuan it needed for the construction of a new LEO megaconstellation, mostly from state funding bodies. In August it launched its first satellite; despite some operational problems in late 2024 when its satellites were struggling to reach their target orbit, by February 2025 it had 76 working satellites in orbit (by then Guowang’s fleet had crept up to 38, following the constellation’s first operational launch in December).

In contrast to Guowang, Qianfan is fundamentally a state project of Shanghai, long a key satellite manufacturing center. The Shanghai government owns the majority of SpaceSail; in 2025, 66% of SpaceSail’s operator SSST was owned by two of the Shanghai government’s investment funds. One of these funds, Shanghai Alliance Investment, together with affiliates of the Chinese Academy of Sciences, own the main Qianfan satellite manufacturer, Genesat, and much of the local space ecosystem that feeds into it.

Despite its murky origins, with its recent growth, Qianfan clearly now has fans in Beijing working to remove barriers to its growth. Although Guowang’s monopoly technically ended in 2023, for two years non-SOEs lacked a regulatory framework to operate. A striking illustration came in January 2025, when a cruise ship demonstrating Qianfan connected only after entering Hong Kong waters, because it did not yet have market access on the mainland. But regulatory changes starting in April 2025 have allowed and actively encouraged growth in the sector. Most notably, a August 2025 MIIT Opinion offered a framework for commercial companies to satellite internet to commercial companies, and Qianfan has now been tested at different locations in mainland China. In August of this year, central government fund Guodiao Phase II Collaborative Development Fund 国调二期协同发展基金股份有限公司 invested 5.76bn RMB in the project, which Curcio describes as an “important endorsement” from the central government.

A key difference with Guowang is its transparency. As one analyst, who wished to remain unnamed, told me, “we know for certain that Qianfan is not a military system, because there’s plenty of information available about it.” That allows it to work with foreign telecomms companies on network development, crucial to its overseas ambitions. Its technology is impressive. According to reports, Qianfan terminals showcased at the 2026 Mobile World Congress in Shanghai, are capable of providing download speeds of up to 450 Mbps and upload speeds of 150 Mbps, above the median in Hong Kong.

The company says it has started to offer basic connectivity services to maritime users, supporting connection of automatic identification systems. In 2025, SpaceSail, which owns it, signed a Memorandum of Understanding (MOU) with Airbus to integrate Qianfan onto the company’s planes — though the company has indicated this might be limited to China-based airlines. Qianfan also has cooperation agreements with Turkey’s state-owned satellite comms monopoly, has been tested in Kazakhstan, and has an MOU with Malaysia’s national satellite operator Measat. Brazil will be the first country after China where its services are widely available, after an approval in March by Brazil’s telecomms agency to provide services to the country.

Stacking Qianfan satellites, Jan 2025. Hainan International Commercial Space Launch Co., Ltd via Andrew Jones on X.

Is two better than one?

The main benefit of having a fairly transparent, commercially-oriented equivalent to Starlink alongside Guowang is that it can expand abroad, competing with Musk’s company. And analysts believe a kind of rivalry has formed, spurring them both on — just as the government had hoped. Dr Lucie Sénéchal-Perrouault, who researches Chinese commercial space activities, told me: “The central government used the competition between Qianfan and Guowang to accelerate their development. They pushed each other to be better”.

Both Qianfan and Guowang have made clear progress this year. As a new leadership cohort was brought in between April and June 2025 — a decade younger than the previous — Guowang seems less cautious and more focused on deployment frequency, and now has 238 working satellites. Qianfan has had a remarkable surge: of its 217 working satellites in orbit, 151 were launched since the start of April. The deal to provide services to Brazil next year seems to be spurring the company to deploy faster, while the increased availability of rockets in China, including a few commercial successes, is making it easier to secure launches.

But though they have accelerated satellite deployment, progress still remains too slow. From Qianfan’s publicly available specs, it is possible to analyze its coverage. Technically, it is already capable of providing continuous global connectivity, meaning it can offer services to Brazil. But with its current number of satellites, bandwidth remains very low. According to calculations from Non-GEO Constellations Analysis Toolkit (NCAT) based on data from the constellation’s ITU filing, Qianfan has upwards of 200 Gbps of aggregate usable capacity available across Brazil — enough for tens of thousands of users. Data from NCAT states that Starlink has roughly 65 times that available in the country — 13 Tbps.

It’s more difficult to estimate Guowang’s coverage because of the limited information available. But the same point applies: the constellation’s low number of satellites likely means that its bandwidth remains low relative to its ambitions, limiting its current applications and strategic utility.

And the low number of satellites launched is a problem for other reasons. ITU rules require constellations to reach certain launch milestones or risk losing access to the orbital slots and spectrum rights they have. Qianfan missed a modest (seemingly self-imposed) goal of 324 satellites by July 2026. Guowang looks likely to miss its target of 400 satellites in orbit by 2027. China will need to launch 10% of the planned total for Guowang — 1299 — by September 2029 or risk losing access to the spectrum. It’s way behind schedule.

The rocket problem

Trying to realise this model of two LEO megaconstellations has one obvious challenge. China’s current annual launch capacity may be as much as 500 satellites a year; each of these megaconstellations is aiming for more than ten thousand satellites in orbit. To build just one of that kind of megaconstellation — let alone keep replenishing it given the 5-7 year life expectancy of such satellites — while also pursuing other space ambitions will require China to increase its launch capacity by an order of magnitude. Building two requires a lot more rockets than that.

The rocket shortage has clearly slowed down both projects. Qianfan in particular has struggled to find rocket companies to launch its satellites: none were launched for six months in 2025 as a result. That’s partly because of how the launch bidding process works: regulation requires three bidders; two public tenders issued in February 2025 to launch Qianfan satellites were canceled because only two entities bid, both subsidiaries of the SOE CASC. Eventually, coinciding with progress in the development of some commercial rocket companies, they got three bidders and in August 2025 awarded contracts to all three - LandSpace, Space Pioneer and CAS Space. Qianfan’s operator contracted the three to launch 94 satellites in seven groups between them, probably to diversify risk given how few times their rockets have been launched. Deployment has been beset by delays but has now finally started, with LandSpace's successful launch on 15th September, carrying ten Qianfan satellites. The other two have faced delays and issues with their rockets; Space Pioneer is all but certain not to be able to deploy Qianfan before 2027.

What happens next?

In its efforts to build its own equivalents to Starlink, the country’s launch capacity – growing, but still in the hundreds, not thousands – is the key constraint on building Chinese megaconstellations, with the tactical and strategic implications that would imply. Since the SOE’s Long March rockets are already in high demand from various state programs, China’s commercial rocket industry is stepping up. The commercial launch sector — driven since opening up in 2014 by subsidies and local state support, while demand remained unclear — now has dozens of rocket companies.1 Success is mixed: most have never launched. But a handful of leading companies with several successful launches under their belt, including LandSpace, are already supplying expendable rockets that have contributed to China’s recent increase in launch capacity. Over the long term, if LandSpace and other commercial companies prove that their reusable models are reliable, launches could shoot up, allowing China to build at least one megaconstellation on a scale comparable to Starlink. For now, that is a capability China cannot yet match.

China’s megaconstellation ambitions have finally given its commercial rocket companies a clear source of demand. This problem was covered in great detail in research by Dr Sénéchal-Perrouault.

China's Starlink Response

Paddy Stephens — Thu, 01 Oct 2026 10:04:50 GMT

Paddy Stephens writes about space, security, and energy for The Wire China, Resilience Media and others. He’s based between London, Nice and Taipei. This is Part 1 of his two-part investigation of China’s Starlink challengers!

The test lacked the cool confidence of the Falcon 9, SpaceX’s dependable workhorse rocket that slows itself down before landing gently on four legs. Nor was it as elegant as SpaceX’s super-heavy Starship rocket booster, which the company has managed to catch on three occasions with two “chopstick” arms. But as the first stage of the Long March 10B slowed itself into a netted metal box on a sea barge in July this year, the state-owned China Aerospace Science and Technology Corporation (CASC) joined SpaceX and Bezos’ Blue Origin to become the third entity in the world to have successfully recovered a rocket booster.

Long March 10B first stage is collected on a sea barge. Credit: Xinhua.

Three soon became four. On its first attempt in December, the booster of the Zhuque-3, a reusable model developed by LandSpace, one of China’s leading commercial rocket companies, ended up as a fireball that smashed spectacularly into the Gobi Desert. Since then, its engineers have clearly been busy. On 18th August, it touched down softly on its extended legs, in size and landing style much like the Falcon 9.

The two tests don’t mean China has cracked rocket reusability — it’ll take multiple successful launches of each of the models to prove their reliability. Then it will take more time still to actualize the falling costs and increased launch capacity that reusability has brought SpaceX. Blaine Curcio, the founder of Orbital Gateway Consulting and an authority on China’s space industry, told me that the Long March 10B test is “a meaningful step forward, but the short term impact [on launch capacity] is not huge”. Nevertheless, with these two tests and other commercial Chinese rocket companies due to test reusable models later this year, the US monopoly on reusable rockets is coming to an end.

These two tests are particularly important for China’s long-term ambitions to build its own equivalents to Starlink, which will require the kind of launch capacity that, as of right now, only SpaceX has. Starlink’s more than 10,0000 satellites account for 70% of the total satellites in low Earth orbit (LEO). China launched 371 satellites in 2025; SpaceX managed 3,169.

One of SpaceX’s Falcon 9 rockets landing in 2021. Image: SpaceX.

How China views Starlink

SpaceX’s Starlink, the world’s largest low Earth orbit satellite constellation, started operational launches in 2019 and conducted a beta launch from the US in 2020. By the end of 2022, the company had over 3,300 working satellites in LEO — half of the global total.

As Ukraine demonstrated the tactical utility of Starlink, SpaceX’s orbital dominance started to worry decisionmakers in Beijing.

The People’s Liberation Army (PLA) was concerned that Starlink gave the US and its partners an edge, with obvious benefits on the battlefield. In May 2022, PLA Daily wrote of the constellation: “[Starlink] will provide high-bandwidth, high-speed military communications worldwide, establishing a robust command-and-control network linking U.S. combat platforms.”

Even in the first year of Russia’s full-scale invasion, it was clear how useful this capability could be on the front line. As researchers affiliated with the PLA and state space apparatus noted in a piece republished in Security Insider 安全内参, a cybersecurity think tank publication, in January 2023, Starlink had already enabled Ukraine to maintain communication despite disruption, support military coordination and send intelligence from civilian mobile phones via military intelligence to combat units. It had also supported drone operations. Analysts from Beijing Institute of Tracking and Telecommunications Technology 北京跟踪与通信技术研究所 (BITTT) , a PLA Aerospace Force research institute, concluded that with its thousands of satellites, “this mega-constellation … poses a major challenge to our existing situational awareness and traditional defense capabilities.”

China had no clear way to disable the rapidly growing megaconstellation either. As PLA Daily highlighted, the system offers low latency and is difficult to jam. With thousands of satellites, it has plenty of redundancy to keep providing coverage even if China ever managed to take out tens or even hundreds of its satellites by kinetic means. And it is able to rapidly respond to cyber or electronic warfare, as the Security Insider paper noted.

Other worries were more far-fetched, such as that Starlink could be used to threaten and attack China’s own satellites, deepening US dominance in orbit. At the end of February 2022, PLA Daily warned that Starlink could be deployed to disable or destroy other countries’ satellites. The BITTT analysts wrote that Starlink satellites are highly maneovurable and “capable of carrying on-orbit operation devices such as robotic arms to achieve the approach and disposal of space-based targets”. This was pure panic. The PLA, convinced that the US military was “deeply involved” in the development of Starlink, has tended to exaggerate the capabilities of the constellation’s satellites – for example, stating that they could conduct reconnaissance and surveillance of ground targets. It is, however, true that Starshield, a spiced-up version of Starlink developed for the US military and announced in December 2022, has some of these capabilities.

Alongside the tactical military advantage Starlink offers the US and its partners, analysts also note that Starlink could offer other strategic benefits to the US to China’s detriment. Satellite internet is rapidly growing globally, offering low-latency connection even in remote areas. That could ultimately reduce the need for terrestrial infrastructure, which in developing countries is often built by Huawei or ZTE. “Longer term, if satellite internet becomes a quasi-competitor to terrestrial internet, then [satellite internet] becomes a really important strategic interest for China – and everyone else,” said Curcio. In the absence of a Chinese satellite internet provider, Chinese companies wishing to use fast satellite internet would be left dependent on Western providers, especially Starlink.

There are two main ways China is responding.

China is pursuing two strategies in response to Starlink. The first is targeting it: slowing its growth, developing the capability to degrade and destroy it if necessary. The second is building its own equivalents, strengthening its own capabilities while nullifying the strategic edge it gives the US and its partners.

It’s worth saying the most extreme form of the first is in tension with the second. If China ever actually tried to physically destroy Starlink, the campaign would generate a huge amount of debris that would threaten some of China’s own assets in space, stifling the country’s space ambitions and ability to maintain its current capabilities by replenishing existing satellites. More on that below.

Option A: How China thinks it could (maybe) stop Starlink

In November 2023, officials from CASC — the state-owned enterprise (SOE) aerospace behemoth behind China’s first reusable rocket — met with Russian military officials to work out a three-tiered proposal about how to stop Starlink. The details were reported in a recent investigation by Der Spiegel, Le Monde and The Insider.

The first proposal was to stop the rapidly growing megaconstellation from getting much bigger, taking advantage of outdated global satellite regulations arguably unfit for a world of Starlink, megaconstellations and superpower competition in LEO. Under this system, ‘paper filings’ – plans for satellite constellations that exist on paper only – become a strategic move to hog orbital slots and spectrum, preventing others from using or filing for them. Claiming sought-after frequency bands and orbital slots through joint filings at the International Telecommunications Union (ITU) would deprive Starlink from using them in its further expansion. China and Russia could also build an international coalition to impose regulatory limits on the constellation’s expansion, making the argument that Starlink’s satellite density increased the risk of collisions in LEO.

These were proposals — it is not clear that China and Russia have actually acted jointly on this, and no joint filings have been submitted to the ITU. But on its own, China is filing to claim spectrum and orbital bands, sometimes with constellations that exist on paper only. At the end of 2025, Chinese spacecraft operators filed plans for two megaconstellations totalling almost 200,000 satellites. If none of these satellites ever launch, the newly-formed, Xiong’an-based “Institute of Radio Spectrum Utilization and Technological Innovation” 无线电频谱开发利用和技术创新研究院 will just lose its rights to the orbital slots and spectrum, while having blocked off these slots to other users into the early 2030s. It seemed to be part of a concerted protest against Starlink: also in December, China lodged an informal complaint at a United Nations Security Council meeting against Starlink for “gobbl[ing] up frequency and orbital resources, increasing the risk of spacecraft collisons [sic] and space debris.”

But if China is trying to stop Starlink’s growth, it’s not succeeding. In July, SpaceX filed an application with the US Federal Communications Commission (FCC) for a 100,000-satellite constellation, presumed to be an updated version of Starlink. Launches continue apace.

Option B: To be able to turn off Starlink when needed

The second level in the proposed escalation ladder against Starlink was for China and Russia to develop the means to turn off Starlink at will. The two discussed developing a joint electromagnetic jamming architecture to block Starlink in certain geographic areas. It’s unclear whether this has actually happened, but Chinese scientists themselves seem to be researching ways to jam Starlink.1 The constellation has so many satellites that if a signal from one is jammed, another can easily provide coverage, meaning simple individual jammers are ineffective. One option would be to use lots of jammers at once: a theoretical paper in November from Zhejiang University and Beijing Institute of Technology found that jammers deployed in a grid formation at 20km altitudes (probably drones) could theoretically cover an average of 38.5 km2 per jammer. But from a back of the envelope calculation based on the area of Taiwan, to jam Starlink across the country would require coordination of almost a thousand aerial jammers.2 That might be possible on paper, but seems difficult to effectively coordinate in a conflict scenario.

There are other options for disrupting access to the system, as the Iranian regime demonstrated earlier this year when protesters turned to Starlink. To connect to the right satellite overhead at any given time, Starlink terminals need an accurate time signal and location, generally supplied by GPS, which is easy to jam. The regime degraded the system in parts of Iran, limiting Starlink access. In any conflict around Taiwan, China would likely employ GPS jamming to hamper US and Taiwanese GPS-dependent missile and other systems, which could degrade connection to Starlink – and other LEO constellations like OneWeb – in areas subject to jamming. Updates, though, are making the system less vulnerable to GPS denial. New V5 terminals, while not designed to be used on the move, rely on Starlink’s own highly jam-resistant position, navigation and timing (PNT) source instead of GPS.

The third and most extreme proposal for joint Sino-Russian collaboration from the investigation was to destroy Starlink. The documents mentioned in the reporting propose starting with cyber warfare to paralyse the system, followed by kinetic attempts to eliminate the satellites themselves.

By itself, China seems to have the capability to take out satellites in LEO. The Secure World Foundation says that China has ‘at least one, and possibly as many as three’ programs to develop direct-ascent anti-satellite weapons, and it destroyed a weather satellite with a ground-launched SC-19 missile in 2007.

But taking out Starlink via kinetic means would be extreme and damaging to China’s own interests. Space debris is already a serious problem — a mission bringing Chinese taikonauts back from the country’s Tiangong space station late last year was delayed after the vessel was hit by suspected debris. Taking out over 10,000 satellites in low Earth orbit would create enough debris to make certain orbits unusable for years, potentially damaging some of China’s over 1,200 satellites in LEO, and making it much harder to launch through these key orbits, thereby hampering China’s own space ambitions.

Some analysts are skeptical that China and Russia would even be able to take out Starlink with kinetic, conventional means. As Starlink continues to grow, it will become increasingly difficult to destroy the whole constellation without some insane escalation like detonating a nuke in space.3

The bottom line is that, while some of the PLA’s writings have exaggerated Starlink’s capabilities, it is clearly a versatile dual-use asset for the US and its partners. Perhaps equally worrying for China is how difficult it is to disable. Attempting to take out thousands of satellites by kinetic means would directly hamper China’s space ambitions. More appealing would be electronic warfare techniques, but the constant iteration of Russian measures and Ukrainian counter-measures shows there is no simple stable solution here either. Cyber warfare may offer other options, but Russian hackers have struggled to cause anything more than brief disruption.

Russia has developed the Volna Kupol Garant (Wave Dome Guarantor), which can jam Starlink connections across an area of about 20 square km (7.7 square miles). But given how large and conspicuous it is, it doesn’t tend to last long on the battlefield. Russia has recently deployed the system away from the frontline to protect infrastructure from Ukrainian Starlink-enabled drone attacks. Given the cooperation between the two countries, China may well have access to similar technology, but it’s not a silver bullet for jamming Starlink in a conflict scenario.

Taiwan does not currently have access to Starlink. The country’s parliament eased foreign ownership restrictions on satellite communications in July, likely to allow Starlink to enter the market. Given Musk’s business interests in China and previous comments about Taiwan, it’s not clear — even if Taiwan did get access to Starlink in peacetime — that access would continue in a conflict. Taiwan already has access to smaller LEO constellations including Eutelsat OneWeb.

Russia is alleged to have developed the capability to detonate a nuke in space as an anti-satellite measure, a scenario US Space Command has wargamed.

Logan Wright on Broken China

Jordan Schneider — Wed, 30 Sep 2026 10:06:04 GMT

Is the Chinese economy well and truly broken? Will China be stuck in low growth for a decade or more? And what would that mean for the industries we cover on ChinaTalk, for China’s military power, and for policymakers around the world?

To discuss, ChinaTalk interviewed Logan Wright, partner at Rhodium Group, my former boss, and, in my humble opinion, the person with the deepest understanding of China’s macroeconomic picture of anyone following the PRC in the West. His new book is Broken China: How the Economic Miracle Shattered and What It Means for the World. Co-hosting is the great Brad Setser of CFR.

We discuss:

Why China’s “economic miracle” has broken down and how the financial system that powered China’s growth for two decades now constrains it.
How the world’s biggest credit expansion turned into a debt trap, with investment generating less and less growth.
Why China’s property collapse, weak domestic demand, and surging exports are actually all part of the same story.
How China’s push into EVs, AI, robotics, and other “new industrial sectors” can’t by itself replace the growth China used to get from property and investment.
The jobs crisis facing China’s 12-13 million graduates per year and the ideological, fiscal, and strategic constraints that make it difficult for Beijing to address.
What a “broken China” means for the US and the rest of the world and why Logan argues there is no longer systemic economic rivalry with China, but only a military and industrial-resilience one.

Listen now on your favorite podcast app, or watch on YouTube!

Why China is “Broken”

Jordan Schneider: Logan, is China’s economy broken? Are we sure?

Logan Wright: The title is not a prediction. It’s an explanation of how we got here. The basic argument is that it’s still underappreciated how central the financial system was to China’s growth over the past two decades, and how much the financial system’s problems now constrain that growth. Those are the parts that are broken.

China does have choices. But those options involve much slower growth in the short term. They cannot maintain the pace of the past two decades. The options they have involve restructuring the fiscal and financial systems: the way they allocate capital, the way they provide countercyclical support, how they raise tax revenue and how they allocate it.

It is not meant to be a dramatic, bearish title. It’s meant to say there’s no point in asking Beijing to fix this model, because they cannot fix this model. If they are truly constrained in their policy options and not willing to fix it, what should the rest of the world do about it?

The other key takeaway is that if growth depends so much on the financial system, and the financial system now constrains growth, then a lot of China’s external messaging — that they’re economically successful and time is on their side — is simply wrong. It will become more obvious that it’s wrong. That raises different questions for Western governments, which I raised in the Foreign Affairs article previewing the book and discuss extensively in its later chapters. It expands our options, and it means we have the capacity to use time on our side as our policy tools become more salient while China’s become less so.

Jordan Schneider: How did we get here? We had China growing at ten, eight, seven percent. Now what’s the latest Rhodium forecast — somewhere in the one to three range?

Logan Wright: At the start of the year we expected growth of 1% to 2.5%. We would argue we’re trending well below that, and we’re likely seeing negative growth in Q2 and Q3, because investment in China is contracting. The question is by what margin. The official data points to it contracting.

Jordan Schneider: That is a dramatic turnabout. We had the excuse of COVID, that things would come back. It’s the fall of 2026 ,and they haven’t. On one hand, it’s a developing economy that’s going to grow slower over time. But going from seven to maybe one and a half is a truly dramatic thing for the Chinese and global economy. Turn the page back to chapters one and two. What broke, such that an economy with so much latent potential is now totally stuck?

Logan Wright: This is a story of contingent outcomes. It was not inevitable. It is the result of distinct policy choices about how the financial system operated.

After the global financial crisis, China saw the largest single-country credit expansion the world has seen in at least the last century, possibly many centuries. China added a third of global GDP to its bank assets in just eight years. We’ve never seen anything remotely like it.

First, there was a deliberate attempt to provide countercyclical support in response to the global financial crisis. The extension after that was essentially path dependence. Growth was strong, it was dependent on investment, and it became too difficult for authorities to shut down without imperiling the overall economy. That facilitated the growth of the shadow banking system, and credit growth became increasingly disconnected from the real economy. A lot of it went into the property sector and local government infrastructure.

After the deleveraging campaign started in 2016 and intensified by 2018, you had a sharp decline in credit growth. This coincides with Xi Jinping’s rise to power and how that changes financial risk, which we should discuss separately. You now have credit that cannot expand at the same rates because the financial system is already overextended. Rather than write down the malinvestment and the losses from the shadow banking system, you have a financial system that provides everlasting life to state-owned producers and local government financing vehicles, but cannot maintain the same pace of credit growth. Credit becomes less and less efficient, generating less growth at a lower quantum of credit. The result is a decline in overall investment.

The number one cause of the rapid decline since 2021 was the collapse of the property sector. It’s critical to understand that as a proximate cause, not the ultimate cause. After the deleveraging campaign hit developers’ shadow financing channels, the property sector was reinflated by borrowing from households in the form of pre-construction sales. Developers put up showrooms and sold properties in advance, for which homebuyers borrowed. When the defaults happened, they happened largely against homebuyers waiting for their apartments. That collapse in confidence weakened construction across the economy. Property was 20% to 25% of China’s GDP at its peak. Construction has fallen almost 80%. Sales have fallen about 60%. Nothing has replaced property as a driver of growth.

Source.

The other underappreciated part is how the slowdown in domestic demand is directly related to the rise in China’s external surplus and supposed competitiveness. It’s a byproduct of falling prices, but also of continued production across industrial sectors that used to find demand in residential construction in China. Now there’s none, so they need to find markets abroad, typically by cutting prices. This has produced persistent deflationary pressure and a weakening of the real exchange rate. The domestic slowdown and the external surplus are very closely linked, but people still treat them as distinct — they argue the economy is K-shaped — when in reality these are the same phenomenon.

Brad Setser: Two questions. This massive credit expansion could be called unique not just because it was so big, but because it didn’t generate an external deficit financed by inflows from abroad, which is unusual given its size. It would be helpful to discuss why such a huge credit expansion, with such heavy investment in property, never spilled over into an external deficit.

There’s also an argument that this credit expansion ended without a crisis. Nothing like the US in 2008, nothing like Spain’s banking sector. There’s never been a moment of catharsis when key parts of the financial sector lose access to financing. Why was it possible to have such a big credit expansion without it ending in something that approximates the financial crisis most countries experience?

Logan Wright: On the crisis moment first. The story in the book is not one of crisis or collapse but of decay: the decay over time of policy tools, of an attempt to forestall immediate crises at the cost of longer-term adjustment. Decay is an easier political choice for China’s leaders — accepting slower growth and inefficient policy tools rather than risking crisis. The path to crisis is not an external shock in China, where everyone assumes the government will provide support. The path to crisis is reform.

And I would argue China had that cathartic moment. It’s exactly what the book leads off with: the protests in the lobby of Evergrande’s headquarters on September 13, 2021. China’s largest firm in its most important industry could no longer meet its obligations to its own employees, who were basically lending money to the firm. They protested and were ultimately not repaid. When China’s largest firm in its most important industry effectively defaults — later on its external obligations, and later on homebuyers — it raises the question of what else is going to provide growth.

Source.

If you told me an economy had a financial crisis at date X and asked what the consequences would be at X plus five years, which is where we are now, I’d say this is entirely consistent with the aftermath of a financial crisis. Highly impaired policy tools that cannot deliver the same countercyclical response. Persistent deflationary pressure. Continued default pressure, as credit risk has migrated from the periphery to the center: from peer-to-peer lending in 2018, to small banks in 2019, to trust companies in 2020, to property developers in 2021, to mortgages in 2022, to local government financing vehicles in 2023. Much slower growth and reliance on external markets because of the inability to generate domestic demand. Everything you’d expect from an impaired, zombified financial system is taking place. The consequences are the same.

On external financing: China’s post-crisis credit expansion was essentially externally financed through steady inflows of deposits into the banking system from exchange rate intervention and reserve accumulation, until about 2013. The change in that pattern in late 2013 and 2014 is exactly what forced increasing reliance on non-core liabilities, with banks competing for wealth management products on higher interest rates. You got huge maturity mismatches between higher-rate short-term financing and a rising pool of assets, because you no longer had steady deposit growth. So there was an external financing component, though I take the broader point that it never spilled over into an external deficit.

Subscribe now

Jordan Schneider: Let me spotlight some of Logan’s prose:

“For years, long-term trends in China’s economy and financial system have appeared unstable. Everyone asks when the crisis is coming. But the more precise question is: how long can these trends continue before China faces the same economic outcomes as a financial crisis? The answers are clearer, as these consequences are already apparent in China’s economy and the decay of the financial and fiscal systems. China has already seen a sharp slowdown in economic growth, led by its vitally important property sector and local government infrastructure investment. Private sector credit remains weak, similar to the effects of a shock to aggregate demand. The downshift in economic growth is structural in nature, as the economy is unlikely to recover to previous growth rates.”

Even though we didn’t get something as dramatic and nationwide as 2008, what you’re living with is a version of permanently suppressed growth like post-bubble Japan in the mid-1990s.

Logan Wright: Japan is probably the closest analogy in recent history, with differences. Japan had zero-bound monetary policy and therefore had to rely on fiscal stimulus to solve the balance sheet recession. China probably still has more space with monetary policy. The constraint is the exchange rate and the stability of the financial system, not getting rates lower. They could get rates lower; it would just introduce different trade-offs.

The constraint China faces is fiscal. You’re already running a fiscal deficit of nine and a half percent of GDP, about two trillion dollars last year. You can run that to 15% or 16% of GDP, but you’re not going to generate much more economic activity unless you fix the underlying system. It’s not that there are no choices to offset short-term pressures. It’s that without resolving the structural issues, you shouldn’t expect a fundamental improvement in the outlook.

And the property crisis was nationwide. As I argue in the book, you had a bank run. It just wasn’t at a bank. You had homeowners wanting their money back in the form of “we want our properties built,” and developers unable to deliver because they had borrowed forward to keep buying land.

Xi’s Rebuttal

Jordan Schneider: All right, Logan — Western PhD who thinks he’s got all the answers. Let’s say I’m Xi Jinping. I’m here to stabilize party rule and make China a world power. I’m not defining national power by ghost cities and second homes for middle-class people in Hebei. I’m defining it by the ability to build everything the world needs, get the world dependent on my manufacturers, and capture the commanding heights of the twenty-first-century technology economy. If that means squeezing consumers so they have slightly worse healthcare and services, if it means stressing people out so they don’t have savings and can’t get a good return on the money in the bank, fine. I inherited an economy working toward goals that weren’t in line with national rejuvenation. Now it is. Look at all the world-leading, technologically vibrant manufacturing powerhouses built over the past fifteen years. What’s wrong with that?

Logan Wright: It’s fair to say, Mr. General Secretary, that we’re not talking about the same things. The simplest response is that if the underlying economy is not growing, your ambitions to develop those industries will face pretty fundamental limits. They already face limits from your own capacity to fund them through the domestic fiscal and financial system.

The limit is that for everything being developed in China, the marginal source of demand is outside China rather than inside. If China remains an investment-led economy — the largest source of investment in the world — it can only grow, and your national rejuvenation can only occur relative to the rest of the world, by displacing others’ export market share, by fighting over a shrinking pie. That pie shrinks further if China’s export prices keep declining and drive disinvestment in the rest of the world, which means your customers, who are largely outside China, are increasingly unable or unwilling to afford everything you’re producing. Those are the limits of the model.

You’ve identified the worldview of He Lifeng and the industrial policy planners well. There’s a fair line of argument in response, which I’m happy to engage with: that I’m measuring the wrong things. That GDP growth doesn’t account for the ways China is developing industrial dependencies and unlocking future growth drivers through industrial clusters, and that there’s a path-dependent process to this industrial development that will be sticky even if external trade defenses go up.

Look at minerals processing. How did China become the dominant force in global metals and minerals processing? It’s a combination of stories. One is cheap credit. Two, state-owned producers willing to keep committing resources to largely unprofitable lines of business, because there’s enough industrial clustering and proximity to downstream consumer electronics and manufacturing that it still makes sense on economies of scale, which makes it very difficult for others to compete. The other is cheap, subsidized power for energy-intensive enterprises like aluminum. Aluminum is basically congealed energy. There’s a reason copper and aluminum smelting have concentrated in China: the downstream consumer electronics and battery businesses benefit from it. Those are measures of comprehensive national power that my book doesn’t fully address.

But from the macroeconomic perspective, those arguments can only go so far, because fundamentally this economy is not growing at nearly the same rates. My further response to Chairman Xi is: are you aware that the United States has outgrown China over the last five years? And has a pretty good chance of outgrowing China over your fourth term, and a near certainty of outgrowing China over your fifth term, if you wanted to be around that long. Western economists are going to point out that China is still number two in the world for a long period of time. You might not care about that. You might downplay those headlines. But it means people will think about China’s inevitable rise differently as long as your domestic economy isn’t expanding.

Relative Shares of Global GDP: China, United States, Japan, the European Union, 1979–2023. Source.

The other problem is that if import growth in China isn’t rising, it’s harder and harder to maintain trade relations with everyone. And if the marginal demand for everything you produce is outside China, how your consumers are feeling and behaving really does matter for the industrial policy project you’ve embarked on. But it’s a fair argument that relative influence, national rejuvenation, and economic statecraft tools aren’t captured solely through a GDP lens or a policy-transmission-efficiency lens. I concede that.

Jordan Schneider: Premier Setser, care to dogpile on Logan?

Brad Setser: I wonder if President Xi sees the world a bit differently. He looks out and says: everybody talks about GDP growth. I care about supply chain dominance. My industrial sector is clearly the world’s number one, and it’s becoming ever bigger relative to the world economy. There’s an ever-expanding set of sectors where Chinese production is central to global supply chains. Everyone says there are limits to how much you can grow through external demand, but those limits have yet to manifest. I’m still getting 10% year-over-year export growth. My exports are growing at two times global trade, because of the incredible productivity and technological advances that have occurred inside China under my rule.

Moreover, I’ve identified a way to make this seemingly unsustainable outcome sustainable. If you don’t want my exports, I’m going to cut off your supplies of the critical inputs you need for your remaining industry. Supply chain dominance allows China to sustain export growth way past what conventional theory would say.

Arguably that’s what has played out. The US is still importing from China; we just import through Vietnam. Europe is reluctant to risk renewed supply chain warfare and is continuing to import more EVs. Perhaps we’ve underestimated how much even a large economy like China can grow without a strong internal growth engine.

Logan Wright: Brad’s response is exactly where the threat from China’s growing export share is manifest today, and it’s a different threat than we were considering five years ago. It’s not that China is going to become the world’s largest economy and maintain these capabilities ad infinitum. It’s that they can use this pattern of export displacement and supply chain dominance to lock it in for a longer period, fighting over a shrinking pie of global demand, because it will be that much more difficult for countries to respond.

The answer is that it’s a lot easier for other developed economies to reinvest and counter this disinvestment pressure with their own policy response than it is for China to maintain the same growth and demand, even if, as Brad says, there’s capacity to extend this model further than we would have expected.

A Billion People Outside the Champions

Jordan Schneider: I had a little Claude chart: percent of Chinese GDP on one axis, productivity growth as the bars, bar width as share of the economy. What it illustrated is that even in the scenario Brad laid out, where the export champions keep kicking ass, you’re still left with a billion Chinese people in industries that don’t seem to have any future. Continue the optimistic Xi case: the world doesn’t block us off, we keep exporting. What are the domestic downsides if, given the credit mess, you’re still stuck throwing money into LGFVs and central SOEs and propping up property, and you don’t have growth from anywhere besides this pretty small slice of economic activity?

Logan Wright: The irony is that you’re saying the most optimistic case for China’s development is one in which they have no control over the pace of development. Domestic controls and policy levers don’t work, so they’re beholden to the global economic cycle and whether external demand rises or falls. They can capture an increasing share of it, but they can’t control the global business cycle. China can grow as fast as its customers grow. It can’t grow relative to the rest of the world. It can’t rise in economic influence over that timeframe as long as it isn’t generating domestic demand.

Domestically, that’s really difficult, and here’s where demographics and employment come to the forefront. Many reporters coming back from China these days tell me: basically, we just want to write stories about the employment crisis. We just want to write stories about jobs, and our editors just want us to write stories about tech. It’s not surprising that’s where the demand function is, but it shows you the disconnect.

There’s a structural problem with leaving the economy unbalanced when your labor force has completely changed. China graduates 12 to 13 million college graduates a year out of a new labor force cohort of 17 to 20 million. It’s an astonishing proportion, but the economy hasn’t readjusted to provide opportunities for them. As that increases, and because birth rates have fallen over the last two decades, the problem becomes more pronounced.

The new development engine Xi is pursuing isn’t very large as a proportion of the economy. We added up all of China’s strategic industries, the new champions, using the 2023 input-output tables: NEVs, all investment in AI data centers and hyperscaler build-out, advanced robotics. We get something like 6.5% of GDP. We could be wrong, but not by a factor of two. There’s not a hidden ten trillion yuan of additional investment out there. Changes in domestic consumption would have a far greater impact on aggregate growth if you could shore up consumption.

Source.

The closer you look at the demographic headwinds, the more acute they are, not only because of the decline in aggregate and working-age population, which hits consumption directly, but because of the structure. An economy that has not seen structural reform will have significant problems generating employment. All of these strategic industries are far more capital-intensive than labor-intensive.

The China scholars engaged in this debate think the constraint on China’s growth model is domestic, and the trade folks think it’s external. Eventually there’ll be pushback. It may be a combination of the two. But one way or the other, none of this looks sustainable. There are far more forces of entropy than inertia here.

Jordan Schneider: Hu Jintao comes to the US in 2006, and George W. Bush, he’s a China guy, asks: what keeps you up at night? I’m stressed about terrorist attacks. Hu Jintao says: creating 25 million new jobs every year. You recount the story of youth unemployment rising past 20% and the statistics disappearing in 2023.

It doesn’t really matter if you have a BYD that costs ten grand if your kid can’t get a job. Remember how freaked out people were in America when unemployment was 6% in 2009 and 2010? The social malaise generated when big swaths of the economy can’t get a job at all, or a job that befits all the hard work of the gaokao, of being the first in your family to go to college — and then they’re stuck delivering packages or going back home — is an enormous social challenge. Xi must appreciate that at some level. How does it cash out in his head?

Source.

Logan Wright: I co-taught with Michael Pettis for years at Beida in his shadow central bank seminar after the global financial crisis. I’ve interacted with a lot of Chinese students and had several as research assistants. The intensity of the social pressures in that hyper-competitive environment, relative to how an American educational environment is structured, is really difficult to describe. You’re exactly right.

How they think about it is a different question. Fundamentally, this is a Marxist worldview of development: we’re building the industrial structure and industrial capacity, and from there, growth has always come from investment; employment has always come from investment. That’s how the model has worked, and you can’t demonstrate that it’s not working now. Yes, there’s a youth unemployment problem, but maybe that’s part of a transition; once these new industries are up and running, things get better. We had these problems when we joined the WTO in the late 1990s. So they downplay it. They don’t think it’s structural; they think it’s cyclical. I can see those arguments percolating within the system. But honestly, I’m not sure they are.

The thing we know least about at this point in the Chinese policymaking process is how information flows from ground-truth economic reality into filtered argumentation at the top. We read Qiushi, the party journal, pretty carefully. These are not reassuring messages that they really get it, or that economists within the system who do get it are having success filtering messages into the party apparatus. Qiushi had an article in mid-August asking, essentially, why are people’s lived experiences of the economy different from our stated growth rates? I urge you to read it. The response basically comes back: well, they haven’t adopted new technologies yet, and if they use more AI in their daily lives it’s going to be better.

Jordan Schneider: Biden tried that argument. It didn’t turn out super well for him.

Logan Wright: The analogy is extremely overextended, but normal people aren’t reading Qiushi. This isn’t a form of campaigning.

Can AI Save It?

Jordan Schneider: The book makes a pretty compelling case that there are no easy outs. All the reforms are hard, and all have serious political downsides, from rearranging power structures to making people unhappy, such that we’re stuck. Take that as a premise: we’re in a low-growth universe, and it’s unclear where the jobs come from. Logan, is AGI going to give us the jobs? Can humanoid robots or DeepSeek V7 save China’s economy?

Logan Wright: You’re asking rhetorical questions. The entire purpose would be to displace manufacturing labor.

Jordan Schneider: Fewer jobs as robots take over, AI doing more work for fewer people.

Logan Wright: Marginal labor productivity can increase while overall labor force participation collapses, which is a likely path.

On AI: we’ve just done a deep dive on China’s AI financing. People assume the state can maintain AI financing indefinitely. It doesn’t really look that way. The frontier models collectively look like they’re generating single-digit billions of dollars in revenue — about $11 billion at this stage for all the frontier models. The hyperscaler build-out is in the range of 930 billion yuan, about $135 billion, this year, and looks like 1.2 trillion yuan, around $160 billion, next year. That’s about 15% to 20% of the US build-out as of now.

The problem is that it depends largely not on the bond market but on the equity market, and on the equity market maintaining extreme valuations. Government guidance funds provide some portion of financing, but just for the hardware, the chip purchases. If China is compute-constrained and increasingly finance-constrained, you’re setting up a situation where the call on industrial policy to fund this new strategic industry keeps rising, for business models that do not look self-sustaining.

So what is this for? That becomes the notion of embedding China’s tech stack in consumer electronics and using it for geopolitical leverage and supply chain dominance, which is a different argument. But it’s not for making money, it’s probably not for making jobs, and it’s probably not for making AGI either.

Jordan Schneider: These are toy companies, and OpenAI and Anthropic are not. They’re making tens, hundreds of billions of dollars.

Logan Wright: They’re not making money.

Jordan Schneider: Right. But they have real revenue to point to, which DeepSeek and Kimi can’t. Alibaba can, but not from the model stuff.

Logan Wright: Moonshot just posted a new number, but it’s all ARR based on four weeks of best performance.

ByteDance has had the largest marginal expansion this year, and my understanding is they’re funding it through an offshore loan, so it’s not necessarily a domestic system. But they’re not generating a lot of revenue right now either, based on what we can see.

From Shadow Banking to Finance as a Utility

Brad Setser: Logan, one of the strengths of your book is how you highlight that the state financial system went through phases of evolution. It hasn’t been one constant set of five state banks providing a constant stream of financing to the same borrowers for twenty-plus years. Some of what we called shadow banking, some people in China thought of as the first steps toward a true market and more competitive banking, but that phased out. Now you’re arguing that weakness in the system has led it to center once again around the state and lenders with a state equity backstop. What’s the next phase? What does China need to do to have a more functioning financial system?

Logan Wright: Their answer right now is that a more functioning equity and bond market will provide meaningful diversification from the state system. That’s how Pan Gongsheng, the PBOC governor, talks about it: we have these new industries, they’re less dependent on the state banking system, therefore it’s okay that loan growth is slowing. Not an unreasonable argument. The problem is the limits of the equity market, the scale of participation in it, and especially the bond market, where there isn’t effective pricing of capital because everything is still priced on implicit state guarantees in some form.

But you’ve put your finger on one of the key stories of the book. Between 2012 and 2016, you had virtually an entire transformation in how the Chinese financial system operated, and almost no one noticed until it was over. You went from a system with a very steady deposit funding base, engaged in financial repression, making loans to state-owned enterprises that were inefficient but where you understood where the money was going, with persistent capital inflows and no external funding problems. By 2016, all of those parameters had changed. The marginal source of growth was a loan to an unregulated entity: a property developer, a third-party asset manager, a local government financing vehicle. The marginal funding came from a wealth management product or some other non-deposit liability. And there was a lot of volatility in the balance of payments, responding not to trade flows but to short-term capital flows and interest rate differentials.

This changed so quickly that you’ve had a recentralization. The trend now is more party governance of the financial system and treating finance as a utility. It’s underappreciated that in 2018 China created for the first time a super-regulator, the Financial Stability and Development Committee and Commission under Liu He, to regulate shadow banking and get all the regulatory bodies under one roof to prevent regulatory arbitrage. Five years later — very quickly in Chinese bureaucratic terms — this was thrown out. Liu He is gone, He Lifeng is the relevant Politburo member for economics and finance, and instead you have party commissions, the Central Financial Commission and the Central Financial Work Commission, conducting the anti-corruption investigation within the financial system. That has reportedly recently ensnared Fang Xinghai, one of the primary Western interlocutors, which sends a chilling signal to the rest of the reformers.

In this environment, if you’re a younger financial technocrat, this is not the time to raise your hand and say, “I have a great idea.” There’s oversight everywhere you look: crackdowns on salaries, oversight of expenses, ongoing anti-corruption investigations. The net result is: lend to state-owned enterprises and go home. The classic line was 3-6-3 — borrow at 3%, lend at 6%, on the golf course by 3. I don’t think anyone’s going to the golf course, but people aren’t spending a lot of time thinking about where money should be allocated.

Look at the last three years. People do not realize the extent to which credit has slowed. In marginal new financing from the state banking system, we’re talking about roughly a ten-percentage-point-of-GDP decline on a flow basis. Credit growth went from an aggregate of 18% a year from 2007 to 2016, to 9% from 2017 to 2024. We are now at five. Bank assets net of government bonds are growing 4.9%. Renminbi loans are at 5%. Loans to consumers are contracting outright. These are the structural problems. These are not conditions, in which consumers are deleveraging and household income growth is slowing, where you can expect any explosion in consumption, outside of a massive fiscal reorganization and a massive change in how the government raises and distributes revenue.

Property, Taxes, and the Fiscal Fix

Brad Setser: Two precise questions. The IMF has said China needs to do a big cleanup of the housing sector: budget government money to assure all the pre-sales are completed, recognize losses, shut down some property companies, recap the banks if necessary. Do you agree? Second, one of the more striking lines in the book, at least to me, is that the taxation system needs to be reorganized and consumption taxed more. That’s striking, given that if you want consumption to grow, you don’t normally tax it.

Logan Wright: The property sector is a bit of a distraction at this point. There’s a lot of official rhetoric, and rhetoric from the IMF, focusing on property. It’s not that it isn’t important, but the story has played out. Further repair of the property sector is going to have a pretty low marginal impact on growth. There’s not much demand for property. Home ownership rates are relatively high. Developers are still impaired. What are you bailing out, and why? It’s a distraction from the fundamental nature of the entire financial system to think that if you solve this one sector, you solve the overhang over the economy.

China claims they’ve already solved the pre-construction sales problem by putting new systems in place to stop developers from using pre-construction financing, and by our accounts this is not a huge overhang anymore. Instead you have a bifurcated market. In interior cities there’s tons of supply, no one’s moving in, there’s not going to be demand, and you’ll have to write off those assets. In coastal cities there’s a mismatch because nothing was built for three years, people are still moving in, and the secondary market is capturing a lot of sales. There’s no real incentive for new construction, and upgrading demand is very patient. So which developers do you bail out, and what do you get in future growth by doing it?

On consumption tax: this is a contradiction that reveals exactly the constraint they face. You can’t change the growth model if tax revenue depends entirely on investment-led growth and trade, which is what’s going on. You have to broaden the tax base, but in broadening it you end up taxing consumption, which you’re trying to encourage. One way around that is luxury taxes or taxes on higher-value items.

The fintech revolution has changed this too. My colleague Allen Feng points out all the time that the effective tax rate of the catering industry has basically collapsed over the last ten years, because small service businesses collect payments on WeChat Pay and Alipay with no tax reporting obligation whatsoever. There are ways to broaden the base just by catching up with changes in the structure of the economy. They will be noticed, they will affect consumption, and they’re arguably long overdue. Property taxes have long been discussed but were too difficult to implement: no national land registry, local enforcement problems, corruption.

You need an extra four to five percentage points of GDP in tax revenue one way or another, or at least a path to get there. Investment-led growth keeps declining, so that’s not a stable source of revenue either. What have they done? They’ve cracked down on overseas individual income tax evasion, which has actually increased individual income tax revenue, but that’s a little over 1% of GDP anyway, so it won’t move the needle. They’ve cracked down on local governments’ tax exemptions to companies in the name of a unified national market. That hasn’t had much impact yet but could pick up. And there are one-off tax treatments hitting listed companies’ revenues this year that are getting noticed but aren’t sustainable. There’s increasing awareness of the revenue problem, even if it isn’t acknowledged. Ultimately, you have to change the entire system, and that’s dramatic for local government investment, because it likely means local government investment and overall growth slow if you change how the tax system incentivizes these forms of output.

How Much Economic Reform Is in the De-Stalinization Speech?

Jordan Schneider: Xi has laid his cards on the table. He’s been around long enough to have danced with the economic reformers and peered over the ledge. We’ve had you on to talk about Huarong and Evergrande, about 2015 and 2016. Every time, he’s internalized the potential downside of disruptive reform that would actually unlock the economy, put credit in the right places, stoke domestic demand, and probably produce more productivity and employment, and he’s said: no, this is too dangerous. As you allude to, he’s led China down the path Yakov Feygin describes in his book, which we covered on a prior episode: the Brezhnev era. Let dead economic dogs lie and keep plodding on. He can comfort himself with the bet on manufacturing, exports, and technologies of the future. As long as he’s running the show, that bet probably isn’t going to change.

So let’s think through the bounds of what whoever succeeds him will be playing with. Not revolution — someone who has won a Joseph Torigian knife fight over the next five to seven years and is thinking about economic priorities slightly differently from Xi. How much operating room will that person have?

Logan Wright: This is one of the critical questions. We pithily describe it as: how much economic reform is in the de-Stalinization speech? And I mean that seriously. Not only has Xi Jinping created new financial risks through policy actions hitting particular sectors, but there’s an inverse relationship between the centralization of power and the credibility of countercyclical policy.

The book argues China’s future is highly contingent. We should not assume these trends continue inevitably; many of them are unsustainable, so they will change. The question is how, and by whom. There’s an optimistic version where Xi Jinping, because he controls enough of the levers of power, abruptly changes his previous pattern of governance. No one in the party stops Xi Jinping from going on television tomorrow and saying: comrades, the fiscal and financial systems we’ve had worked for the last ten to fifteen years, but they won’t work for the next ten to fifteen. We’re going to tax the wealthy, break up SOEs, dismantle local government financing platforms, shore up migrant workers’ pensions with the proceeds, sell a bunch of assets. That means slower investment growth, but everyone will be better off and it’s a more sustainable future. There is no physical obstacle to that speech. That’s the optimistic case.

Why won’t he do it now? First, he doesn’t really believe it. Second, the East is rising and the West is declining, and the East can’t have a problem this fundamental. There’s an ideological question. But the other reason is that as long as strategic competition is a meaningful constraint on Xi’s or his successor’s policy, those considerations will prevent China from acknowledging these problems. So one of the critical issues is how you mitigate strategic competition, and how China sees paths out of it in terms of the space it has to operate.

On Feygin’s work: thank you for mentioning it, because it’s really underappreciated. I set out about three years ago to answer the question of how the Soviets internalized their economic decline. I kept looking in the 1980s, and others finally directed me to the 1960s, because that’s when the debate actually happened. Yakov’s is the only book with deep work from the Soviet archives chronicling the Soviet economic reform debate, and he reaches exactly the conclusion you describe: Brezhnev chose decay. He chose decay, the scientific-technical revolution, and cybernetics, because the Soviets thought technology would work for them — we’re better planners, so we’ll use technology to allocate resources effectively. Kevin Rudd has recently published on how China thinks about AI the same way: AI is the way we get out of relying on markets, because AI will tell us how to allocate resources. It’s ideologically where they’re coming from.

The key point of the Soviet history is that the Soviets regularly relaxed the assumption of strategic competition. That gave way to periods of détente. The Cold War was persistent, but competition was more intense at some times than others. If China relaxes strategic competition with the West on multiple fronts, there will be ample opportunity to reconsider a lot of these options. It’s worth considering what the future of export controls looks like: I don’t think any export control practitioner would have imagined AI model distillation as their key threat, but doesn’t that present a more fundamental threat than many of the things they’re trying to prevent? That’s where the scope is open, under Xi Jinping or his successor.

To sum up something I think about all the time: the longer Xi is in power and the more the status quo persists, the more probable a break from these patterns becomes. But if there were a succession in 2027, or a clear delegitimization of Xi in 2027, I’d argue a break is less likely — a successor would more likely continue this, because there’s no obvious pushback from the rest of the world yet, no obvious message that the model is failing. That’s another theme of the book: this will probably not fail on its own. But China will notice when things fail. They’ll notice when they cannot export at the same rates. They noticed when local governments went bankrupt, on Xi Jinping’s trip to Gansu in 2024, and panicked. This is a responsive system, and it will respond to the model changing. It’s grinding and decaying, but it has not been forced to change yet.

Jordan Schneider: You focus on the external signals. I wonder if the internal ones come first. I think it comes down to unemployment. You could be the leader of China and see the unemployment statistics and say: this is not tenable, I need to change it tomorrow. Even if it’s a Xi Jinping heart attack in 2027 — a Li Qiang or whoever, we can go down the list, I had a tier ranking made of who wins the fight in 2027 — I could see at least half of those guys becoming a nerd about unemployment numbers. Not caring about unemployment is something that only happens after a long time in power. Even Xi in 2013 or 2014 would probably be more freaked out about it. Not that there’s going to be a revolution tomorrow, but this is the core issue. What has been his core economic focus? Anti-poverty and commanding heights. His conception now is that anti-poverty is solved, so we don’t have to worry about the social stuff. You could still have risen to the top of the State Council today, look at the median city-dwelling Han Chinese national, and say: this is not good long term. I need to reorient.

Logan Wright: You may be right that the internal pushback is more severe. Many of the China scholars I trust think that’s the case — that Europeans will never push back against Chinese trade faster than Chinese workers will call for change. The problem is: what do they ask for, other than jobs? There’s a big gap between “we have a jobs problem” and “we have a structural economic issue that requires a multi-year fix.” It’s one thing to call for more subsidies to improve the cost of living. It’s quite another to call for structural change. There are people thinking about it within the system, for sure.

One of the lowest-hanging fruits they have is equalizing the rural and urban pensions of migrant workers. It costs roughly 130,000 renminbi per person, for 250 to 300 million people — about a 30%-of-GDP problem. But it can be financed, and you don’t have to do it all at once. You can have the central bank buy special bonds, a trillion yuan a year. You can start with the young, so they have a social safety net, see a future and can spend more aggressively. But they’re not doing any of this. It’s been discussed: Liu Shijin called for 10 trillion yuan in these kinds of bonds in 2024, and it’s been part of the new-urbanization dialogue for some time. Instead, the fiscal pressures manifest as: we can’t do it; we need to tell the companies to do it. So they tell Meituan to do it, and it happens in dribs and drabs. Not the reassuring reform you’d want.

So there are pressure points. I just don’t know how they produce the kind of change we’re talking about in how the model operates. The dilemma is that you have to envision Xi Jinping changing his mind. That was the zero-COVID problem as well. China is a one-party state with a technocratic elite. How could they have such a botched exit from zero-COVID in November and December 2022? The answer is simple: Xi Jinping wasn’t allowing it, until he did. Then there’s no explanation and no discussion; the world moves on, and we’re not criticizing that decision anymore. That’s harder to do with these kinds of pressure points.

Jordan Schneider: The other issue: you spotlight Zhu Rongji as a guy who made tough decisions, there was pain, and then China got multiple points of GDP growth for five or ten years. The other playbook is the 1980s, where market liberalization goes haywire with inflation and things almost completely collapse. The payoff matrix Zhu Rongji expected on the other side was double-digit growth. What do we get on the other side of the Logan platform? Three percent growth, two or three years later? Is that going to make up for the near-term pain and all the political fights it kicks off? You need a real hold on the party to force this through, because you’re forcing people with power, money, and influence to take haircuts.

Logan Wright: That’s a very good point. It’s hard to marshal support for “we’re going to get back to roughly the growth rate we were at, at a more sustainable rate, less dependent on the rest of the world.” But it’s not just China’s leaders who should think about that. American leaders, and the rest of the world, should think about what happens after this model changes: how do you create the political space to facilitate that accommodation once you’ve seen the path out?

That’s exactly why my book is not a prediction. It’s an explanation of how we got here. The future is highly contingent. It depends on exactly the choices you’re describing, whether from Xi in response to a failing model or from the next generation of Chinese leaders, and also on how the rest of the world accommodates it, and whether there’s a path out of or to mitigate strategic competition on the other side.

What It Means for the Rest of the World

Jordan Schneider: I remember doing a report at Rhodium on US-China relative GDP out to 2050. For China to be number one, it needed to grow two to three percentage points faster than the US for a few decades. If we’re in the world you’re describing, where that’s not even close to happening — the US grew what, four percent last year, in nominal terms, and China’s stuck at half a percent —

Logan Wright: It’s nominal dollars. And that can happen with a massive depreciation of the dollar and under almost no other circumstances.

Jordan Schneider: Taking that as a premise: what’s the takeaway for policymakers around the world if China is maybe growing as a share of global manufacturing but not as a share of global GDP over the next decade-plus?

Logan Wright: This is what I address in the Foreign Affairs article. The takeaway is that you don’t have a systemic economic rivalry with China in the sense of worrying about China overtaking the Western democracies in economic power over time. What you worry about is that China’s concentration of industries and manufacturing power is exceeding the red lines of resilience that we want in Western democracies for our defense industrial bases and our economic security, especially in light of China’s weaponization of supply chains, which has made previous conceptions of economic efficiency far less relevant in a world where trade is fragmenting.

The net result: China has not been successful with this model, and is under more pressure than everyone thinks to maintain access to open markets for its manufactured goods. The rest of the world has a lot more leverage over China than it is currently using. That leverage is best delivered collectively, because China is less vulnerable to trade restrictions from any individual country, including the United States. But China’s tools of restricting supply will become less salient over time than tools that threaten China’s access to Western, or G7, demand, which will become more salient as long as China cannot grow relative to the rest of the world without displacing export share abroad. As long as China can’t grow faster than its customers, and the marginal demand for everything it produces is outside China, they face these strategic headwinds.

It’s not a short-term versus long-term problem. There’s a military and security threat from China. There’s an industrial and technological competitiveness threat. But there is no longer really a systemic economic rivalry. And that makes those first two threats deterrable, not inexorable.

Jordan Schneider: Brad, do you buy it?

Brad Setser: I would not go so far as to say there’s no systemic economic rivalry. There are ways of measuring economic competition other than the competition to have the world’s largest GDP. There are many areas of the global economy where the US and China compete, and they’re competing to demonstrate that their respective models give them more leverage than the other, even though those models generate different sources of leverage. At least for now, the perception is that China’s supply chain controls have been more potent than the US’s demand-side restrictions. Until that perception changes, the notion of rivalry will persist.

To conclude: China has been growing faster than its markets. Its exports have clearly been growing faster than its markets, and its imports less fast than reported demand growth — though you’d probably say, accurately, that reported demand growth is overstated. Is that pattern worth preserving if it means China is stagnant in the rest of its economy? Are we, in some sense, better off with an unbalanced Chinese economy than a more balanced one?

Logan Wright: That’s a very valid question. Your summary is accurate: so far, political coalitions around demand-side restrictions have not been as effective as China’s supply-side restrictions. I’d argue that’s a tactical problem rather than a fundamental imbalance between the two tools. The other pushback I get on the Foreign Affairs piece, which is valid, is that there’s a difference in political objectives. In Western democracies, where you’re balancing consumer and producer welfare to a greater extent, it’s not always politically sustainable to use demand-side restrictions for strategic leverage. That’s a tactical problem too; there are ways to design around it.

But it goes to your question: maybe an unbalanced model, with global imbalances expanding, is still more beneficial for global consumers in aggregate than it is harmful to security interests and global producers. I think that argument breaks down pretty quickly, as it’s breaking down in Europe right now, based on China’s weaponization of supply chains. If China were still using more carrots than sticks, and had framed its goals slightly differently, it would be a more salient question.

The interesting thing in global imbalances right now is how much the delta in trade this year, and in Chinese surpluses relative to the US, is tied to the AI trade. Dollar recycling from China is no longer done by a sovereign reserve manager but by commercial banks acting at its behest, and it’s increasingly going into overseas portfolio assets, equities and bonds, while the US capital account surplus is increasingly funded by those same inflows rather than Treasuries and agencies. So you have Chinese hardware exports funding data center build-outs, and the recycling of those dollars in Asia back into hyperscaler bonds and tech equities, which has kept this model afloat. It goes to your question: isn’t this working out? If we saw a big disruption in those flows, we’d see a big disruption in those engines of growth at the same time.

I’d argue the domestic foundations of this in China are not sustainable, for the reasons we’ve discussed, so it will end one way or another. The question we grapple with in global imbalances is whether they end abruptly or you can smooth your way out. Historically, they’ve ended abruptly. Are the asymmetries between supply-side and demand-side restrictions tactical or structural? I’d argue there are still clear pathways to escalate pressure on the fastest-growing sources of Chinese exports without extensively impacting Western consumer prices, while delivering meaningful restraint on China’s model. The question is to what end. What industries are you trying to support? Where does it make sense to reinvest and develop industrial policy? That’s a much broader set of questions. In Germany, it’s probably easier to answer them today than in the United States. But there you are.

Brad Setser: Congratulations, Logan. It’s hard to find a better summary of how China’s financial system has evolved anywhere.

Jordan Schneider: I remember reading Arthur Kroeber’s China’s Economy: What Everyone Needs to Know close to ten years ago, and this is the sequel, explaining why growth has fallen off a cliff and won’t come back anytime soon. You’ve done it in a way that is accessible, gripping, and provocative.

What’s the theme of ChinaTalk nowadays? Chinese national power, where it’s going, and how emerging technologies fit into that. Your book, for folks who don’t have a PhD in Chinese central bank dynamics or haven’t been playing in the markets for twenty years, contextualizes how these emerging technologies relate to the broader Chinese economic system and to the party’s credibility with the population that it has the mandate of heaven. As you alluded to with the journalists, it’s fun to focus on the shiny stuff. There is real development and real delta in those industries, in a way that 80% to 90% of the Chinese economy just does not have. It’s totally stuck and will continue to be stuck for a really long time. To think that AGI brought to you by DeepSeek is going to unstick that 90% and unwind all their problems is the bet Brezhnev made in the 1960s: hoping you roll snake eyes eight times in a row and solve all your problems with a technology that, even if it hits the third standard deviation of what people hope for, will not solve the problems you needed to solve.

Logan Wright: I agree. The issue is that there’s been very little rethinking of the strategy at high levels, even after the headwinds emerged. Maybe these AI companies develop world-leading technologies; maybe we embed them in consumer electronics and advanced manufacturing; therefore, we can continue. But nobody is stepping back and asking why we’re trying to do that, and what it looks like if we succeed. There’s a lot of thinking about incremental progress in technologies that will be world-leading, without thinking about what happens if they succeed and what it’s in pursuit of. That won’t happen until there’s real external pressure. We’re starting to see that pressure mount — I’ll be interested in the debates in Europe — but China is still very complacent about the possibility of a much larger pushback.

Jordan Schneider: Is it less fun to be an analyst of a country growing at half a percent a year?

Logan Wright: I joke about that. I’ve been doing this professionally for 21 or 22 years, and there’s something new and interesting every year, even following the same economy. From interviewing Chinese pig farmers extensively in 2007 to understand the pork cycle and how it fed into CPI, to tracking trade pushback and AI financing this year, to different patterns of global imbalances — there’s always something new. We won’t run short of topics.

The other thing is that China tends to put new words on the same old concepts. “Involution” becomes the catchword because you can’t say we already tried supply-side structural reform. Instead, we’re using a word identified with Indonesian agricultural policy, because we need a new word for it. As one of my analyst friends says: “they’re just making up new slogans, so they’re not reforming.”

Jordan Schneider: It’s a long quest to expand your Chinese vocabulary, and if it takes Xi to come up with new four-character phrases, we can’t complain.

Logan, you’re tempting me back into a ChinaEconTalk renaissance. We’ll re-launch the old spin-off.

Logan Wright: As I joked with someone earlier today, when I got into this field, people just covered Chinese politics, and the economics was all very exotic. Then it became completely mainstream, and now we’re back to exotic again, because people aren’t focusing on it nearly as much.

Xizus

Jordache — Tue, 29 Sep 2026 10:04:09 GMT