DeepSeek: A Tragedy Foretold?

Does the dream of Chinese open-source AI have a future?

We released a new podcast yesterday interviewing Kevin Xu of

Interconnected

on DeepSeek. Check it out on Apple Podcasts, Spotify, or your favorite podcast app.

By open sourcing its models, DeepSeek opened the door for any curious developer to learn from and build upon its cost-saving innovations.¹

This is the open-source ideal: free exchange of ideas in the global researcher’s sandbox that allows clever and creative ideas to compound. Proponents of OS models argue that it can accelerate science and innovation, improve transparency, distribute governance, and enhance market competition.

Still, the debate on open versus closed source rages in the AI community. It’s been hard for anyone to provide a definitive near-term commercial reason for why a firm would open source its AI models.²

In the Chinese tech space, this pragmatic sentiment is common. Baidu founder Robin Li 李彦宏 put it bluntly:

Open source is a kind of IQ tax 智商税...

When you rationally consider what value a large model can bring to you and at what cost, you should always choose a closed-source model…Many times, a model may seem useful, but when you calculate the costs, it’s not cost-effective so customers abandon it.

This is why I said that open-source models cannot beat closed-source models.

Alibaba maintains its open-source Qwen, but makes money by upselling APIs, cloud services, and computing infrastructure to customers. Kai-fu Lee’s 李開復 start-up, 01.AI, released Yi-34B as a way to “give back” to the Chinese developer community, but the company ultimately targets state-of-the-art proprietary models as the basis for its commercial offerings.

DeepSeek, on the other hand, has no near-term profit strategy.

In a 2023 interview, CEO Liang Wenfeng 梁文锋 very clearly stated that there is no commercial rationale for DeepSeek’s dream of “limitless research.” This sentiment brings to mind a quote from OpenAI’s Sam Altman in 2019, when he said, “I have no idea how we may one day generate revenue.”

High-Flyer Quant says it isn’t in it for the returns, either. The quantitative hedge fund financing DeepSeek recently emphasized that High-Flyer’s AI model research will not be used for stock trading: “It has nothing to do with finance. … What we care about is long-term social value.”

We should take these statements of principle at face value — this isn’t a government front, since the way DeepSeek has moved is so antithetical to traditional Chinese government-backed industry. It’s more accurate to say that DeepSeek’s employees, largely composed of young homegrown talent, are driven by something other than money-making. Is that madness, one interviewer asked? Liang’s response was telling:

Liang: I’m unsure if it’s madness, but many inexplicable phenomena exist in this world. Take many programmers, for example — they’re passionate contributors to open-source communities. Even after an exhausting day, they still dedicate time to contributing code.

Waves: There is a sense of spiritual reward in it.

Liang: It’s like walking 50 kilometers — your body is completely exhausted, but your spirit feels deeply fulfilled.

Waves: Do you think curiosity-driven madness lasts long-term?

Liang: Not everyone can stay passionate their entire life. But most people, in their younger years, can wholeheartedly dedicate themselves to something without any materialistic aims.

These young Chinese developers’ intense passion to work on open-source projects is sometimes called “the open source ethos” 开源情怀, which Kevin Xu from Interconnected explains well:

Most engineers are thrilled if their open-source projects — a database, a container registry, etc. — are used by a foreign company, especially a Silicon Valley one. They’d tack on free labor on top of already free software, to fix bugs, resolve issues, all day all night. It’s all for the validation and approval.

Implicit in this “zeal” or “calling” is an acute awareness that no one in the West respects what they do because everything in China is stolen or created by cheating. They are also aware that Chinese firms have been taking for free lots of open source tech to advance, but they want to create their own, contribute, and prove that their tech is good enough to be taken for free by foreign firms — some nationalism, some engineering pride.

In his more recent interview, Liang shared a similar insight. It’s easier to recruit engineers to solve tough problems if they have a thirst to prove themselves, he explained: “Top talents in China are underestimated because there’s so little hardcore innovation happening at the societal level, leaving them unrecognized.”

That is now changing. Thanks to recent open-source models, DeepSeek has earned global recognition and respect from engineers around the world. But will China’s government see it the same way?

Beijing’s approach to open source

The Chinese government has already expressed some support for open source 开源 development.

In 2018, a (since-deleted) white paper and the formation of the China AIOSS Development Alliance 中国人工智能开源软件发展联盟 brought open-source AI into the spotlight. An open ecology would be achieved, the white paper asserts, by cultivating OS communities and talent, promoting standards, establishing funding mechanisms, improving the intellectual property rights regime, and strengthening security reviews.

Top-level policy plans³ reference open source when discussing software and technology ecosystems, but not AI specifically. Two separate five-year plans also advocate for the benefits of open source.⁴

This positive political stance toward its domestic open-source community stems from a government and industry desire to reduce Chinese dependence on foreign software.

The government’s push for open source in the early 2000s — including the creation of several OS software alliances and a locally developed “Red Flag Linux” 中科红旗 — was a way to limit the impact of Microsoft Windows operating systems. Later, the Ministry of Industry and Information Technology designated Gitee as China’s national “independent, open-source code hosting platform” to replace GitHub, which it has struggled to censor. In the Chinese chip industry, open-source projects are being explored as a way to reduce dependence on closed-source Western ecosystems.⁵

Huawei exemplifies how geopolitics can color support for Chinese open source. In 2019, the president of Huawei’s consumer software group warned that without its own OS community, “All of China’s software communities will be in great danger the moment something happens beyond our control.” Huawei is obviously referring here to the struggle with US sanctions.

How it could all go wrong

So, will the Chinese government allow DeepSeek’s team to continue with the excitement of their nerd show-and-tell?

Unfortunately, potential liabilities from AI technology could push the government away from open source despite all the positive rhetoric.

Operating systems can’t disseminate information and power to the public in the way that AI can.

An open-source AI model grants the public broad access, usage, and customizability — which can’t easily be moderated or rescinded. There are general AI safety risks.

And beyond that, with the prospect of future advancements of AI, an outspoken chatbot might not be the only threat on the government’s radar. In our recent show with Miles Brundage, Jordan puts it this way:

We’re not far from a world where, until systems are hardened, someone could download something or spin up a cloud server somewhere and do real damage to someone’s life or critical infrastructure. It’s not just about reading outputs about Tiananmen — it’s about democratizing power by giving individuals access to an incredibly powerful technology that has the potential to cause real social damage. China does not let civilians buy guns — once open-source AI really gets weapons-grade, and one person can shut the lights off in a city, is that really something the CCP will allow to proliferate without any control? I don’t see that as a world state that government officials in Beijing, or the West for that matter, will accept.

Today, China’s generative AI regulations lack specific guidance for open-source providers. As regulators attempt to balance the country’s need for control with its ambition for innovation, DeepSeek’s team — driven by curiosity and passion rather than near-term profit — might be in a vulnerable spot.

The uncertain future of DeepSeek and open source

At least for now, though, the Chinese government sees the benefits of open-source AI.

To be clear, the only government statement⁶ we have on DeepSeek (for now) endorses open source:

DeepSeek’s greatest strength lies in its open-source approach, which empowers researchers worldwide…This highlights the importance of transcending a narrow, competitive mindset. China’s technological progress can and should contribute to humanity on a broader scale. Ultimately, what is good for China can be good for the world, and vice versa. As long as Chinese technology aligns with global progress, no barriers—no matter how high or deep—can halt its development.

And DeepSeek’s success has inspired more dialogue in China about the advantages of open source. Chinese AI startup MiniMax released several open-source models with the hope that “there will be encouragement for good work and criticism for bad work, and people outside will be able to contribute.” Chinese analysts pointed out that cost-effective open-source models support widespread access and adoption, including to countries in the Global South. See a viral post by an engineer refuting Dario’s export control essay:

Dario raises a critical question: What would happen if China gains access to millions of high-end GPUs by 2026–2027? His answer is this—if China cannot obtain this computing power, the U.S. will enter a phase of "unipolar AI dominance" and could solidify its advantage long-term through AI’s self-reinforcing mechanisms. However, if China does acquire these resources, the U.S. may face a prolonged “AI arms race.”

But my perspective is this: Whether unipolar or bipolar, AI development has already irreversibly entered a phase of global diffusion. The U.S. will not monopolize AI, China will not be contained, and nations like Europe, Japan, India, and others will not remain absent. Variables such as export controls, model competitions, and capital flows may influence the pace of the race, but they cannot halt the world’s march toward more advanced forms of AI.

DeepSeek is not an endpoint, but a signal—its significance lies not in “defeating” anyone, but in proving that the world has entered an irreversible era of large-scale AI competition.

When thinking about DeepSeek, two things can be true at once:

• DeepSeek likely chose to open source its models for the same reason developers from around the world choose to open source: out of genuine faith in the value of an open, global research community — to show off their accomplishments and inspire others to build upon their work.

• At the same time, as AI models become more powerful, governments might have an incentive to step in and take command.

[Jordan: Beyond direct government intervention, DeepSeek’s accomplishments will open doors for corporate partnerships bearing orders of magnitude more compute than DeepSeek runs today. DeepSeek has been slept on domestically in China — it has no outside funders, no fancy returnees with foreign degrees and experience at top Western labs, no government contracts, and no popular consumer apps. This will change. Just like what happened to OpenAI with Microsoft post-ChatGPT, China’s top hyperscalers — ByteDance, Alibaba, Tencent, and Huawei — will come knocking. Given Liang’s commitment to hardcore innovation and admission that compute access is holding him back, he might not pass up the opportunity of extra compute for R&D. This could scale up DeepSeek to really push the frontier of innovation with the major AI players in the West — but doing so could also unleash the same sort of forces that led OpenAI to splinter and caused its top researchers to scatter to the winds.]

Does open-source AI have a future in China? Will the lure of hyperscaler compute pull DeepSeek away from its open-source ideals? Will the government intervene for the sake of security and control?

Subscribe to ChinaTalk as we continue to follow this story.

We’re looking for advertisers. Respond to this email to connect.

---

Jordan: So, R1 is really good at poetry. Let’s close with its meditation on DeepSeek and its future with open-source models. I fed it this article (initially it refused, telling me in Chinese “Sorry, I haven’t learned how to think about these types of questions, I’m good at math, coding, logical topics, so please let’s chat about those things.” “对不起，我还没有学会如何思考这类问题，我擅长数学、代码、逻辑类的题目，欢迎与我交流.” Then I got ChatGPT to summarize the piece above, fed it back in, told it to write an award-winning contemporary poem, and after a few rounds it came out with this.

The Unlicensed Sky

1. The definition of open-source AI is a work in progress. DeepSeek fulfills generally accepted definitions of open source by releasing its code, model, and technical report, but it did not, for instance, release its data. Its code is MIT-licensed, allowing free use, modification, and commercialization.

2. Meta makes sense from a ‘commoditize your complement’ perspective but this logic doesn’t apply to pure-play AI labs like DeepSeek.

3. See the 13th Five-Year National Informatization Plan and the Software and Information Technology Services Industry Development Plan.

4. The 14th Five-Year Plan for the Development of Software and Information Technology Services targets the construction of two to three OS communities with international influence by 2025. The 14th Five-Year Plan for National Informatization articulates specific steps to advance domestic OS. These steps include deepening drivers of innovation, developing OS frameworks for AI, and fostering OS communities centered around domestic champion enterprises.

5. Instruction-set architectures (ISAs) are the interface between the hardware of a chip and the software running on a computer. Intel and ARM, American and British companies, respectively, have long offered the closed-source ISAs that most chips in the world use, allowing these firms to reap high profits. Importantly, Chinese firms, as proprietary systems subject to American export controls, risk losing access to these fundamental licenses if relations between Washington and Beijing further deteriorate.

The economics of open source remain challenging for individual firms, and Beijing has not yet rolled out a “Big Fund” 大基金 for open-source ISA development, as it has for other segments of the chip industry. Seeds of interest, however, are sprouting on private and public grounds: besides the giants like Huawei and Alibaba investing in open-source ISAs for MCUs and CPUs, lesser-known firms like VeriSilicon 芯原股份, Jiangsu Yunyong 云涌科技, Bluetrum 中科蓝讯, C*Core Technology 国芯科技, and many others have ongoing research projects leveraging the open-source RISC-V, Linux, and Khronos ecosystems to develop solutions for IoT applications, natural language processing, neural networks, self-driving cars, and more. For example, VeriSilicon’s ongoing digital signal processor project spent 242 million RMB from 2020 to 2023, using RISC-V systems to develop image-recognition chips not dependent on closed-source Western technology. Another firm, Beken 博通集成, reported receiving a 3.5 million RMB government subsidy for its project in develop a high-security platform chip for the “national secret algorithms” 国密算法 (essentially, encryption standards) that the PRC National Cryptography Administration requires certain businesses to implement.

6. Published in WeChat by Zhejiang Publicity 浙江宣传, a provincial CCP authority.