DeepSeek's Secret to Success

How it broke the China tech mold

Guest piece by

JS Tan

, a PhD Candidate at MIT’s international development program, researching the political economy of innovation in the US and China with a focus on cloud computing. He was previously a software engineer and writes on substack here.

As of December 2024, DeepSeek was relatively unknown.

Then its base model, DeepSeek V3, outperformed leading open-source models, and R1 broke the internet. No easy feat when operating with less compute than western labs.¹

Their breakthroughs raise two key questions:

1. How did DeepSeek outcompete Chinese AI incumbents, who have thrown far more money and people at building frontier models?

2. What does DeepSeek’s success tell us about China’s broader tech innovation model?

DeepSeek’s success is not just a product of technical ingenuity, but also deeply rooted in its unique approach to labor relations. Chinese tech firms are known for their grueling work schedules, rigid hierarchies, and relentless internal competition. DeepSeek’s flat management structure, in contrast, focuses on empowering its workers with autonomy and creating a collaborative environment.

DeepSeek is hardly a product of China’s innovation system. The company is neither a state-led project nor a direct beneficiary of China’s AI-focused industrial policies. Rather, it was self-funded by a former hedge-fund manager and emerged from the periphery of China’s tech landscape. While DeepSeek makes it look as though China has secured a solid foothold in the future of AI, it is premature to claim that DeepSeek’s success validates China’s innovation system as a whole.

How DeepSeek Broke the Tired 996 Playbook

To appreciate why DeepSeek’s approach to labor relations is unique, we must first understand the Chinese tech-industry norm.

Perhaps the most notable aspect of China’s tech sector is its long-practiced “996 work regime” — 9 a.m. to 9 p.m., six days a week. This workplace culture emerged during the rise of China’s digital economy in the mid-2000s and solidified during the hyper-competitive years that followed. Employees are kept on a tight leash, subject to stringent reporting requirements (often submitting weekly or even daily reports), and expected to clock in and out of the office to prevent them from “stealing time” from their employers. Employers set demanding key performance indicators (KPIs) and practice “stack ranking,” a performance management system where employees are ranked against each other.²

Since the mid-2010s, these grueling hours and draconian management practices were a staple of China’s tech industry. The long hours were considered a basic requirement to catch up to the United States, while the industry’s punitive management practices were seen as a necessity to squeeze maximum value out of workers.

Indeed, speed and the ability to rapidly iterate were paramount during China’s digital growth years, when companies were focused on aggressive user growth and market expansion. The primary goal was to quickly and continuously roll out new features and products to outpace competitors and capture market share. This relentless pursuit of expansion demanded a workforce that functioned like a well-oiled machine. As a result, employees were treated less as innovators and more as cogs in a machine, each performing a narrowly defined role to contribute to the company’s overarching growth objectives.

Such labor relations can be seen at Pinduoduo, a rising challenger to Alibaba’s dominance in e-commerce. The company is infamous for requiring an extreme version of the 996 work culture, with reports suggesting that employees work even longer hours, sometimes up to 380 hours per month. Management uses digital-surveillance tools — including location-tracking systems — to measure employee productivity. Even bathroom breaks are scrutinized, with employees reporting that prolonged absences can trigger disciplinary action. Those who fail to meet performance benchmarks risk demotion, loss of bonuses, or even termination, leading to a culture of fear and relentless pressure to outperform each other.

Since the late 2010s, however, China’s internet-user growth has plateaued, and key digital services — such as food delivery, e-commerce, social media, and gaming — have reached saturation. Those developments have put the efficacy of this model under strain.

But instead of focusing on developing new value-added digital innovations, most firms in the tech sector, even after public backlash about the 996 working schedule, have doubled down on squeezing their workforce, cutting costs, and relying on business models driven by price competition. This approach comes at a cost: stifling creativity, discouraging independent problem-solving, and ultimately hindering China’s ability to engage in long-term innovation-based competition.

New Approach to Talent

DeepSeek’s approach to labor relations represents a radical departure from China’s tech-industry norms. Since its founding in 2023, the company has eschewed the hierarchical and control-heavy management practices standard across China’s tech sector. Instead, it has built a workplace culture centered on flat management, academic-style collaboration, and autonomy for young talent.

The team size is deliberately kept small, at about 150 employees, and management roles are de-emphasized. Research groups are formed based on specific goals, with no fixed hierarchies or rigid roles. Team members focus on tasks they excel at, collaborating freely and consulting experts across groups when challenges arise. This approach ensures that every idea with potential receives the resources it needs to flourish. Liang Wenfeng 梁文峰, the company’s founder, noted that “everyone has unique experiences and comes with their own ideas. They don’t need pushing. … When an idea shows potential, we allocate resources from the top down.” To that end, DeepSeek actively avoids the performative aspects of traditional tech workplaces. There are no weekly reports, no internal competitions that pit employees against each other, and famously, no KPIs.

On the human capital front: DeepSeek has focused its recruitment efforts on young but high-potential individuals over seasoned AI researchers or executives. Many of DeepSeek’s researchers, including those who contributed to the groundbreaking V3 model, joined the company fresh out of top universities, often with little to no prior work experience. Said one headhunter to a Chinese media outlet who worked with DeepSeek, “they look for 3-5 years of work experience at the most. Any more than 8 and you’re just a ‘pass’ for them.” Liang explains the bias towards youth: “We need people who are extremely passionate about technology, not people who are used to using experience to find answers. Real innovation often comes from people who don't have baggage.” While other Chinese tech firms also prefer younger candidates, that’s more because they don’t have families and can work longer hours than for their lateral thinking.

Heavy emphasis is placed on educational background and competition achievements. The company is known to reject candidates who’ve achieved anything but gold in programming or math competitions. And beyond a cultural commitment to open source, DeepSeek attracts talent with money and compute, beating salaries offered by Bytedance and promising to allocate compute for the best ideas rather than to the most experienced researchers.

This hiring practice contrasts with state-backed firms like Zhipu, whose recruiting strategy has been to poach high-profile seasoned industry recruits — such as former Microsoft and Alibaba veteran Hu Yunhua 胡云华 — to bolster its credibility and drive tech transfer from incumbents.

[Jordan: this strategy has worked wonders for Chinese industrial policy in the semiconductor industry. Poaching experienced talent from TSMC and Samsung has been integral to SMIC, Huawei and CXMT’s success. But AI engineering is in a unique moment where young lateral thinking often trumps talent trained in the pre-transformer era.]

DeepSeek’s success highlights that the labor relations underpinning technological development are critical for innovation. While many of China’s tech giants have focused on squeezing maximum output from overworked employees, DeepSeek has demonstrated the transformative potential of a supportive and empowering workplace culture. By breaking away from the hierarchical, control-driven norms of the past, the company has unlocked the creative potential of its workforce, allowing it to achieve results that outstrip its better-funded competitors.

An Outlier in China’s Innovation Landscape

Like its approach to labor, DeepSeek’s funding and corporate-governance structure is equally unconventional. Unlike many of its peers, the company didn’t rely on state-backed initiatives or investments from tech incumbents. Instead, its former hedge fund founder essentially bankrolled the company. The company’s origins are in the financial sector, emerging from High-Flyer, a Chinese hedge fund also co-founded by Liang Wenfeng. DeepSeek itself emerged from High-Flyer’s pivot into AI after the 2021 regulatory crackdown on speculative trading. As a result of this setup, DeepSeek’s research funding came entirely from its hedge fund parent’s R&D budget. This unique funding arrangement means that the company could operate independently of the constraints often associated with state or corporate funding. In this way, DeepSeek is a complete outlier.

Once again, let’s contrast this with the Chinese AI startup, Zhipu. Zhipu is not only state-backed (by Beijing Zhongguancun Science City Innovation Development, a state-backed investment vehicle) but has also secured substantial funding from VCs and China’s tech giants, including Tencent and Alibaba — both of which are designated by China’s State Council as key members of the “national AI teams.” In this way, Zhipu represents the mainstream of China’s innovation ecosystem: it is closely tied to both state institutions and industry heavyweights.

DeepSeek, by comparison, has remained on the periphery, carving out a path free from the institutional expectations and rigid frameworks that often accompany mainstream scrutiny. Its funding model — self-financed by its founder rather than reliant on state or corporate backing — has allowed the company to operate with a level of autonomy rarely seen in China’s tech sector.

Tech Transfer vs. Indigenous Innovation

This brings us to a larger question: how does DeepSeek’s success fit into ongoing debates about Chinese innovation? And how must we update our perspectives on Chinese innovation to account for DeepSeek?

The debate around Chinese innovation often flip-flops between two starkly opposing views: China is doomed versus China is the next technology superpower. As I see it, this divide is about a fundamental disagreement on the source of China’s growth — whether it relies on technology transfer from advanced economies or thrives on its indigenous ability to innovate.

• Those who believe China’s success depends on access to foreign technology would argue that, in today’s fragmented, nationalist economic climate (especially under a Trump administration willing to disrupt global value chains), China faces an existential risk of being cut off from critical modern technologies. From this perspective, isolation from the West would deal a devastating blow to the country’s ability to innovate.

• On the other hand, those who believe Chinese growth stems from the country’s ability to cultivate indigenous capabilities would see American technology bans, sanctions, tariffs, and other barriers as accelerants, rather than obstacles, to Chinese growth. In this view, such restrictions compel Chinese firms to innovate, upgrade, and develop homegrown technological solutions, ultimately strengthening China’s self-reliance and long-term competitiveness.

[See also Nancy Yu’s piece on China’s industrial policy.]

So far, this debate has primarily unfolded in the context of advanced manufacturing sectors, from solar PV to batteries, and, more recently, electric vehicles. In the early stages — starting in the US-China trade wars of Trump’s first presidency — the technology transfer perspective was dominant: the prevailing theory was that Chinese firms needed to first acquire fundamental technologies from the West, leveraging this know-how to scale up production and outcompete global rivals. So the initial restrictions placed on Chinese firms, unsurprisingly, were seen as a major blow to China’s trajectory. China’s dominance in solar PV, batteries and EV production, however, has shifted the narrative to the indigenous innovation perspective, with local R&D and homegrown technological advancements now seen as the primary drivers of Chinese competitiveness.

When it comes to China’s tech industry, its success is portrayed as a result of technology transfer rather than indigenous innovation. Part of the reason is that AI is highly technical and requires a vastly different type of input: human capital, which China has historically been weaker and thus reliant on foreign networks to make up for the shortfall. Scholars like MIT professor Huang Yasheng attribute the rise of China’s tech sector to the many collaborations it has had with other countries. Even Chinese AI experts think talent is the primary bottleneck in catching up.

Indeed, China’s post-2000s ICT sector built its success on the back of overseas technical know-how. Many of China’s early tech founders either received education or spent considerable time in the United States. Chinese tech firms privilege employees with overseas experience, particularly those who have worked in US-based tech firms. In the generative AI age, this trend has only accelerated: Alibaba, ByteDance, and Tencent each set up R&D offices in Silicon Valley to increase their access to US talent. This reliance on international networks has been especially pronounced in the generative AI era, where Chinese tech giants have lagged behind their Western counterparts and depended on foreign talent to catch up.

Is DeepSeek the exception or the new rule?

This is where DeepSeek diverges from the traditional technology transfer model that has long defined China’s tech sector. Instead of relying on foreign-trained experts or international R&D networks, DeepSeek’s exclusively uses local talent. Liang himself also never studied or worked outside of mainland China. The DeepSeek story shows that China always had the indigenous capacity to push the frontier in LLMs, but just needed the right organizational structure to flourish. Much like China’s advancements in solar manufacturing, batteries, and electric vehicles, DeepSeek symbolizes a critical turning point in tech/AI: China is no longer merely playing catch-up, but is now competing on equal footing with the leading innovators in the West.

While I hope the “tech transfer vs. indigenous innovation” perspective is helpful in thinking about China’s innovation system, I must admit that it is somewhat of a false dichotomy. As development economists would remind us, all technology must first be transferred to and absorbed by latecomers; only then can they innovate and create breakthroughs of their own. Thus, tech transfer and indigenous innovation are not mutually exclusive — they’re part of the same sequential progression. First, technology must be transferred to and absorbed by latecomers; only then can they innovate and create breakthroughs of their own.

If we are to claim that China has the indigenous capabilities to develop frontier AI models, then China’s innovation model must be able to replicate the conditions underlying DeepSeek’s success. But this is unlikely: DeepSeek is an outlier of China’s innovation model.

Unlike solar PV manufacturers, EV makers, or AI companies like Zhipu, DeepSeek has thus far received no direct state support. In fact, its success was facilitated, in large part, by operating on the periphery — free from the draconian labor practices, hierarchical management structures, and state-driven priorities that define China’s mainstream innovation ecosystem.

Can China’s tech industry overhaul its approach to labor relations, corporate governance, and management practices to enable more firms to innovate in AI? The real test lies in whether the mainstream, state-supported ecosystem can evolve to nurture more companies like DeepSeek — or whether such firms will remain rare exceptions. The answer to this will define the long-term competitiveness of China’s AI firms.

For now, though, all eyes are on DeepSeek.

In December 2022, JD.com AI-research executive He Xiaodong 何晓冬 told local media,

In order to say goodbye to Silicon Valley–worship, China’s internet ecosystem needs to build its own ChatGPT with uniquely Chinese innovative characteristics, and even a Chinese AI firm that exceeds OpenAI in capability. This is an essential question for the development of China’s AI industry.

DeepSeek made it — not by taking the well-trodden path of seeking Chinese government support, but by bucking the mold completely.

The parallels between OpenAI and DeepSeek are striking: both came to prominence with small research teams (in 2019, OpenAI had just 150 employees), both operate under unconventional corporate-governance structures, and both CEOs gave short shrift to viable commercial plans, instead radically prioritizing research (Liang Wenfeng: “We do not have financing plans in the short term. Money has never been the problem for us”; Sam Altman: “We have no idea how we may one day generate revenue. We have made a soft promise to investors that, ‘Once we’ve built a generally intelligent system, basically we will ask it to figure out a way to generate an investment return for you’”).

But now that DeepSeek has moved from an outlier and fully into the public consciousness — just as OpenAI found itself a few short years ago — its real test has begun. How will it fare? Can High-Flyer cash and Nvidia H800s/A100 stockpiles keep DeepSeek running at the frontier forever, or will its growth aspirations pressure the company to seek outside investors or partnerships with conventional cloud players? Does Liang’s recent meeting with Premier Li Qiang bode well for DeepSeek’s future regulatory environment, or does Liang need to think about getting his own crew of Beijing lobbyists? Will Liang receive the treatment of a national hero, or will his fame — and wealth — put a months-long Jack Ma–style disappearance in his future? If an organizational vision crisis arises (à la the Altman vs. Musk feud), will Liang be able to steer DeepSeek through it?

In short, how long can DeepSeek buck the mold?

new year mood music

1. We’ll leave it to Anthropic CEO Dario Amodei to characterize their chip situation. “DeepSeek also does not show that China can always obtain the chips it needs via smuggling, or that the controls always have loopholes. I don't believe the export controls were ever designed to prevent China from getting a few tens of thousands of chips. $1B of economic activity can be hidden, but it's hard to hide $100B or even $10B. A million chips may also be physically difficult to smuggle. It's also instructive to look at the chips DeepSeek is currently reported to have. This is a mix of H100's, H800's, and H20's, according to SemiAnalysis, adding up to 50k total. H100's have been banned under the export controls since their release, so if DeepSeek has any they must have been smuggled (note that Nvidia has stated that DeepSeek's advances are "fully export control compliant"). H800's were allowed under the initial round of 2022 export controls, but were banned in Oct 2023 when the controls were updated, so these were probably shipped before the ban. H20's are less efficient for training and more efficient for sampling — and are still allowed, although I think they should be banned. All of that is to say that it appears that a substantial fraction of DeepSeek's AI chip fleet consists of chips that haven't been banned (but should be); chips that were shipped before they were banned; and some that seem very likely to have been smuggled. This shows that the export controls are actually working and adapting: loopholes are being closed; otherwise, they would likely have a full fleet of top-of-the-line H100's. If we can close them fast enough, we may be able to prevent China from getting millions of chips, increasing the likelihood of a unipolar world with the US ahead.”

2. Broadly the management style of 赛马, ‘horse racing’ or a bake-off in a western context, where you have individuals or teams compete to execute on the same task, has been common across top software companies. See this recent feature on how it plays out at Tencent and NetEase.