Why is AI in China developing so quickly? The answer lies inside the laboratories.

Title: Notes from Inside China’s AI Labs
Author: Nathan Lambert
Translation: Peggy, BlockBeats

Author: Lu Dong BlockBeats

Source:

Reprint: Mars Finance

Editor’s note: Chinese AI laboratories are becoming an increasingly unavoidable force in the global large model competition. Their advantages are not only in talent, engineering strength, and rapid iteration but also stem from a very practical organizational approach: talk less about concepts, build more models; emphasize team execution over individual stars; rely less on external services and prefer to master core technology stacks themselves.

After visiting several leading Chinese AI labs, author Nathan Lambert found that the Chinese AI ecosystem is not exactly the same as in the U.S. The U.S. places more emphasis on original paradigms, capital investment, and the influence of top scientists; China, on the other hand, excels at quickly catching up within existing directions through open source, engineering optimization, and large investments from young researchers, rapidly pushing model capabilities to the forefront.

What’s most worth noting is not whether Chinese AI has already surpassed the U.S., but that two different development paths are forming: the U.S. is more like a frontier race driven by capital and star labs; China is more like an industrial competition propelled by engineering capability, open source ecosystems, and technological self-control awareness.

This means that future AI competition will not only be about model rankings but also about organizational ability, developer ecosystems, and industry execution. The true change in Chinese AI lies in its shift from merely copying Silicon Valley to participating in global frontiers in its own way.

Below is the original text:

Sitting on a new-style high-speed train from Hangzhou to Shanghai, I look out the window and see the distinct ridges of mountains, dotted with wind turbines, forming silhouettes under the sunset. The mountains serve as the background, while in front of me are vast fields interwoven with high-rise buildings.

I return to China with great humility. Going to such an unfamiliar place, yet being warmly welcomed—this is a very warm and human experience. I am fortunate to meet many people in the AI ecosystem—people I had only known from afar before; and they greet me with bright smiles and enthusiasm, reminding me once again that my work and the entire AI ecosystem are fundamentally global.

Chinese Researchers’ Mindset

Chinese companies building language models can be seen as very suitable as “fast followers” of this technology. They are built on China’s long-standing educational and work culture traditions, and also have construction approaches for tech companies that differ somewhat from the West.

If you only look at output—namely, the latest and largest models, and the intelligent workflows supported by these models—and at input factors like excellent scientists, large-scale data, and accelerated computing resources, then Chinese and American labs seem broadly similar. The real long-term differences lie in how these factors are organized and shaped.

I have always believed that one reason Chinese labs are very good at catching up and staying near the frontier is that their culture aligns very well with this task. But before directly communicating with people, I thought it was not appropriate to attribute this intuition to any significant influence. After talking with many humble, open, and excellent scientists in top Chinese labs, many of my ideas became clearer.

Building the best large language models today largely depends on meticulous work across the entire tech stack: from data, to architecture details, to reinforcement learning algorithms. Every part of the model can be improved, and how to combine these improvements is a complex process. In this process, work by some very smart individuals may need to be set aside to maximize overall model performance in multi-objective optimization.

American researchers are obviously also very skilled at solving individual component problems, but they tend to have a “speak for themselves” culture. As scientists, actively advocating for one’s work often leads to more success; and the current culture is also promoting a new path to fame—becoming a “top AI scientist.” This can lead to direct conflicts.

There are widespread rumors that the Llama organization collapsed under political pressure after these interests were embedded into hierarchical structures. I’ve also heard from other labs that sometimes it’s necessary to “appease” top researchers, so they stop complaining that their ideas weren’t included in the final models. Whether or not this is entirely true, the message is clear: self-awareness and career advancement desires can indeed hinder the construction of the best models. Even a small cultural difference between the U.S. and China can have meaningful impacts on final output.

Part of this difference relates to who is building these models in China. In all labs, a very direct reality is that a large proportion of core contributors are still students. These labs are quite young, which reminds me of our organizational approach at AI2: students are regarded as peers and are directly integrated into large language model teams.

This is very different from top U.S. labs. In the U.S., companies like OpenAI, Anthropic, Cursor, etc., generally do not offer internships. Other companies like Google nominally offer internships related to Gemini, but many worry that their internships might be isolated from core work.

In summary, this slight cultural difference may enhance model-building capabilities in the following ways: people are more willing to do less glamorous work to improve the final model; newcomers to AI development, unaffected by previous hype cycles, can adapt more quickly to modern techniques; some Chinese scientists explicitly see this as an advantage; lower self-awareness makes organizational expansion easier because people are less likely to “game the system”; large talent pools are well suited to solve problems that have already been conceptually validated elsewhere, and so on.

This more favorable capacity for building current language models contrasts with a known stereotype: that Chinese researchers produce less of the “from 0 to 1” creative academic research that opens new fields.

In several more academic-oriented lab visits during this trip, many leaders mentioned they are cultivating a more ambitious research culture. Meanwhile, some technical leaders we spoke with doubted whether this reshaping of scientific research could be achieved in the short term, as it would require redesigning education and incentive systems—an overhaul too large to happen under current economic conditions.

This culture seems to be training a group of students and engineers very skilled at “big language model construction games.” Of course, their numbers are also extremely sufficient.

These students told me that China is experiencing a talent outflow similar to the U.S.: many who previously considered an academic path now plan to stay in industry. The most interesting comment came from a researcher who originally wanted to become a professor; he said he wanted to be a professor to stay close to the education system, but then commented that education has already been “solved” by large language models—“Why do students still need to come talk to me?”

Fresh eyes entering the field of large language models is an advantage. Over the past few years, we’ve seen key paradigms in large language models constantly evolve: from expanding MoE, to reinforcement learning, to supporting intelligent agents. Excelling in any of these areas requires rapidly absorbing vast background information, including broader literature and internal technical stacks.

Students are accustomed to doing this kind of work and are willing to set aside all preconceptions about “what should be effective” with humility. They dive in wholeheartedly, dedicating their lives just to improve models.

These students are also surprisingly direct, without some of the philosophical distractions that can sidetrack scientists. When I ask them about the economic impact of models or long-term societal risks, Chinese researchers with complex views and a desire to influence these issues are noticeably fewer. They see their role as building the best models.

This subtle but very perceptible difference becomes most apparent when engaging in long conversations with an elegant, intelligent researcher who can clearly express themselves in English: when asking about more philosophical questions regarding AI, these fundamental issues hang in the air, and the researcher shows a simple confusion. For them, it’s a category error.

One researcher even cited Dan Wang’s famous judgment: compared to the lawyer-led U.S., China is governed by engineers. When discussing these issues, he used this analogy to emphasize their desire to build. In China, there is no systematic path to cultivate Chinese scientists’ star influence like the mainstream podcasts Dwarkesh or Lex.

I tried to ask Chinese scientists to comment on the future economic uncertainties caused by AI, or the moral debates about how models should behave; these questions ultimately revealed their backgrounds and education (already edited 1). They are extremely focused on their work, but grew up in a system that does not encourage discussion or expression about how society should be organized or changed.

Looking from a broader perspective, especially in Beijing, I felt it resembled the Bay Area: a competitive lab just a few minutes’ walk or drive away. After landing, I stopped by Alibaba’s Beijing campus on the way to the hotel. Over the next 36 hours, we visited Zhipu AI, Dark Side of the Moon, Tsinghua University, Meituan, Xiaomi, and 01.ai.

In China, taking Didi is very convenient. If you choose an XL model, you’re often assigned a small electric van with a massage chair. When asking researchers about talent competition, they said it’s very similar to what we experience in the U.S. Job hopping is normal, and where people choose to go depends largely on where the atmosphere is best at the moment.

In China, the large language model community feels more like an ecosystem than warring tribes. In many private conversations, I heard almost only respect for peers. All Chinese labs are very wary of ByteDance and its popular models like Doubao, because it’s the only leading closed-source lab in China. Meanwhile, all labs highly respect DeepSeek, considering it the most tasteful research-oriented lab in execution. In the U.S., private exchanges with lab members often spark sparks quickly.

What impressed me most about the humility of Chinese researchers is that they often shrug at the business level, saying that’s not their problem. In the U.S., it seems everyone is obsessed with industry trends at various levels—from data vendors, to compute, to funding.

Differences and Similarities Between China’s AI Industry and Western Labs

Today, building an AI model is so interesting because it’s no longer just about gathering talented researchers in one building to create an engineering miracle. It used to be more like that, but to sustain AI operations, large language models are becoming hybrids: involving building, deployment, funding, and driving adoption of this creation.

Top AI companies exist within complex ecosystems. These ecosystems provide funding, compute, data, and more resources to continuously push the frontier forward.

In Western ecosystems, the way of integrating the various input factors needed to create and sustain large language models has been relatively well conceptualized and mapped out. Anthropic and OpenAI are typical representatives. Therefore, if we can identify clear differences in how Chinese labs think about these issues, we can see which companies might bet on meaningful distinctions in the future. Of course, these futures will also be strongly influenced by funding and/or compute constraints.

Here are the main “AI industry level” insights I gathered from conversations with these labs:

First, early signs of domestic AI demand are emerging. A widely discussed hypothesis is that China’s AI market will be smaller because Chinese companies are generally reluctant to pay for software, thus never unlocking a huge inference market capable of supporting labs.

But this judgment only applies to SaaS-related software spending. Historically, the SaaS ecosystem in China has always been small. On the other hand, China still has a large cloud market.

A key and unresolved question is whether Chinese enterprise AI spending will resemble SaaS—smaller scale—or cloud—more foundational expenditure. This is even being discussed within Chinese labs. Overall, I feel AI is trending more toward the cloud market, and no one is truly worried that markets around new tools won’t grow.

Second, most developers are heavily influenced by Claude. Although Claude is officially banned in China, most Chinese AI developers are very enamored with Claude and how it has changed their software development approach. Just because China has historically been less willing to buy software doesn’t mean I believe a huge inference demand wave won’t emerge.

Chinese technologists are very pragmatic, humble, and motivated. This impression is even stronger than any historical habit of “not paying for software.”

Some Chinese researchers mention they use their own tools, like Kimi or GLM command-line tools, but everyone mentions using Claude. Surprisingly, few mention Codex, which is apparently rapidly gaining popularity in the Bay Area.

Third, Chinese companies have a strong sense of technological ownership. Chinese culture is combining with a roaring economic engine, producing unpredictable results. A deep impression I took away is that the numerous AI models reflect a pragmatic balance among many local tech companies. There is no overarching plan.

This industry is defined by respect for ByteDance and Alibaba, which are seen as major incumbents likely to win many markets through strong resources. DeepSeek is a respected technical leader but far from a market leader. They set directions but lack the economic structure to dominate markets.

This leaves companies like Meituan or Ant Group. Westerners might be surprised why they are also building these models. But in fact, they clearly see large language models as core to future tech products, and thus need a strong foundation.

When fine-tuning a powerful general model, feedback from open-source communities makes their tech stacks more robust, while they can also keep internal fine-tuned versions for their products. The “open first” mindset in this industry is largely pragmatic: it helps models get strong feedback, benefits open source communities, and empowers their own missions.

Fourth, government support is real but its scale remains unclear. People often assert that the Chinese government is actively helping to promote open large language model competitions. But it’s a relatively decentralized system with many layers, each lacking a clear manual on what they should do.

Different districts in Beijing compete to attract tech companies’ offices. The “help” offered to these companies almost certainly includes removing bureaucratic hurdles like licensing. But how far can this help go? Can different government levels attract talent? Can they help smuggle chips?

Throughout the visits, many mentions of government interest or assistance occurred, but the information was far from enough to confidently report details or form a clear worldview on how government might influence China’s AI trajectory.

Of course, there’s no sign that China’s top leadership is directly influencing specific technical decisions on models.

Fifth, the data industry is far less developed than in the West. We’ve heard that Anthropic or OpenAI spend over $10 million per environment, with cumulative annual expenditures reaching hundreds of millions of dollars to push reinforcement learning frontiers. So we wonder: are Chinese labs also buying similar environments from U.S. companies, or is there a mirrored domestic ecosystem supporting them?

The answer isn’t “no data industry,” but rather that, based on their experience, the quality of the data industry is relatively poor, so many prefer to build their own environments or data internally. Researchers spend a lot of time creating reinforcement learning training environments, while larger companies like ByteDance or Alibaba can have internal data annotation teams to support this. All of this echoes the “build rather than buy” mentality mentioned earlier.

Sixth, there is a very strong demand for more NVIDIA chips. NVIDIA’s compute power is the gold standard for training, and progress is limited without more compute. If supply were sufficient, they would obviously buy more. Other accelerators, including Huawei’s, have received positive reviews for inference. Countless labs can use Huawei chips.

These points paint a very different AI ecosystem. Applying Western lab operation models to Chinese counterparts often results in category errors. The key question is whether these different ecosystems will produce substantively different model types or whether Chinese models will always be interpreted as similar to U.S. frontier models from 3 to 9 months ago.

Conclusion: Global Equilibrium

Before this trip, I knew very little about China; now I feel I’ve just begun to learn. China is not a place that can be fully described by rules or formulas, but one with very different driving mechanisms and chemical reactions. Its culture is so ancient, so profound, and still deeply intertwined with how domestic tech development unfolds. I have much more to learn.

Many parts of the current U.S. power structure regard their existing view of China as a key psychological tool in decision-making. After formal and informal face-to-face exchanges with nearly every top Chinese AI lab, I find that China possesses many qualities and instincts that are difficult for Western decision-making to model.

Even when I directly asked these labs why they open-source their most powerful models, I find it hard to fully connect the “ownership mentality” with “sincere support for the ecosystem.”

These labs are very pragmatic; they are not necessarily absolute open-source advocates, and not every model they build is open-sourced. But they have deep intentions to support developers, foster ecosystems, and see openness as a way to better understand their own models.

Almost every major Chinese tech company is building its own general large language model. We’ve already seen Meituan, Xiaomi, and others release open weights models. Western counterparts usually only purchase services.

These companies build large language models not just to gain presence in trending new areas but from a deep, fundamental desire: to control their own tech stacks and develop the most important technologies today. When I look up and see cranes on the horizon, it clearly aligns with China’s broader construction culture and energy.

The warmth, charm, and sincerity of Chinese researchers are very close and approachable. The harsh geopolitical discussions we’re used to in the U.S. do not seem to penetrate their world at all. The world could use more of this simple positivity. As a member of the AI community, I am now more concerned about cracks forming between members and groups around national labels.

If I said I wouldn’t want U.S. labs to be the clear leaders in every part of the AI tech stack, I’d be lying. Especially in the open model space I’ve invested so much time in, I am American—this is an honest preference.

At the same time, I hope the open ecosystem can thrive globally, as it can create safer, more accessible, and more useful AI for the world. The current question is whether U.S. labs will act to seize this leadership position.

As I finish this article, more rumors about government orders affecting open models are circulating. This could further complicate the synergy between U.S. leadership and the global ecosystem—and it does not make me more confident.

Thanks to all the excellent people I’ve spoken with at Dark Side of the Moon, Zhipu AI, Meituan, Xiaomi, Tongyi Qianwen, Ant Lingguang, 01.ai, and other institutions. Everyone has been so enthusiastic and generous with their time. As my thoughts continue to take shape, I will keep sharing observations about China, including broader cultural aspects and the AI field itself.

Clearly, these insights are directly related to the unfolding story of AI frontiers.