Editor's Note: Chinese AI labs are becoming an increasingly undeniable force in the global competition for large-scale AI models. Their advantages stem not only from a large pool of talent, strong engineering capabilities, and rapid iteration, but also from a very practical organizational approach: less talk of concepts, more model building; less emphasis on individual stars, more emphasis on team execution; less reliance on external services, and a preference for mastering the core technology stack themselves.

After visiting several leading AI labs in China, author Nathan Lambert found that the AI ​​ecosystem in China is not entirely the same as in the United States. The United States places greater emphasis on original paradigms, capital investment, and the personal influence of top scientists; China, on the other hand, is better at quickly catching up in existing areas, rapidly pushing model capabilities to the forefront through open source, engineering optimization, and the investment of a large number of young researchers.

The most noteworthy point is not whether China's AI has surpassed that of the United States, but rather that two different development paths are emerging: the United States is more like a cutting-edge competition driven by capital and star labs, while China is more like an industrial competition driven by engineering capabilities, open-source ecosystem, and technological self-control.

This means that future AI competition will not only be a battle of model rankings, but also a competition of organizational capabilities, developer ecosystems, and industry execution. The real change in China's AI is that it is no longer simply replicating Silicon Valley, but participating in the global forefront in its own way.

The following is the original text:

Sitting on the new high-speed train from Hangzhou to Shanghai, I looked out the window and saw undulating mountain ridges dotted with wind turbines, their silhouettes silhouetted against the setting sun. The mountains formed a backdrop, while before me stretched vast fields interspersed with clusters of high-rise buildings.

I returned from China with utmost humility. To receive such a warm welcome in such an unfamiliar place was a truly heartwarming and human experience. I had the privilege of meeting many people in the AI ​​ecosystem—people I had only known from afar before; their radiant smiles and enthusiasm reminded me once again that my work, and the entire AI ecosystem itself, is global.

The Mindset of Chinese Researchers

Chinese companies building language models are well-suited to be "fast followers" of this technology. They build upon China's long-standing tradition of education and work culture, while also adopting a slightly different approach to building technology companies compared to the West.

If we only look at the outputs—the latest and largest models, and the agent-based workflows they support—and then at the inputs—such as top scientists, massive amounts of data, and accelerated computing resources—then Chinese and American labs appear largely similar. The real, long-standing differences lie in how these elements are organized and shaped.

I've always thought that one reason Chinese labs are so adept at catching up and staying close to the forefront is that their culture is a perfect fit for the task. But without speaking directly with people, I didn't feel comfortable attributing this intuition to some significant influence. After talking with many excellent, humble, and open-minded scientists in leading Chinese labs, many of my thoughts have become clearer.

Building the best large language models today largely depends on meticulous work across the entire technology stack: from data and architectural details to the implementation of reinforcement learning algorithms. Each part of the model has the potential to improve performance, and combining these improvements is a complex process. In this process, the work of some very bright individuals may have to be put on hold to maximize the overall model's performance in multi-objective optimization.

American researchers are clearly also very adept at solving problems involving individual components, but the US has a stronger culture of "speaking up for oneself." As a scientist, you tend to be more successful when you actively seek attention for your work; and contemporary culture is also driving a new path to fame: becoming a "top AI scientist." This creates direct conflict.

Widely rumored, the Llama organization collapsed under political pressure after these interests were embedded in a hierarchical structure. I've also heard from other labs that sometimes it might be necessary to "appease" a top researcher and get them to stop complaining that their ideas weren't incorporated into the final model. Whether this is entirely true or not, the point is clear: self-awareness and the desire for career advancement can indeed hinder people from building the best models. Even a small, directional cultural difference like this between the US and China can have a meaningful impact on the final output.

Part of the difference relates to who in China is actually building these models. A stark reality across all the labs is that a large proportion of the core contributors are still students. These labs are quite young, which reminds me of our organizational structure at Ai2: students were treated as peers and directly integrated into the large language model team.

This is very different from top labs in the US. In the US, companies like OpenAI, Anthropic, and Cursor don't offer internships at all. Other companies like Google may nominally offer internships related to Gemini, but many people worry that their internships might be isolated from the truly core work.

In summary, these slight cultural differences may enhance model-building capabilities in the following ways: people are more willing to do less glamorous work to improve the final model; those new to AI development may be less affected by previous AI hype cycles and thus adapt more quickly to new modern technological methods. In fact, a Chinese scientist I spoke with explicitly considered this an advantage; lower self-awareness makes organizational structures easier to scale to some extent because people are less likely to try to "play with the system"; a large pool of talent is well-suited for solving problems that already have proof-of-concept elsewhere, and so on.

This tendency to build contemporary language models contrasts with a known stereotype: that Chinese researchers produce fewer of the more creative, groundbreaking "from 0 to 1" academic studies.

During several more academic lab visits during this trip, many leaders mentioned that they were cultivating a more ambitious research culture. At the same time, some technical leaders we spoke with doubted whether this reshaping of scientific research methods was possible in the short term, as it requires a complete redesign of the education and incentive systems, a transformation too large to be achieved under current economic conditions.

This culture appears to be training a group of students and engineers who are highly skilled at "large language model building games." And, of course, their numbers are extremely high.

These students told me that China is experiencing a similar brain drain to the United States: many people who previously considered academia are now planning to stay in industry. One of the most interesting comments came from a researcher who originally wanted to be a professor. He said he wanted to be a professor because he wanted to be closer to the education system; but he then commented that education has already been solved by the big language model—"Why would students still come to me to chat!"

Students entering the field of large language models with fresh perspectives is an advantage. Over the past few years, we've seen key paradigm shifts in large language models: from extending MoE to extending reinforcement learning, and then to supporting intelligent agents. Doing any of these things well requires an extremely rapid absorption of vast amounts of background information, including both broader literature and the technology stack within one's own company.

Students are used to doing this kind of thing and are willing to humbly let go of all preconceived notions about "what should work." They dive headfirst into it, dedicating their lives to it, just to gain the opportunity to improve the model.

These students were also remarkably direct, without any of the philosophical minutiae that would distract scientists. When I asked them how they viewed the economic implications or long-term social risks of their models, there were significantly fewer Chinese researchers with complex perspectives and a desire to exert influence on these issues. They saw their role simply as building the best possible models.

This difference is subtle and easily dismissed. But it's most readily apparent when you're having a long conversation with an elegant, intelligent researcher who can speak clearly in English: when you ask more philosophical questions about AI, these fundamental questions hang in the air, and the other person displays a simple bewilderment. To them, it's a categorical error.

One researcher even cited Dan Wang's famous observation: compared to the lawyer-dominated United States, China is governed by engineers. He used this analogy to emphasize their desire for development when discussing these issues. In China, there is no systematic path to cultivating the star power of Chinese scientists, unlike mainstream podcasts like Dwarkesh or Lex.

I tried to get Chinese scientists to comment on the economic uncertainties caused by AI, the problems beyond simple AGI capabilities, or the ethical debates about how models should behave; these questions ultimately revealed the backgrounds and education of these scientists (edited 1). They were extremely focused on their work, but they grew up in a system that did not encourage discussion and expression of how society should be organized and how it should change.

Looking at it from a broader perspective, especially in Beijing, it felt a lot like the Bay Area: a competitive lab might be just a few minutes' walk or taxi ride away. After landing, I stopped by Alibaba's Beijing campus on my way to the hotel. Over the next 36 hours, we visited Zhipu AI, Lunar Dark Side, Tsinghua University, Meituan, Xiaomi, and 01.ai.

Taking Didi Chuxing (a ride-hailing app) is very convenient in China. If you choose the XL model, you'll often be assigned to a small electric van with a massage chair. We asked researchers about the talent war, and they said it's very similar to what we experience in the US. Researchers frequently change jobs, and where people choose to go largely depends on where the atmosphere is best at the moment.

In China, the large language model community feels more like an ecosystem than a warring tribe. In many private conversations, I heard almost exclusively respect for peers. All Chinese labs are wary of ByteDance and its popular Doubao model, as it's the only cutting-edge closed-source lab in China. Meanwhile, all labs deeply respect DeepSeek, considering it the most research-savvy lab at the executive level. In the US, sparks often fly quickly when you have private conversations with lab members.

What struck me most about the humility of Chinese researchers was their tendency to shrug off business matters, saying it wasn't their problem. In contrast, in the US, everyone seemed engrossed in various industry trends across different ecosystems, from data sellers to computing power to financing.

Differences and similarities between China's AI industry and Western laboratories

What makes building an AI model so interesting today is that it's no longer just about bringing a group of brilliant researchers together in one building to create an engineering marvel. That was certainly more of a thing in the past, but to sustain the AI ​​business, large language models are becoming a hybrid: it involves building, deploying, funding, and driving the adoption of that creation.

Leading AI companies exist within complex ecosystems. These ecosystems provide funding, computing power, data, and more to continuously drive cutting-edge development.

In the Western ecosystem, the integration of the various inputs required to create and maintain large language models has been relatively well conceptualized and mapped. Anthropic and OpenAI are typical examples. Therefore, if we can find significant differences in the way Chinese labs think about these issues, we can see what meaningful differences different companies might be betting on in the future. Of course, these futures will also be strongly influenced by funding and/or computing power constraints.

I have summarized the biggest takeaways from my exchanges with these laboratories at the "AI industry level" as follows:

First, there are early signs of domestic demand for AI.
One widely discussed hypothesis is that the Chinese AI market will be smaller because Chinese companies are generally unwilling to pay for software, and therefore will never be able to unlock a large enough inference market to support labs.

However, this assessment only applies to software spending corresponding to the SaaS ecosystem. And the SaaS ecosystem in China has historically been very small. On the other hand, China clearly still has a huge cloud market.

A key, yet unanswered question is: will Chinese companies' spending on AI resemble the SaaS market—smaller in scale—or the cloud market—basic spending? This question is being discussed even within Chinese labs. Overall, I feel AI is moving closer to the cloud market, and no one is truly worried that the market surrounding new tools won't grow.

Second, most developers are heavily influenced by Claude.
Despite Claude being nominally banned in China, most AI developers in China are fascinated by it and how it has changed the way they build software. Just because China has historically been less willing to buy software doesn't mean I'd assume there won't be a massive surge in inference demand there.

Chinese technical personnel are very pragmatic, humble, and motivated. This struck me more strongly than any historical habit of "not buying software."

Some Chinese researchers mentioned using their own tools for building, such as Kimi or GLM's command-line tools, but everyone mentioned using Claude. Surprisingly, very few people mentioned Codex, which is clearly gaining popularity rapidly in the Bay Area.

Third, Chinese companies have a technology ownership mentality.
Chinese culture is merging with a booming economic engine, producing some unpredictable results. One profound impression I've gained is that the sheer number of AI models reflects a pragmatic balance among many tech companies here. There isn't a single overarching plan.

This industry is defined by a respect for ByteDance and Alibaba, large incumbents perceived as having the resources to win significant market share. DeepSeek is a respected technology leader, but far from a market leader. They set the direction, but lack the economic structure to win the market.

This has left companies like Meituan or Ant Group. Westerners might be surprised that they are also building these models. But in reality, they clearly see large language models as the core of future technological products, and therefore they need a strong foundation.

When companies fine-tune a robust, general-purpose model, the feedback from the open-source community strengthens their technology stack, while also allowing them to retain internally tweaked versions of their products. This "open-first" mentality in the industry is largely defined by pragmatism: it helps models receive strong feedback, gives back to the open-source community, and empowers their own mission.

Fourth, government support is real, but the scale is unclear.
It is often asserted that the Chinese government is actively supporting the Open Large Language Modeling Competition. However, this is a relatively decentralized government system comprised of many layers, and each layer lacks a clear operating manual specifying its own responsibilities.

Different neighborhoods in Beijing compete to attract tech companies to set up offices there. The "assistance" offered to these companies almost certainly includes removing bureaucratic red tape, such as license renewals. But how far can this assistance go? Can different levels of government help attract talent? Can they facilitate chip smuggling?

Throughout the visit, there were indeed many mentions of government interest or assistance, but the relevant information was far from sufficient for me to report details in an assertive manner, nor was it enough for me to form a confident worldview about how the government could actually change the trajectory of AI development in China.

Of course, there is absolutely no indication that the highest levels of the Chinese government influenced any technical decisions regarding the model.

Fifth, the data industry is far less developed than that of the West.
We've previously heard that Anthropic or OpenAI spend over $10 million on a single environment, with cumulative annual expenditures reaching hundreds of millions of dollars to advance the frontiers of reinforcement learning. Therefore, we're curious whether Chinese labs are also purchasing similar environments from American companies, or whether there's a mirror-like domestic ecosystem supporting them.

The answer isn't that there's no data industry at all, but rather that, based on their experience, the quality of data in the industry is relatively poor. Therefore, in many cases, it's better to build the environment or data in-house. Researchers spend a significant amount of time creating reinforcement learning training environments themselves, while larger companies like ByteDance and Alibaba can have in-house data annotation teams to support this. All of this echoes the "build it yourself rather than buy it" mentality mentioned earlier.

Sixth, the demand for more Nvidia chips is extremely strong.
Nvidia's computing power is the gold standard for training, and everyone's progress is limited by the lack of more computing power. If the supply were sufficient, they would obviously buy it. Other accelerators, including but not limited to Huawei's, have received positive reviews for inference. Countless labs have access to Huawei chips.

These key points depict a very different AI ecosystem. Quickly applying the operating methods of Western labs to Chinese counterparts often leads to misclassification. The crucial question is whether these different ecosystems will produce substantially different types of models; or whether Chinese models will always be interpreted as similar to cutting-edge US models from 3 to 9 months ago.

Conclusion: Global equilibrium

Before this trip, I knew so little about China; and as I left, I felt I was only just beginning to learn. China isn't a place that can be expressed by rules or recipes, but rather a place with very different dynamics and chemistry. Its culture is so ancient, so profound, and still so completely intertwined with the way technology is built domestically. I still have so much to learn.

Many parts of the current U.S. power structure use their existing views of China as a key psychological tool in decision-making. After having formal or informal face-to-face exchanges with almost every leading AI lab in China, I have found that China possesses many qualities and instincts that are difficult to model using Western decision-making methods.

Even if I directly ask these labs why they are releasing their most powerful models, I still find it difficult to fully connect the intersection between "ownership mentality" and "genuine support for the ecosystem".

The labs here are very pragmatic; they aren't necessarily open-source absolutists, and not every model they build will be released publicly. But they have a deep intention to support developers, support the ecosystem, and use openness as a way to further understand their own models.

Almost every major Chinese tech company is building its own general-purpose big language model. We've seen platform service companies like Meituan and large consumer tech companies like Xiaomi release open weighted models. In contrast, their American counterparts typically only purchase services.

These companies are building large language models not to make a name for themselves in trendy new things, but out of a deep and fundamental desire: to control their own technology stack and develop the most important technologies of the moment. When I look up from my laptop and see clusters of cranes on the horizon, it clearly resonates with China's broader culture and energy of construction.

The human touch, charm, and genuine warmth of Chinese researchers are incredibly endearing. On a personal level, the kind of brutal geopolitical discussions we're accustomed to in the US completely lack their presence. The world could have more of this simple positivity. As a member of the AI ​​community, I'm now more concerned about the growing rifts between members and groups centered around nationality labels.

It would be a lie to say I don't want American labs to be clear leaders in every part of the AI ​​technology stack. Especially in the area of ​​open models, where I've devoted a lot of time, being American is an honest preference.

At the same time, I hope that the open ecosystem itself can thrive globally, because it can create safer, more accessible, and more useful AI for the world. The question now is whether US labs will take action to assume this leadership position.

As I finish writing this, more rumors are circulating about the impact of the executive order on the open model. This could further complicate the synergy between U.S. leadership and the global ecosystem—and it does little to give me more confidence.

I'd like to thank all the wonderful people I had the opportunity to speak with at Lunar Dark Side, Zhipu, Meituan, Xiaomi, Tongyi Qianwen, AntLight, 01.ai, and other organizations. Everyone was so enthusiastic and generous with their time. As my ideas take shape, I will continue to share my observations about China, including broader cultural aspects and the field of AI itself.

Clearly, this knowledge will be directly related to the unfolding story of cutting-edge AI developments.