Editor's Note: China's AI labs are becoming an increasingly hard-to-ignore force in the global race of large models. Their advantage lies not only in abundant talent, strong engineering, and fast iteration but also in a very practical organizational approach: less talk about concepts, more focus on models; less emphasis on individual stars, more emphasis on team execution; less reliance on external services, more inclination towards mastering the core technology stack themselves.
After visiting several top AI labs in China, author Nathan Lambert found that the Chinese AI ecosystem is not exactly the same as that of the United States. The U.S. places more emphasis on original paradigms, capital investment, and the personal influence of top scientists; China, on the other hand, excels at rapidly catching up in existing directions, pushing model capabilities to the forefront quickly through open source, engineering optimization, and the involvement of numerous young researchers.
What is most worthy of attention is not whether Chinese AI has surpassed the U.S., but that two different development paths are emerging: the U.S. resembles more of a cutting-edge competition driven by capital and star labs, while China is more of an industry competition driven by engineering capabilities, open-source ecosystems, and technological self-awareness.
This implies that the future AI competition is not just a competition of model rankings but also a competition of organizational capabilities, developer ecosystems, and industrial execution capabilities. The real change in Chinese AI lies in its transformation from merely replicating Silicon Valley to 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 the undulating mountain ridges, with wind turbines dotted on the hills forming silhouettes under the setting sun. The mountains formed the backdrop, while vast fields and clusters of high-rises intertwined in front of my eyes.
I return from China with great humility. Going to such a foreign place and receiving such a warm welcome is a very warm and human experience. I was fortunate to meet many people in the AI ecosystem, people I had only known from afar before; and they greeted me with bright smiles and warmth, reminding me once again that my work and the entire AI ecosystem itself are global.
Mindset of Chinese Researchers
Chinese companies that are building language models can be said to be very suitable as "fast followers" in this technology. They are built on China's long-standing educational and work culture traditions, while also having a slightly different way of building technology companies compared to the West.
If we only look at the output, that is, the latest and largest models, and the workflows these models support; and then look at the input factors, such as top-notch scientists, massive amounts of data, and accelerated computing resources, then Chinese labs and American labs appear broadly similar. The real long-term difference lies in how these factors are organized and shaped.
I have always believed that one of the reasons Chinese labs are very good at catching up and staying near the frontier is that they are culturally very well-suited to this task. But without directly interacting with people, I felt uncomfortable attributing this intuition to any significant influence. After talking to many excellent, humble, and open scientists in the top Chinese labs, many of my ideas became clearer.
Today, building the best large language models depends largely on meticulous work throughout the entire tech stack: from data, to architectural details, to the implementation of reinforcement learning algorithms. Each part of the model has the potential to bring some improvement, and how these improvements are combined is a complex process. In this process, work done by some very clever individuals may need to be set aside to allow the overall model to be maximized in multi-objective optimization.
American researchers are obviously also very good at solving individual component problems, but the U.S. has more of a "speak up for yourself" culture. As a scientist, when you actively seek attention for your work, you often succeed more; and contemporary culture is also promoting a new path to fame, that is, becoming a "top AI scientist." This can lead to direct conflicts.
There have been widespread rumors that the Llama organization once collapsed under political pressure after these interest appeals were embedded in a hierarchical organization. I have also heard from other labs that sometimes it may be necessary to "appease" a top researcher to stop them from complaining that their ideas were not incorporated into the final model. Whether or not this is entirely true, the meaning is clear: self-awareness and career advancement desires can indeed hinder people from building the best models. Even a small directional cultural difference between the U.S. and China could have a meaningful impact on the final output.
Part of this difference has to do with who exactly is building these models in China. In all labs, a very direct reality is that a large proportion of the core contributors are students. These labs are quite young, which reminds me of how we are organized at Ai2: students are seen as peers and are directly integrated into the large language model team.
This is very different from the top labs in the U.S. In the U.S., companies like OpenAI, Anthropic, and Cursor do not offer internships at all. Companies like Google nominally offer internships related to Gemini, but many people are concerned that their internships will be isolated from the real core work.
In summary, this slight cultural difference may enhance model-building capabilities in the following ways: people are more willing to do the less glamorous work to improve the final model; those newly involved in AI development may be less influenced by previous AI hype cycles, allowing them to quickly adapt to new modern technical approaches. In fact, a Chinese scientist I spoke with made this point very clear; lower self-consciousness makes organizational structures somewhat easier to scale, as people are less likely to try to "game the system"; a large talent pool is well-suited to solving problems that have already been conceptually validated elsewhere, and so on.
This tendency towards a more conducive environment for building current language models contrasts with a known stereotype: people often believe that Chinese researchers produce less of the more creative, "from 0 to 1," academic research that can open up new fields.
In several more academic-focused lab visits during this trip, many leads mentioned that they are fostering this more ambitious research culture. Meanwhile, some of the tech leads we spoke with questioned whether this reshaping of scientific research methodology could be achieved in the short term, as it requires a redesign of the education system and incentive structure, which is a monumental task that is difficult to bring about under the current economic equilibrium.
This culture seems to be training a cohort of students and engineers who excel at the "big language model building game." Of course, their numbers are also abundant.
These students told me that China is also experiencing a talent outflow similar to the United States: many who previously considered an academic path are now intending to stay in the industry. One of the most intriguing statements came from a researcher who initially wanted to become a professor; he said he wanted to become a professor to be close to the education system, but then he went on to comment that education has already been solved by large language models — "why should students still come to me for a chat!"
Students entering the field of large language models with fresh eyes have an advantage. Over the past few years, we have seen a continuous evolution in the key paradigms of large language models: from scaling MoE, to scaling reinforcement learning, to supporting agents. To do any of these well requires rapidly absorbing a large amount of background information, including both a broader set of literature and the internal tech stack of the company.
Students are accustomed to doing these kinds of things and are willing to humbly set aside all preconceived notions of what "should work." They dive in headfirst, dedicating their lives to the opportunity to improve the model.
These students are also remarkably direct and lack some of the philosophical digressions that might distract scientists. When I asked them about how they view the economic impact of the model or long-term societal risks, Chinese researchers with complex views and a desire to influence these issues were noticeably scarce. They see their role as building the best model.
This difference is very subtle and easily denied. However, it is most easily felt when you engage in a long conversation with an elegant, intelligent, and English-proficient researcher: when you ask some more philosophical questions about AI, these fundamental questions hang in the air, and the other party reveals a simple confusion. For them, it is a categorical error.
One researcher even quoted Dan Wang's famous judgment: compared to the lawyer-led United States, China is governed by engineers. When discussing these issues, he used this analogy to emphasize the desire they have for what they want to build. In China, there is no systematic path to cultivate Chinese scientists' star power like super mainstream podcasters such as Dwarkesh or Lex.
I tried to get Chinese scientists to comment on the future economic uncertainty caused by AI, the issues beyond simple AGI capabilities, or the ethical debates on how models should behave; these questions ultimately showed me the growth and educational backgrounds of these scientists. They are extremely focused on their work, but they grew up in a system that does not encourage discussion and expression of how society should be organized or changed.
Zooming out, especially in Beijing, gave me a feeling like the Bay Area: a competitive lab possibly just a few minutes' walk or drive away. After getting off the plane, on the way to the hotel, I casually went to the Alibaba Beijing campus. Over the next 36 hours, we visited Zhipe AI, the dark side of the moon, Tsinghua University, Meituan, Xiaomi, and 01.ai.
It's very convenient to use DiDi for transportation in China. If you choose the XL vehicle type, you are often assigned to an electric small van with a massage chair. We asked researchers about talent wars, and they said it is very similar to what we experience in the United States. Researcher job-hopping is common, 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 warring tribes. In many off-the-record conversations, I heard almost only respect for peers. All Chinese labs are wary of ByteDance and its popular Bean Model because it is the only cutting-edge closed-source lab in China. At the same time, all labs hold DeepSeek in high regard, considering it the lab with the most research taste at the execution level. In the United States, when you have off-the-record conversations with lab members, sparks often fly.
One of the most impressive aspects of Chinese researchers' humility is that they often shrug on a business level, saying it's not their problem. In the United States, it seems like everyone is obsessed with various industrial trends at various ecosystem levels, from data vendors to computing power, and then to funding.
The Differences and Similarities Between the Chinese AI Industry and Western Labs
Today, what makes building an AI model so fascinating is that it is no longer just about bringing a group of brilliant researchers together in the same building to collectively create an engineering marvel. It used to be more like this in the past, but to sustain an AI business, large language models are becoming a hybrid: it involves building, deploying, financing, and driving the adoption of this creation.
Top AI companies exist within a complex ecosystem. These ecosystems provide funding, computing power, data, and other resources to continuously drive the frontier forward.
In the Western ecosystem, the integration of various input elements needed to create and maintain large language models has been relatively well conceptualized and mapped out. Anthropic and OpenAI are typical representatives. Therefore, if we can identify a significantly different way of thinking about these issues in Chinese labs, we can see meaningful differences that different companies may bet on in the future. Of course, these futures will also be heavily influenced by funding and/or computing power constraints.
I have summarized below some of the key "AI industry-level" takeaways I obtained from interacting with these labs:
First, early signs of domestic AI demand have emerged.
There is a widely debated assumption that the Chinese AI market will be smaller because Chinese companies are generally not willing to pay for software, thus never unleashing a vast reasoning market big enough to support a lab.
However, this judgment only applies to software spending corresponding to the SaaS ecosystem. The SaaS ecosystem has historically been small in China. On the other hand, China still obviously has a huge cloud market.
A key and as-yet-unanswered question is: Will Chinese enterprise spending on AI be more like the SaaS market, i.e., smaller in scale; or more like the cloud market, i.e., foundational spending. This question is also being discussed within Chinese labs. Overall, I feel that AI is moving closer to the cloud market, and no one is really worried that the market around new tools cannot grow.
Second, most developers are deeply influenced by Claude.
Although Claude is nominally banned in China, most AI developers in China are very enamored with Claude and how it has changed the way they build software. Just because China has been less willing to purchase software in the past does not mean that I would believe China will not see a huge surge in inference demand.
Chinese tech professionals are very pragmatic, humble, and driven. This feeling I get is stronger than any historical habit of "not spending money on software."
Some Chinese researchers will mention using their own tools for building, such as Kimi or GLM's command-line tool, but everyone will mention using Claude. Surprisingly, few people mention Codex, which is rapidly gaining popularity in the Bay Area.
Third, Chinese companies have a technology ownership mentality.
Chinese culture is blending with a roaring economic engine, producing some unpredictable outcomes. One profound feeling I left with is the multitude of AI models reflecting a pragmatic balance many tech companies here embody in reality. There is no grand plan.
This industry is defined by a reverence for ByteDance and Alibaba. They are considered large incumbents expected to win many markets with their strong resources. DeepSeek is a respected technology leader but far from a market leader. They set direction but lack the economic structure to win markets.
This leaves companies like Meituan or Ant Group. Westerners might be surprised why they are also building these models. In reality, they clearly see large language models as the core of future tech products, hence needing a strong foundation.
As they fine-tune a powerful general model, the open-source community's feedback on the model makes their tech stack more robust, while they can also keep internal fine-tuned versions for their products. The "open-first" mindset in this industry is largely defined by pragmatism: it helps the model gain strong feedback, gives back to the open-source community, and empowers their own mission.
Fourth, government support is indeed real but the scale is unclear.
It is often asserted that the Chinese government is actively aiding the open large language model competition. However, this is a relatively decentralized government system composed of many levels, and each level does not have a clear playbook dictating what they should do.
Different districts in Beijing compete to attract tech companies to set up offices there. The "help" provided to these companies almost certainly includes cutting through bureaucratic red tape such as licensing processes. But how far can this help go? Can different government levels help attract talent? Can they help smuggle chips?
Throughout the entire visit, there were indeed many mentions of government interest or assistance, but the 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 on how the government could truly change China's AI development trajectory.
Of course, there is also no indication whatsoever that the top echelons of the Chinese government are influencing any technical decisions regarding the models.
Fifth, the data industry in China is far less developed than in the West.
Previously, we heard that Anthropic or OpenAI would spend over $10 million for a single environment, with annual cumulative expenditures reaching the scale of hundreds of millions of dollars to advance the forefront of reinforcement learning. Therefore, we are very curious to know if Chinese labs are also purchasing the same environments from American companies, or if there is a mirrored domestic ecosystem supporting them.
The answer is not a complete "lack of data industry," but rather that, based on their experience, the quality of the data industry is relatively poor, so many times it is better to internally build environments or data. Researchers themselves spend a lot of time creating reinforcement learning training environments, while larger companies like ByteDance and Alibaba can have internal data labeling teams to support this. All of this echoes the mentality mentioned earlier of "building instead of buying."
Sixth, there is a strong desire for more NVIDIA chips.
NVIDIA's computing power is the gold standard for training, and everyone's progress is limited by a lack of more computing power. If supply is sufficient, they will clearly purchase more. Other accelerators, including but not limited to Huawei, have received positive evaluations in terms of inference. Countless labs can use Huawei chips.
These points paint a very different AI ecosystem. Trying to quickly apply the operational model of Western labs to their Chinese counterparts often leads to categorization errors. The key question is whether these different ecosystems will produce model types with substantial differences; or whether Chinese models will always be interpreted as similar to U.S. frontier models from 3 to 9 months ago.
Conclusion: Global Balance
Before this trip, I knew too little about China; and upon leaving, I feel like I have just begun to learn. China is not a place that can be expressed by rules or formulas, but a place with very different mechanisms and chemical reactions. Its culture is so ancient, so profound, and still completely intertwined with the way domestic technology is being developed. There is still much for me to learn.
Many parts of the current power structure in the United States view their existing views on China as a key psychological tool in decision-making. After formal or informal face-to-face exchanges with almost every top AI lab in China, I found that China has many qualities and instincts that are difficult to model using Western decision-making methods.
Even when I directly ask these labs why they open-source their strongest models, I still find it challenging to fully connect the intersection of the "ownership mindset" and "sincere ecosystem support."
These labs are very practical and not necessarily absolutists in terms of open-source philosophy, not every model they build is open-sourced. But they have deep intentions in supporting developers, supporting the ecosystem, and using openness as a way to further understand their own models.
Almost every major Chinese tech company is building its own large language model. We have seen platform service companies like Meituan and large consumer tech companies like Xiaomi release open-weight models. Similar companies in the United States usually only buy services.
These companies are building large language models not to make a splash in hot new things, but out of a deep and fundamental desire: to control their own tech stack and develop the most important technology of the moment. When I look up from my laptop and always see clusters of cranes on the horizon, this clearly aligns with China's broader culture of construction and development energy.
The warmth, charm, and genuine kindness of Chinese researchers are very endearing. On a personal level, the kind of brutal geopolitical discussions we are used to in the United States have not seeped into them at all. This world could have more of this simple positivity. As a member of the AI community, what worries me now is that cracks are emerging around nationality labels, members, and groups.
If I say I don't want American labs to be clear leaders in every part of the AI tech stack, I would be lying. Especially in the open model field where I spend a lot of time, I am American, and this is an honest preference.
At the same time, I hope the open ecosystem itself can thrive globally, as this can create a safer, more accessible, and more useful AI for the world. The current issue is whether American labs will take action to take up this leadership position.
As I finish writing this article, more rumors about executive orders affecting the open model are circulating. This could further complicate the collaborative relationship between American leadership and the global ecosystem—this does not make me more confident.
I thank all the excellent individuals I had the privilege to talk to at Moon's Dark Side, Zhi Spectrum, Meituan, Xiaomi, Tongyi Qianwen, Ant Group's Lingguang, 01.ai, and other institutions. Everyone has been so enthusiastic and so generous with their time. As my thoughts take shape, I will continue to share observations about China, including broader cultural aspects and the AI field itself.
Obviously, this knowledge will be directly related to the unfolding story of the frontier of AI development.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
