Notes from inside China's AI labs
Original author: Nathan Lambert
Compiled by: Peggy, BlockBeats
Editor’s Note: Chinese AI labs are becoming an increasingly hard-to-ignore force in the global large model competition. Their advantage lies not just in abundant talent, strong engineering, and rapid iteration, but also in a highly practical organizational approach: talk less about concepts, build more models; emphasize team execution over individual stars; rely less on external services and prefer to own their core technology stack.
After visiting several leading Chinese AI labs, author Nathan Lambert found that China’s AI ecosystem differs from that of the United States. The U.S. places greater emphasis on original paradigms, capital investment, and the personal influence of top scientists; China, by contrast, excels at rapidly catching up in existing directions through open-source initiatives, engineering optimizations, and the contributions of a large number of young researchers, quickly pushing model capabilities to the frontier.
What matters most is not whether China’s AI has surpassed that of the United States, but that two distinct development paths are emerging: the U.S. resembles a frontier race driven by capital and star labs, while China resembles an industrial competition fueled by engineering capabilities, an open-source ecosystem, and a mindset of technological self-reliance.
This means that future AI competition will not only be about model leaderboards, but also about organizational capability, developer ecosystems, and industrial execution. The real transformation in China’s AI lies in its shift from merely copying Silicon Valley to participating in global frontiers in its own way.
The following is the original text:
Sitting on the high-speed train from Hangzhou to Shanghai, I gazed out the window and saw distinct ridgelines dotted with wind turbines, silhouetted against the setting sun. The mountains formed the backdrop, while before me stretched vast fields interwoven with clusters of high-rise buildings.
I returned from China with great humility. Being welcomed so warmly to such a foreign place was a profoundly warm and deeply human experience. I had the privilege of meeting many people from the AI ecosystem whom I had previously only known from afar; they greeted me with radiant smiles and open hearts, reminding me once again that my work—and the entire AI ecosystem itself—is truly global.
The mindset of Chinese researchers
Chinese companies building language models are well-positioned to be rapid followers of this technology, grounded in China’s long-standing educational and work culture traditions, and employing slightly different approaches to building tech companies compared to the West.
If you look only at outputs—such as the latest and largest models, and the agent-based workflows these models enable—and at input factors like top-tier scientists, large-scale data, and accelerated computing resources, Chinese and U.S. labs appear largely similar. The true, enduring differences emerge in how these elements are organized and shaped.
I have always believed that one reason Chinese labs are so adept at catching up and staying near the frontier is that their culture aligns very well with this task. But before speaking directly with people, I felt it wasn’t appropriate to attribute this intuition to any significant influence. After speaking with many outstanding, humble, and open scientists at China’s top labs, many of my ideas became much clearer.
Building the best large language model today depends heavily on meticulous work across the entire technology stack: from data, to architectural details, to the implementation of reinforcement learning algorithms. Each component of the model can offer incremental improvements, and combining these improvements into a cohesive whole is a complex process. During this process, the work of some highly intelligent individuals may need to be set aside to maximize the overall model’s performance in multi-objective optimization.
American researchers are also clearly very skilled at solving individual component problems, but America has a stronger culture of "speaking up for oneself." As a scientist, when you actively advocate for attention toward your own work, you tend to be more successful; and contemporary culture is promoting a new path to fame—becoming a "top AI scientist." This leads to direct conflict.
It is widely rumored that the Llama organization collapsed under political pressure after these interests were embedded into a hierarchical structure. I’ve also heard from other labs that sometimes it’s necessary to “pacify” a top researcher to get them to stop complaining that their ideas weren’t included in the final model. Whether or not this is entirely true, the message is clear: self-awareness and the desire for career advancement can indeed hinder the creation of the best models. Even a small cultural difference in this regard between the U.S. and China could have a meaningful impact on the final outcome.
Part of the difference relates to who in China is building these models. In all labs, a straightforward reality is that a large proportion of core contributors are still students. These labs are quite young, which reminds me of how we organize things at Ai2: students are treated as peers and directly integrated into large language model teams.
This is very different from top labs in the United States. In the U.S., companies like OpenAI, Anthropic, and Cursor simply do not offer internships. Other companies, such as Google, nominally offer internships related to Gemini, but many worry that their internships might be isolated from truly core work.
In summary, this subtle cultural difference may enhance model-building capabilities in several ways: people are more willing to do less glamorous work to improve the final model; newcomers to AI development may be less influenced by previous AI hype cycles, allowing them to adapt more quickly to modern techniques. In fact, a Chinese scientist I spoke with explicitly viewed this as an advantage: lower levels of self-consciousness make organizational structures somewhat easier to scale, as fewer people try to “game the system”; a large talent pool is well-suited to solving problems that already have proof-of-concept solutions elsewhere, and so on.
This better reflects the current capabilities of language models, contrasting with a known stereotype: that Chinese researchers produce less of the highly creative, field-defining "from 0 to 1" academic research.
During several more academically oriented lab visits on this trip, many leaders spoke of cultivating a more ambitious research culture. Meanwhile, some technical leaders we spoke with expressed skepticism that this reshaping of scientific research could be achieved in the short term, as it requires redesigning educational and incentive systems—a transformation too large to occur under current economic equilibria.
This culture seems to be training a large cohort of students and engineers who are exceptionally skilled at the "large language model building game." Of course, their numbers are also extremely abundant.
These students told me that a similar brain drain is also occurring in China: many who previously considered pursuing academic careers now plan to stay in industry. One of the most interesting comments came from a researcher who originally wanted to become a professor because he hoped to be close to the education system; he then added that education has already been solved by large language models—“Why would students still come to talk to me!”
Students entering the field of large language models bring a fresh perspective, which is an advantage. Over the past few years, we have seen key paradigms in large language models continuously evolve: from scaling MoE, to scaling reinforcement learning, to supporting agents. Doing any of these things well requires rapidly absorbing vast amounts of background information, encompassing both the broader literature and the company’s internal tech stack.
Students are accustomed to doing this and are willing to set aside all preconceptions about what "should work" with humility. They dive in wholeheartedly, dedicating their lives to gaining opportunities to improve the model.
These students were also refreshingly direct, with none of the philosophical digressions that might distract scientists. When I asked them about the economic impact or long-term social risks of the models, there were far fewer Chinese researchers with nuanced perspectives who wanted to influence these issues. They saw their role simply as building the best possible models.
This difference is subtle and easily denied. But it becomes most apparent during a long conversation with an elegant, intelligent researcher who can express themselves clearly in English: when you ask deeper, more philosophical questions about AI, these fundamental issues hang in the air, and the other person reveals a simple confusion. To them, it is a category error.
One researcher even cited Dan Wang’s famous characterization: whereas the United States is governed by lawyers, China is governed by engineers. In discussing these issues, he used this analogy to emphasize their desire to build. In China, there is no systematic pathway to cultivate the star influence of Chinese scientists in the way that super-mainstream podcasts like Dwarkesh or Lex do.
I tried to get Chinese scientists to comment on the economic uncertainties sparked by AI, issues beyond simple AGI capabilities, or moral debates about how models should behave; these questions ultimately revealed the scientists’ upbringing and educational backgrounds (edited). They are extremely focused on their work, but they grew up in a system that does not encourage discussion or expression about how society should be organized or changed.
Looking at the bigger picture, especially Beijing, it feels a lot like the Bay Area: a competitive lab where the next breakthrough could be just a few minutes away by foot or taxi. After landing, I stopped by Alibaba’s Beijing campus on the way to my hotel. Over the next 36 hours, we visited Zhipu AI, Moonshot AI, Tsinghua University, Meituan, Xiaomi, and 01.ai.
Taking a DiDi in China is very convenient. If you choose the XL vehicle option, you’re often assigned an electric minivan with massage seats. We asked researchers about the talent war, and they said it’s very similar to what we’ve experienced in the U.S. It’s normal for researchers to switch jobs, and where people choose to go depends largely on where the atmosphere is currently the best.
In China, the large language model community feels more like an ecosystem than warring tribes. In many private conversations, I’ve heard almost nothing but respect for peers. All Chinese labs hold ByteDance and its popular Doubao model in high regard, as it is the only leading closed-source lab in China. Meanwhile, all labs deeply respect DeepSeek, viewing it as the lab with the most refined research taste in execution. In the U.S., private conversations with lab members often spark immediate tension.
What impresses me most about Chinese researchers' humility is that they often shrug and say it’s not their problem, even on the business level. In the U.S., it seems everyone is obsessed with industry trends across every layer of the ecosystem—from data providers to computing power to fundraising.
Differences and similarities between China's AI industry and Western labs
Today, building an AI model is so fascinating because it’s no longer just about gathering a group of brilliant researchers in the same building to create an engineering marvel. It used to be more like that, but to sustain AI businesses, large language models are now becoming hybrids: they involve building, deploying, financing, and driving adoption of this creation.
Leading AI companies exist within complex ecosystems. These ecosystems provide funding, computing power, data, and other resources to continuously push the frontier forward.
In the Western ecosystem, the integration of various inputs required to create and maintain large language models has been relatively well conceptualized and mapped out. Anthropic and OpenAI are typical examples. Therefore, if we can identify clear differences in how Chinese labs think about these issues, we can discern potential meaningful distinctions that different companies may bet on in the future. Of course, these futures will also be strongly influenced by funding and/or compute constraints.
Here are the key takeaways at the AI industry level from my discussions with these labs:
First, early signs of domestic AI demand have already emerged.
There is a widely discussed hypothesis that China’s AI market will be smaller because Chinese companies are typically unwilling to pay for software, thus never unlocking a large enough inference market to support labs.
But this judgment applies only to software spending corresponding to the SaaS ecosystem, which has historically been small in China. On the other hand, China clearly still has a massive cloud market.
A key and still unanswered question is whether Chinese enterprises’ spending on AI will resemble the SaaS market—smaller in scale—or the cloud market—foundational spending. This is even a topic of internal discussion within Chinese labs. Overall, I feel AI is moving closer to the cloud market, and no one truly worries that the market around these new tools won’t grow.
Second, most developers are deeply influenced by Claude.
Although Claude is officially banned in China, most Chinese AI developers are deeply fascinated by Claude and how it has transformed the way they build software. Just because China has been less willing to purchase software in the past doesn’t mean I believe China won’t experience a massive surge in demand for reasoning capabilities.
Chinese technical professionals are extremely practical, humble, and motivated. This impression on me is stronger than any historical habit of "not paying for software."
Some Chinese researchers mention using their own tools for building, such as the command-line tools for Kimi or GLM, but everyone mentions using Claude. Surprisingly, few mention Codex, even though Codex is rapidly gaining popularity in the Bay Area.
Third, Chinese companies have a mindset of technological ownership.
Chinese culture is combining with a roaring economic engine, producing some unpredictable outcomes. One lasting impression I took away is that the sheer number of AI models reflects a pragmatic balance among many tech companies here. There is no overarching plan.
This industry is defined by respect for ByteDance and Alibaba — large incumbents seen as having the resources to dominate many markets. DeepSeek is a respected technology leader, but far from a market leader. They set the direction, but lack the structural capacity to win the market economically.
This leaves companies like Meituan or Ant Group. Westerners might be surprised that they are also building these models. But in reality, they clearly view large language models as central to future technology products and therefore need a strong foundation.
When they fine-tune a powerful general-purpose model, feedback from the open-source community strengthens their tech stack, while they can retain internal fine-tuned versions for their own products. This “open-first” mindset in the industry is largely defined by pragmatism: it helps models gain 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 claimed that the Chinese government is actively helping to open up the large language model competition. But this is a relatively decentralized government system composed of many levels, and each level does not have a clear set of operational guidelines defining exactly what it should do.
Different neighborhoods in Beijing compete to attract tech companies to set up their offices there. The "assistance" offered to these companies almost certainly includes streamlining bureaucratic processes such as removing licensing hurdles. But how far can this assistance go? Can different levels of government help attract talent? Can they help smuggle chips?
Throughout the visit, there were indeed many references to government interest or assistance, but the information was far from sufficient for me to report details with certainty or to form a confident worldview on how the government could truly alter the trajectory of AI development in China.
There is also absolutely no indication that China's top leadership is influencing any technical decisions of the model.
Fifth, the data industry is far less developed than in the West.
We had heard that Anthropic or OpenAI spends over $10 million on a single environment, with cumulative annual expenditures reaching hundreds of millions of dollars to push the frontiers of reinforcement learning. Therefore, we are curious whether Chinese labs are also purchasing the same environments from U.S. companies, or whether a mirrored domestic ecosystem supports them.
The answer is not that there is “no data industry” in the absolute sense, but rather that, based on their experience, the quality of the data industry is relatively poor; thus, it is often better to build internal environments or data themselves. Researchers frequently spend significant time creating reinforcement learning training environments, while larger companies like ByteDance and Alibaba can afford internal data labeling teams to support this. All of this echoes the earlier mentioned mindset of “building rather than buying.”
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 additional computing power. If supply were sufficient, they would clearly purchase more. Other accelerators, including Huawei among others, have received positive reviews for inference. Countless laboratories have access to Huawei chips.
These points paint a very different AI ecosystem. Applying the operational model of Western labs directly to their Chinese counterparts often results in category errors. The key question is whether these differing ecosystems will produce fundamentally distinct types of models, or whether Chinese models will always be interpreted as resembling U.S. frontier models from three to nine months prior.
Conclusion: Global Equilibrium
Before this trip, I knew very little about China; by the time I left, I felt I had only just begun to learn. China is not a place that can be described by rules or formulas, but rather one with very different dynamics and chemical reactions. Its culture is so ancient and profound, yet still completely intertwined with the way technology is built domestically. I have so much more to learn.
Many components of the current U.S. power structure treat their existing view of China as a key psychological tool in decision-making. After having formal and informal face-to-face conversations with nearly every leading AI lab in China, I’ve found that China possesses many qualities and instincts that are difficult for Western decision-making frameworks to model.
Even if I directly ask these labs why they openly release their most powerful models, I still struggle to fully connect the dots between a mindset of ownership and genuine support for the ecosystem.
The lab here is very pragmatic and not necessarily an absolutist when it comes to open source—not every model they build is openly released. However, they have a strong intention to support developers, nurture the ecosystem, and use openness as a way to better understand their own models.
Almost every major Chinese tech company is building its own general-purpose large language model. We have already seen platform-based service companies like Meituan and large consumer tech companies like Xiaomi release open-weight models. In contrast, their American counterparts typically only purchase services.
These companies build large language models not to gain visibility in trendy new phenomena, but out of a deeper, more fundamental desire: to control their own technology stack and develop the most important technology of today. Whenever I look up from my laptop and see clusters of cranes on the horizon, it clearly aligns with China’s broader culture and energy of construction.
The humanity, charm, and sincere warmth of Chinese researchers make them incredibly approachable. On a personal level, the harsh geopolitical debates we’re accustomed to in the U.S. simply don’t permeate them. The world could use more of this simple positivity. As a member of the AI community, I’m now more concerned that fractures are emerging between members and groups over nationality labels.
If I said I didn’t want U.S. labs to be clear leaders in every part of the AI technology stack, I would be lying. Especially in the field of open models, where I’ve invested a lot of time—I’m American, and this is an honest preference.
At the same time, I hope the open ecosystem itself can thrive globally, as this can create AI that is safer, more accessible, and more useful for the world. The current question is whether U.S. labs will take action to assume this leadership role.
As I finished writing this article, more rumors were circulating about how executive orders could impact open models. This could further complicate the synergy between U.S. leadership and the global ecosystem—and it hasn’t made me any more confident.
I thank all the outstanding individuals I had the privilege to speak with at Moonshot, Zhipu, Meituan, Xiaomi, Qwen, Ant Lingguang, 01.ai, and other organizations. Everyone was so enthusiastic and generously gave their time. As my ideas take shape, I will continue sharing observations about China, encompassing both broader cultural dimensions and the AI field itself.
Clearly, this knowledge is directly relevant to the story unfolding with the latest advancements in AI.
Click to learn about the open positions at BlockBeats
Welcome to the official BlockBeats community:
Telegram subscription group: https://t.me/theblockbeats
Telegram group: https://t.me/BlockBeats_App
Official Twitter account: https://twitter.com/BlockBeatsAsia