AI researcher Nathan Lambert recently visited multiple major Chinese AI labs, including Moonshot AI, Zhipu, Meituan, Xiaomi, Alibaba’s Qwen, Ant Ling, and 01.ai, and wrote up this in-depth observation log. He admits that the trip helped him re-understand China’s AI ecosystem; this record is not just a researcher’s travel journal, but a first-hand diagnostic report on China’s AI—from technical culture to industrial structure.
China AI’s core strengths: culture, talent, and a pragmatic mindset
Where China’s researchers’ competitiveness comes from: a tendency to bury their heads in hard work
Lambert believes there’s a key factor that’s often overlooked behind why Chinese labs can quickly catch up to—and even match—frontier AI: research culture and organizational atmosphere.
Compared with U.S. researchers who generally have a strong desire to perform individually—tending to speak up for their research results and build personal brands through media and social communities—Chinese researchers more often place themselves behind model quality. They’re more willing to work on tasks that quietly and genuinely improve model performance, and they’re also more able to accept that their ideas may be discarded during multi-objective optimization.
Lambert points out that even in the U.S., there have been reports of labs needing to “pay top researchers to make them stop complaining that their proposals weren’t adopted,” which symbolizes that real organizational friction exists behind Western labs.
This cultural difference produces a clear effect at the organizational level: lower self-awareness makes it easier for organizational structures to expand upward; researchers at different levels can collaborate more effectively instead of defending their own interests.
Students integrating into LLM development teams as one of the main forces
Another phenomenon that impressed Lambert is that core contributors in many labs still account for a very high proportion of current students. These students aren’t treated in a differentiated way; instead, they directly integrate into LLM development teams. This contrasts sharply with the U.S. ecosystem shaped by OpenAI and Anthropic, which almost don’t provide internship opportunities—or even when they do, interns are separated from core work:
Students’ advantage lies in “having no baggage.” They haven’t gone through the inertia assumptions left behind by the earlier AI waves, so they can absorb new technologies quickly—from MoE expansion to reinforcement learning to agent development. Each paradigm shift is, for them, a brand-new starting point; there’s no need to discard old cognition.
Competition or cooperation? Revealing China’s “engineer-led governance” ecosystem
Lambert noticed that when he tried to discuss with Chinese researchers long-term social risks of AI, economic impacts, or moral debates about model behavior, the conversations often fell into a kind of silence. He understood that it wasn’t that the other side was deliberately avoiding; it’s simply that these issues truly weren’t within their scope of thinking.
He cited a scholar, Dan Wang, to explain his stance on “China governed by engineers, the U.S. governed by lawyers”: “Their job is to make the models good; the other issues are left to others.”
To Lambert, this makes China’s AI community more like a “community” rather than competing tribes. Across labs there is generally mutual respect; they hold big-name players like ByteDance in high esteem, highly praise DeepSeek’s research taste and execution strength, yet they don’t have the kind of gunpowder-heavy competitive tension found in U.S. labs.
China AI limitations and disadvantages: chip, data, and a creativity gap
Nvidia chips are a common bottleneck for every lab
Due to U.S. export controls, insufficient compute supply from Nvidia is the common limitation faced by all of China’s labs. Lambert observed that nearly every lab explicitly said that if compute supply were sufficient, they would expand procurement without hesitation.
Domestic accelerators like Huawei’s are viewed positively on the inference side and many labs have already used them in large quantities; but on the training side, Nvidia remains an irreplaceable golden standard, and this gap is difficult to fully fill in the short term with other approaches.
The data industry is the biggest weakness, and building in-house becomes the mainstream choice
Compared with Anthropic and OpenAI investing hundreds of millions of dollars each year to buy reinforcement learning training environments, China’s external data industry still shows a clear gap in quality. Lambert observed that most labs believe purchased data available in the market tends to be low-quality, so they would rather allocate resources to build their own training environments, and researchers themselves spend a large amount of time constructing these environments.
While large enterprises like ByteDance and Alibaba have internal data teams that can provide support, as analyst Zephyr from Citrini put it, this is still the biggest weakness in China’s AI ecosystem.
(Chinese startup Moonshot’s robot claims to be Claude, and the distilled Anthropic model slips up)
Pragmatism behind open source, not ideology
When facing outside questions about why companies like Meituan and Xiaomi need to build and open general-purpose large models, Lambert believes there’s a very pragmatic business logic behind it: open source can get feedback from external communities and improve model quality; at the same time, companies can keep internally fine-tuned versions for their own products, retaining control of the core technology stack.
This “technical ownership” mindset drives almost all major Chinese tech companies to build foundation models in-house rather than relying on external services—very different from what U.S. AI companies choose.
Delphi Ventures founder’s perspective: strong execution, but lacking creativity
Delphi Ventures co-founder José Maria Macedo also recently conducted in-depth visits to China’s AI ecosystem and, from an investor perspective, provided another layer of observation that contrasts with Lambert’s technical lens.
Macedo believes Chinese founders generally have impeccable resumes and astonishing execution ability, but comparatively, the entrepreneurial drive to go “from zero to one” is rarer: “They’re better at taking existing directions and producing excellent upgraded versions rather than proposing entirely new problems the market hasn’t even realized yet.” He attributes this to the way the education system has long reinforced a mindset of training “problem solvers” rather than “question askers.”
(Top talent is everywhere, yet can’t make something like OpenAI? An investor’s two-week deep dive reveals China’s real AI problems)
The U.S. should still strive to lead the open ecosystem
Lambert admits that China is a place that can’t be understood by simply applying Western frameworks: “Its culture is too old and too deep, and the way it interweaves with the technology ecosystem produces a unique kind of chemistry.”
As an American, he hopes U.S. AI labs with open-model-first approaches can continue to stay ahead; but he’s even more concerned that if the U.S. limits the development of open models through administrative orders, it would weaken its own leading position in the global open AI ecosystem, tilting the balance of this competition toward an unpredictable direction.
This article In-depth field visit to China’s AI labs: researchers reveal that the “chip and data gaps” are the key difference between China and the U.S. first appeared on Lian News ABMedia.
