Notes from inside China's AI labs

Nathan Lambert wrote this on interconnects on 7 May 2026, on the train back from Hangzhou to Shanghai after a trip through China's AI labs. In Beijing alone he stopped at Alibaba's campus on the way from the airport and then, in 36 hours, visited Z.ai, Moonshot AI, Tsinghua University, Meituan, Xiaomi and 01.ai; his thanks also name Qwen and Ant Ling. Most conversations were off the record. His open-source-ai-reading-list describes it as a piece on how Chinese labs themselves talk about building models and how the Chinese industry differs from the American one.

His thesis is that the ingredients and the outputs look the same in both countries. Both have excellent scientists, large data and accelerated compute, and both ship large models for agentic work. The differences are in how the work is organized and in the people doing it, and those differences make Chinese labs, in his words, "the perfect fast-followers". He had suspected this before the trip and says he did not feel entitled to claim it until he had talked to the researchers.

Culture and who builds the models

Building a top LLM now means careful work across data, architecture and RL implementation, and fitting it together means shelving some brilliant individual contributions for the sake of the whole model. American researchers are rewarded for promoting their own work, and a "star AI scientist" career path exists. Lambert repeats the rumor that the Llama organization collapsed under these politics, and a story of labs paying off a top researcher to stop complaining that his idea did not make the final model. A small cultural shift away from that, he argues, shows up in the result.

Many core contributors at every Chinese lab are active students, treated as peers, much like Ai2 where Lambert works. OpenAI, Anthropic and Cursor offer no internships, and Gemini internships at Google risk being walled off from real work. He sums up the advantages in four points: more willingness to do unglamorous work that improves the final model; newcomers carry no habits from earlier AI hype cycles, so they adopt new techniques faster (one Chinese scientist named this as a strength); less ego lets org charts grow a bit larger before they break; and there are many people well suited to reproducing something already shown to work elsewhere.

The flip side is the stereotype that Chinese research produces fewer field-founding, 0-to-1 ideas. Leaders at the more academic labs talk about building that culture, while some technical leaders doubt it can happen soon, because it would take redesigning the education and incentive systems. Students described the same drift from academia to industry as in the US. One who had wanted to be a professor said education is solved by LLMs: "why would a student talk to me!"

Questions about AI's economic effects, long-term risk, or how models should behave often met polite confusion. Lambert calls it a category error to them. One researcher cited Dan Wang's line that China is run by engineers and the US by lawyers. China has no equivalent of Dwarkesh or Lex to turn scientists into public figures, and Lambert attributes the missing opinions to an education in which debating how society should be arranged is not encouraged.

Respect between labs

Beijing reminded him of the Bay Area: competitors a short Didi ride apart, researchers switching labs on "the best current vibes". The tone differed. Off the record, Chinese researchers expressed only respect for their peers, where in the US "sparks fly quickly". All of them fear ByteDance, whose Doubao makes it the only frontier closed lab in China, and all of them name DeepSeek as the lab with the best research taste. Researchers also tended to shrug at business questions as someone else's problem, where their American counterparts talk constantly about data vendors, compute deals and fundraising.

Six industry observations

On domestic demand, a common hypothesis says China's AI market will stay small because Chinese companies don't pay for software. Lambert thinks that holds for SaaS, which has always been tiny in China, but not for cloud, which is large. The labs themselves debate which one AI spending will resemble, and he leaned toward cloud.

Most developers he met were "Claude-pilled". Claude is nominally banned in China, yet every researcher mentioned building with it; some also use the Kimi or GLM CLIs, and Codex, popular in the Bay Area, barely came up. How Claude reaches Chinese companies at scale is the subject of anthropic-threat-report-september-2026.

Chinese companies want to own their technology. Lambert saw no master plan, just an equilibrium. ByteDance and Alibaba are expected to win most markets on resources, and DeepSeek sets the technical direction without being set up to win economically. That is why Meituan (LongCat) and Ant Group (Ling) train their own models. They see LLMs as central to future products, release the general model to get feedback from the open community, and keep fine-tuned versions in-house. Their openness is practical.

Government aid is real, but its size is unclear. Government in China is spread across many levels without a shared playbook; Beijing neighborhoods compete to host tech offices, and help probably includes cutting permit red tape. Whether it extends to recruiting talent or smuggling chips, Lambert heard too little to say. He saw no hint of top-level government influence over technical decisions.

The data industry is weak. Where Anthropic and OpenAI are reported to spend $10M or more on single RL environments and hundreds of millions a year in total, Chinese labs found local vendors poor and build environments and data in-house, with researchers writing environments themselves and ByteDance and Alibaba running internal labeling teams.

Every lab wants more Nvidia chips and is limited by not having them. Huawei and other accelerators were spoken of well for inference, and many labs have access to Huawei hardware.

The open question

Lambert warns that mapping Western lab behavior onto Chinese labs will often be a category error, and leaves open whether these ecosystems will produce different kinds of models or whether Chinese models will always look like the US frontier of three to nine months earlier. Even after asking directly why labs release their best models, he could not fully connect their ownership mentality with their genuine support for open development. The companies are not absolutists, but they are deliberate about supporting developers and learning from open release.

He ends with his own position: he wants American labs to lead every part of the stack, open models especially, and he wants open models to thrive globally. He worries more about fissures along national lines within the research community, and notes rumors of executive orders on open models as he finished writing, the thread picked up in six-months-to-live-for-open-models.

How it fits with the other pages

The trip report lines up with the Z.ai interview six months earlier. In the zai-playbook, Zixuan Li describes many active PhD students on GLM, a core team of 100 to 200, and a lab more focused on catching up than on AI risk debates, which matches what Lambert found. It also fits Kevin Xu's chinese-open-source-a-definitive-history on the state, which Xu says arrived late and followed the community. Xu's history explains the labs' openness as a two-decade culture and an overseas growth strategy, while Lambert reports a pragmatic openness he could not fully explain; the two are compatible but not the same account. On data, Lambert's finding that labs build everything in-house sits awkwardly with the vertical-data advantage claimed in chinas-structural-advantage-in-open-source-ai.

The limits are ones Lambert states himself: a single trip, off-the-record talks, and by his own account little knowledge of China beforehand. There are no figures on compute, revenue or headcount, and the cultural contrast rests on impressions from conversations.