Interconnects: China's AI labs say Nvidia access is the ceiling

Nathan Lambert's notes from a trip through Beijing's AI labs are useful mostly because they narrow the story. He visited Moonshot AI, Zhipu, Meituan, Xiaomi, 01.ai and Tsinghua, and the consistent thing every lab told him was that Nvidia compute access is the primary bottleneck. Not talent, not data, not organizational will. Chips.

The cultural read is the other half of the piece and it is more speculative. Lambert describes Chinese labs as more willing to do non-flashy work to improve the final model, less interested in philosophical debates about AI's societal impact (one framing he heard treats those debates as a category error), and structured to lean heavily on students as core contributors in a way US labs like OpenAI, Anthropic and Cursor do not, because those labs simply don't offer internships. He also reports more collegiality across Chinese labs than the US mutual-suspicion norm, with DeepSeek in particular credited internally for the best research taste in execution.

If you are a Western operator, the practical takeaways are narrow but sharp. Chinese developers reportedly use Claude widely despite restrictions, so demand for frontier Western models is real even where distribution is awkward. Meituan and Xiaomi are building their own models to keep stack control rather than renting one, which suggests the buy-versus-build calculus in Chinese enterprise looks different than the SaaS-lite version Western analysts tend to assume. And the compute story reinforces that US export controls are the one lever actually shaping the ceiling.

The honest caveat is that this is one Western researcher's short visit, and a commenter identifying as an employee at a visited lab, Mindful, pushed back that internal tribalism exists but was not visible to guests. Lambert himself hedges the government-role question, saying he has far too little to report the details as assertive. What the reporting does not give you is hard financials, timelines, or any real read on how these labs will price and monetize.

Worth reading as a corrective to both the China-is-years-behind and the China-is-about-to-overtake postures. The picture in the middle is a set of labs that are executing well, are supply-constrained on chips, and are watching what Anthropic and OpenAI ship as closely as anyone.