A researcher reads traces of Nvidia chips China may not buy in the weights of Chinese AI models
According to a former OpenAI employee, DeepSeek-V4-Pro and Kimi K3 among others were initialised on Blackwell systems, though that does not prove where they were located.
Chinese AI labs have probably trained their most advanced open models on Nvidia Blackwell chips, which the United States does not allow to be exported to China. That is the claim of Will DePue, a former OpenAI researcher who now runs the research group Unemployment Labs, The Register and TechCrunch report. According to DePue, a model's weights carry a trace of the GPU on which it was initialised: the starting values are laid out by a pseudorandom algorithm in a pattern that depends on the chip, and that pattern survives training. He says he searched more than six trillion combinations of chip layout and seed value.
In this way he found traces of Blackwell in six models, including DeepSeek-V4-Pro, Moonshot AI's Kimi K3 and MiniMax-M3; Ant Group is also said to use such chips. Four models match the Grace Blackwell systems GB200 and GB300. DePue acknowledges limitations: the method does not show where the chips were located. TechCrunch notes that training in a data centre outside China, for instance in Malaysia, does not necessarily breach export rules, even though those rules are meant to prevent exactly that kind of access.
The report comes at a sensitive moment. Nvidia chief Jensen Huang stepped down this week after an affair with Alice Huang, who is linked to the Singapore company Megaspeed, which is under investigation for smuggling chips to China. According to The Register, at least four bills to tighten export controls are awaiting consideration in the US Congress.