Research claims China trained top AI models on banned Nvidia chips

In short: Researchers say some of China’s best-known AI models show evidence they were trained using Nvidia chips that the US restricts from being shipped to China.

What's going on

A new report from Will DePue, a former OpenAI researcher who now leads the nonprofit Unemployment Labs, looked at several openly released AI models from Chinese companies including DeepSeek, Moonshot AI, MiniMax, and Ant Group.

DePue analyzed the models’ “weights,” which are the internal settings a model learns during training (like the knobs and dials that shape how it answers). He says those weights can still carry “fingerprints” of the computer chips used to train the model. Based on that, he reported evidence that some models, including Moonshot’s Kimi K3 releases and DeepSeek’s V4 models, were trained on Nvidia Blackwell GPUs.

Blackwell is Nvidia’s newer, more powerful line of GPUs, or graphics processing units (specialized chips that act like high-speed engines for AI training). The US government restricts exporting Blackwell chips to China and to Chinese-owned firms.

This research adds to earlier public claims from US officials. The White House technology adviser Michael Kratsios said earlier this year that Moonshot used export-banned Nvidia chips, citing US intelligence sources.

What to watch

It is still not clear whether laws were broken. The chips could have been used outside mainland China, such as in offshore data centers, which can work like renting a powerful computer room in another country. The story also lands amid wider scrutiny of Nvidia’s China exposure and alleged chip smuggling routes through places like Malaysia.

Source: TechCrunch AI