Report links Chinese AI model training to restricted Nvidia Blackwell chips

Research examining openly released Chinese AI models suggests that restricted Nvidia Blackwell processors were used in their training, according to a November 13 report from TechCrunch. The findings raise questions about access to advanced computing hardware, but do not establish that export laws were broken.

What model weights may reveal

TechCrunch reports that Will DePue, a former OpenAI researcher who leads the nonprofit Unemployment Labs, analyzed model weights released by DeepSeek, Moonshot AI, MiniMax and Ant Group. Weights are the learned parameters that underpin a model’s behavior.

According to the report, DePue identified evidence of training configurations associated with Blackwell hardware. The models cited include Moonshot’s Kimi K3 releases and DeepSeek’s V4 models. These are reported research findings, rather than confirmation of where particular chips were installed or how they were obtained.

Access does not establish a violation

The report explicitly cautions that the findings do not mean the models were trained entirely on Nvidia hardware. It also leaves open whether the processors were accessed through offshore data centers, rather than shipped into China.

TechCrunch cites estimates from research nonprofit Epoch AI that approximately two million advanced chips entered China during 2024 and 2025 through arrangements concealing their destination. Separately, Epoch estimates that Malaysia imported 275,000 Nvidia AI accelerators over those two years. Those figures are estimates, not a count of chips tied to the models DePue examined.

For the wider PC and technology audience, the central issue is Nvidia’s role in AI infrastructure. Evidence about training hardware can inform the debate over export restrictions, while questions about location, ownership and legal compliance remain separate.