Top Chinese AI labs built their newest models on Nvidia's Blackwell chips

A researcher claims that six Chinese AI models, including DeepSeek-V4-Pro and Flash, Moonshot’s Kimi K3, and MiniMax-M3, were initialized on Nvidia Blackwell hardware, despite export restrictions barring these chips from China. The method searches released model weights for traces of their original random initialization: because the starting values depend on both GPU type and random seed, the researcher says those traces can identify the hardware. After testing more than six trillion chip-and-seed combinations, he reports strong matches for Blackwell, including four models linked to Grace Blackwell systems. He says tests on physical B300 and GB200 machines reproduced starting weights for DeepSeek-V4-Flash and Kimi K3, while an H100 did not. The technique also reportedly matched 24 models to hardware their developers had identified.

The researcher argues that the matches are far too strong to be chance and says senior researchers at OpenAI and Anthropic independently verified the findings. Based on model details and release timelines, he estimates that Kimi K3 and DeepSeek-V4-Pro each used 2,000-3,000 Blackwell GPUs, with clusters costing roughly $150-$200 million apiece; the labs’ combined restricted-chip spending may exceed $500 million. The evidence identifies the hardware used, not its location: the chips could have been smuggled into China or accessed abroad. The researcher concludes that export controls are failing and warns that labs may now be able to remove such hardware signatures from future models.