Study Finds Chinese AI Models Trained on Banned Nvidia Blackwell Chips

A forensic analysis of model weights has found that six open-source models released since April 2026 by DeepSeek, Moonshot AI, MiniMax, and Ant Group exhibit numerical signatures matching Nvidia Blackwell hardware, while earlier models correspond to the previous Hopper generation. The conclusion derives entirely from the models' own weights rather than insider leaks, and was validated on physical Blackwell chips for DeepSeek-V4-Flash and Kimi K3, providing reproducible technical evidence for export control enforcement.

The study, led by former OpenAI researcher Will DePue and reviewed by senior researchers at OpenAI and Anthropic, stresses that the evidence only proves access to Blackwell compute, possibly through overseas data center rentals, and does not directly equate to smuggling or illegality. However, the affair also brought attention to Nvidia CEO Jensen Huang's departure due to personal ties with a Megaspeed executive suspected of chip smuggling, spotlighting enforcement gaps and Nvidia's compliance responsibility.

The finding marks the first technical evidence that Chinese frontier labs can still use the latest Nvidia hardware despite export controls, challenging the effectiveness of U.S. policy. If compute access channels cannot be effectively blocked, the iteration speed of Chinese open-source models may continue to exceed expectations, impacting the global AI competition landscape and regulatory direction.