A forensic analysis of public model weights identified six Chinese AI models initialized on Nvidia Blackwell chips subject to export restrictions. The models came from DeepSeek, Moonshot AI, MiniMax and Ant Group and were analyzed by former OpenAI researcher Will DePue through his lab, Unemployment Labs. The method relies on initial weight values and examined more than six trillion combinations of chip layouts and random seeds, enabling the exact reproduction of certain rows from DeepSeek and Kimi K3 on real Blackwell hardware. DePue estimates that the Kimi K3 and DeepSeek-V4-Pro training runs each used about 2,000 to 3,000 Blackwell GPUs, costing $150 million to $200 million per cluster and more than $500 million for the five final training runs combined. The report nevertheless states that its findings establish neither a legal violation nor how the chips were obtained, since they may have been rented from data centers outside China.
Source: Read the original article
