Ex-OpenAI researcher finds Blackwell signatures in open-weight models

Former OpenAI researcher Will Depue describes a forensic method for identifying the hardware used to initialize AI models. Because GPU type and random-number seed produce distinctive starting weight patterns, he argues, traces of those patterns can remain in released model weights after training. Searching more than six trillion chip-and-seed combinations, he says he found Blackwell signatures in six Chinese models, including DeepSeek-V4-Pro and V4-Flash, Moonshot’s Kimi K3, and MiniMax-M3; four matched NVIDIA Grace Blackwell systems. He reports validating the method against 24 models whose developers had disclosed their hardware. Recreating the recovered starting weights on a B300 matched DeepSeek-V4-Flash bit for bit, while a GB200 reproduced Kimi K3 with 0.999995 correlation. The reported statistical matches range from 49 to 1,385 standard deviations above chance.

Depue estimates that Kimi K3 and DeepSeek-V4-Pro each used roughly 2,000-3,000 Blackwell GPUs, with clusters costing about $150-$200 million apiece; he says the identified labs’ total hardware spending would exceed half a billion dollars. He presents the findings as evidence that export controls on advanced NVIDIA chips are being circumvented, but the weight analysis cannot establish where the chips were located: they may have been smuggled into China or used by the labs abroad. He also notes that labs may be able to remove such signatures from future models.