Last week I wrote about a PLA-linked university buying two Nvidia H200s through a Chinese reseller, and I asked what BIS would actually…
Last week I wrote about a PLA-linked university buying two Nvidia H200s through a Chinese reseller, and I asked what BIS would actually do about that enforcement gap. A new forensic report suggests the gap is far larger than a couple of chips.
Will DePue, who helped build Sora, GPT-4o, o3 and GPT-5 at OpenAI and now runs the nonprofit Unemployment Labs, says he found a Blackwell hardware fingerprint in the published weights of six Chinese models released since April: DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash. They come from DeepSeek, Moonshot AI, MiniMax and Ant Group.
He says earlier models from the same four labs carry Hopper fingerprints instead.
The method is clever. A chip splits weight initialization across its processor cores, which number 132 on H100 and 148 or 152 on Blackwell parts, and that leaves a statistical trace in the checkpoint. The trace points directly to the hardware that initialized a model, which supports but does not prove that the same hardware trained it.
DePue reproduced exact matches on real Blackwell hardware, published his code on GitHub, and correctly identified all 24 control models. He says senior researchers at OpenAI and Anthropic confirmed it independently.
His revised estimate is roughly 12,000 Blackwell GPUs across five runs, within a wide range of 7,000 to 22,000, or over half a billion dollars of hardware at about $70,000 per GPU, with no export licenses reported. He is clear that he can't prove where the chips were, who owned them, or that any rules were broken.
Still, suspicion has become something anyone can check, and I'd like to know whether enforcement moves up to the scale this evidence points to.
$NVDA
Sources:
Unemployment Labs: https://unemploymentlabs.com/checkpoint-forensics
The Decoder: https://the-decoder.com/former-openai-researcher-says-he-found-nvidia-blackwell-fingerprints-hidden-inside-six-chinese-ai-models/
Not financial advice. For informational and educational purposes only.
