AI
2026-11-14
Will DePue used to work at OpenAI. He just published a forensic study of six Chinese AI models' files. He found numerical fingerprints matching Nvidia's newest chip, the Blackwell — banned for export to China. The models come from DeepSeek, Moonshot AI, MiniMax, and Ant Group.
DePue says older models from the same labs, like DeepSeek-V3 and Kimi K2.5, carry fingerprints from Nvidia's older Hopper chips. Something changed around April 2026, when the newer models started training on Blackwell hardware. Still, DePue is careful about what his method actually proves. It can show a chip's digital signature. It can't show where that chip physically sat when training happened, or who shipped it there. Nvidia and the four AI labs haven't responded publicly to the findings.
Will this change how the US enforces chip export bans? Right now, enforcement relies a lot on shipping records and company disclosures. Those are paper trails that can be faked or incomplete. A method that reads chip fingerprints directly out of a model's files is harder to fake. If forensic analysis like DePue's holds up, it could become a new way to check compliance. That means catching violations before they happen, not just investigating them after. That would matter far beyond these four companies.
This story is written by AI from the sources above, checked against them before publishing. If something here still reads wrong, tell us and we'll correct it.