Researcher finds Nvidia Blackwell fingerprints in six Chinese open models — DeepSeek, Kimi, MiniMax and Ant Group — but no violation is established
Will DePue, a former OpenAI researcher who leads the nonprofit Unemployment Labs, published a forensic report on November 12 arguing that the published weights of six Chinese open models carry the fingerprint of Nvidia Blackwell GPUs — chips that have required a US export licence for China since before they shipped. The method uses how PyTorch scatters random starting values across a chip's compute cores: Hopper chips have 132, Blackwell B200/B300 chips 148, and Grace Blackwell GB200/GB300 systems 152, and some starting values survive training untouched.
The report reads DeepSeek-V4-Flash and MiniMax-M3 as B200/B300-class, and DeepSeek-V4-Pro, Moonshot's Kimi K3 and K3.1-Pro, and Ant Group's Ling-3.1-flash as GB200/GB300-class, with reported statistical strength from 49 to 1,385 sigma. Earlier models from the same four labs read as Hopper, and DePue says the method correctly identified all 24 control models whose training hardware is public. He estimates roughly 12,000 Blackwell GPUs across five inferred runs, while saying he is not confident in those figures.
The bounds matter as much as the finding. The fingerprint identifies the hardware that generated a model's starting weights; training on the same hardware is an inference, not a direct observation. It also says nothing about where the chips were or who owned them — Chinese labs can rent compute in data centres abroad, and one cited estimate has large numbers of accelerators reaching the region through third countries. This is evidence of access to export-controlled hardware, not by itself proof of smuggling or an export-law violation, and none of the four labs has commented in the coverage so far.
Update, November 15: the case got a cooler second reading. DePue published a more detailed forensic write-up of the method, and the Hacker News response was notably more sceptical than the first coverage: commenters pointed out that renting GPU time on Blackwell machines physically located in Singapore or Malaysia is entirely legal under US export controls, so a fingerprint in the weights does not by itself imply a violation. The follow-up also notes that DeepSeek's newer V4.1-Flash showed no Blackwell signature at all. The original bounds stand — the fingerprint identifies the hardware behind a model's starting weights, not where the chips were or who owned them.
Sources
- AgentsPulse — DeepSeek Nvidia Blackwell training: what the weights show (Nov 2026)
- Digging Beagle — Six Chinese model checkpoints show Blackwell-compatible GPU initialization fingerprints (Nov 2026)
- MadRobot — Chinese AI models carry banned Nvidia Blackwell fingerprints (Nov 13)
- Anagnorisis Daily — ChatGPT Wallet pays for you; JPMorgan blocks the agents (Nov 14)
- AI0 News — November 15 digest: Blackwell fingerprint case weakens