A forensic report published on 12 November claims six of China’s newest AI models carry the numerical fingerprints of Nvidia Blackwell GPUs — chips the United States bars from export to China without a licence. The report, from former OpenAI researcher Will DePue at the nonprofit Unemployment Labs, drew wide attention on X, and TechCrunch recapped it on 13 November.

The method reads the models’ own published weights. The random-number seed and thread layout used when a model’s initial values were created leave a pattern tied to the GPU’s streaming-multiprocessor count — 148 for Blackwell B200/B300 chips, 152 for GB200/GB300. The six models named are DeepSeek-V4-Flash and V4-Pro, Moonshot’s Kimi K3 and K3.1-Pro, MiniMax-M3 and Ant Group’s Ling-3.1-flash, all released since April 2026. For two of them the match was confirmed on physical Blackwell hardware: a rented GB200 reproduced Kimi K3’s weights to a correlation of 0.999995.

The report is careful about what that proves. The fingerprint identifies the hardware that created a model’s starting weights, and supports — without proving — the inference that Blackwell also did the training. It says nothing about where the chips were, how access was obtained, or whether any export control was violated: Blackwell systems can be rented from data centres outside China. The only alternative hardware the report identifies, specially configured AMD chips, is also export-restricted, and was ruled out for DeepSeek-V4-Flash and called very unlikely for the other five.

The report also flags its own limits: it is the work of a single researcher, not a peer-reviewed paper, and no independent reproduction had been published at the time of writing. Notably, DeepSeek’s latest model, V4.1-Flash, showed no fingerprint at all — a null result, not proof of which hardware trained it. And the scale estimate of about 12,000 Blackwell GPUs across five runs comes with the author’s own warning that he is “not confident in these estimates”.

Sources