Report says Chinese AI models used banned Nvidia chips

On November 12, former OpenAI researcher Will DePue published a forensic analysis saying six open-weight models from four leading Chinese AI labs have numerical fingerprints that match Nvidia ↘ 4.34% Blackwell GPUs. U.S. rules bar those chips from export to China. The Unemployment Labs report names DeepSeek-V4-Flash and DeepSeek-V4-Pro, Moonshot AI's Kimi K3 and Kimi K3.1-Pro, MiniMax-M3, and Ant Group's Ling-3.1-flash. It also says earlier models from the same companies match Nvidia's older Hopper generation.unemploymentlabs.com

The authors are careful about what the evidence shows. They write that the fingerprint "identifies how the weights were created and supports, without proving, the inference that Blackwell also performed the training". On X, DePue said senior researchers at OpenAI and Anthropic had checked the result independently, but the report does not name any outside reviewers. The work has not been peer-reviewed.cryptobriefing.com+2

How the fingerprint works

The method rests on a model's starting weights. DePue says the way those initial values are laid out depends on the exact GPU configuration and the random seed, and that traces of this pattern survive in the published model. He searched about six trillion combinations of seed and chip layout. The report says the strongest match chance alone would produce is about 8 sigma, while the six models score between 49 and 1,385 sigma.unemploymentlabs.com

The team also tested its matches on physical chips. On a B300, unmodified PyTorch using DeepSeek's recovered seed reproduced 445 rows of V4-Flash bit for bit. A rented GB200 reproduced all 13 preserved rows of Kimi K3. Changing either seed by one produced no matches. The authors say a specially configured AMD Advanced Micro Devices, Inc. ↗ 5.86% MI300X could in theory produce similar layouts, but they call that explanation very unlikely. They also say their method agreed with all 24 cases in which a developer had publicly disclosed the hardware it used.unite.ai+1

Limits and context

According to the report, the newest readable models from Alibaba ↘ 7.05%, Zhipu AI, Tencent, Xiaomi and StepFun all still read as Hopper. The report estimates that each of the largest final training runs used thousands of Blackwell GPUs, and that the clusters together cost more than $500 million. It describes these figures as rough estimates.unemploymentlabs.com

The evidence does not show where the chips were, who owned them, or how the labs got access. Blackwell systems can be rented outside China, and the report says its results "are not a finding that any law was broken".cryptobriefing.com

The report adds to existing scrutiny of how Chinese labs get Nvidia chips. In July, White House science adviser Michael Kratsios said Moonshot had "accessed GB300s in Thailand." Nvidia and the four labs had not commented at the time of publication.cryptobriefing.com