Researchers at Unemployment Labs allege that four Chinese AI companies, DeepSeek, Moonshot AI (Kimi), MiniMax and Ant Group, used Nvidia’s Blackwell chips to build six models released since April 2026. Blackwell chips need a US export licence to reach China, and the researchers say none has been reported granted.
None of this is a finding by a court or a regulator. It is the researchers’ own reading of the companies’ public model files, and the authors say it does not show where the chips were, how the labs got access or whether any export rule was broken. The labs’ side is not in the report.
The allegation is not new; what the report adds is a technical method. A senior US administration official made the same charge against DeepSeek on 23 February: its newest model, the official said, ran on Blackwells that were probably clustered in Inner Mongolia, and Washington does not ship them to China. The official did not say how the government got the information, and Nvidia declined to comment. On 22 July, Michael Kratsios, director of the White House Office of Science and Technology Policy, said in a post on X that Moonshot AI had bought servers built on GB300 Blackwell chips and used GB300s in Thailand, probably for training. The Unemployment Labs report notes those earlier claims came without technical evidence.
The team says it found the match by reading the weights. Before training starts, a model’s billions of weights are filled with starting values by a random-number routine, and the way those values are laid out depends on how many processing units the GPU has. The researchers recover that layout from the published weights and read off the chip: Hopper parts have 132 units, Blackwell’s B200 and B300 have 148, and the GB200 and GB300 have 152.
Two models land on 148 units: DeepSeek-V4-Flash and MiniMax-M3. Four land on 152: Kimi K3, Kimi K3.1-Pro, DeepSeek-V4-Pro and Ant’s Ling-3.1-flash. Each lab’s older releases look different. DeepSeek’s V3 (December 2024) and Moonshot’s Kimi K2.5 (January 2026) read as Hopper, as do MiniMax-M2.7 and Ling-3.0-flash.
Measured in sigma, the report’s strongest test for each of the six models scores between 49 and 1,385, which it says rules out coincidence. For DeepSeek-V4-Flash and Kimi K3 the team also reproduced the layout on physical Blackwell chips with unmodified PyTorch. The team says its method agreed with all 24 cases where a developer had publicly named the hardware, and that a test on DeepSeek-V4-Flash points to Nvidia rather than AMD, whose MI300X can be configured to mimic the layout. The authors call the V4-Flash evidence conclusive.
That is where the proof stops. What the authors establish is a shared starting pattern in the weights. A model can be set up on one machine and trained on another, and Blackwell systems can be rented from data centres outside China, so the pattern alone cannot say who held the chips or how they were obtained. The authors add that Hopper readings for earlier models do not by themselves show those releases complied with export controls.
Not every Chinese lab shows the pattern. Alibaba’s Qwen, Zhipu, Tencent, Xiaomi and StepFun still read as Hopper, and Meituan describes LongCat-2.0 as built on AI ASIC superpods, where the team found no Nvidia pattern. Its rough estimate is about 4,500 Blackwell GPUs for Kimi K3 and 3,000 for DeepSeek-V4-Pro, which it admits rests on assumptions about efficiency and run length. The report was published on 12 November and revised the same day with lower GPU estimates.
