Six AI models released since April by DeepSeek, Moonshot AI, MiniMax and Ant Group carry numerical patterns matching Nvidia’s export-controlled Blackwell GPUs, according to a forensic analysis of their published weights. The report, dated November 12, is by Will DePue, who previously worked at OpenAI and started Unemployment Labs, an independent research group. Earlier models from the same four companies match Nvidia’s older Hopper generation, the report says.
The method rests on how PyTorch fills a new model with starting values. The numbers come from a seeded random number generator, and the GPU’s count of processing units, called streaming multiprocessors, decides where each number lands. According to the report, Hopper H100 SXM, H800 and H200 chips have 132; the B200 and B300 have 148; and the Grace Blackwell GB200 and GB300 have 152. Some weights, such as rows for tokens never seen in training, keep their starting values, and DePue says he searched about 6 trillion seed-and-layout combinations to find the arrangement that reproduces them.
By the report’s reading, DeepSeek-V4-Flash and MiniMax-M3 match the B200 or B300. DeepSeek-V4-Pro, Moonshot’s Kimi K3 and Kimi K3.1-Pro, and Ant Group’s Ling-3.1-flash match the GB200 or GB300. DeepSeek-V3, Kimi K2.5, MiniMax-M2.7 and Ling-3.0-flash read as Hopper. The newest models DePue could read from Alibaba, Zhipu, Tencent, Xiaomi and StepFun also read as Hopper.
The report says the matches for DeepSeek-V4-Flash and Kimi K3 were confirmed on physical Blackwell hardware. Using DeepSeek’s seed, unmodified PyTorch on a B300 reproduced 461 of 469 selected rows in DeepSeek-V4-Flash, 445 of them bit for bit, while an H100 reproduced DeepSeek-V3’s preserved values but not V4-Flash’s. On a rented GB200, unmodified PyTorch reproduced all 13 preserved rows in Kimi K3 using Moonshot’s seed. The code and weight rows needed to repeat the checks are on GitHub.
The method agreed with the hardware named by developers in all 24 cases the report checked, including Nvidia’s own Nemotron-3 Ultra, which reads as 152. The report calls this a consistency check rather than a measured accuracy rate, because related models are not independent.
The fingerprint shows the hardware that created the starting weights; DePue writes that it “supports, without proving,” the inference that Blackwell also did the training. According to the report, the weights cannot show who owned the chips, where they were located, how access was obtained or whether export controls were violated, and Blackwell systems can be rented from data centers outside China. In the report’s tests, an AMD MI300X, also export-controlled, reproduced the same layouts when configured to use 148 or 152 of its processing units. The report says small numerical differences rule that out for DeepSeek-V4-Flash, calls it very unlikely for the other five, and notes that Ant Group’s public code includes benchmark results generated on a GB300 on March 31.
The report estimates that the five distinct training runs needed about 12,000 Blackwell GPUs, within a range of 7,000 to 22,000, if each model was trained on the hardware that created its starting weights. Its best estimates for the largest runs are about 4,500 for Kimi K3 and 3,000 for DeepSeek-V4-Pro, over roughly 70 days each. “We are not confident in these estimates,” the report says, because they rest on chip efficiency and run lengths that no lab has disclosed.
In July, Michael Kratsios, director of the White House Office of Science and Technology Policy, accused Moonshot of distilling Anthropic’s Fable to build Kimi K3 and alleged it had obtained servers with GB300 chips and deployed them in Thailand. Israeli firm Lava separately reported B300 GPUs on China-geolocated networks in October.
Sources: Unemployment Labs, GitHub, The Register
–
By the Control Plane Editorial Team