Will DePue, who leads the nonprofit research group Unemployment Labs and previously worked at OpenAI, published a forensic report on November 12 arguing that six recent models from four Chinese AI companies were initialized on Nvidia Blackwell GPUs. The companies are DeepSeek, Moonshot AI, MiniMax and Ant Group. Blackwell-class chips need a US export license for China, and the report says none has been reported. TechCrunch, The Register and The Decoder covered the findings on November 13.
The method rests on a quirk of how models start life. Before training, a model's billions of weights receive random starting values generated from a seed by PyTorch, which spreads that work across many GPU threads. The number of threads depends on how many streaming multiprocessors the chip has: 132 on Hopper parts such as the H100, H200 and H800, 148 on the B200 and B300, and 152 on the GB200 and GB300. The same seed therefore places values in different positions on different chips. Some weights, such as rows for tokens never seen in training, keep their starting values almost untouched, and trained weights retain a faint echo. The report searched roughly six trillion seed and layout combinations and then tested each match on weights held out of the search.
According to the report, DeepSeek-V4-Flash and MiniMax-M3 match the 148-unit layout, while DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash match the 152-unit layout. Earlier releases from the same labs, including DeepSeek-V3, Kimi K2.5, MiniMax-M2.7 and Ling-3.0-flash, read as Hopper. The newest models the author could read from Alibaba, Zhipu, Tencent, Xiaomi and StepFun also read as Hopper, and Meituan's LongCat-2.0 showed no Nvidia fingerprint, consistent with the company's statement that it trained on ASIC superpods. The report puts its strongest evidence for each Blackwell model between 49 and 1,385 standard deviations, against roughly 8 expected from chance across the whole search, and says the method agreed with all 24 models whose developers had publicly named their hardware.
The report also estimates scale, with heavy caveats. It puts the five training runs at about 12,000 Blackwell GPUs combined, with Kimi K3 near 4,500 and DeepSeek-V4-Pro near 3,000, and at roughly $70,000 per GPU it prices the largest clusters at $200 million to $300 million and the five together above $500 million. The author says he is not confident in these figures, which depend on assumed efficiency, token counts and run lengths that no lab has published, and the stated range for the total is 7,000 to 22,000 GPUs. A revision on November 12 replaced lower first-draft numbers of 2,000 to 3,000 GPUs and $150 million to $200 million for the largest cluster.
The limits are spelled out in the report itself. The fingerprint identifies the hardware that generated the starting weights, not every chip that later helped train a model, and it cannot show where the chips sit or who owns them. Blackwell systems can be rented from data centers outside China, so the evidence does not by itself establish smuggling or an export-control violation. An AMD MI300X restricted to 148 or 152 units can reproduce the same layouts, so the author checked for small numerical differences between vendors: for DeepSeek-V4-Flash the published weight matched Nvidia's result in 363 of 377 differing values and AMD's in none. He calls the AMD explanation very unlikely for the other five models but could not run the same test on them.
Independent checking is partial. The Institute for Progress posted a GitHub repository that replays part of the study. It found that the downloaded weights matched the hashes recorded by the author, reproduced the starting-weight patterns for DeepSeek-V3-Base on an H200 and for DeepSeek-V4-Flash and MiniMax-M3 on a B200, and detected the starting-value trace in the trained weights of DeepSeek-V4-Flash. It could not test Kimi K3, which needs a 152-unit GPU it lacked, and said most of the remaining models lacked public samples or settings for a GPU replay. DePue says senior researchers at OpenAI and Anthropic have confirmed the result; that statement comes from him, as relayed by The Register and The Decoder, and no such confirmation was seen directly.
TechCrunch set the report against the week's Nvidia news, including Jensen Huang's departure and scrutiny of Megaspeed, and recalled that a White House adviser said in July that Moonshot AI had used banned Nvidia chips. The Register noted four bills in Congress aimed at tightening limits on China's access to advanced AI hardware, one of which targets the rental loophole. None of the coverage reviewed included a response from the four labs, and the report asks that findings be attributed to DePue.