Six AI models released by DeepSeek, Moonshot AI, MiniMax and Ant Group since April carry numerical patterns that match Nvidia’s Blackwell GPUs, according to an analysis by Will DePue published November 12 on Unemployment Labs. The same four labs’ earlier releases line up with Hopper, Nvidia’s previous generation. Blackwell chips need a U.S. export license for China, and the analysis says none has been reported.

Everything here is one analyst’s inference from public files, and DePue draws a hard line around it. His report says the weights cannot show the chips’ location, the route by which the labs got access, or any breach of export controls. It adds that Blackwell capacity can be rented in data centers outside China, so the evidence concerns access to hardware and, by itself, does not show smuggling.

The six models are DeepSeek-V4-Flash and DeepSeek-V4-Pro, Kimi K3 and Kimi K3.1-Pro from Moonshot, MiniMax-M3, and Ant Group’s Ling-3.1-flash. DeepSeek-V4-Flash and MiniMax-M3 match the 148-core B200 or B300 layout. The other four match the 152-core GB200 or GB300 layout.

The method rests on how PyTorch fills a new model with random starting values. A generator called Philox splits that job across GPU threads, and the thread count depends on how many processing units the chip has. Hopper parts such as the H100, H200 and H800 have 132, so the placement of values differs from Blackwell’s 148 or 152. Some weights that training never touched keep their starting values, and trained weights keep a weaker trace.

DePue tested the claim on physical hardware. An H100 reproduced all 57,344 preserved values in DeepSeek-V3, and a B300 reproduced 461 of 469 selected rows in DeepSeek-V4-Flash. A rented GB200 reproduced all 13 preserved rows of Kimi K3. He also reports that his reading agrees with all 24 cases where a developer has named the hardware behind a model.

On statistics, DePue searched roughly 6 trillion seed and layout combinations, where a chance match should top out near 8 sigma. His strongest result for each of the six models runs from 49 to 1,385 sigma, and his weakest single test hits 20.

The timeline is the sharpest part. DeepSeek reads as Hopper through V3 (December 2024) and Blackwell from V4 in April 2026. Moonshot flips between Kimi K2.5 in January and K3 in July. MiniMax flips between M2.7 in April and M3 in June, and Ant Group between Ling-3.0-flash in August and Ling-3.1-flash in October. Not every lab moved: the newest models DePue could read from Alibaba, Zhipu, Tencent, Xiaomi and StepFun all read as Hopper.

DePue also sizes the runs, and he says he is not confident in the numbers. For Kimi K3 and DeepSeek-V4-Pro he estimates about 4,500 and 3,000 Blackwell GPUs over two to three months, and about 12,000 across five runs. At roughly $70,000 per GPU, a price derived from a Wall Street Journal report on chips destined for ByteDance in Malaysia, the largest clusters would cost $200 million to $300 million. He revised those figures upward on November 12 after the first version assumed better GPU efficiency than disclosed Blackwell runs achieved.

The caveats cut both ways. The fingerprint points to the chip that generated the initial values, not to every chip that later trained the model, since a lab could create weights on one machine and shift them elsewhere. AMD’s MI300X, also export-restricted, can mimic the layouts if configured oddly. DePue excluded it for DeepSeek-V4-Flash, where 363 of 377 telling values matched Nvidia’s arithmetic and none matched AMD’s. For the other five he calls it very unlikely without proving it.

Earlier allegations that DeepSeek and Moonshot used Blackwell came from American officials and the press, including Reuters in February and White House OSTP director Michael Kratsios in July. They offered no technical evidence. DePue’s report supplies some, though it settles a question about hardware, not about law.

Anyone weighing this as a policy story should note the gap it leaves open: the owner and location of the chips. The public code reruns the core check on a laptop in seconds, so independent replication will likely arrive well before any answer to that.

Will DePue, Unemployment Labs, November 12, 2026.