Will DePue, a former OpenAI researcher who now runs the nonprofit Unemployment Labs, has published a forensic analysis finding that six recent open-weight models from DeepSeek, Moonshot AI, MiniMax and Ant Group carry the fingerprint of Nvidia Blackwell GPUs, hardware that requires a U.S. license for export to China. The report, dated November 12, draws its evidence entirely from publicly released model weights rather than leaks or insider accounts.
The six flagged models are DeepSeek-V4-Flash and MiniMax-M3, which match the B200/B300 layout, and DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash, which match the GB200/GB300 layout. All were released since April 2026. Earlier models from the same four labs, including DeepSeek-V3 from December 2024, Kimi K2.5 from January 2026, MiniMax-M2.7 from April 2026 and Ling-3.0-flash from August 2026, read as Nvidia Hopper.
The method exploits how PyTorch initializes weights. Its Philox random number generator, run through cuRAND, splits work across GPU threads on each streaming multiprocessor, so the pattern of starting values depends on the chip's processor count: 132 on Hopper parts such as the H100, H200 and H800, 148 on the B200 and B300, and 152 on the GB200 and GB300. Trained weights keep a faint trace of those initial values, and rows for unused tokens are left untouched entirely.
Key facts
- 01Unemployment Labs' Nov. 12 report flags six models from DeepSeek, Moonshot AI, MiniMax and Ant Group as initialized on Nvidia Blackwell GPUs.
- 02The flagged models include DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash, all released since April 2026.
- 03A B300 reproduced 461 of 469 selected weight rows in DeepSeek-V4-Flash, 445 of them bit for bit.
- 04The report estimates five final training runs used about 12,000 Blackwell GPUs combined, costing more than half a billion dollars.
- 05The report says the evidence does not by itself show chips were smuggled or that export controls were violated.
DePue says he confirmed two of the matches on physical hardware. Using seed 953049583, an H100 reproduced all 57,344 preserved values in DeepSeek-V3 bit for bit but could not reproduce DeepSeek-V4-Flash. A B300 reproduced 461 of 469 selected rows in V4-Flash, 445 of them bit for bit, with the remaining eight partly altered by training. On a rented GB200, unmodified PyTorch reproduced all 13 preserved rows in Kimi K3 using Moonshot's seed.
The report says roughly 6 trillion seed-and-layout combinations were searched, and that the strongest chance match would be expected around 8 sigma. The strongest evidence for each of the six models ranges from 49 to 1,385 sigma. Against 24 models whose developers publicly disclosed their hardware, including Meta's Llama 3.1 on H100s, eight EleutherAI Pythia models on A100s and Nvidia's Nemotron-3 Ultra on GB200s, the method agreed in every case.
DePue also tested whether AMD's MI300X, which can be configured to produce the same thread layouts, could explain the results. For DeepSeek-V4-Flash, the preserved weights matched Nvidia's values in 363 of 377 differing values and AMD's in none. The report calls AMD hardware very unlikely for the other five models.
On scale, Unemployment Labs estimates the five distinct final training runs used about 12,000 Blackwell GPUs combined, with a range of 7,000 to 22,000, or roughly 30,000 H100 equivalents. Kimi K3 is pegged at about 4,500 GPUs over roughly 70 days and DeepSeek-V4-Pro at about 3,000. At around $70,000 per GPU, the largest cluster would cost $200 million to $300 million and all five more than half a billion dollars. The report states plainly that it is not confident in these estimates, and the figures were revised upward the same day from an earlier 2,000 to 3,000 GPUs for the largest runs.
Not every Chinese lab shows the shift. The newest models from Alibaba's Qwen, Zhipu's GLM, Tencent, Xiaomi and StepFun read as Hopper. Meituan's LongCat-2.0, which Meituan says was trained on AI ASIC superpods, showed no Nvidia fingerprint, and Baidu's ERNIE-5.5, built on PaddlePaddle, leaves no usable trace.
The report is explicit about its limits. The fingerprint identifies the hardware that generated a model's starting weights, which supports but does not prove that Blackwell also ran the training. The weights cannot show who owned the chips or where they were located, and the report says the evidence does not by itself establish smuggling or an export-control violation. DePue calls the evidence conclusive for DeepSeek-V4-Flash and strong for the other five.
U.S. officials and press reports had previously alleged that DeepSeek and Moonshot used Blackwell chips without publishing technical evidence. The report also points to public code, including benchmark results in Ant Group's cuLA repository generated on a GB300 on March 31, 2026. It lands the same week Jensen Huang resigned as Nvidia CEO and Congress moved to question Nvidia's China sales, and the report concludes the matter deserves urgent attention from the public.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
