
Forensics
Report Finds Blackwell’s Fingerprint in Six Chinese Models
A former OpenAI engineer says six models from DeepSeek, Moonshot, MiniMax and Ant Group were started on Nvidia’s export-controlled Blackwell chips; the method can’t show where they were.
BEIJING, Nov. 15 — Six of China’s newest artificial-intelligence models were started on Nvidia’s Blackwell chips, which need a United States license for export to China, according to a forensic report published on Thursday by Will DePue, a former OpenAI engineer. The models come from DeepSeek, Moonshot AI, MiniMax and Ant Group.The SixDeepSeek-V4-Flash and V4-Pro, Moonshot’s Kimi K3 and K3.1-Pro, MiniMax-M3 and Ant Group’s Ling-3.1-flash, all released since April. The report says K3.1-Pro starts from the same seed as K3 and may be a later checkpoint of it.
The evidence comes from the models’ own published weights, not from leaks or insiders, the report says. It reads a numerical trace of the chip that set each model’s starting values, a trace that, in the report’s words, “supports, without proving, the inference that Blackwell also performed the training.”
The report lands in a week of questions in Washington about how Nvidia’s restricted chips reach China. Jensen Huang resigned as Nvidia’s chief executive on Nov. 11, and the chairmen of three House committees have pressed Nvidia on chip smuggling. Elsewhere in this edition, the Journal traces one route. Earlier claims that DeepSeek and Moonshot used Blackwell came from American officials and press reports, the report notes, “without publishing technical evidence.”
How it works
Before a model is trained, its billions of weights need starting values. In PyTorch, the software behind most large models, those values come from a pseudorandom generator fed a number called the seed; the same seed and settings always give the same numbers. On a graphics chip the work is split among threads, and the number of threads depends on the chip. In the configurations the report examined, a Hopper H100 SXM, H200 or H800 uses 270,336 for the job, a Blackwell B200 or B300 303,104, and a Grace Blackwell GB200 or GB300 311,296, by the report’s count. The same seed on a different chip puts the same numbers in different places.Streaming MultiprocessorsThe thread count follows the chip’s processor cores: 132 on the Hopper configurations examined, 148 on the B200 and B300, 152 on the GB200 and GB300, at 2,048 threads each.
Some weights, such as those for words a model never meets in training, keep their starting values. “I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models,” Mr. DePue said, as quoted by The Register. Across that many tries, by the report’s statistics, chance alone would reach about 8 sigma, a measure of distance from random variation; the six models’ strongest results run from 49 to 1,385.
The method gets known answers right, the report says. It agrees with all 24 cases in which a developer has named its hardware, which the report calls “a consistency check rather than a measured accuracy rate.” Nvidia’s own Nemotron-3 Ultra, whose technical report names the GB200 for one phase of training, reads as 152 cores, “the same as Kimi K3.” For DeepSeek-V4-Flash the report went further: on a B300, unmodified PyTorch and DeepSeek’s seed, 953049583, reproduced “461 of the 469 selected rows,” 445 of them bit for bit. An H100 did not.
Earlier models from the same four labs match Nvidia’s older Hopper generation, the report finds; all six Blackwell readings are in models released since April. DeepSeek’s models read as Hopper through V3, released in December 2024; Moonshot’s through Kimi K2.5, in January. The newest models it could read from Alibaba’s Qwen, Zhipu, Tencent, Xiaomi and StepFun “all read as Hopper-trained.”Hopper in ChinaThe A100 and H100 class was controlled from October 2022, but the A800 and H800, with the same core counts, could be exported until October 2023, the report says. The fingerprint cannot tell an H800 from an H100 SXM or H200.
What it can’t say
The report says its findings “do not establish where the chips were located, how access was obtained, or whether export controls were violated.” Blackwell systems can be rented from data centers outside China, it notes, and some labs may have used them through American cloud providers, The Register reported.
An AMD MI300X, also restricted from export to China, reproduces both layouts when limited to 152 or 148 of its units, the report says, a setup it calls “very unusual.” It ruled AMD out for DeepSeek-V4-Flash: of 377 values where the two makers’ arithmetic differs, 363 matched Nvidia’s and none AMD’s. It calls AMD “very unlikely” for the other five. Its conclusion is “strong evidence” for six models and “conclusive” for V4-Flash alone, though Mr. DePue, as The Register quoted him, called the evidence “conclusive” for all six.
The report puts the chips for five final training runs at about 12,000, costing more than half a billion dollars at roughly $70,000 apiece, and adds: “We are not confident in these estimates.”
Another researcher, Bojie Li, wrote on his blog that he rebuilt the method “from scratch, without his code,” and scanned 265 open models; he and the report find the same chip class for 65 of them, and he found nothing that contradicts it. His readings add two Chinese models in the B200 and B300 class: Alibaba’s Qwen3.8-Flash-Next, where the report found no signature, and LLaDA-MoE-v2, from Renmin University and Ant. He wrote no AMD code; for all but one of the Blackwell-class readings, he wrote, “an MI300X set up this way would look the same.”
The Register reported that Mr. DePue claims multiple senior researchers, “both independent and inside frontier AI labs like OpenAI and Anthropic,” have confirmed the result. None was named, and Daily Cortex could not verify it.
What comes next
As of 5:27 a.m. Eastern on Sunday, Daily Cortex had found no response from the five companies in the report, The Register’s story or Nvidia’s newsroom.
In Congress, bills to tighten the rules are already waiting. The Register named four: the AI Overwatch Act, the Chip Security Act, the MATCH Act and the Cloud Security Act, which The Register says aims to close the route through American cloud providers. CNBC reported that President Trump was due to meet Xi Jinping during the Nov. 18–19 APEC summit in Shenzhen. A U.S. official said on a press call on Friday, CNBC reported, “we do not negotiate American export controls.”
Sources
- Unemployment Labs, “Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs,” Nov. 12, 2026.
- Unemployment Labs (GitHub), “checkpoint-forensics: code and preserved rows to repeat the checks,” Nov. 13, 2026.
- The Register, “Chinese AI labs are training models on Nvidia’s Blackwell despite export controls, says ex-OpenAI engineer,” Nov. 13, 2026.
- Bojie Li (01.me), “Which GPU created each open model? Reading the weights of 265 LLMs,” Nov. 13, 2026.
- Office of Representative Josh Gottheimer, “Gottheimer, Moolenaar Introduce Bipartisan Bill to Close Loophole in Advanced AI Chip Export Controls,” June 26, 2026.
- Nvidia Newsroom, “Latest news (checked 5:27 a.m. Eastern, Nov. 15, 2026).”
- Nvidia, Form 8-K, “Current report, Item 5.02 (the board accepted Jen-Hsun Huang’s resignation, Nov. 11),” Nov. 12, 2026.
- House Committee on Foreign Affairs, “Chairmen Mast, Moolenaar, and Crawford Send Second Request to Jensen Huang and NVIDIA to Brief Congress on the Chinese Intelligence Asset,” Nov. 12, 2026.
- CNBC, “Trump doesn’t plan to raise Nvidia, spy issue during China trip: U.S. officials,” Nov. 13, 2026.