Will DePue says he has found a hidden fingerprint in published AI models that points to the class of NVIDIA chip used to create them. In a report from Unemployment Labs, the independent group he started after leaving OpenAI, he argues that six recent models from DeepSeek, Moonshot AI, MiniMax and Ant Group were first filled on NVIDIA’s Blackwell chips, which need a US licence for export to China. None of those licences have been reported.
I have a major result to share.
November 12, 2026
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
The finding is about how a model is started, not a photograph of a server room. An AI model’s “weights” are the billions of numbers that determine how it answers a question. Before training can begin, those numbers need starting values. In the software most labs use, called PyTorch, those values come from a formula that begins with a single starting number, the seed. Use the same seed and you get the same numbers every time, which is handy when you want to repeat an experiment.
To fill billions of numbers quickly, the work is split across many small workers on the GPU, the specialised chip that trains the model. How many workers there are depends on how many processing units the chip has. NVIDIA’s previous generation, Hopper, has 132 of those units. The newer Blackwell B200 and B300 have 148, and the Grace Blackwell GB200 and GB300 parts, used in NVIDIA’s flagship liquid-cooled racks, have 152. The same seed therefore lays the numbers out in a different pattern on each class of chip. Traces of that pattern, DePue says, can survive training and still be read in a published model.
The random numbers a model starts with aren’t truly random.
November 12, 2026
They come from a formula that starts from one starting number called the seed, and using the same seed gives the same numbers every time. Engineers set it up this way so they can repeat an experiment.
What he matched
He searched more than six trillion combinations of chip layout and seed, then compared the results with open-weights models the companies had released themselves. The report is not based on leaks or insider accounts. For DeepSeek-V4-Flash, he says a physical Blackwell B300 running ordinary PyTorch reproduced the model’s untouched starting numbers bit for bit, while a Hopper H100 reproduced none of them, and changing the seed by one broke the match. A rented GB200, he says, reproduced Kimi K3’s preserved rows to a correlation of 0.999995.
The six models, and what the report says they match, are set out below. Earlier models from the same labs still read as Hopper.
| Lab | Model | Released | Chip the model was started on |
|---|---|---|---|
| DeepSeek | DeepSeek-V3 | Dec 2024 | NVIDIA Hopper |
| DeepSeek-V4-Flash | Apr 2026 | NVIDIA Blackwell B200/B300 | |
| DeepSeek-V4-Pro | Apr 2026 | NVIDIA Blackwell GB200/GB300 | |
| Moonshot AI | Kimi K2.5 | Jan 2026 | NVIDIA Hopper |
| Kimi K3 | Jul 2026 | NVIDIA Blackwell GB200/GB300 | |
| Kimi K3.1-Pro | Nov 2026 | NVIDIA Blackwell GB200/GB300 | |
| MiniMax | MiniMax-M2.7 | Apr 2026 | NVIDIA Hopper |
| MiniMax-M3 | Jun 2026 | NVIDIA Blackwell B200/B300 | |
| Ant Group | Ling-3.0-flash | Aug 2026 | NVIDIA Hopper |
| Ling-3.1-flash | Oct 2026 | NVIDIA Blackwell GB200/GB300 |
DePue estimates that the largest runs, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. At a cost he derives from a Wall Street Journal report on Blackwell chips being installed in Malaysia, those clusters would come to about US$150 million to US$200 million each, and more than half a billion dollars across the labs, not counting experimental compute. Those figures are his estimates, not a public inventory.
What the fingerprint does not show
What the method cannot tell you is as important as what it can. It identifies the hardware that generated the starting numbers. It supports, without proving, the idea that the same chips also did the training. A lab could, in principle, start a model on one system and finish it on another, or rent Blackwell machines in a data centre outside China. The weights cannot show who owned the chips, where they sat, or whether an export-control rule was broken. DePue says an unusually configured AMD MI300X could theoretically produce a similar layout; he has ruled that out for DeepSeek-V4-Flash and calls it very unlikely for the other five models.
Of course, this tells us nothing about where these chips physically are. They could have been smuggled into China or the labs are using them abroad through gaps that Congress has not closed. In either case, it’s clear that the export bans are deeply failing.
November 12, 2026
US officials and press reports had already alleged that DeepSeek and Moonshot were using Blackwell chips, without publishing this kind of technical evidence. The report says MiniMax and Ant Group had not been publicly tied to Blackwell before. Ant Group’s public code includes benchmark results generated on a GB300, dated 31 March 2026.
Not every Chinese lab shows the same switch. The newest models DePue could read from Alibaba, Zhipu, Tencent, Xiaomi and StepFun still read as Hopper. Meituan’s LongCat-2.0 left no NVIDIA fingerprint; the company says it was trained on “AI ASIC superpods.”
He says the method agreed with all 24 cases where a developer had already named the hardware behind a specific model, including NVIDIA’s own Nemotron-3 Ultra, which reads as the same 152-unit layout as Kimi K3. The statistical matches, he writes, range from 49 to 1,385 sigma, far above the roughly 8 sigma he would expect from the best accidental hit in a search this large. A sigma, here, is a measure of how far a result sits above ordinary chance.
DePue says he is highly confident, and that senior researchers, both independent and at labs including OpenAI and Anthropic, have confirmed the result. Those researchers have not been named, and neither company has issued a public statement. NVIDIA and the four Chinese labs had also not commented publicly at the time of writing. Dylan Patel, who runs the research firm SemiAnalysis, replied that the conclusion was already visible in code the labs had published. DePue said some of what is obvious to specialists is not common knowledge.
He plans to publish a GitHub repository and a full paper with the implementation details, and notes that once the method is public, labs can erase many of these signatures. Reproduction scripts, the report says, are available to reviewers on request.
MTS, an AI news account, also posted a short summary of the report. It called DePue’s work strong forensic evidence that the four labs had trained their newest models on Blackwell GPUs. That’s a bit stronger than the report itself. DePue says the fingerprint is of how the models were started, and that this supports, without proving, the use of the same chips for training.
SITUATION DETECTED: Former OpenAI researcher Will DePue says he has found strong forensic evidence that Chinese AI labs DeepSeek, Moonshot AI, MiniMax, and Ant Group all trained their newest frontier models on NVIDIA Blackwell GPUs, despite U.S. export controls.
November 12, 2026
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
