Former OpenAI Researcher Finds Traces of Banned Blackwell Chips in Six Chinese Models
Former OpenAI researcher Will Depue says he found traces of Nvidia Blackwell chips, which the US does not allow to be sold to China, in six Chinese models. Before training, PyTorch fills the weights with random numbers, and the pattern depends on the GPU's core count: older Hopper chips have 132, while Blackwell has 148 or 152. This proves access to the hardware, but not smuggling, since compute can be rented abroad.
AI-processed from The Decoder; edited by Hamidun News
The US has banned the sale of Nvidia's most powerful accelerators to China: chips from the Blackwell generation. Officials and journalists have long suspected that Chinese labs got their hands on them anyway, but no public evidence existed. Now such evidence may have surfaced. Former OpenAI researcher Will Depue claims he found a Blackwell "fingerprint" in six Chinese models released since April 2026. The fingerprint is a pattern of random numbers inside the model's weights. It depends on the number of compute cores in the GPU: the previous Hopper generation (H100 and H800) has 132, while Blackwell has 148 or 152. If the weights show a "148" or "152" pattern, the model was initialized on Blackwell.
Depue worked on Sora, Sora 2, GPT-4o, o3 and GPT-5 before founding Unemployment Labs, a nonprofit research group. He says senior researchers at OpenAI and Anthropic have independently verified the method. So far, those confirmations rest only on his word, but he has published the verification code on GitHub, so others can reproduce the result.
How to Identify a GPU From a Model's Weights
A neural network is billions of numbers, called weights. Before training, PyTorch fills them with random values generated from a single starting number, the seed. To do this quickly, the GPU splits the work across its cores. Picture several crews sorting one enormous deck of cards: with 132 crews, the deck ends up in one order, and with 148 or 152 crews, in a different one. Different chips have different core counts, so the pattern differs too.
Training then changes the weights millions of times, but some cells, such as those for rare words and symbols, barely move and keep their original values. Even in heavily trained parts of the model, a faint statistical echo of the initial fill remains. Depue reconstructed this pattern in reverse and reproduced the results exactly for DeepSeek-V4-Flash and Kimi K3 on real Blackwell chips.
Which Models Are Under Suspicion
According to his data, six Chinese models carry the Blackwell signature:
- DeepSeek-V4-Flash and MiniMax-M3 point to B200 or B300 (148 cores);
- DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant Group's Ling-3.1-flash point to 152 cores, that is, flagship GB200 or GB300 liquid-cooled racks.
Earlier models such as DeepSeek-V3 and Kimi K2.5 show traces of older Hopper hardware. Models from Alibaba (Qwen3.8), Zhipu, Tencent, Xiaomi and StepFun still show Hopper as well, while Meituan's LongCat-2.0 points to Huawei chips.
To test the method, Depue ran 24 control models with known hardware, including Meta's Llama 3.1, which was trained on H100s. He says all 24 were identified correctly and that a chance match is statistically ruled out. He rejected the one alternative, AMD's MI300X accelerator, which is also under restrictions and could be artificially "throttled" to produce the same pattern, for DeepSeek-V4-Flash based on minor differences in rounding.
The scale is striking. The report estimates that the five training runs it examined required about 12,000 Blackwell GPUs, roughly the computing power of 30,000 H100s. Kimi K3 was trained for about 70 days on 4,500 chips, and DeepSeek-V4-Pro on about 3,000. At around $70,000 per chip, one such cluster costs $200–300 million, and all five together cost more than half a billion dollars. This matches an estimate from Dylan Patel of SemiAnalysis: in an interview with Sequoia Capital, he said DeepSeek-V3's architecture was tailored to Hopper, while DeepSeek-V4 was built for Blackwell and Huawei's Ascend chips.
What the Fingerprint Does Not Prove
The caveats matter here. The method shows only which hardware initialized the weights. The training itself most likely ran on the same systems, but that is an inference, not a direct measurement. The data does not say where the GPUs physically sit or who owns them. Chinese companies may rent capacity in data centers outside China, which does not necessarily violate export rules. The report therefore shows access to top-tier hardware, not smuggling or a breach of US law.
For the industry, the takeaway cuts two ways. First, policymakers will have to decide what to do about cloud rentals, since restrictions on chip sales do not close that route. Second, model auditing has a new tool. Open weights reveal not only an architecture but also traces of the hardware it was trained on. If the method holds up under independent scrutiny, it could be used to check labs' claims about their compute against reality.
- Nov 21, 2026Anthropic mathematician says Claude helped prove the Riemann hypothesis in six weeks
- Nov 21, 2026Startup Osmo Trains AI on Smells Using 10 Million Human Ratings and Generates Fragrances From Text
- Nov 21, 2026North Korea Shows First Chatbot "Source of Wisdom": Science Only, No Politics
- Nov 21, 2026OpenAI Opens ChatGPT Spaces to Plus Subscribers: Shared Documents with AI
Key Questions Answered
How to Identify a GPU From a Model's Weights?
A neural network is billions of numbers, called weights. Before training, PyTorch fills them with random values generated from a single starting number, the seed. To do this quickly, the GPU splits the work across its cores. Picture several crews sorting one enormous deck of cards: with 132 crews, the deck ends up in one order, and with 148 or 152 crews, in a different one. Different chips have different core counts, so the pattern differs too.
Which Models Are Under Suspicion?
According to his data, six Chinese models carry the Blackwell signature: DeepSeek-V4-Flash and MiniMax-M3 point to B200 or B300 (148 cores); DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant Group's Ling-3.1-flash point to 152 cores, that is, flagship GB200 or GB300 liquid-cooled racks. Earlier models such as DeepSeek-V3 and Kimi K2.5 show traces of older Hopper hardware. Models from Alibaba (Qwen3.8), Zhipu, Tencent, Xiaomi and StepFun still show Hopper as well, while Meituan's LongCat-2.0 points to Huawei chips.
What the Fingerprint Does Not Prove?
The caveats matter here. The method shows only which hardware initialized the weights. The training itself most likely ran on the same systems, but that is an inference, not a direct measurement. The data does not say where the GPUs physically sit or who owns them. Chinese companies may rent capacity in data centers outside China, which does not necessarily violate export rules. The report therefore shows access to top-tier hardware, not smuggling or a breach of US law.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.