A statistical trace inside published AI models may offer clues about the chips used to initialize them. Will DePue, an AI researcher and founder of Unemployment Labs, says he found Nvidia Blackwell signatures in six Chinese models released since April 2026. His findings point to access to the hardware, but cannot show where the GPUs were located or who owned them.
How model weights can retain a hardware trace
An AI model is made up of billions of numerical values called weights. Before training begins, PyTorch fills those values with random numbers generated from a starting value, or seed. The process spreads work across a GPU’s compute cores, and DePue’s analysis says differences in core counts can leave distinct statistical patterns in the initial values.
According to the report, Nvidia’s older Hopper chips, including the H100 and H800, have 132 compute cores. Blackwell chips use 148 or 152 cores, depending on the model: B200/B300 or GB200/GB300. The different counts produce different patterns during initialization.
Training repeatedly adjusts model weights, but some values associated with unused words or symbols retain their original random values. DePue says a faint statistical echo of the initial pattern can remain even in trained models. He reports reproducing the results for DeepSeek-V4-Flash and Kimi K3 on real Blackwell chips, and has published verification code on GitHub.
Which models show the Blackwell pattern
DePue’s analysis assigns DeepSeek-V4-Flash and MiniMax-M3 to the 148-core pattern associated with B200 or B300 chips. It links DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ant Group’s Ling-3.1-flash to the 152-core pattern associated with liquid-cooled GB200 or GB300 rack systems.
Earlier models, including DeepSeek-V3 and Kimi K2.5, show Hopper signatures, according to DePue. He says the method correctly identified all 24 control models with known hardware, including Meta’s Llama 3.1 (H100), and that the statistical result is unlikely to be explained by chance.
The report also describes models that do not show a Blackwell signature. Alibaba’s Qwen3.8 and models from Zhipu, Tencent, Xiaomi, and StepFun show Hopper patterns. Meituan’s LongCat-2.0 points to Huawei hardware.
The estimated computing resources are substantial
DePue estimates that the five training runs he examined used about 12,000 Blackwell GPUs combined, equivalent in computing power to roughly 30,000 older H100 chips. His estimates put Kimi K3 at 4,500 Blackwell processors over about 70 days, and DeepSeek-V4-Pro at about 3,000 processors.
At about $70,000 per chip, DePue estimates that each of those clusters would cost $200 million to $300 million. The five systems together would cost more than half a billion dollars. These are estimates in the report, not a direct accounting of purchases or ownership.
The findings are consistent with comments from Dylan Patel, chief analyst at SemiAnalysis. In an interview with Sequoia Capital, Patel said DeepSeek-V3’s architecture was tailored for Hopper, while DeepSeek-V4’s was built for Blackwell GPUs and Huawei’s Ascend chips.
What the fingerprint does—and does not—show
The method identifies the hardware that initialized the weights. DePue says the training likely took place on the same systems, but the fingerprint alone does not establish that. The report says an AMD MI300X, also restricted, could theoretically be artificially throttled to produce a similar pattern; DePue says he ruled it out for DeepSeek-V4-Flash using small rounding differences between chips.
Most importantly, the data does not reveal the GPUs’ location or ownership. Chinese companies could rent computing capacity at data centers outside China. So, while the analysis is presented as technical evidence of access to Nvidia’s export-controlled Blackwell hardware, it does not by itself prove physical smuggling or a violation of US export law.
DePue says senior researchers at OpenAI and Anthropic independently checked and confirmed his findings. That review lends support to the method, while the limits of the evidence still matter: a model can indicate a hardware signature without showing the path by which its developers obtained computing capacity.