Researcher Claims Nvidia Blackwell Signatures Appear in Chinese AI Models

Former OpenAI researcher Will DePue says he found technical evidence that Chinese AI companies DeepSeek, Moonshot AI, MiniMax, and Ant Group are using Nvidia’s export-controlled Blackwell GPUs. The Decoder reports that DePue’s method reads chip-specific signatures directly from published language model weights. He says six Chinese models released since April 2026 carry Blackwell-generation patterns, while some earlier models show traces of Nvidia’s older Hopper generation.

The reported method focuses on how PyTorch initializes model weights before training begins. Because different GPU generations split initialization work across different numbers of compute cores, DePue says the process leaves a characteristic statistical pattern. According to the report, Hopper chips such as H100 and H800 use 132 compute cores, while Blackwell chips use 148 or 152 cores depending on the model family.

DePue says some unused or lightly altered parts of trained models retain enough of the original random initialization to identify the hardware pattern. He also says he reproduced results for DeepSeek-V4-Flash and Kimi K3 on real Blackwell chips and published verification code on GitHub.

The analysis maps DeepSeek-V4-Flash and MiniMax-M3 to Nvidia B200 or B300 chips, which have a 148-core pattern. It maps DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ant Group’s Ling-3.1-flash to the 152-core pattern associated with Nvidia’s liquid-cooled GB200 or GB300 rack systems. DePue says the method correctly identified all 24 control models with known hardware, including Meta’s Llama 3.1 on H100.

DePue’s report estimates the five training runs he examined required about 12,000 Blackwell GPUs combined, comparable to roughly 30,000 older H100 chips in computing power. The Decoder reports that Kimi K3 alone required an estimated 4,500 Blackwell processors over about 70 days, while DeepSeek-V4-Pro required about 3,000. At a unit price of about $70,000, the analysis estimates the systems together cost more than half a billion dollars.

The fingerprinting method indicates which hardware initialized the model weights, and the report says training likely ran on the same systems. But the evidence does not reveal where the GPUs are located or who owns them. Because Chinese companies can rent compute capacity in data centers outside China, the findings suggest access to Nvidia’s top hardware but do not by themselves prove smuggling or a violation of US export law.