Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models
The Decoder
A forensic analysis claims six recent Chinese AI models show technical evidence of being trained on Nvidia's restricted Blackwell hardware.
Former OpenAI researcher Will DePue conducted a forensic investigation indicating that leading Chinese AI labs have utilized Nvidia's export-controlled Blackwell GPUs. According to the analysis, models from DeepSeek, Moonshot AI, MiniMax, and Ant Group carry the distinct signature of Nvidia's Blackwell generation.
The method tracks initial random values left in model weights, which mirror the core counts of specific GPU architectures. The report maps these models to 148-core or 152-core Blackwell systems, estimating that the training runs required thousands of processors and hundreds of millions of dollars.
While the findings suggest access to advanced hardware, the analysis notes it does not prove physical smuggling or a direct violation of U.S. export law, as companies can potentially rent computing capacity outside of China. Senior researchers at OpenAI and Anthropic reportedly independently verified the findings.
Why it matters: This analysis provides the first technical evidence suggesting that leading Chinese AI labs are successfully utilizing U.S. export-controlled Blackwell hardware despite ongoing restrictions.



