THE CRUNCH
Will DePue, a former OpenAI researcher who worked on Sora, GPT-4o and GPT-5 before founding the research group Unemployment Labs, has published a forensic analysis claiming that six Chinese AI models were initialised on Nvidia's export-controlled Blackwell GPUs. The models named include DeepSeek-V4-Pro, DeepSeek-V4-Flash, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ant Group's Ling-3.1-flash, from DeepSeek, Moonshot AI, MiniMax and Ant Group. DePue says senior researchers at OpenAI and Anthropic have independently checked and confirmed the findings.
The method exploits how models are built. Before training, a model's weights, the billions of numerical values that make it up, are filled with pseudorandom numbers generated on the training GPU. Because different chips split that work across different numbers of compute cores, the initial pattern acts as a hardware fingerprint, and traces of it survive training into the published model. Hopper chips such as the H100 use 132 cores, while Blackwell parts use 148 or 152 depending on the model. DePue says he searched more than six trillion chip-and-seed combinations, reproduced the results exactly on real Blackwell hardware for two models, and correctly identified all 24 control models with known hardware.
The analysis maps DeepSeek-V4-Flash and MiniMax-M3 to B200 or B300 chips, and DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash to the 152-core GB200 or GB300 rack systems. DePue estimates the five training runs he examined needed about 12,000 Blackwell GPUs, roughly the computing power of 30,000 older H100s, and cost more than half a billion dollars in total at around $70,000 per chip. Earlier models such as DeepSeek-V3 and Kimi K2.5 show only Hopper signatures, and labs including Alibaba, Zhipu, Tencent, Xiaomi and StepFun still show only Hopper signatures, while Meituan's LongCat-2.0 points to Huawei hardware.
There are caveats. The method shows which hardware initialised the weights, not where the GPUs are or who owns them, so it does not by itself prove smuggling or a breach of US export law; Chinese firms can legally rent capacity at data centres outside China. The only alternative technical explanation DePue identifies is an AMD MI300X, itself restricted, artificially throttled to mimic the pattern, which he says he ruled out for DeepSeek-V4-Flash using rounding differences between the chips.