A former OpenAI researcher claims to have discovered the first technical evidence that six Chinese super intelligence (SI) models were trained on Nvidia's restricted Blackwell chips. Will DePue states that his analysis of model weights reveals hardware signatures distinct from the older Hopper generation, confirming long-standing suspicions about access to advanced compute in China.
What Happened
DePue’s report identifies a characteristic pattern in the initial random values of model weights, which are generated by PyTorch before training begins. An SI model consists of billions of numerical values called weights. Because Nvidia's Blackwell chips (B200/B300 or GB200/GB300) have different core counts (148 or 152) compared to Hopper chips (132), the distribution of these values leaves a statistical fingerprint. DePue says he reproduced these results for DeepSeek-V4-Flash and Kimi K3 on real Blackwell hardware and published the verification code on GitHub.
The analysis maps DeepSeek-V4-Flash and MiniMax-M3 to B200 or B300 chips, while DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ant Group's Ling-3.1-flash show signatures of the liquid-cooled GB200 or GB300 systems. In contrast, earlier models like DeepSeek-V3 and Kimi K2.5, along with recent releases from Alibaba, Zhipu, Tencent, Xiaomi, and StepFun, display Hopper signatures. Meituan's LongCat-2.0 points to Huawei hardware.
DePue estimates that the five training runs examined required approximately 12,000 Blackwell GPUs. Kimi K3 alone reportedly used 4,500 Blackwell processors over 70 days, while DeepSeek-V4-Pro used about 3,000. At an estimated unit price of $70,000, these clusters would cost between $200 million and $300 million each, totaling more than half a billion dollars for all five systems.
Why It Matters
This development provides concrete data to support previous suspicions held by US officials and news outlets regarding Chinese access to Nvidia's top-tier hardware. While the findings align with analysis from SemiAnalysis chief analyst Dylan Patel, who noted that DeepSeek-V4's architecture was built for Blackwell GPUs, DePue’s work offers a method to verify hardware usage through public model weights.
However, the fingerprint does not prove a legal violation. The method identifies the hardware used to initialize weights, and while training likely occurred on the same systems, the data does not reveal the physical location or ownership of the GPUs. Chinese companies can legally rent compute capacity at data centers outside China, meaning this evidence shows access to Nvidia's top hardware but does not, by itself, prove physical smuggling or a breach of US export laws.
The Bottom Line
DePue’s analysis suggests that major Chinese SI models are increasingly leveraging Nvidia’s Blackwell architecture, marking a shift from the Hopper generation. While this indicates significant investment in restricted hardware, the question of whether this access complies with export controls remains unresolved without further evidence of where the compute occurred.