Former OpenAI researcher Alex DePue says his code can spot the tiny statistical imprint left by the GPU that first fills a model’s weights. He applied the method to six Chinese language models released after April 2026 and found patterns that match Nvidia’s Blackwell generation. The finding gives the first technical proof that those models were trained on Blackwell chips, which are the newest, most powerful GPUs Nvidia sells.
DePue’s test compares the random seed pattern that each GPU core writes when a model is built. Blackwell chips have either 148 or 152 compute cores, while the older Hopper chips use 132. The six models – DeepSeek‑V4‑Flash, DeepSeek‑V4‑Pro, MiniMax‑M3, Kimi K3, Kimi K3.1‑Pro and Ant Group’s Ling‑3.1‑flash – all show the 148‑ or 152‑core signature. Earlier Chinese releases like DeepSeek‑V3 and Kimi K2.5 still carry Hopper’s 132‑core pattern.
What the numbers mean
DePue estimates the five training runs he examined needed about 12,000 Blackwell GPUs in total – roughly the power of 30,000 older H100 chips. One model, Kimi K3, would have required around 4,500 Blackwell GPUs over 70 days; DeepSeek‑V4‑Pro about 3,000 GPUs. At about $70,000 per GPU, each cluster costs $200‑$300 million, and all five together exceed $500 million.
Context and limits
The fingerprint only shows which hardware initialized the model’s weights. It does not reveal where the GPUs were located or who owned them. Chinese firms can rent foreign data‑center capacity, so the evidence does not prove illegal export or smuggling. DePue also ruled out AMD’s MI300X chips as an alternative source because of tiny rounding differences.
Why it matters
If the models really ran on Blackwell GPUs, they likely have higher performance than earlier Chinese AI offerings, affecting users who rely on these services for faster responses. The finding also shows Chinese firms can access cutting‑edge US hardware, raising questions about export controls. However, the analysis does not prove where the hardware is located or whether any laws were broken.
