What changed
Will DePue, a former OpenAI researcher, claims he found technical evidence that six Chinese AI models contain signatures of Nvidia's Blackwell GPUs. His analysis, verified by senior researchers at OpenAI and Anthropic, indicates that models from DeepSeek, Moonshot AI, MiniMax, and Ant Group may be using Blackwell hardware.
What this means in practice
DePue's investigation focused on the patterns left in AI models during training. These patterns, created by the way GPUs distribute computational tasks, can reveal the type of hardware used. Older Hopper chips have 132 compute cores, while newer Blackwell chips have 148 or 152 cores. The analysis found that models like DeepSeek-V4-Flash and Kimi K3 show the 148 or 152-core pattern of Blackwell chips. The report estimates that training these models required thousands of Blackwell GPUs, costing hundreds of millions of dollars.
Limitations and availability
The findings do not prove illegal activity, as the data does not reveal ownership or location of the GPUs. Chinese companies could be using rented computing power outside China. The fingerprint method also rules out AMD MI300X chips as a possible explanation. The data only shows which hardware initialized the weights, not the entire training process. The analysis does not confirm physical smuggling or legal violations, as the evidence is based on patterns in published models.
