Researcher Will DePue pushed back against skepticism over his forensic study linking six Chinese AI models to export-controlled Nvidia Blackwell GPUs, saying he has not yet tested Huawei Ascend random number generators on DeepSeek’s September V4.1-Flash release and cannot rule Nvidia hardware in or out. DePue said a search of 4.3 billion random seeds across GPU layouts from 16 to 512 processors and PyTorch’s CPU generator returned zero matches on V4.1-Flash’s untouched weight rows, showing DeepSeek changed its weight initialization scheme after releasing V4 in April. He argued that returning zero matches across more than 6 trillion combinations—as his method also did on OpenAI’s gpt-oss-120b and xAI’s Grok-1—proves the technique does not produce false Blackwell hits by chance.

DePue also rejected claims that Chinese labs could have initialized model weights on rented Blackwell GPUs before training them on different clusters. He said his forensic method matched the exact disclosed training hardware across all 24 open-weights models where developers stated what chips they used, including DeepSeek-V3 and Moonshot AI’s Kimi K2 on Nvidia H800s and Nvidia’s Nemotron models on GB200s. For Kimi K3, at least 64 separate processes on Grace Blackwell chips generated the starting weights, and transferring them to another cluster would have required saving and moving a checkpoint exceeding 5 TB just to generate random numbers a local training cluster produces in seconds.