Former OpenAI researcher finds signs of restricted Nvidia Blackwell chips in Chinese artificial intelligence models

Chinese artificial intelligence labs including DeepSeek, Moonshot AI, MiniMax, and Ant Group have built models using Nvidia's restricted Blackwell processors despite American export controls, according to a forensic analysis by Will DePue, a former OpenAI engineer who founded Unemployment Labs. DePue examined public open-weights models and concluded that six of them, such as DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3, were initialized on Blackwell hardware, with four showing matches to liquid-cooled Grace Blackwell GB200 and GB300 systems. [1]

The findings rely on reverse-engineering model parameters to detect unique hardware patterns left behind during the initialization of starting weights. Using pseudorandom number algorithms, PyTorch, and Nvidia's cuRAND library, DePue evaluated more than six trillion combinations of chip layouts and seed numbers to identify the underlying processors. However, DePue noted that the forensic signature only proves the models were initialized on Blackwell systems and does not reveal physical chip locations, leaving open the possibility that Chinese groups accessed the restricted hardware through American cloud providers. [1][2]

The revelations add fuel to an escalating political debate in Washington over closing loopholes in technology trade restrictions. Lawmakers are currently weighing multiple proposals to curb overseas AI access, including the AI Overwatch Act and the Cloud Security Act, amid broader scrutiny surrounding former Nvidia chief executive Jensen Huang and separate investigations by authorities into suspected chip-smuggling networks. [1]