Former OpenAI engineer Will DePue published what he calls forensic evidence that DeepSeek, Moonshot AI, MiniMax and Ant Group trained models on Nvidia Blackwell GPUs despite US export controls, The Register reports.
His claim: when a model’s weights are first initialized, the layout depends on the exact GPU and seed, leaving “a sort of fingerprint of the hardware,” and traces survive training into the published model. “This sounds insane,” he wrote, “but I discovered that almost every AI model carries a hidden signature of the GPU that trained it.”
It lands in the middle of the Jensen Huang fallout and a push in Congress for tighter controls, with at least four bills pending, including the AI Overwatch Act.
My take
Clever if it holds up, and it’s the kind of thing that needs independent replication before anyone hangs policy on it. Either way, export controls that depend on chips staying put were always going to leak.
God Bless, and Use the Source!