Chinese AI labs are training models on Nvidia's Blackwell despite export controls, says ex-OpenAI engineer

Export controls may not be keeping Nvidia's newest training hardware out of Chinese AI development. A forensic report cited by The Register claims models from DeepSeek, Moonshot AI, MiniMax, and Ant Group carry computational "fingerprints" consistent with training on Nvidia's Blackwell GPUs, according to a former OpenAI engineer. The finding does not by itself reveal how the chips were accessed, how many were used, or whether the evidence conclusively identifies Blackwell hardware. Still, it matters for builders tracking model capabilities and infrastructure constraints: if Chinese labs can obtain or remotely access restricted accelerators, hardware controls may slow frontier training without fully preventing it. The report also raises a practical question for policymakers, whether enforcement should focus more heavily on supply chains, cloud access, and compute tracking rather than chip shipments alone.