Ex-OpenAI engineer says Chinese labs run Nvidia Blackwell chips despite export bans

A former OpenAI engineer has told The Register that several Chinese AI research labs are training large language models on Nvidia’s Blackwell GPUs, despite the chips being subject to U.S. export controls source. The forensic analysis cited by the engineer found the Blackwell hardware fingerprint in models released by DeepSeek, Moonshot AI, MiniMax and Ant Group.

The engineer’s report suggests that the labs have either obtained the chips through indirect channels or are circumventing the export regime.

If the claim is accurate, it would indicate that current export-control mechanisms are insufficient to prevent advanced AI hardware from reaching sanctioned end-users. The situation could affect multinational chip vendors, compliance teams, and AI developers who rely on clear regulatory guidance for sourcing hardware.

Our read: The allegation underscores a tangible risk that sophisticated AI compute resources can slip past export barriers, potentially reshaping competitive dynamics and prompting tighter scrutiny of supply-chain transparency for high-end GPUs.