Research from Unemployment Labs links Nvidia Blackwell chips to AI models from DeepSeek and Moonshot AI. The findings do not establish that the models were trained entirely on Nvidia hardware, and the chips could have been used offshore.
Will DePue, a former OpenAI researcher who leads Unemployment Labs, analysed openly released model weights from DeepSeek, Moonshot AI, MiniMax and Ant Group. According to TechCrunch, he found evidence of Blackwell training configurations in models including DeepSeek V4 and Moonshot AI Kimi K3.
Nvidia Blackwell GPUs are restricted from export to China or Chinese-owned firms. The report notes that the chips could have been located in an offshore data centre, rather than in China.
For companies building with these AI models, the research offers a view into the computing infrastructure behind them. It also highlights a distinction worth keeping when assessing suppliers: a model developer's home country does not necessarily reveal where its training hardware operates.
Companies assessing AI suppliers cannot infer their training infrastructure from the developer's location alone.
Access to offshore computing capacity matters when judging how restrictions affect a model developer's ability to build.
The location of the chips and the full mix of hardware used for training remain unproven.
Written automatically by the Made in 10 Newsroom from the source above.
Chinese AI projects worth over $3.2 billion specify restricted Nvidia chips, according to a Reuters investigation. The figure covers whole projects, not confirmed chip purchases or deliveries.
Sierra has released Personal Agent Protocol (Poppy) 1.0, with Shopify enabling it for every store. The standard controls how AI agents identify themselves and get permission to act, but its specification remains a draft.
Google is adding a Google Wallet connection for every US user of the Gemini app. Access is read-only, but Google may train AI on transaction summaries if users keep Gemini activity on.