hn.today

Evidence Chinese AI labs built their newest models on Nvidia's Blackwell chips

twitter.com6 points0 comments
Screenshot of Evidence Chinese AI labs built their newest models on Nvidia's Blackwell chips

Will Depue says he identified hardware fingerprints in the starting weights of six Chinese AI models, including DeepSeek-V4-Pro and V4-Flash, Moonshot’s Kimi K3, and MiniMax-M3. His method searches for GPU-specific patterns left by the random values used to initialize model weights, then compares those patterns with publicly released model files. He reports searching more than six trillion combinations of hardware configuration and seed, and says four models match Nvidia Grace Blackwell systems, including GB200 and GB300. As validation, a B300 reportedly reproduced DeepSeek-V4-Flash’s untouched starting weights exactly, while a GB200 reproduced Kimi K3’s with near-perfect correlation. An H100 did not match, and changing the seed by one broke the result. Depue also says the method correctly identified hardware for 24 models whose developers had disclosed what they used.

He estimates that Kimi K3 and DeepSeek-V4-Pro each required roughly 2,000-3,000 Blackwell GPUs, costing about $150-200 million per cluster, and argues that the labs’ total use of restricted chips could exceed half a billion dollars. The evidence, he says, is statistically far stronger than chance and has been independently checked by senior OpenAI and Anthropic researchers. The fingerprints indicate which hardware initialized the models, but do not show where the chips were located: they may have been brought into China illicitly or used by the labs abroad. Depue argues that either possibility exposes gaps in export controls.

Read on twitter.com0 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in AI

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.