Ex-OpenAI researcher traces six Chinese AI models to banned Nvidia chips

Will DePue joined OpenAI in 2023, where he worked on its Sora video models and on post-training for o3 and GPT-4o. He published the analysis through Unemployment Labs, and says it leans entirely on the weights the Chinese labs released publicly — no leaks, no insider accounts. On X, he said senior researchers at OpenAI and Anthropic independently verified the result, though the report names no outside reviewers.

Six models are named: DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ant Group's Ling-3.1-flash. Four of them match Nvidia's GB200 or GB300 Grace Blackwell systems, per the report, and two match B200 or B300 chips. DePue estimates the two largest runs, Kimi K3 and DeepSeek-V4-Pro, each used about 2,000 to 3,000 Blackwell GPUs — clusters he prices at $150 million to $200 million apiece, with all five final training runs together topping $500 million.

The method hinges on the starting values a model's weights get before training begins. DePue says the layout of those values depends on the exact GPU and the random seed used, and that traces of the pattern survive into the published model. He searched more than six trillion combinations of chip layout and seed to find matches.

Then he tested the matches on real hardware. With the recovered seed, a Blackwell B300 reproduced DeepSeek-V4-Flash's untouched starting weights bit for bit, and a rented GB200 reproduced all 13 preserved rows of Kimi K3. Nvidia's older H100 reproduced none of V4-Flash's rows, and neither did AMD's MI300X; shifting the seed by one broke every match. The report puts chance's strongest possible match at about 8 standard deviations, while the real matches range from 49 to 1,385.

The report is explicit about its limits. It says the method identifies the hardware that generated the starting weights, not every chip that later helped train the model. It shows nothing about where the chips were, who owned them or how they were obtained, and notes Blackwell systems can be rented in data centers outside China. In the report's words, the results are not a finding that any law was broken.

Washington has been circling this for months. In July, White House science adviser Michael Kratsios said Moonshot had accessed GB300s in Thailand, likely to train its AI models, pointing at Kimi K3. On November 12, The Information reported that companies tied to Alice Huang invested in DeepSeek and promised it Nvidia chips, a day after Senators Elizabeth Warren and Chris Coons pressed the Trump administration to preserve Nvidia's export records. Nvidia, DeepSeek, Moonshot, MiniMax and Ant Group had not commented at publication, and DePue says more results from the same method are coming.