A former OpenAI researcher links six Chinese models to NVIDIA Blackwell chips, subject to export restrictions
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
Ant Group, MiniMax, DeepSeek and Moonshot AI appear in an analysis by Will DePue that finds the trace of NVIDIA Blackwell chips in six of their models. Hopper, the previous generation, is the one that leaves its mark on the earlier models from those companies, and everything comes from weights already published. DePue warns that this does not say where the chips were or whether export rules were broken.
Will DePue, a former OpenAI researcher, released a forensic analysis through Unemployment Labs on Thursday, November 12, according to which six models launched by DeepSeek, Moonshot AI, MiniMax and Ant Group since April 2026 carry a numerical trace consistent with Blackwell, the NVIDIA chips that can only be exported to China with a US license. The report adds that no such license has been reported. Earlier models from those same companies, by contrast, line up with Hopper, the previous generation.
The list is made up of Ling-3.1-flash, by Ant Group; MiniMax-M3; Moonshot AI's Kimi K3.1-Pro and Kimi K3; and DeepSeek-V4-Pro and DeepSeek-V4-Flash. The method relies only on the weights the companies have made public, the numbers that set how a model responds. Before training, those numbers get starting values whose order depends on how many processing units the chip that generates them has, and part of that order survives. With DeepSeek-V4-Flash and Kimi K3, DePue says he reproduced those values on real Blackwell chips, and on the 24 models whose developer has publicly said what hardware it used, the method gives the same answer as the one declared.
The report itself warns that it does not establish where the chips were, how the companies got access to them or whether any export rule was broken, and it notes that Blackwell systems can be rented in data centers outside China. The trace points to the chip that generated the starting values; it does not prove that all of the training was done on it. As the only alternative, it names AMD's MI300X, also restricted, which would only leave that same trace in a highly unusual setup. A check of small numerical differences rules it out for DeepSeek-V4-Flash; for the other five that check was not conclusive, although DePue sees AMD as very unlikely there. The report calls the evidence conclusive only for DeepSeek-V4-Flash and describes it as strong for the remaining five.
The newest models DePue could examine from StepFun, Xiaomi, Tencent, Zhipu (GLM) and Alibaba (Qwen) still show the Hopper trace. Before this report, US officials and the press had already pointed at DeepSeek and Moonshot AI over the use of Blackwell, but no one had put technical proof on the table. DePue estimates that the two largest training runs, those of Kimi K3 and DeepSeek-V4-Pro, needed thousands of Blackwell chips for two to three months, although he says he is not confident in those estimates. The work is signed by him alone and, according to DiarioBitcoin, no published external review of his method is on record. According to Trending Topics, neither NVIDIA nor the four Chinese companies had commented on the report yet.
Why it matters · analysis and opinion
If the method holds up when others examine it, chip export controls would have a way to be checked from the outside, just by looking at what companies publish. That changes a debate about China's access to the most advanced hardware that until now rested on accusations without technical evidence. The work itself sets out its limits: it cannot tell a purchased chip from one rented outside China, and it does not work with closed models, which do not publish their weights. The next step is for other researchers to repeat the tests on the same models. For anyone using these open models, nothing changes for now in how they are downloaded or used.
Source: Unemployment Labs · Written with the help of AI: how we make the news
