The thesis and the forensic method
Former OpenAI researcher Will DePue, now head of the nonprofit research group Unemployment Labs, claims to have produced the first technical evidence that DeepSeek, Moonshot AI, MiniMax, and Ant Group have access to Nvidia Blackwell GPUs, hardware subject to U.S. export controls. According to a report by The Decoder, U.S. officials and media outlets had already hypothesized such use by DeepSeek and Moonshot AI, but without publishing evidence. DePue states that senior researchers at OpenAI and Anthropic independently verified his method; this is his assertion, not a widely confirmed fact.
According to the source's reconstruction, the analysis is based on the published language models. Before training, the PyTorch framework fills billions of weights with random numbers generated from an initial seed; because the GPU distributes the work across its compute cores, and different cards have a different number of cores, the filling leaves a characteristic fingerprint. Hopper chips such as the H100 and H800 use 132 cores, while Blackwell chips use 148 or 152 depending on the model. Training then changes those values millions of times, but some positions remain at the initial random values, preserving a statistical trace of the starting configuration.
Which models and which chips
According to the report, six models published after April 2026 show the signature of the Blackwell generation, while earlier models such as DeepSeek-V3 and Kimi K2.5 show traces of the older Hopper generation. The analysis associates DeepSeek-V4-Flash and MiniMax-M3 with B200 or B300 chips (148 cores). DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ant Group's Ling-3.1-flash would instead show the 152-core pattern of GB200 or GB300 rack systems. DePue claims to have statistically ruled out chance and to have reproduced the results exactly for DeepSeek-V4-Flash and Kimi K3 on real Blackwell chips, publishing the verification code on GitHub.
The collected data indicate that other Chinese labs, including Alibaba with Qwen3.8 and the organizations Zhipu, Tencent, Xiaomi, and StepFun, would still show only Hopper signatures, while Meituan's LongCat-2.0 would point to Huawei hardware. The source also notes a convergence with the analysis by Dylan Patel of SemiAnalysis, according to which DeepSeek-V3's architecture was still tailored to Hopper GPUs, while DeepSeek-V4's would be designed for Blackwell and Huawei's Ascend chips.
Estimated costs and limits of the evidence
The report estimates that the five training runs examined required a total of about 12,000 Blackwell GPUs, equivalent in compute power to about 30,000 H100s. Kimi K3 alone would have used about 4,500 processors over a period of seventy days, and DeepSeek-V4-Pro about 3,000. At a unit price of about $70,000, according to the source each cluster would have cost between $200 million and $300 million, and the five systems together more than half a billion.
Questions nevertheless remain open. The technique, the source specifies, shows only which hardware initialized the weights, although it is highly likely that training took place on the same systems. The only alternative explanation would be an AMD MI300X card, also subject to restrictions, deliberately limited to produce the same pattern; DePue states that he ruled out AMD for DeepSeek-V4-Flash thanks to small rounding differences between the chips. The data do not reveal where the GPUs are located or who owns them: because Chinese companies can rent computing capacity in data centers outside China, the source emphasizes that the technical evidence demonstrates access to Nvidia's flagship hardware but does not in itself prove physical smuggling or a violation of U.S. export rules.
