Will DePue’s forensic report appendix shows an exhaustive search recovered no Nvidia Blackwell hardware signature from DeepSeek’s newest model, DeepSeek-V4.1-Flash, and confirmed that both V4.1-Flash and LongCat were trained on Huawei Ascend chips instead. While DePue matched six earlier models from DeepSeek, Moonshot AI, MiniMax and Ant Group to export-controlled Blackwell GPUs, the analysis does not establish where the chips were physically located or whether US export rules were violated, as Chinese developers can rent GB300 capacity remotely through cloud providers.
Technical reviewers noted that Chinese labs had already signaled Blackwell use publicly in their code repositories. DeepSeek’s open-source FlashMLA repository explicitly listed support for Nvidia SM100 and SM103 Blackwell architectures and removed Hopper support at the end of September, while Ant Group’s GitHub contains GB300 benchmark logs dated March 31 and MiniMax’s models shifted from Hopper on M2.7 in April to Blackwell on M3 in June. Analysts also noted that the estimated 2,000 to 3,000 Blackwell GPUs used for Kimi K3 and DeepSeek-V4-Pro represent roughly 6 MW of compute, far smaller than frontier US training clusters.