GoKawiilResearch firm says Chinese AI models were trained on restricted Nvidia Blackwell chips
Independent researcher UnemploymentLabs says mathematical analysis of model weights shows DeepSeek V4, Kimi K3.1, MiniMax-M3 and Ling-3.1-Flash were trained using Nvidia Blackwell GPUs, chips subject to US export restrictions on China. The group bases its claim on patterns tied to the number of shader modules in Blackwell hardware and the cuRAND randomization process used during weight initialization in PyTorch-based training.
GoKawiil's interpretation of the reporting above, not reported fact.
The researchers themselves caution that their method cannot confirm where the chips were located, how they were obtained, or whether any export rule was actually broken, since renting foreign-based Blackwell hardware may not be illegal. This suggests the finding highlights a gap in enforcement of US chip controls rather than proving a violation, and could intensify scrutiny of how Chinese labs access high-end Nvidia compute. It also points to growing interest in forensic techniques for tracing hardware provenance in AI model training.
- UnemploymentLabs claims DeepSeek V4, Kimi K3.1, MiniMax-M3 and Ling-3.1-Flash show signs of Blackwell GPU training
- The method relies on shader module counts and cuRAND weight-initialization patterns, and only works on open-weight models
- Researchers stress this does not prove export-control violations, since renting Blackwell chips abroad may not be illegal
Source: tomshardware.com — Bruno Ferreira, 2026-11-14
Published there as: “Researchers allege DeepSeek, Kimi, and other Chinese AI models were trained using banned Nvidia Blackwell GPUs”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.