Six Model Weights Carry Nvidia’s Fingerprints
On 14 November, nonprofit Unemployment Labs — led by former OpenAI researcher Will Depue — published a checkpoint-forensics report on six Chinese models released since April: DeepSeek V4-Flash and V4-Pro, Moonshot’s Kimi K3 and K3.1-Pro, MiniMax M3 and Ant Group’s Ling-3.1-Flash. It found initialisation fingerprints specific to the Blackwell architecture: V4-Flash and M3 matched the B200/B300 compute profile, while the other four aligned with GB200/GB300 Grace-CPU pairings. The report estimates roughly 12,000 Blackwell GPUs went into Chinese frontier training, with about $500 million in chip and cluster cost even counting final-stage training alone. The authors are explicit about the limit of their evidence — weights prove the hardware was used, not where it sat, so lawful offshore rental is fully consistent with the finding — but it lands in the middle of the Alice Huang/Megaspeed affair and hands Washington its first technical exhibit for the “cloud loophole” argument.
