Six Chinese Frontier AI Models Show Blackwell GPU Fingerprints, Export Control Enforcement Faces New Test

On November 12, 2026, a forensic report released by Unemployment Labs concluded that six models—DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3, and Ling-3.1-flash—used export-controlled NVIDIA Blackwell GPUs during the training initialization phase, with an estimated 12,000 GPUs involved.

Factual Reconstruction

The report reached this conclusion by analyzing the numerical distribution patterns of model weight initialization, with confidence ranging from 49 to 1,385 standard deviations. The DeepSeek-V4 series and MiniMax-M3 match B200/B300 chip signatures, while the Kimi K3 series and Ling-3.1-flash match GB200/GB300 signatures. Previously, older versions of models from each lab all showed traces of Hopper-generation hardware; the new versions collectively switched between April and October 2026.

All evidence came from publicly available model weights, not leaks or insider information. The researchers recovered the models' initial values and searched for hardware configurations and initialization settings that could reproduce these values. DeepSeek-V4-Flash and Kimi K3 were also confirmed on physical Blackwell chips.

Mechanism Breakdown

The method identifies hardware fingerprints through the numerical distribution of weight initialization. Hopper fingerprints cannot distinguish among H800, H100 SXM, or H200, whereas Blackwell fingerprints can distinguish between two configurations: 148 SM and 152 SM. The report validated the method on 24 models with publicly disclosed hardware information, and the method's conclusions were fully consistent with developer announcements.

Among the six models, the strongest evidence reached 1,385 sigma, and even the weakest single test reached 20 sigma, far exceeding the expected upper bound for random matching. Training-scale estimates indicate that each run requires 1,500 to 3,500 Blackwell GPUs operating for one and a half to three months.

Industry Impact

New models from four labs collectively show Blackwell traces, while the latest models from other Chinese labs such as Alibaba Qwen, Zhipu GLM, Tencent, Xiaomi, and StepFun still show Hopper signatures. No NVIDIA fingerprints were detected in Meituan LongCat-2.0.

This finding provides a verifiable technical path at the intersection of evaluation transparency and chip export controls: model weights themselves can become carriers of hardware usage traces.

Strategic Assessment

(The following is analysis, not fact.) If similar forensic methods are widely adopted, model publishers may face greater disclosure pressure, and the concealment of hardware acquisition channels will face further challenges; historical precedents show that once technical traces are made public, they often accelerate the iteration of regulatory tools, but may also push parties to explore non-NVIDIA alternatives.