Forensic Analysis Finds Blackwell GPU Fingerprints in Six Chinese Frontier AI Models

Six Chinese frontier models released after April 2026, including DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3, and Ling-3.1-flash, exhibit hardware fingerprints matching NVIDIA Blackwell GPUs.
The analysis recovered initial weight values to trace hardware configuration, finding that DeepSeek-V4-Flash and MiniMax-M3 align with B200/B300 chips (148 SM), while DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ling-3.1-flash align with GB200/GB300 chips (152 SM).
Statistical evidence for the Blackwell match ranges from 49 to 1,385 sigma, far exceeding the 8 sigma threshold for accidental coincidence across 6 trillion seed-layout combinations.
Estimates suggest each model required 500 to 3,000 Blackwell GPUs for 2 to 3 months of training, with total final training costs exceeding $500 million for the five models included in the cost analysis.

Methodology and Verification

The forensic method uses the deterministic nature of the Philox PRNG algorithm in PyTorch and NVIDIA cuRAND to reverse-engineer GPU architecture from preserved initial weights.
Verification on 24 models with publicly disclosed hardware confirmed the method's accuracy, including bit-for-bit reproduction of DeepSeek-V3 on H100 and V4-Flash on B300.
AMD MI300X was ruled out for DeepSeek-V4-Flash based on numerical differences in 377 preserved values, where 363 matched NVIDIA and 0 matched AMD.

The analysis identifies the hardware used for initialization but does not prove the entire training process occurred on that hardware or identify the physical location or owner of the chips.
Models initialized on CPU, such as certain Qwen3 variants, or using non-NVIDIA frameworks like Baidu's PaddlePaddle, do not yield GPU fingerprints.
The report notes that while the evidence strongly suggests access to export-controlled hardware, it does not directly prove export control violations or the method of acquisition.