Chinese frontier AI models carry the fingerprint of export-controlled GPUs
Web page · https://unemploymentlabs.com/checkpoint-forensics
First seen by Oriel on .
AI summary
The report uses published model weights to identify Blackwell GPU initialization fingerprints in six recent models from four Chinese AI companies, while emphasizing that this does not prove export-control violations.
- DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash match NVIDIA Blackwell configurations.
- The method reconstructs Philox-generated initial weights by testing random seeds and GPU thread layouts, then validates matches on held-out weights.
- B200/B300 matches have 148 streaming multiprocessors, while GB200/GB300 matches have 152; Hopper models match 132 but cannot distinguish H800, H100 or H200.
- The report estimates final training runs used roughly 500 to 3,000 Blackwell GPUs for two to three months, with the largest clusters costing $150 million to $200 million.
- The technique requires published weights and suitable PyTorch GPU fingerprints, so closed models, some PaddlePaddle models and CPU-initialized models cannot be identified reliably.
#gpu forensics #nvidia blackwell #ai model training #export controls
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Points
Score reported by the source.
- Hacker News: 450, latest observation ; +437 over observed period since
Comments
Comment count reported by the source.
- Hacker News: 194, latest observation ; +194 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hacker News | 50474883 | https://unemploymentlabs.com/checkpoint-forensics Read discussion |