Unemployment Labs Detects Blackwell GPU Fingerprints in Six Chinese AI Models
·2026.11.14 12:51
Key point
Analysis of model weights reveals initialization patterns from NVIDIA Blackwell GPUs in models from DeepSeek, Moonshot AI, MiniMax, and Ant Group.
Details
- Unemployment Labs identified GPU initialization fingerprints in six models from four Chinese organizations, indicating the use of NVIDIA Blackwell hardware (B200/B300 or GB200/GB300) for weight initialization.
- The method exploits the Philox pseudorandom number generation algorithm in PyTorch, where the number of streaming multiprocessors (SMs) determines the placement of initial weight values; Blackwell GPUs have 148 or 152 SMs, distinct from Hopper's 132 SMs.
- Statistical analysis of 6 trillion seed and hardware combinations yielded matches of 49 to 1,385 sigma, far exceeding the 8 sigma expected by chance, with physical reproduction tests on B300 and GB200 hardware confirming the findings.
- The study estimates a central figure of 12,000 Blackwell GPUs used across five independent training runs, with a revised range of 7,000 to 22,000, though it cannot determine the physical location or ownership of the hardware.
- The technique cannot distinguish between specific Blackwell models (e.g., B200 vs. B300) or prove export control violations, as weights may be initialized on one cluster and trained on another.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.