Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models
Reported by Startup Fortune, Glitchwire, Crypto Briefing and 5 more outlets8 sources ↓
Will DePue says patterns in publicly released weights from six Chinese AI models match Nvidia Blackwell GPUs. The analysis identifies hardware used to initialize weights, but does not establish where the GPUs were or who owned them.
Key points
- The six models are from DeepSeek, Moonshot AI, MiniMax and Ant Group
- DePue says the weight patterns identify Blackwell hardware used to initialize the models
- The analysis does not establish where the GPUs were, who owned them or whether export laws were broken
- DePue says real Blackwell hardware reproduced starting weights for DeepSeek-V4-Flash and Kimi K3
Sources
- Startup FortuneRead ↗Breaking News as Researcher Says Six Chinese AI Models Carry the Fingerprint of Banned Nvidia Chips
- GlitchwireRead ↗OpenAI Researcher Claims GPU Fingerprints Prove Chinese Labs Trained on Export-Banned Blackwell Chips
- Crypto BriefingRead ↗Former OpenAI researcher says six Chinese AI models were built on banned Nvidia chips
- techtimes.co.ukRead ↗Former OpenAI Researcher Uncovers Hidden Nvidia Chip Fingerprints in Six Chinese AI Models
- The DecoderRead ↗Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models
Details
Will DePue, a former OpenAI researcher who published the analysis through Unemployment Labs, examined publicly released weights from six models: DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ant Group’s Ling-3.1-flash. His analysis attributes four models to Nvidia systems with 152 compute cores and two to systems with 148, corresponding to different Blackwell hardware configurations.
The analysis looks for statistical traces of the random values used to initialize model weights before training. DePue says the patterns depend on the chip configuration and random seed, and reports reproducing starting weights for DeepSeek-V4-Flash and Kimi K3 on Blackwell hardware.
The estimate of GPU use varies by report. Crypto Briefing says Kimi K3 and DeepSeek-V4-Pro each used about 2,000 to 3,000 Blackwell GPUs, while The Decoder reports estimates of 4,500 for Kimi K3 and 3,000 for DeepSeek-V4-Pro, and about 12,000 across five runs. These are estimates from DePue’s analysis, not independently confirmed counts.
The method indicates which hardware initialized the weights; it does not establish where the GPUs were, who owned them or how they were accessed. The articles note that Chinese companies can rent computing capacity at data centers outside China. The findings therefore do not by themselves show that export controls were violated.
Written from coverage as of
Images
- The Decoder
- TechCrunch
- Tom's Hardware
- Crypto Briefing
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh