Researchers claim Chinese AI models used banned Nvidia Blackwell chips
Independent researchers allege that DeepSeek and other Chinese artificial intelligence models were trained using Nvidia Blackwell GPUs despite U.S. export controls.
New fingerprinting technique identifies hardware
UnemploymentLabs, an independent research outfit, has published evidence linking specific chips to open-weight models. The team analyzed DeepSeek V4, Kimi K3.1, MiniMax-M3, and Ling-3.1-Flash for training hardware origins. Scientists used a statistical method based on shader module counts to match the models to Nvidia silicon. This approach correlates weight initialization patterns with specific hardware configurations without revealing physical locations.
Blackwell chips have unique counts
Nvidia Hopper chips contain 132 shader modules while Blackwell B series units hold 148. The newer Blackwell GB variants feature 152 shader modules for distinct identification. Researchers initialized model weights using cuRAND software with specific seed values to generate consistent adjustments. Changing the thread count per shader module from 2,048 broke the statistical correlation between test data and models.
Legal status remains unclear
The study notes that renting Blackwell hardware abroad is currently not illegal for Chinese companies. Recent deals between Tencent and Oracle show U.S. firms are not strictly prohibited from such transactions. The analysis cannot prove export controls were broken or identify where the chips were accessed. Researchers also could not determine VRAM capacity or memory bandwidth characteristics of the specific cards used.
Reported by one outlet
Only one outlet has published this. Nothing here has been checked against a second report, so read it as that outlet's account and follow the link below for the original.
Reported by
1 independent outlet. Headline as published. Links open the original report.