The method exploits the weight initialization step. Random initialization is split across the chip's streaming multiprocessors (SMs), and each chip family has a different count: 132 on Hopper, 148 on Blackwell B and 152 on Blackwell GB. The researchers recreated the process with the same chips and software and found the resulting weight distribution precisely matched that of the Chinese models. When they changed the SM count, the match disappeared. This let them identify the chip family for each model, as well as an approximate number of cards.
Models examined include DeepSeek V4, Moonshot's Kimi K3.1, MiniMax-M3 and Ant Group's Ling-3.1-Flash. The method works only on open-weight models and does not show whether training ran entirely on Nvidia chips. The chips may have been in a data center outside China, for example in Malaysia; according to Tom's Hardware, it is currently not illegal for Chinese companies to rent Blackwell hardware abroad.
The research comes a day after Jensen Huang resigned as Nvidia CEO amid a scandal tied to a company accused of smuggling chips into China. Epoch AI estimates that about 2 million advanced chips entered China in 2024 and 2025 through schemes that hid their final destination.