What changed
Researchers from UnemploymentLabs have identified a strong correlation between the training of several open-weight Chinese AI models and the use of Nvidia Blackwell chips.
The analysis, based on statistical examination of model weights, suggests that models such as DeepSeek V4, Kimi K3.1, MiniMax-M3, and Ling-3.1-Flash may have been trained using Blackwell hardware.
The study notes that while the math checks out, it does not confirm the exact location or method of access to the chips.
What this means in practice
The findings highlight the potential use of Nvidia Blackwell chips in training Chinese AI models, despite U.S. export controls. The researchers emphasize that their method relies on models with open weights and the specific hardware and software stacks used during training.
They simulated the initialization process using the same chips, software, and thread counts, and found that the statistical distribution of weight values matched those of the Chinese models.
Changing the SM and thread counts disrupted the correlation, allowing them to identify the chip family used.
Limitations and availability
The analysis is limited to models with open weights and depends on both hardware and software stacks. It cannot determine the exact variant of Blackwell chips used or the number of cards involved.
Additionally, the study ruled out the possibility of AMD cards like the MI300X producing the same weight value patterns. The researchers also note that the method does not account for VRAM capacity or memory bandwidth, which distinguish individual cards.
The findings do not prove that export controls were violated, as it remains legal for Chinese companies to rent Blackwell hardware abroad.
