What changed
A new research effort claims that the leading open-weight AI models developed by Chinese companies were trained on Nvidia chips that the U.S. government has restricted from shipping to China. The analysis of model weights by Will DePue of Unemployment Labs found evidence that models like Moonshot’s Kimi K3 and DeepSeek’s V4 were trained on Blackwell GPUs, which are banned for export to China.
What this means in practice
The findings suggest that despite U.S. export controls, Chinese AI labs have accessed powerful Nvidia chips, possibly through offshore data centers. This has implications for the U.S. strategy to maintain its technological edge against China. The use of these chips could have contributed to the rapid advancement of AI capabilities in China, even as export bans are intended to prevent such access.
Limitations and availability
The supplied announcement does not specify whether the models were entirely trained on Nvidia chips or if other hardware was used. It also does not clarify the exact extent of the chip smuggling or the role of Megaspeed and its subsidiary Speedmatrix in facilitating this. The U.S. export controls are designed to prevent Chinese access to the most powerful chips, but the use of offshore data centers has allowed Chinese firms to bypass these restrictions. The situation highlights the challenges in enforcing export controls and the potential impact on Nvidia’s business if smuggling is confirmed.
