Chinese AI labs are training models on Nvidia's Blackwell despite export controls, says ex-OpenAI engineer - The Register
Chinese AI companies are training models on Nvidia's restricted Blackwell GPUs despite US export controls, according to a former OpenAI engineer.
US export controls are not stopping AI labs in China from using restricted Nvidia chips, according to a report from a former OpenAI engineer. Will DePue published forensic evidence showing that "DeepSeek, Moonshot AI, MiniMax and Ant Group have gained access to Nvidia's Blackwell GPUs, despite restrictions on exporting these chips to China, and are using that hardware to create their AI models."
Security concerns
The findings come amid heightened security concerns surrounding Nvidia and its CEO. Jensen Huang resigned following revelations of his relationship with Chinese businesswoman Alice Huang, who founded Singapore-based Megaspeed International in 2023. The company is reportedly under investigation by US authorities for allegedly smuggling Nvidia chips to China.
Huang's departure coincides with a renewed push to tighten US chip export controls. House Speaker Mike Johnson supports the idea, and there are at least four bills awaiting consideration that aim to limit China's access to leading edge AI tech: the AI Overwatch Act, the Chip Security Act, Match Act, and the Cloud Security Act.
US Senators Jim Banks and Elizabeth Warren cited Jensen Huang's "relationship with a suspected Chinese spy" as justification for passing the AI Overwatch Act, which would codify existing chip export controls. The Trump administration has stated it's not sure whether Alice Huang was a spy, an allegation denied by Beijing and currently under investigation by the FBI.
Evidence of chip usage
Only a few weeks ago, Lava, an Israel-based security company, said it had found "32 Nvidia B300 GPUs, part of Nvidia's Blackwell Ultra generation, through publicly exposed monitoring endpoints with IP addresses geolocated to China." These GPUs are subject to US export controls.
DePue's analysis took a different approach, examining the published weights of various models from Chinese AI labs. Model weights must be initialized with some value, often using a pseudorandom number generation algorithm called Philox. Because this process is not truly random, it's possible to select a chip and seed value, generate initial weights, and compare that pattern to a model's published weights.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models," said DePue. "In doing so, I found conclusive evidence that six Chinese AI models – including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3 – were initialized on Nvidia Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in Nvidia's flagship liquid-cooled rack-scale systems."
Policy implications
Noting that this is the first time MiniMax and Ant Group have been publicly linked to usage of Nvidia's Blackwell, DePue also cited several limitations to his approach. One is that usage of Blackwell GPUs doesn't prove where the GPUs were located, so it's possible some of these Chinese labs used this hardware through US providers – a loophole the Cloud Security Act aims to close.
The findings come as OpenAI and Anthropic have each complained about Chinese labs copying their models through distillation, a practice also noted by the US Cybersecurity and Infrastructure Security Agency. With multiple bills pending in Congress, the US government appears poised to take more aggressive measures to prevent China from accessing advanced AI technology.
Related
