Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

Report says leading Chinese open models carry traces of export-banned Nvidia Blackwell training

A new analysis of open-weight models claims that frontier releases from Chinese labs such as DeepSeek and Moonshot carry training fingerprints from Nvidia Blackwell chips, which are barred from export to China. The report stresses that this does not prove a violation, since the hardware could sit in offshore data centers, but it pushes a growing enforcement gap into the open.

Published
研究称中国头部开源模型留下 Blackwell 训练痕迹,出口管制再受质疑
Image source: nvidia.com

A new analysis of open-weight models has put Washington's chip export controls back in the spotlight, with reports suggesting that frontier models released publicly by Chinese labs may have been trained on Nvidia Blackwell chips that are barred from export to China.

The work is attributed to Will DePue, a former OpenAI researcher who leads the nonprofit research group Unemployment Labs. According to coverage of the finding, he examined the weights of openly released models from labs including DeepSeek, Moonshot AI, MiniMax and Ant Group, looking for traces of the hardware used to train them.

The method is the heart of the story. Model weights retain configuration signatures of the accelerators used in training, much like a fingerprint left on the finished product, so the training cluster can be inferred even when a lab never discloses its hardware. DePue says the most closely watched models in the set carry Blackwell fingerprints.

The models named in coverage include Moonshot's Kimi line and DeepSeek's latest open releases — precisely the systems that have been cited as evidence that Chinese labs can close the gap despite restricted compute.

The policy context is unambiguous. Blackwell is a focal point of U.S. export controls and is banned from sale to China and Chinese-owned firms; the controls are designed to slow Chinese frontier work by keeping the most advanced compute out of reach.

The reporting is careful about limits. A fingerprint in the weights does not mean a model was trained entirely on Nvidia hardware, and it does not establish that any law was broken, because the chips could sit in an offshore data center. That is the loophole critics keep raising: the rules govern who owns hardware and where it is installed, not who uses it.

That is why the implication goes beyond any single lab. If advanced compute can be reached reliably through cross-border cloud rental and overseas facilities, an export regime drawn around physical location will keep leaking.

Three things to watch: whether the analysis is independently reproduced, whether Nvidia and regulators respond on the record, and whether the next round of rules reaches offshore compute rental and cloud access.

Why it matters

If the fingerprinting holds up under independent review, the debate shifts from whether a rule was broken to whether the rule works at all, putting offshore data centers and cross-border cloud rental at the center of the next round of controls.

NvidiaAI芯片出口管制
Back to realtime news

Nearby Updates

All