How did Unemployment Labs identify the chips?

It counted the streaming multiprocessors (SM) reflected in the initial parameters created during PyTorch training, which lets it infer which AI chip was used with high confidence.

Does this prove Chinese companies used NVIDIA Blackwell chips?

No. Unemployment Labs says the method is inferential and does not definitively prove use, and the chips may not have been located inside China.

Why wasn't Baidu's ERNIE-5.5 analyzed?

It was built with PaddlePaddle rather than PyTorch, and the method cannot analyze models developed with other frameworks.

Large Language ModelsAI Regulation & PolicyGIGAZINE AI1d agoNov 13, 2026, 22:01 JST

Unemployment Labs: Chinese AI models show NVIDIA Blackwell traces

Send on LINE

3 Key Points

What happened
Unemployment Labs analyzed Chinese AI models by counting the streaming multiprocessors (SM) of the chips used to create their initial parameters in PyTorch. It found DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash used NVIDIA Blackwell chips during the export-control period.
Why it matters
The group cautions that the method is inferential and does not prove use, nor that the chips were located inside the country.

WHO IT HITSPolicymakers and export-control officials who monitor chip flows may need to weigh this forensic evidence. AI developers and cloud providers relying on PyTorch may also see their chip choices become traceable.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Get these free every morning

Context & Analysis

The finding arrives against a backdrop of loosening restrictions: the U.S. government approved sales to Chinese firms in May 2026, and China was reported to be moving toward approving imports. Unemployment Labs' analysis points to the earlier period of tight controls, suggesting American chips were in use then.

The method relies on a quirk of PyTorch: when initial parameters are created, they leave a fingerprint tied to the number of streaming multiprocessors on the chip. By counting those, the group says it can identify the chip with high confidence. It applied this to a list of models spanning DeepSeek, Moonshot AI, MiniMax and Ant Group, reading a shift from Hopper-generation to Blackwell-generation chips.

Separately, reports on November 12, 2026 said a company tied to Alice Huang, whose relationship with Jensen Huang became public, had promised to deliver Blackwell-generation chips to DeepSeek. Unemployment Labs stresses its method cannot analyze models built on other frameworks, such as Baidu's ERNIE-5.5 on PaddlePaddle, and that chip location is not established.

FAQ

How did Unemployment Labs identify the chips?: It counted the streaming multiprocessors (SM) reflected in the initial parameters created during PyTorch training, which lets it infer which AI chip was used with high confidence.
Does this prove Chinese companies used NVIDIA Blackwell chips?: No. Unemployment Labs says the method is inferential and does not definitively prove use, and the chips may not have been located inside China.
Why wasn't Baidu's ERNIE-5.5 analyzed?: It was built with PaddlePaddle rather than PyTorch, and the method cannot analyze models developed with other frameworks.

GIGAZINE AIRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.