Analysis finds signs of Blackwell chips in 6 Chinese AI models

A forensic analysis of the weights of models says 4 Chinese AI labs used NVIDIA Blackwell chips to make 6 models since April 2026. The labs include DeepSeek and Moonshot AI. The US restricts the export of these chips to China. Their models before April 2026 agree with the Hopper generation of chips. The analysis does not show where the chips were or how the labs got them. It does not show if an export rule was broken.

How it works

Before training, the weights of a model, its adjustable numbers, must have start values. Software makes them with a random number method, and the layout of the chips changes the result. The analysis recovers the start values and finds the layout. For 2 models, it tested the same result on Blackwell chips.

The numbers

For 24 models where the developer named the hardware, the method agreed with all of them. Sigma measures how far a result is above random variation. Random variation gives about 8 sigma at most. The 6 models show 49 to 1,385.

What is not known

The analysis says its result supports training on Blackwell chips, but it shows only how the weights were made. Special AMD hardware is the only alternative. Other labs, such as Alibaba and Zhipu, use Hopper in their new models.

Insight: Chips are controlled going in, and models going out

The analysis says 6 models agree with chips that the US restricts. In October 2026, China drafted export controls on 23 AI technologies. They cover model weights, not API access, which is use of a model through software. Each side controls a different part, and the draft has no controls on API use, which reads as an open door for buyers of Chinese models.