Study Finds Chinese AI Models May Have Been Trained on Restricted Blackwell Chips
A study found technical traces of Nvidia Blackwell chip use in six open-weight AI models released by Chinese companies including DeepSeek and Moonshot AI. The findings do not establish that the models relied entirely on those chips or that the companies violated US export controls.
Nvidia’s Blackwell GPUs, which the US restricts from export to China, may have been used to train some Chinese open-weight AI models. In a report, Will DePue, a former OpenAI researcher who now leads the nonprofit research group Unemployment Labs, said six models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group show technical traces consistent with a Blackwell training environment.

File photo; the original image is credited to Reuters.(VietVance)
The research team inferred the training hardware from characteristics of the models’ weights. According to the report, DeepSeek-V4-Flash and MiniMax-M3 show traces of Blackwell B200 or B300 chips, while DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-Flash show characteristics of GB200 or GB300 chips. The researchers estimated that Kimi K3 and DeepSeek-V4-Pro would each require thousands of Blackwell GPUs running continuously for about two to three months.
The report said some models were trained using a combination of Hopper and Blackwell chips. The technical traces do not prove that the models relied entirely on Blackwell chips, nor are they sufficient to establish that the companies violated US export controls. Chinese developers may use computing capacity remotely through data centers in countries such as Malaysia, without bringing the chips directly into China. Malaysia is among the markets permitted to import Blackwell chips.
Source: VietVance


