Study: Chinese AI models show signs of access to export-controlled Nvidia chips
A forensic analysis of published model weights indicates that four Chinese AI companies began building their newest models on Nvidia Blackwell GPUs, which are restricted from export to China. The study does not say where the chips are located or whether export rules were broken.
A report published online, based on technical analysis of publicly released artificial intelligence (AI) models, concludes that the Chinese companies DeepSeek, Moonshot AI, MiniMax and Ant Group have gained access to Blackwell graphics processors (GPUs) made by the U.S. firm Nvidia. Exporting these chips to China requires a U.S. license, and the report says none have been reported. The finding suggests that several leading Chinese developers may be training their newest models on this hardware.
How it was determined
The analysis relies solely on the models' own published weights, not on leaks or insider accounts. Weights are the billions of numbers that determine how a model processes information. Before training, they are given starting values by a pseudorandom number generator called Philox. How those values are laid out depends on the number of processing blocks on the specific chip, so the layout reveals what kind of chip created the weights. The researchers say they also confirmed the match on physical Blackwell chips for two models, DeepSeek-V4-Flash and Kimi K3.
Results
Six models released since April 2026 match Blackwell, while earlier models from the same companies match the older Hopper generation. They are DeepSeek-V4-Flash and V4-Pro, Kimi K3 and K3.1-Pro, MiniMax-M3 and Ant Group's Ling-3.1-flash. Not every Chinese company has switched: the newest models from Alibaba (Qwen), Zhipu (GLM), Tencent, Xiaomi and StepFun match Hopper. Meituan says its LongCat-2.0 was trained on other AI-specific processors, and no Nvidia fingerprint was found in it.
Based on model size, the authors estimate each run required roughly 1,500 to 3,500 Blackwell GPUs for one and a half to three months, about 12,000 across the five models. That could cost more than half a billion dollars, though the estimates are rough.
Limitations
The study provides technical evidence of access, but does not establish where the chips were located, how access was obtained, or whether export controls were violated. The only alternative the authors identified is specially configured AMD hardware, which is also export-restricted; they rule it out for DeepSeek-V4-Flash and consider it very unlikely for the other five.
