Forensic analysis conducted by Unemployment Labs suggests that several Chinese AI models were trained using thousands of NVIDIA Blackwell GPUs. By reverse engineering the published weights of models from developers including DeepSeek, Moonshot, MiniMax, and Inclusion AI, researchers identified hardware imprints that match the specific architecture of NVIDIA's latest Blackwell and Hopper chips.
The study found that individual training runs utilized clusters ranging from 3,000 to 4,500 Blackwell GPUs, with combined footprints across participating labs reaching upwards of 12,000 chips. When accounting for models such as DeepSeek-V4-Flash, DeepSeek-V4-Pro, MiniMax-M3, Kimi K3, and Ling-3.1-flash, the researchers estimated a cumulative usage of between 7,000 and 22,000 Blackwell GPUs.
Researchers identified the hardware by analyzing the initialization of model weights. Because each GPU architecture has a fixed number of processing cores and a specific streaming multiprocessor (SM) layout, the hardware dictates a unique grid pattern when placing initial random numbers into a model. By comparing these patterns against the SM layouts of NVIDIA's Hopper and Blackwell series, the lab was able to identify the specific chip families used during the training initialization phase.
While the findings indicate access to advanced computing hardware, the report noted that the data does not establish whether export control regulations were violated. NVIDIA's Blackwell GPUs are subject to U.S. export controls that prohibit sales to China, though the report clarified that Chinese labs can often access such hardware remotely via GPU clusters located abroad.
"Notably, our findings provide technical evidence of access to advanced computing hardware, but do not establish where the chips were located, how access was obtained, or whether export controls were violated," the Unemployment Labs report stated.
The analysis was limited to models with open weights. The researchers noted that some models, such as Alibaba’s Qwen3, left no trace because they were initialized on ordinary processors rather than GPUs, which prevents the creation of a hardware fingerprint.
