Forensic analysis by Uemployment Labs claims some Chinese AI labs accessed NVIDIA's advanced Blackwell GPUs. The report examined five models, including DeepSeek-V4-Pro, Kimi K3 and Ling-3.1-flash, using GPU core layouts to distinguish Blackwell from Hopper. Researchers reverse engineered model weights to match hardware fingerprints linked to Blackwell chips.
The lab recovered thread counts for processors that initialized training runs. Data suggests individual runs used clusters of 3,000 to 4,500 Blackwell GPUs, with combined footprints reaching 12,000 chips. Researchers analyzed published weights to determine the seed generating initial weights, then compared static weight patterns against streaming multiprocessor layouts of Hopper and Blackwell GPUs.
The report estimates 7,000 to 22,000 Blackwell GPUs were used cumulatively by DeepSeek, Moonshot, MiniMax and Inclusion AI. Blackwell GPUs are export-controlled and cannot be sold to China, though remote access via GPU clusters is possible. The evidence does not evaluate whether export control regulations were violated, but comes as US legislation seeks to target remote access.
