Analysis of open weights: six models from DeepSeek, Moonshot, MiniMax, and Ant Group released since April carry fingerprints of export-controlled Blackwell GPUs
First reported by Unemploymentlabs ·
Your AI model's performance may be capped if its underlying hardware differs from the latest generation used by leading Chinese competitors.
A forensic analysis of publicly released AI models has revealed that six advanced models from Chinese companies DeepSeek, Moonshot AI, MiniMax, and Ant Group, launched since April 2026, exhibit numerical patterns consistent with NVIDIA's export-controlled Blackwell GPUs. These patterns differ from the hardware fingerprints found in earlier models from the same companies, which align with NVIDIA's older Hopper generation. The investigation, based on the models' published weights, suggests these leading Chinese AI developers are utilizing Blackwell hardware for training their latest AI systems. While the analysis provides technical evidence of access to advanced computing hardware, it does not pinpoint the location of the chips, the method of acquisition, or confirm any violations of export controls. The use of Blackwell GPUs, estimated to cost hundreds of millions of dollars for training runs, highlights a significant investment in cutting-edge hardware for AI development.
The findings strongly suggest that Chinese AI developers have circumvented U.S. export controls to acquire and utilize NVIDIA's Blackwell GPUs. This hardware is crucial for training state-of-the-art AI models, and its restricted access to China indicates a significant technological advantage for those who possess it. The ability to replicate the Blackwell fingerprint in models from multiple firms implies a coordinated or widespread acquisition of this restricted technology.
This development has critical implications for the global AI landscape, potentially shifting the competitive balance and raising concerns about the effectiveness of U.S. technology export policies. Companies unable to access such advanced hardware may find their models falling behind in performance and capability, impacting their market position and innovation potential.
AI-written summary. May contain errors.