Chinese frontier AI models carry the fingerprint of export-controlled GPUs (unemploymentlabs.com)

Recent forensic analysis of AI models from leading Chinese firms DeepSeek, Moonshot AI, MiniMax, and Ant Group has uncovered evidence that these companies are using NVIDIA’s advanced Blackwell GPUs for model training, circumventing export restrictions on this high-performance hardware. Specifically, six models released since April 2026 exhibit numerical patterns that strongly correlate with Blackwell chips, indicating a shift from earlier models trained on older Hopper generation GPUs. This suggests that Chinese AI developers are gaining modern computing capabilities, intensifying the competitive landscape in AI development, especially amid ongoing geopolitical tensions regarding technology exports. The significance of these findings lies in the potential breach of U.S. export controls, which restrict the sale of these cutting-edge GPUs to China, aiming to limit that country’s advancements in AI. The analysis, based on model weights and the seed and thread configurations used during training, shows that substantial resources (estimated between 500 and 3,000 Blackwell GPUs for each model) are being employed, invoking concerns over how these companies may have acquired the hardware. With training costs potentially reaching hundreds of millions of dollars, this development could shift the balance of power in the global AI race, as Chinese companies leverage these powerful resources to enhance their AI capabilities.