RC RANDOM CHAOS

Forensic Analysis Reveals Chinese AI Labs Training on Blackwell GPUs

· via Hacker News

Original source

Chinese frontier AI models carry the fingerprint of export-controlled GPUs

Hacker News →

Researchers analyzing published model weights have identified hardware fingerprints showing that four major Chinese AI companies—DeepSeek, Moonshot AI, MiniMax, and Ant Group—are training frontier models on NVIDIA’s Blackwell GPUs, chips subject to U.S. export restrictions to China. The forensic method examines initialization patterns in published weights to determine which processor architecture generated them, then validates findings against known hardware configurations. Six models released since April 2026 match Blackwell specifications, while earlier releases from the same companies match the older Hopper generation.

The statistical evidence is strong: matches range from 49 to 1,385 sigma above random chance, and the method correctly predicted hardware in 24 cases where developers had publicly disclosed it. The two largest training runs are estimated to have required 3,000 to 4,500 Blackwell GPUs each running for two to three months, suggesting total cluster costs of $200 million to $300 million for individual models and over half a billion dollars across all five training runs. However, the analysis establishes only that these companies possess access to the hardware—it does not determine where chips are located, how access was obtained, or whether export controls were actually circumvented.

Adoption remains uneven: established Chinese labs including Alibaba (Qwen), Zhipu (GLM), and Tencent continue releasing models trained on Hopper-generation hardware. One company, Meituan, appears to rely entirely on custom AI ASICs rather than NVIDIA chips. The findings provide the first technical evidence confirming earlier U.S. government and press allegations about Blackwell access among Chinese AI developers.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.