Urgent.News

What's breaking now, across thousands of outlets.

AI

Chinese AI models reveal clues pointing to restricted Blackwell chips

The clues are hidden in the random numbers that form an AI model’s starting point—and hint at a new way to trace its origins

Recent reports by independent researcher Will DePue suggest that some of China's top AI models may have been trained on Nvidia's restricted Blackwell chips, which are banned from export to China. DePue discovered clues hidden in the random numbers that form an AI model's starting point, which may provide a new way to trace the origin of these models.

The six models from four Chinese AI companies, including DeepSeek, Moonshot AI, MiniMax, and Ant Group, appear to have originated on Nvidia's Blackwell chips. Prior to Blackwell, these models were built on Nvidia's Hopper generation chips, which Chinese companies have been allowed to purchase at times. DePue developed a forensic technique based on the unique signatures left behind by different processors.

By analyzing the billions of numbers that determine how AI models respond to prompts, DePue found that the weights in these models were initialized with a specific seed that depended on the chip used. Since the initialization is not truly random, the same seed generates different numbers on different hardware. DePue believes that further analysis could reveal more about the origin of these models.

If his findings are accurate, it would confirm that the U.S. hardware export restrictions have a side door, allowing some of China's top labs to access the best hardware.

Written by urgent.news from Scientific American's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

This story

This is one outlet's version. Read the fullest account.

Every outlet that reported this, and when →

Read the original at scientificamerican.com →

More in AI

More from Thursday 19 November →