The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard
Minor World

Nvidia Blackwell chips fingerprinted in Chinese AI models

By The Daily Byte ·14 Nov 2026 ·2 min read
Close-up of an Nvidia Blackwell GPU package showing the Nvidia logo on the central chip

Six of China’s leading AI models carry the numerical fingerprint of Nvidia Blackwell chips, which need a US licence for export to China, a forensic analysis published on 12 November finds. The report, by former OpenAI researcher Will DePue of the non-profit Unemployment Labs, covers DeepSeek-V4-Flash and DeepSeek-V4-Pro, Moonshot AI’s Kimi K3 and Kimi K3.1-Pro, MiniMax-M3 and Ant Group’s Ling-3.1-flash, all released since April 2026.

How the fingerprint works

Before training, a model’s weights are given starting values by a random-number generator in PyTorch. Where each number lands depends on the chip’s processing units, or streaming multiprocessors: 132 on Nvidia’s older Hopper chips, 148 on Blackwell B200 and B300 chips, and 152 on the larger GB200 and GB300. Traces of that arrangement survive in the published weights. The matches run from 49 to 1,385 standard deviations above chance, where about eight would be expected, and two were confirmed on physical Blackwell hardware. It also matched the hardware in all 24 cases where a developer had disclosed it.

What the report does not show

The author is careful about the limits. The report says the evidence shows access to advanced computing hardware, but does not establish where the chips were, how access was obtained, or whether export controls were violated. The chips could have been rented abroad. Earlier models from the same labs, and the newest from Alibaba, Zhipu, Tencent, Xiaomi and StepFun, still read as Hopper.

The scale of the training runs

The report estimates the five largest training runs together needed about 12,000 Blackwell GPUs: about 4,500 for Kimi K3 and 3,000 for DeepSeek-V4-Pro. At roughly $70,000 per chip, the largest cluster would cost $200–300 million. White House technology adviser Michael Kratsios said earlier this year that Moonshot had used export-banned chips.

Sources: Unemployment Labs and TechCrunch.