Six Chinese AI models match NVIDIA Blackwell GPU fingerprints
Original title: Chinese frontier AI models carry the fingerprint of export-controlled GPUs
The analysis estimates each run needed roughly 500 to 3,000 Blackwell GPUs for two to three months.
A forensic study of published weights says six models from DeepSeek, Moonshot AI, MiniMax and Ant Group match NVIDIA Blackwell hardware, while earlier models from those labs match Hopper. DeepSeek-V4-Flash and MiniMax-M3 match B200 or B300, and four others match GB200 or GB300. The authors say the fingerprint shows how weights were initialized and supports, without proving, Blackwell training; they confirmed two matches on physical Blackwell chips. They do not establish chip location, access method or an export-control violation. Newer readable models from several other Chinese labs still match Hopper.
Unconfirmed: The authors say specially configured AMD hardware is a possible alternative and was excluded only for DeepSeek-V4-Flash.
- Reported evidence ranges from 49 to 1,385 sigma for the six models.
- The method agreed with public hardware disclosures for 24 models.
- DeepSeek-V3's estimate of 2,000 to 2,500 chips matches DeepSeek's disclosed 2,048.
- Alibaba, Zhipu, Tencent, Xiaomi and StepFun models examined still read as Hopper.
- Meituan's LongCat-2.0 showed no NVIDIA fingerprint.
- Vendors
- NVIDIA
- Organizations
- DeepSeekMoonshot AIMiniMaxAnt Group
- Countries
- ChinaUnited States
From the original article
14 min read at unemploymentlabs.com
Our forensic analysis of publicly released AI models finds evidence that DeepSeek, Moonshot AI, MiniMax and Ant Group have gained access to NVIDIA’s Blackwell GPUs, despite restrictions on exporting these chips to China, and are using that hardware to create their AI models.
By firloop · unemploymentlabs.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.