Blackwell Chips Found in 6 Chinese AI Models

Bold gradient illustration of a computer chip with a fingerprint, with the words "BLACKWELL FINGERPRINT FOUND."

Every AI model carries a hidden signature of the hardware that created it — and on 12 November 2026, a forensic report said that signature points to Nvidia's most advanced chips inside six of China's leading AI models. The chips are Blackwell GPUs, which US export controls ban from sale to China. The author calls the evidence conclusive, but says it does not prove any law was broken.

Here is how the fingerprint works, what was found, and where the limits are.

How a GPU leaves a fingerprint

When engineers start training a model, its starting weights are generated from a random seed. A pseudorandom generator called Philox produces those numbers, split across the GPU's parallel processor cores, known as streaming multiprocessors, or SMs.

Different chips have different SM counts: Hopper 132, Blackwell B200/B300 148, Grace Blackwell GB200/GB300 152. Because the seed and chip layout together decide the pattern of starting weights, each combination leaves a distinct trace — a fingerprint of the hardware used at the start.

Traces of that pattern survive training: enough of the starting pattern remains in the published model for a forensic check to find.

What the report found

The report, "Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled Nvidia Blackwell GPUs", was published on 12 November by Will DePue, a former OpenAI researcher who now runs Unemployment Labs. His evidence comes entirely from the models' published weights, not from leaks or insiders.

Six open-weight models released since April 2026 carry Blackwell fingerprints: DeepSeek-V4-Flash and MiniMax-M3 match B200/B300 chips (148 SMs); DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant Group's Ling-3.1-flash match Grace Blackwell systems (152 SMs).

The shift is abrupt: the same labs' earlier models match Hopper. Ant Group's Ling-3.0-flash, released in August, reads as Hopper; Ling-3.1-flash, two months later, matches Grace Blackwell. The newest readable models from Alibaba, Zhipu, Tencent, Xiaomi and StepFun still read as Hopper — the pattern is specific to these six.

Why the author calls it conclusive

DePue searched roughly six trillion combinations of chip layout and seed number, testing each match against held-back weights. A chance match should reach only about 8 sigma; the real matches ranged from 49 to 1,385 sigma.

Then he checked on real hardware. With the recovered seed, a rented Blackwell B300 reproduced 461 of 469 selected rows of DeepSeek-V4-Flash's starting weights, 445 bit for bit. An H100 reproduced none. A rented GB200 reproduced Kimi K3's starting values to a correlation of 0.999995.

The method was also tested against models whose hardware is public: for all 24 open-weight models whose developers named their chips, including Meta's Llama 3.1, it agreed in every case. Senior researchers at OpenAI and Anthropic independently verified the results, according to the report.

What it does not prove

The fingerprint identifies the hardware used at initialization. It supports, without proving, that Blackwell hardware also did the training. It does not establish where the chips were, how access was obtained, or whether export controls were violated.

The method also has limits. It needs published weights, so it applies only to open-weight releases. And DeepSeek-V4.1-Flash showed no fingerprint despite an exhaustive search — a null result, not proof of other hardware.

Blackwell chips need a US licence for export to China, and none has been reported. But Chinese companies can legally rent Blackwell hardware at data centres outside China. The only other hardware that could produce the pattern, the authors say, is specially configured AMD hardware, also export-restricted; they excluded it for V4-Flash and consider it very unlikely for the other five.

Why it matters

US officials and press reports had already alleged that DeepSeek and Moonshot AI were using Blackwell chips, but without publishing technical evidence. This report is the first forensic, weight-based evidence — reproducible maths on public weights, with the seed values published so others can check.

The authors estimate each of five final runs needed roughly 500 to 3,000 Blackwell GPUs for two to three months, costing more than $500 million combined — estimates, not procurement records.

Key takeaways

  • A forensic report published 12 November 2026 says six open-weight models from DeepSeek, Moonshot AI, MiniMax and Ant Group carry the fingerprint of Nvidia Blackwell GPUs.
  • The fingerprint comes from the random seed and GPU layout used to generate a model's starting weights — Hopper has 132 SMs, B200/B300 have 148, Grace Blackwell has 152.
  • The matches ran from 49 to 1,385 sigma against about 8 expected by chance, and a rented B300 reproduced DeepSeek-V4-Flash's starting weights almost exactly.
  • The evidence shows access to Blackwell-class hardware at initialization — not proof of smuggling, location, or a broken law. Renting Blackwell hardware abroad is legal under current rules.
  • DeepSeek-V4.1-Flash showed no fingerprint despite an exhaustive search — a null result, not proof of other hardware.

Conclusion

The clever part of this report is that it does not rely on a leak or a source. It relies on maths that anyone with the published weights can repeat. That is also its limit: it shows where the starting numbers came from, not how the chips got there or whether any rule was broken.

FAQ

How can a GPU leave a fingerprint in an AI model?

A model's starting weights are generated from a random seed and spread across a GPU's processor cores. Different chips have different core counts — Hopper 132, Blackwell B200/B300 148, Grace Blackwell 152 — so each combination leaves a distinct pattern.

Does the report prove China broke US export controls?

No. The report states its results are not a finding that any law was broken. They show access to Blackwell-class hardware at initialisation, but not where the chips were, how access was obtained, or whether export controls were violated.

Which models were found to carry the fingerprint?

Six: DeepSeek-V4-Flash and DeepSeek-V4-Pro, MiniMax-M3, Kimi K3 and Kimi K3.1-Pro, and Ant Group's Ling-3.1-flash — all released since April 2026. Earlier models from the same labs match Hopper.

Researched with AI assistance and reviewed for accuracy.

Sources

Follow on WhatsApp

Follow Daily Tech & AI on WhatsApp for one tech and AI update every day: Daily Tech & AI WhatsApp channel.

Comments