Six Chinese AI Models Carry the Fingerprint of Export-Restricted Nvidia Blackwell GPUs, Report Finds

A forensic analysis published on 12 November 2026 says six frontier AI models from China's leading labs carry the fingerprint of Nvidia's export-restricted Blackwell GPUs in their published weights. The report, by former OpenAI researcher Will DePue at Unemployment Labs, names 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. Blackwell chips require a US licence for export to China, and the report says no such licences have been reported.

Checkpoint forensics is the method behind the claim. A checkpoint is a saved copy of a model's weights — the billions of numbers that make up a trained model. When training begins, those weights are filled with pseudo-random starting values, generated by an algorithm called Philox through Nvidia's cuRAND library, with the work split across the graphics card's streaming multiprocessors, or SMs. Change the number of SMs and the same starting seed lands the same numbers in different positions. The pattern left behind is, in effect, a fingerprint of the hardware that created the starting weights.

How the fingerprint was found

The statistical strength the report claims is striking. DePue's team searched more than six trillion combinations of chip layout and seed number, matching patterns in the models' published open weights — evidence from weights the companies themselves published, with no leaks or insider accounts. Across that search, a chance match should reach only about 8 sigma — how far a result sits from random noise. The real matches ranged from 49 to 1,385 sigma, and even the weakest individual hold-out test reached 20 sigma.

The team also confirmed the matches on physical hardware. Running unmodified PyTorch with the recovered seed, a Blackwell B300 reproduced 461 of 469 selected rows in DeepSeek-V4-Flash, 445 of them bit for bit, while an H100 reproduced none — and changing the seed by one broke the match. A rented Grace Blackwell GB200 reproduced Kimi K3's starting values to a correlation of 0.999995.

The six models — and the clean switch from Hopper

ModelLabReleasedFingerprint read
DeepSeek-V4-FlashDeepSeekApril 2026Blackwell B200/B300 (148 SM)
DeepSeek-V4-ProDeepSeekApril 2026Grace Blackwell GB200/GB300 (152 SM)
MiniMax-M3MiniMaxJune 2026Blackwell B200/B300 (148 SM)
Kimi K3Moonshot AIJuly 2026Grace Blackwell GB200/GB300 (152 SM)
Ling-3.1-flashAnt GroupOctober 2026Grace Blackwell GB200/GB300 (152 SM)
Kimi K3.1-ProMoonshot AINovember 2026Grace Blackwell GB200/GB300 (152 SM)

What makes the table persuasive is the pattern across it. Every one of these labs' earlier models reads as Hopper: DeepSeek-V3 (December 2024), Kimi K2 and K2.5, MiniMax-M2.7 (April 2026) and Ling-3.0-flash (August 2026). Then, within months, each lab's newest model flips to Blackwell. One detail: Kimi K3.1-Pro starts from the same seed as K3 and may be a K3 checkpoint with further post-training.

Rows of Nvidia graphics processing units in a data centre server rack, green indicator lights glowing
AI-generated illustration (generated 14 Nov 2026)


The report also draws a line the other way. The newest checked models from Alibaba (Qwen), Zhipu (GLM), Tencent, Xiaomi and StepFun all still read as Hopper. So this is not a claim that every Chinese model runs on Blackwell — it is a claim about four specific labs, and the first time MiniMax and Ant Group have been publicly linked to Blackwell use. US officials and press reports had already alleged that DeepSeek and Moonshot were using Blackwell chips, but without publishing technical evidence. The report also estimates the training clusters at roughly 500 to 3,000 Blackwell GPUs per model — rough, order-of-magnitude figures, not audited numbers.

What the report does NOT prove

First, location. A fingerprint identifies the hardware that created a model's starting weights. It says nothing about where that hardware was sitting. The chips could have been accessed through cloud providers outside China, which would leave exactly the same fingerprint. A pending bill, the Cloud Security Act, is written to close that remote-access route — itself a sign lawmakers already know the route exists.

Second, procurement. The report does not show how any lab obtained access — smuggling, third-country routing, cloud rental, or any other path. It contains no shipping records, no invoices, no customs data. The report notes that Blackwell parts require a US licence for export to China and that none has been reported, but that is a policy observation, not a legal finding. No court, regulator or government agency has found that any law was broken, and none of the four labs has confirmed the claims.

Third, training. The fingerprint "identifies how the weights were created and supports, without proving, the inference that Blackwell also performed the training", in the report's own words. Initialisation is only the first step; the bulk of training could, in principle, have run on other hardware. The authors consider one alternative — specially configured AMD hardware, also export-restricted — but call it very unlikely, and the distinction between initialisation and training remains.

There are further limits in the method itself. It only works on open-weight models; closed models cannot be tested at all. If weights were initialised on a CPU, or by a different framework such as Baidu's PaddlePaddle, no Nvidia GPU trace exists. And the report records a null result that cuts against any simple narrative: DeepSeek-V4.1-Flash, the lab's newest model, showed no fingerprint "despite an exhaustive search". That is not a clean bill of health, and it does not prove the model was trained on Huawei hardware — the report links no Huawei chip to DeepSeek — but a missing fingerprint proves nothing either way.

Close-up of a computer chip with visible circuits under magnification, blue and silver tones
AI-generated illustration (generated 14 Nov 2026)


For readers of INORYVA's earlier coverage, this report sits beside two different DeepSeek stories that should not be confused with it. The DeepSeek Moonshot distillation report concerned an allegation that the two labs distilled OpenAI's models — a claim about training data, not hardware. The US bans DeepSeek federal systems story concerned a government order removing DeepSeek from federal systems — a policy decision, not a forensic finding.

This report also makes no link to Jensen Huang's resignation from Nvidia, and none should be invented: the two stories landed on the same day, 12 November, but the report is about model weights published months earlier. For context on forensic evidence, the Cursor Claude Code extension malware story shows a tampered extension caught by examining what it actually did; and the flagship AI model comparison covers what these models can do, rather than what hardware started them.

What to watch next

The most useful next step is independent reproduction. Engineer Peter Liu has already published a CPU-only script that recovers the same SM peaks and seeds for DeepSeek-V4-Flash, Kimi K3 and MiniMax-M3 without renting Blackwell hardware, and the report publishes the seeds themselves — DeepSeek's is 953049583, shared by V3 and V4-Flash; Moonshot's is 20256586. Anyone with the public weights can run the check. Watch also for responses from the four labs and from Nvidia, and for any statement from the US Commerce Department, which enforces the export controls.

A researcher examining lines of code and data charts on multiple monitors in a dark office
AI-generated illustration (generated 14 Nov 2026)


Key Takeaways

  • Will DePue's Unemployment Labs report, published 12 November 2026, says six Chinese frontier models — DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ant Ling-3.1-flash — carry initialisation patterns matching export-restricted Nvidia Blackwell GPUs.
  • The method, checkpoint forensics, reads Philox random-number patterns in published open weights: 148-SM layouts match B200/B300 chips and 152-SM layouts match GB200/GB300 systems, with matches of 49 to 1,385 sigma against an 8-sigma chance ceiling.
  • Earlier models from the same labs read as Hopper, and the newest models from Alibaba, Zhipu, Tencent, Xiaomi and StepFun still read as Hopper — so the claim is about four labs, not all Chinese AI.
  • The report does NOT prove where the chips were located, how access was obtained, or whether export controls were violated — it says so explicitly, and no regulator or court has found a violation.
  • DeepSeek-V4.1-Flash showed no fingerprint despite an exhaustive search — a null result that proves nothing about what hardware trained it, including no link to Huawei chips.
  • Cluster-size estimates of hundreds to thousands of GPUs are the authors' rough estimates, not audited figures.

Conclusion

The Blackwell fingerprint report is the most concrete evidence yet in the long argument over Chinese labs and export-controlled chips, because it is reproducible: the weights are public, the seeds are published, and anyone can run the check. But its strength is also its limit. It proves which hardware started six models, not where that hardware sat, how it was reached, or whether any law was broken — and the report says that in its own words.

Sources

 

Comments