A forensic analysis of published AI model weights has found that six of China’s leading AI models were started on NVIDIA Blackwell graphics processors, the advanced chips that United States export controls bar from China. The analysis, published by the research group Unemployment Labs, examined the open weights of models from DeepSeek, Moonshot AI, MiniMax and Ant Group, and says the hardware fingerprint in each one matches Blackwell chips that require a US export licence, with no such licence reported. The finding turns months of allegations about diverted chips into testable evidence, and it lands as Congress weighs new controls.
How a fingerprint survives training
Every AI model starts as a set of initial values, laid down before training begins. Most labs generate those values with a pseudorandom generator called Philox, fed by a number known as a seed. The framework PyTorch spreads that work across thousands of parallel threads on the graphics processor, and the thread count depends on the chip.
A Hopper chip uses 270,336 threads, a Blackwell B200 or B300 uses 303,104, and a Grace Blackwell part uses 311,296. The same seed on a different chip scatters the same numbers into different positions, and traces of that arrangement survive in the weights. By searching more than six trillion seed-and-layout combinations, the researchers matched the pattern to a chip class.
Six models, four labs
The report names six models released since April 2026. DeepSeek’s V4-Flash matches the B200 or B300 layout, and its V4-Pro matches the Grace Blackwell layout. Moonshot AI’s Kimi K3 and K3.1-Pro both match Grace Blackwell, as does Ant Group’s Ling-3.1-flash, while MiniMax’s M3 matches the B200 or B300.
The same labs’ earlier models read as Hopper, the previous generation. DeepSeek’s V3, Moonshot’s K2.5 and MiniMax’s M2.7 all carry the older chip’s fingerprint, which dates the switch to Blackwell to the spring and summer of 2026. Other Chinese labs have not switched: the newest models from Alibaba, Zhipu, Tencent, Xiaomi and StepFun still read as Hopper.

The scale of the runs
The report puts the five training runs at roughly 12,000 Blackwell graphics processors between them. The two largest, Kimi K3 and DeepSeek-V4-Pro, needed about 4,500 and 3,000 chips each over two to three months. At roughly $70,000 per processor, the largest clusters would cost $200 million to $300 million, and the five runs together more than half a billion dollars.
The authors stress these are estimates, built on assumptions about efficiency and run length that the weights cannot confirm. Halving a run’s length would double its chip count, and the figures cover only the training runs, not experiments or earlier work. The report says the labs’ total spending on export-controlled graphics processors is likely in the billions of dollars.
What the evidence does not prove
The report is careful about its own limits. A fingerprint shows which hardware created a model’s starting weights, not where that hardware sat, who owned it, or how the lab obtained it. The weights alone cannot prove that chips were smuggled into China, or that any export-control law was broken.
Blackwell systems can be rented from data centres outside China, and the report says its evidence concerns access to the hardware, not the route by which it was reached. The authors also tested the one serious alternative, a specially configured AMD chip, and ruled it out for DeepSeek-V4-Flash, where 363 of 377 differing values matched NVIDIA and none matched AMD.
Why it lands now
The finding arrives in the middle of a scandal that has already cost NVIDIA its chief executive. Jensen Huang resigned after reports that he had a relationship with a Chinese businesswoman whose company, Megaspeed, is under investigation over alleged chip smuggling. Congress has since summoned Huang and Commerce Secretary Howard Lutnick to testify in December.
Lawmakers are also pressing four bills that would tighten the rules, including one aimed at the remote-rental route the report leaves open. Separately, a security firm reported finding 32 Blackwell B300 processors through publicly exposed monitoring systems with internet addresses geolocated to China. Neither development proves smuggling, but both point the same way: the controls are leaking.
Our Read
What makes this report different from the allegations that preceded it is that it can be checked. The authors published the seeds and the preserved weights, and they reproduced the match on rented hardware: a Blackwell B300 regenerated DeepSeek-V4-Flash’s starting values, with 445 rows matching bit for bit, while a Hopper H100 could not.
That reproducibility is the story’s strength, and it is also its limit. The method only works on models whose weights are public, so closed systems stay out of reach. The policy question is sharper than the smuggling question: if leading labs can simply rent Blackwell capacity abroad, then the export ban’s weak point is not the border but the cloud.
Three things to watch: the promised code and paper, which will let researchers repeat the test; any response from NVIDIA or the four labs named; and whether the rental loophole becomes the centre of the December hearings. Until then, the strongest claim in the report is the narrowest one — six models carry Blackwell’s fingerprint, and nobody has shown otherwise.
Why It Matters
Export controls only work if they can be enforced, and enforcement has so far rested on allegations, investigations and paper trails. A forensic test that reads the answer out of the models themselves changes that: it gives regulators, lawmakers and rival labs a check that does not depend on anyone’s word.
If the finding holds up under independent testing, the debate shifts from whether the rules are being broken to how they are being broken — and whether a ban that stops at the border can survive a world where the world’s best chips can be rented by the hour.
Background
The United States has restricted exports of NVIDIA’s advanced AI chips to China since 2022, allowing only earlier-generation hardware such as the H200 under licence. Officials and press reports have long alleged that Chinese labs were using Blackwell chips anyway, but until now the claims rested on shipment records and unnamed sources rather than evidence drawn from the models themselves.
Source: Unemployment Labs report and Chosun Biz report.


0 Comments