How a model file gave away the chips that made it
15 Nov 2026 · News
A public report says six recent Chinese AI models were started on Nvidia Blackwell chips, which need a US license to ship to China. The evidence sits inside the models' own published files, and anyone can check it. We did.
On 12 November Unemployment Labs published a forensic report by Will DePue. It finds the fingerprint of Blackwell GPUs in six models from DeepSeek, Moonshot AI, MiniMax and Ant Group released since April. Earlier models from the same labs read as Nvidia's older Hopper chips. The code is public, and on this desk's ordinary CPU it reproduced the author's table exactly. This is the strongest kind of evidence in AI right now, because it does not ask you to trust anyone.
How a file remembers its chip
An AI model is billions of numbers. Before training starts, each number gets a random starting value. The randomness is fake on purpose. Give PyTorch the same seed and it produces the same numbers every time.
Where those numbers land depends on the chip. On Nvidia GPUs, PyTorch splits the job across 2,048 threads for each of the chip's processing units, called SMs. The count of SMs decides which thread writes which position. Hopper chips have 132 SMs. A B200 or B300 has 148, and a GB200 or GB300 has 152. Same seed, different chip, different layout.
Most of those starting values get rewritten in training. A few never do. A model's vocabulary includes some tokens that never show up in the training text, so the rows of numbers for those tokens never change. They still hold the original random fill. Find the seed, try every SM count, and only one count rebuilds those rows. The report says the author found the seeds by brute force across all four billion possibilities.
We reran it
The repo's README says the check "runs on a laptop in a few seconds." It does. We ran it twice on the box, once with the rows stored in the repo and once after downloading them fresh from Hugging Face at the pinned versions. Both runs printed the same table:
| Model | Rows rebuilt | SMs | Chip family |
|---|---|---|---|
| DeepSeek-V3 (Dec 2024) | 8 of 8 | 132 | Hopper |
| DeepSeek-V4-Flash | 461 of 469 | 148 | Blackwell B200/B300 |
| Kimi K3 | 13 of 13 | 152 | Blackwell GB200/GB300 |
| MiniMax-M3 | 17 of 17 | 148 | Blackwell B200/B300 |
| Ling-3.1-flash | 384 of 384 | 152 | Blackwell GB200/GB300 |
The matches are near perfect, a correlation of 0.98 or higher, and 0.99999 or higher for four of the five. Shift the seed by one and the match drops to about zero. The report adds that the author repeated the check on physical H100, B300 and GB200 chips with stock PyTorch. We have no GPU, so we did not repeat that part.
What it does not show
The report spells out its own limits. The fingerprint shows the hardware that created the starting weights. It does not show where the chips sat, who owned them or how a lab got access. Blackwell systems can be rented in data centers outside China, so this is not proof of smuggling or of a broken export rule. A lab could also create weights on one cluster and train them on another.
One other chip can fake the layout. An AMD MI300X has 304 units but can be split into two groups of 152 or capped at 148. The author tested one and found it reproduces the Blackwell layouts. A second test on tiny numerical differences ruled AMD out for DeepSeek-V4-Flash, 363 matches to Nvidia and none to AMD in the values where the two differ. For the other five models the report calls AMD very unlikely but not excluded. The MI300X is also restricted from export to China.
The method only works on open weights made with PyTorch on a GPU. It found nothing in DeepSeek-V4.1-Flash or in Meituan's LongCat-2.0, and closed models cannot be checked at all. A blank result says nothing either way.
Then and now
In October we wrote about an outside safety report that still went through the lab, because METR's own post said OpenAI could redact and reshape the writeup. At the end of October it was OpenAI picking the referees for its own physics proof. Both times the outside check ran through the lab, on the lab's terms.
This check needed no one's permission. It runs on files the labs chose to publish, and a stranger can rerun it in a few seconds. Washington's claims about Chinese labs and Blackwell chips have so far rested on officials' statements and press reports. The report notes that those claims came without published technical evidence. Now the evidence is a script.
Our call
By 31 December 2027, a Commerce Department export enforcement action, a congressional committee report or a court filing will cite weight-level hardware forensics like this one as evidence of what chips made a model. A test anyone can rerun is hard for a government to ignore, and most open models built with PyTorch on GPUs carry the data to run it.