Restricted chips left their fingerprint in six Chinese AI models
The real question is what that fingerprint actually shows.

Random numbers that keep a return address
Before it learns anything, an AI model is a huge table of randomly drawn numbers. Training then adjusts them one by one over weeks. But some never move: the ones reserved for words the model will never see, not once. They stay exactly as they were on day one, like dust under a piece of furniture nobody shifts.
Will DePue, a former OpenAI engineer who left this year to start a nonprofit called Unemployment Labs, took a close look at that dust. In a report published on November 12, he claims he can read from it which chip model each AI was born on. And his conclusion is an awkward one for Washington: six recent Chinese models were born on Nvidia Blackwell chips, the generation the US bars from sale to China without a license.
A chip that isn't sold in Beijing
Blackwell is Nvidia's latest data-center line: the B200, the B300, and their versions paired with the company's own processor, the GB200 and GB300. A US Commerce Department note dated May 31, 2026 reminds readers that since November 17, 2023, supplying this class of advanced chip to a company whose parent is Chinese requires a license, even when that company operates outside China. No license has been reported for these models, DePue notes.
The suspicion isn't new. In July, the White House accused Moonshot, the lab behind Kimi K3, of accessing GB300 servers in Thailand. As we wrote at the time, no public technical evidence had been presented. What changes here is the source: it isn't a leak or an informant, but the weights the labs published themselves.
Same draw, 132 or 152 columns
The mechanism comes down to an engineering detail. To fill billions of cells quickly, PyTorch, the software used to build most large models, splits the random draw into thousands of small parallel tasks: 2,048 per compute core. And the number of cores varies from chip to chip: 132 on the previous generation's Hopper (H100, H800, H200), 148 on the B200 and B300, 152 on the GB200 and GB300.
Think of dealing a deck of cards around a table. With 132 players or with 152, the cards come off the deck in the same order, but they don't land in front of the same people. Same here: one starting seed produces the same numbers, but in different places depending on the chip. PyTorch's code confirms it: the split depends on the chip's core count.
So DePue searched for the seed and core count that reproduce the frozen values. He says he went through roughly 6 trillion combinations.
DeepSeek: one seed, two machines
The cleanest case is DeepSeek, which reuses the same seed from one model to the next. DeepSeek-V3, released in December 2024, matches 132 cores: Hopper, which fits the H800s the lab had said it used. DeepSeek-V4-Flash, released in April 2026, matches 148 cores.
DePue then ran the counter-test on real machines. On a B300, unmodified PyTorch regenerates 461 of the 469 tested rows of V4-Flash, 445 of them bit for bit. On an H100, it regenerates those of V3. Neither chip reproduces the other's model, and shifting the seed by a single unit yields nothing. Kimi K3 gave the same result on a rented GB200.
The full list is six models from four labs: DeepSeek-V4-Flash and V4-Pro, Kimi K3 and K3.1-Pro from Moonshot, MiniMax-M3, and Ling-3.1-flash from Ant Group, Alipay's parent company. At each lab, the previous model still read as Hopper. Not everyone switched: the latest models from Alibaba, Zhipu, Tencent, Xiaomi and StepFun still read as Hopper.
To test his method, DePue ran it on 24 models whose builders have published their hardware. It got each one right: 132 cores for Meta's Llama 3.1, trained on H100s, and 152 for Nvidia's Nemotron-3 Ultra, one phase of which ran on GB200s. He says himself that these control cases aren't all independent, and calls it a consistency check rather than a measured success rate.
The check runs in a few seconds on a laptop. We re-ran it, re-downloading the weights from the four labs' official repositories, and it landed on the same values. Only the right core count makes the numbers line up; every other one stays at noise level.
DeepSeek wasn't exactly hiding it, either: the engine it published in October for running its models only targets Blackwell chips and Huawei's Ascend.
What the fingerprint doesn't say
First, this is a published method, not a proven one. The check we re-ran is DePue's, with his code. Nobody else has published the counter-test on real Blackwell chips so far. DePue claims on X that researchers at OpenAI and Anthropic verified his results, but neither company has spoken publicly, and his report doesn't mention it.
Second, another chip can mimic the signature: an AMD MI300X, configured in a very unusual way to use only 148 or 152 of its units. DePue ruled it out for DeepSeek-V4-Flash thanks to tiny rounding differences between the two manufacturers. For the other five models, he considers it very unlikely but can't exclude it. That chip is also subject to export controls.
The fingerprint also points to the machine that created the starting weights, not necessarily the ones that did all the training. A model can be born on one chip and grow up on others. DePue nonetheless estimates that roughly 12,000 Blackwell chips would have gone into the five training runs, for more than half a billion dollars. He writes that he doesn't trust those estimates, and he revised them upward on the very day of publication.
Above all, the method can't say where the chips were or who owned them. Blackwell servers can be rented in data centers outside China, and renting computing power isn't buying a chip. That's exactly the gap a US bill, the Cloud Security Act, aims to close, according to The Register. The report is blunt about it: its evidence points to access to the hardware, not to smuggling or a breach of export rules.
Chance kept the machine's name. It didn't keep its address.
Was this article useful?
Topics covered:
Frequently asked questions
How can you tell which chip an AI model was born on?
Which Chinese models carry the trace of Blackwell chips?
Is this evidence of chip smuggling?
Is Will DePue's method reliable?
Have OpenAI and Anthropic verified these results?

Alexandre Noto
Co-founder & Tech Expert
Alexandre has been in tech for over 20 years. Entrepreneur, software architect and AI enthusiast, he translates complex concepts into accessible explanations. At Declic Media, he is the technical voice that makes AI understandable for everyone.
All articles by Alexandre →