Open Weights Are Now Evidence
A researcher traced six Chinese open models to export-banned Nvidia Blackwell chips using only their published weights. It proves no crime, but it turns every weight release into a permanent disclosure.
OPC Community
OPC Community

A former OpenAI researcher read the hardware out of six Chinese models. The chips are the small story.
There are 469 rows of numbers inside DeepSeek-V4-Flash that training barely touched. Last week a researcher regenerated 445 of them, bit for bit, on a chip DeepSeek isn't allowed to buy.
His name is Will DePue. He used to work at OpenAI and now publishes as Unemployment Labs. His report names six models: DeepSeek V4-Flash and V4-Pro, Kimi K3 and K3.1-Pro, MiniMax-M3, and Ant Group's Ling-3.1-flash. He argues all six were initialized on Nvidia Blackwell systems, which are export-banned for China.
The timing could not have been worse. Jensen Huang is out at Nvidia amid allegations tied to Megaspeed, a customer suspected of routing chips into China. Ron Wyden and Richard Neal told the Commerce Secretary to preserve every record on Nvidia's China approvals. So of course everyone read the report as the smoking gun in a smuggling case.
It isn't one. The report says so itself. What it actually shows is stranger, and it will outlive this news cycle by years.
The Report Doesn't Prove a Crime
DePue didn't bury his caveat. The findings, he writes, "do not establish where the chips were located, how access was obtained, or whether export controls were violated."
The headlines skated right past that sentence, and they shouldn't have. A Chinese lab renting Blackwell capacity in a data center outside China breaks no American law today. The Remote Access Security Act would treat that rental as an export. It passed the House 369 to 22 on January 12, and it has been sitting in the Senate for more than 300 days.
So the fingerprints fit two stories equally well. In the first, somebody trucked $150 million of GPUs into Inner Mongolia. That's what a Trump administration official told Reuters about DeepSeek back in February, and DeepSeek denied it. In the second, Moonshot signed a cloud contract in a country where signing one is perfectly legal.
The report can't tell those apart, and Washington isn't going to wait until someone can. Here's my bet: a finding that proves no crime becomes the best argument for passing the law that creates one.
The politics come later, though. Start with how he did it.
How a Random Number Generator Became a Witness
When PyTorch initializes a model on a GPU, it fills the weights using a random number generator called Philox. How Philox spreads its work across threads depends on how many streaming multiprocessors the chip has. Change the chip and you change the layout, and the same seed spits out different numbers.
Training then overwrites most of that. Most, not all. Some rows never get a meaningful gradient. Others keep a faint statistical echo of where they started.
DePue searched trillions of seed and layout combinations, picked the best match on one set of weights, then tested it on weights he had held out. Before pointing the method at anything contested, he validated it on 24 models whose hardware was already known.
You can't wave the results away. Unmodified PyTorch on a B300 reproduced DeepSeek-V4-Flash's surviving rows exactly. A rented GB200 reproduced Kimi K3's starting values to a correlation of 0.999995. He puts the two biggest clusters at 2,000 to 3,000 Blackwell GPUs each and about 12,000 across all six models, a figure he himself calls rough.
Nobody leaked anything. No insider talked. He downloaded files the labs published themselves, and he did arithmetic.
Think about what follows from that. Every lab that has ever released weights also released a partial record of how those weights were made. They didn't know they were doing it. And the record sits in a format anyone with a GPU and some patience can now interrogate.
“A weight file is a deposition its author never agreed to give.”
The Fix Is Easy and the Past Is Permanent
The take going around is that Chinese labs have a liability they can't patch. That's half right.
For future releases, the patch is trivial. Initialize on CPU, or swap in a custom generator, and this particular trace is gone. Someone may already have done it: a later update to the report found no fingerprint in DeepSeek-V4.1-Flash after an exhaustive search.
For everything already shipped, there is no patch. Kimi K3 is on thousands of hard drives, and you can't recall a torrent. Whatever those weights say about where they came from, they'll keep saying it to every analyst, journalist, and congressional staffer who runs the script.
The second-order problem is worse. DePue found one channel because he went looking for one. Training is a long physical process full of numerical quirks, and I'd bet heavily that initialization isn't the last place hardware leaves a mark. A lab can scrub the trace it knows about. It can't scrub the one somebody discovers in 2028, inside a file it shipped in 2026.
So releasing weights means something different than it did two weeks ago. It used to be purely a distribution call: give the model away, win the developers, set the default. Now every release is also a permanent disclosure, and nobody knows how much it discloses.
Transparency Got Punished First
If you like open models, this next part should bother you.
The method only works on published weights. Nobody can audit a closed model behind an API this way. So the labs under the spotlight are exactly the ones that shared their work, and sharing is the reason they're exposed.
A Chinese lab that trained on the same rented Blackwells and kept its weights private faces zero forensic risk. Any general counsel in Beijing or Hangzhou can do that math in about a minute.
Open weights have been the Chinese labs' strongest card outside China. They're why an indie hacker in Lisbon runs DeepSeek on a rented box, and why Kimi shows up in half the agent stacks I see. The whole strategy assumed that releasing weights cost you nothing but the weights.
I expect the response within two release cycles. Drops will slow down, checkpoints will get scrubbed, and at least one major lab will hold its flagship back as API-only while it publishes a smaller model for goodwill. Nobody will call it a retreat. They'll call it a "staged release."
If that happens, we lose the best free models we have, and no ban will have taken them. A measurement technique will.
What Builders Owe This Story
Say you ship on Kimi or DeepSeek weights. Your honest status is that you have no legal problem. No statute makes it unlawful to download or run a model because of the chips that trained it. Anyone telling you to rip out your stack this week is selling something.
You do have two real problems.
The first is a question nobody has ever asked you. Enterprise buyers already send security questionnaires about data residency and subprocessors. After a week in which a chip CEO fell and two sets of congressional letters went out, "where did your model's training compute come from?" is going on that list. In October you could answer "I don't know" and nobody blinked. Try it now.
Getting ahead of it is cheap. Write down which weights you run, which version, and what is publicly known about their provenance, including DePue's caveat that nothing here establishes a violation. That's one paragraph. Mostly the buyer wants to see that you thought about it.
The second problem is supply. If I'm right that releases slow down or go closed, the weights you depend on today are the best open version of that lineage you'll get for a while. Mirror them. Pin the hash. And keep your prompts and evals portable enough that swapping models costs you a week instead of a quarter.
Then there's the thing I keep turning over. This audit cuts both ways. The method that reads Blackwell out of Kimi K3 will read the hardware out of any open model, from any country, for whatever reason someone later decides matters. Today that reason is export controls. Nothing says it stays that way.
For three years the argument for open weights was that you could inspect them. People meant safety, bias, backdoors. Nobody meant the supply chain.
DePue regenerated 445 rows that training left alone. The labs forgot those rows were there. The file didn't.
Frequently Asked Questions
Does Will DePue's report prove Chinese AI labs violated US export controls?
No. The report itself says its findings "do not establish where the chips were located, how access was obtained, or whether export controls were violated." The fingerprints fit smuggled GPUs and a legal cloud rental outside China equally well, because renting Blackwell capacity abroad breaks no American law today. The Remote Access Security Act, which would treat that rental as an export, passed the House but is still sitting in the Senate.
How can model weights reveal which GPU a model was trained on?
When PyTorch initializes a model on a GPU, it fills the weights using a random number generator called Philox, and the layout of that work depends on the chip's number of streaming multiprocessors. Training overwrites most of those starting values, but some rows never get a meaningful gradient and others keep a faint statistical echo. DePue searched seed and layout combinations until he could regenerate those surviving values, matching 445 of 469 rows in DeepSeek-V4-Flash bit for bit on a B300.
Is it legal to keep building on DeepSeek or Kimi weights, and what should builders do now?
The article's position is that you have no legal problem: no statute makes it unlawful to download or run a model because of the chips that trained it. The practical risks are enterprise buyers asking where your model's training compute came from, and open releases slowing down or going API-only. Write a one-paragraph provenance note on the weights and versions you run, mirror the weights and pin the hash, and keep your prompts and evals portable so swapping models is cheap.
Explore OPC City Chapters
Join the OPC Community
Connect with solo founders worldwide. Get daily insights, curated opportunities, and peer support.
Join free