Hacker news

  • Top
  • New
  • Past
  • Ask
  • Show
  • Jobs

Chinese frontier AI models carry the fingerprint of export-controlled GPUs (https://unemploymentlabs.com)

456 points by firloop 8 days ago | 198 comments | View on ycombinator

flowerlad 7 days ago |

Chinese AI companies are free to use any Nvidia GPU they want as long as the GPU is not physically located within China. It is entirely consistent with US export controls for Chinese AI companies to remotely use GPUs located in Singapore or Malaysia.

“In facilities across the region, Megaspeed rents Nvidia chips to Chinese tech giants such as Alibaba, an arrangement that does not violate current US export control regulations.” See [1].

[1] https://www.atlanticcouncil.org/dispatches/beyond-the-spy-sc...

x313 7 days ago |

The author of this piece is a SWE-turned-influencer "China hawk" from OpenAI (i.e. the primary advocate of export bans along with Anthropic), and this isn't an academic paper despite the theming.

I have no idea whether this is legit or not (there's no link to any code), I'm mentioning this because all the comments here are assuming truth without verifying the contents.

GenerWork 7 days ago |

It's insane how many articles have dropped in the past 24 hours that all have a similar tone. At this point it's hard to see this as anything other than a broad coordination from someone who wants to destroy Jensen/Nvidia.

martinwright 7 days ago |

I definitely care about this much less than it seems I'm supposed to.

I think the frontier models are definitely in a healthy dynamic where it seems the near-but-not-quite performance of open-weight models are what is pushing the most innovation and price competition from OpenAI, Anthropic and co

loneboat 7 days ago |

> We recover traces of the models’ initial values, and then search for the hardware configurations and initial configurations that reproduce them

From my (very cursory, naive, layman's) glance at the summary, they're identifying and tracing some characteristics that specific hardware imparts into a model's initialized weights. But then concluding that this indicates the models were also "trained on" that same hardware.

But is that a valid assumption? Is it possible that a model was initialized on one system configuration, while the bulk of the training was performed on another?

In fact, it would seem reasonable for the "okay, I'm setting up this new model on my workstation, prepping it for training next month" pipeline step to be far removed from the "okay, now deploy the [initialized, but not yet trained] model across our giant fleet of beefy training machines".

int32_64 7 days ago |

I have run some large open models on neoclouds and the KYC has been minimal, even if the USG stops all GPU exports it will be extremely difficult to crack down on Chinese renting GPUs on Western servers.

There will be a black market of renting IDs to KYC and gain access to the chips, and then the hosts will monitor their servers for what is being trained, it will be a cat and mouse game.

We are headed for an increasingly unfree computation landscape and the politicians are going to make demands that will be hard to deliver.

basiccalendar74 7 days ago |

as the post mentions, the models could have been trained on 1. GPUs within China or 2. GPUs outside China for example datacenters in Singapore.

The 2nd option is perfectly legal and allowed by US government.

cadamsdotcom 8 days ago |

What if the gov tolerated gpu exports on purpose?

Black & grey market devices marked up many times cost buyers heavily, while still suppressing motivation to develop sovereignty.

If nvidia selling to China tolerated for the realpolitik and gets undone by incoming powers to create an appearance of change, is it really the right thing to do? Is it the smart thing to do?

It's an interesting thought experiment, though, we will probably never know.

culi 7 days ago |

> The newest models we could read from Alibaba (Qwen), Zhipu (GLM), Tencent, Xiaomi and StepFun all read as Hopper-trained. Meituan says its LongCat-2.0 was trained entirely on “AI ASIC superpods,” and we found no NVIDIA fingerprint in the model weights it’s released.

Yeah makes sense. They are in transition. This article doesn't really tell us anything new (or interesting). There's nothing wrong w Chinese labs using GPUs that are already in the country. An export ban just means reduced import of new GPUs

saidnooneever 8 days ago |

i dont know what is expected of export controls... they are never effective. they also find western chips in iranian drones etc. there is a huge gray area of resellers who resel to resellers etc. its pretty much uncontrollable....

stratos123 7 days ago |

Weird detail: some of these results are based on initial patterns in weights surviving through training. But isn't muon, which is used for many of these models, well-known to be able to consistently erase patterns in weights, unlike previous optimizers [1]?

This doesn't really call the overall results of the report in the question (the strongest finding is the perfectly matching weights for dead-weight rows which weren't touched by an optimizer at all), but it does suggest that the explanation in the twitter thread[2] of how this works is way too simplistic. Are the patterns in question somehow "special" and not erasable even with muon?

[1] https://docs.modula.systems/examples/weight-erasure/

[2] https://twitter.com/willdepue/status/2120992457630326841

coldtea 4 days ago |

For the things people are worried, it would be more dangerous when they didn't carry those fingerprints anymore - because it would mean they've caught up and switched to their own.

octoberfranklin 7 days ago |

Title is deceptive; at most they've shown it carries the fingerprint of "a 148-CU GPU".

treebeard901 7 days ago |

If you could rent compute to what is essentially a competitor and monitor everything about their code and models in the process, since you're still in control of the hardware, why would a company or in this case a country not take that competitive advantage?

Numerlor 6 days ago |

You don't need to do any fingerprinting, kernels for those GPUs have been in open-sourced code from the Chinese labs. It's not like they're hiding it

Alien1Being 7 days ago |

Not too many years before the headline reads

"Western AI models using GPUs export controlled by China"

The decline of the USA and the West has been precipitous.

yumraj 7 days ago |

Not doubting that they have access to these.

But, could the fingerprints be due to distillation from Blackwell trained models rather than training?

nittanymount 5 days ago |

one important thing is hidden in the notes under appendix2, this is not even mentioned in the conclusion.

> We recovered no signature from DeepSeek-V4.1-Flash (after an exhaustive search),

DS v4.1 is the latest version, no signature found in this model.

sam0x17 7 days ago |

Just further proof that export controls rarely restrict the ones they are designed to restrict

mark_l_watson 7 days ago |

This is all so stupid.

US businesses need two things:

1. access to the very best frontier models for hard problems

2. good enough open models from China, the US (I like poolside.ai, but lots of good choices), and even other countries like France that make smaller investments in Digital Sovereignty.

If the ‘super rich fucks’ in the US who have been making risky AI bets succeed in convincing the US government to curtail the use of small efficient models for easier problems, that will just hurt our economy.

JohnsonZou 7 days ago |

Awesome job, most interesting analysis I've read these days.

netvarun 7 days ago |

@dang curious, is this post getting heavily downvoted or flagged?

I get that the geopolitical angle is sensitive, especially with the backdrop of today's big news - but I thought this was a really well-researched technical deep dive, right up HN's alley. (TIL: you could fingerprint gpu hardware just from an llm weights!)

nullc 7 days ago |

Cute trick but it's not exactly a secret, e.g. deepseek recently published their own inference engine -- and it only has kernels for SM100 (blackwell) and Huawei Ascend 950, and all its published performance results using blackwell.

And these were completely openly disclosed as headline properties of the software: https://github.com/deepseek-ai/DeepInfer/tree/main/#performa...

No one found it remarkable a couple weeks ago, ... but now the fact that DS is using blackwell is apparently some super secret requiring clever RNG init stats???

lynguist 7 days ago |

Ah, that means the recent huge open-source AI advances, that we all benefit from, directly and indirectly, come from Jensen Huang's China boycott circumvention.

Which means, in future, we maybe cannot expect such large capability increases.

brcmthrowaway 7 days ago |

Are (ex-)OpenAI/Anthropic employees leading the charge here to protect their portfolios?

Cameri 8 days ago |

So export controls are more like export friction

demo4567 4 days ago |

Then reduce the price of GPUs and RAMs. So we can own them or shut up and gtfo.

K0balt 3 days ago |

This article states the obvious?

It’s perfectly legal to use Nvidia GPUs to train Chinese models. It’s called cloud compute, maybe the author heard of it?

This seems like a Jensen hit piece to me, making it seem like shady Nvidia practices are responsible for Chinese models.

GPU pearl clutching is ensuring that the USA loses the market in the long term. Robotics = local models. Robots will outnumber humans in 30 years. We are making sure that China throws their full ability into catching up their fab technology, and they have logistical and ecosystem advantages that the USA and Europe can never hope to match, not to mention the power production.

We’d be much better off if they kept buying our GPUs and couldn’t justify the enormous investment to catch up their fabs.

syngrog66 5 days ago |

and Jensen's "cousin" -- a woman who looks strangely just like him -- happens to be CEO of AMD. makes one wonder exactly just how much nepotism was involved in this clearly coordinated circular financing and sanctions evasion operation.

piokoch 7 days ago |

US frontier AI models carry the fingerprint of data acquired from illegal sources like Anna Archive, libgenesis, etc. and, (un?)surprisingly, nobody cares.

Only Anthropic was stupid enough to let be caught red-handed but get away with some ridiculously low penalty.

elonfboy 7 days ago |

WAIT! Are you saying the giant plagiarism machine was built with…stolen IP?!?!

glennfiedler 7 days ago |

Ladies and gentlemen, we are now officially in Huang-gate.

flakiness 7 days ago |

Such a great domain name by the way.

sandos 7 days ago |

My BS-meter just went through the roof:

"Specially configured AMD hardware, which is also restricted from export to China, is the only alternative we have identified; we have excluded it for DeepSeek-V4-Flash"

This sounds as if they are using something VERY fuzzy to "match" things, and if they are ALSO matching against completely different hardware which this makes it sounds like.... then I dont know if this means anything.

gmerc 5 days ago |

htrp 7 days ago |

clickbait written as research....

yogthos 7 days ago |

oh no, anyways

rubberstamp 7 days ago |

Land of the free*.

*conditions apply

ftchd 7 days ago |

Well, d'uh...

duty_lines 6 days ago |

The "not physically located within China" framing is doing a lot of work here, and under the EAR it isn't the operative test. Jurisdiction follows the item: a US-origin GPU stays subject to the EAR wherever it is re-exported, and a foreign-made one is caught if it exceeds the de minimis US-content threshold or falls under the foreign direct product rule. So the question isn't where the box sits, it's who the re-export is going to and what the exporter knows about end use — "knowledge" in the regulation includes the conscious disregard / high probability standard, which is how a reseller's log of "customer declines to state workloads" becomes evidence rather than a shield.

Two other mechanics people miss: releasing controlled technology to a foreign national is itself a "deemed export", so staffing matters as much as shipping; and the control lists are keyed to ECCN, not to marketing names, so the classification question is per-part-number and per-performance-threshold, not per-product-line.

None of that makes enforcement easy across borders — the jurisdiction argument above is exactly why the extraterritorial reach fights keep landing in the same place: the US can reach the item and the US-person nexus far more reliably than it can reach a foreign reseller's intent.

apishare_api 7 days ago |

[flagged]

black_13 5 days ago |

[dead]

exabrial 8 days ago |

I guarantee you this isn't an "accident". Whatever is being sent over there is loaded with its own special spyware lol

fibonacci112358 8 days ago |

[flagged]