Researcher says Chinese open models carry Nvidia Blackwell fingerprints
Will DePue's report says six models from four Chinese labs were started on Blackwell chips. It says it cannot show where the chips were or that any law was broken.
Key takeaways
- A report by former OpenAI researcher Will DePue says six open-weight models from Ant Group, DeepSeek, MiniMax and Moonshot AI carry traces of Nvidia Blackwell chips.
- The method reads how a model's starting weights were laid out, which depends on the chip's number of processing units.
- The report says the method matched disclosed hardware in all 24 cases where a developer had stated it, which it describes as a consistency check, not an accuracy figure.
- It does not show where the chips were, who owned them or that any export law was broken; Blackwell systems can be rented outside China.
- The named companies and Nvidia had not commented as of Storm Media's report; DeepSeek has denied an earlier, similar claim by a US official.
Six recent open-weight AI models from four Chinese developers show numerical traces of Nvidia's Blackwell chips, which need a US license for export to China, according to a report published on Nov. 12 by Will DePue, a former OpenAI researcher who runs the nonprofit research group Unemployment Labs. The developers are Ant Group, DeepSeek, MiniMax and Moonshot AI. TechCrunch, Tom's Hardware, Storm Media and CASRAI have covered the report.
DePue says his evidence comes solely from the published model weights, with no leaked material or inside sources. The report is self-published and has not been peer reviewed, and CASRAI notes that the findings have not been independently verified.
What does the report claim?
According to the report, two models, DeepSeek-V4-Flash and MiniMax-M3, match the layout of Nvidia's B200 or B300 chips. Four others match the GB200 or GB300: DeepSeek-V4-Pro, Ling-3.1-flash from Ant Group, and Moonshot's Kimi K3 and Kimi K3.1-Pro. Older releases from these labs, including DeepSeek-V3 and Moonshot's K2 models, read as Nvidia's previous Hopper generation. Storm Media and CASRAI report that DePue calls the DeepSeek-V4-Flash result conclusive and the remaining five strong.
Not every Chinese lab has switched. The report says the latest models it could test from StepFun, Tencent, Xiaomi, Zhipu and Alibaba's Qwen team still point to Hopper, and that Meituan's LongCat-2.0 showed no Nvidia fingerprint at all.
How does the method work?
Before training, a model's weights are filled with starting values generated from a seed number. On Nvidia hardware, that work is split across the chip's processing units, called streaming multiprocessors, and their number decides where each value lands. According to the report and Tom's Hardware, a Hopper chip has 132 of these units, a B200 or B300 has 148, and a GB200 or GB300 has 152. Part of the starting pattern can survive training. The report says a specially configured AMD MI300X, also restricted from export to China, can produce the same layouts; it says it ruled this out for DeepSeek-V4-Flash and considers it very unlikely for the other five.
DePue says he searched about 6 trillion combinations of seed and chip layout and tested matches on weights left out of the search. Under the report's statistical model, the strongest match expected by chance would reach about 8 sigma; the best result for each of the six models lies between 49 and 1,385 sigma. It says the method agreed with the developer's own statement in all 24 cases where a developer had disclosed its hardware, which it describes as a check on consistency, not a measured accuracy figure, because several of those models are related, and that the DeepSeek-V4-Flash and Kimi K3 results were also reproduced on physical Blackwell chips.
What does the report not show?
The report is explicit about its limits. The fingerprint identifies the hardware that generated a model's initial weights, it says, but cannot tell where those chips sat, who owned them or whether Blackwell handled all of the training. The report also says this evidence on its own is not proof of smuggling or of a breach of export rules. The report and Storm Media note that Blackwell capacity can be rented in data centers located outside China, and TechCrunch says the hardware may have been located abroad. Tom's Hardware notes that, as far as it can tell, renting Blackwell hardware abroad is not currently illegal for Chinese companies. Storm Media, citing The Information, reports that the Commerce Department is finalizing a rule to limit Chinese firms' remote use of advanced chips housed in data centers in other countries.
DePue also estimates the size of the training runs: roughly 4,500 Blackwell GPUs for Kimi K3, 3,000 for DeepSeek-V4-Pro and about 12,000 across five runs, within a range of 7,000 to 22,000. The estimate covers five runs because the report says Kimi K3.1-Pro may be a further-trained checkpoint of K3. The report says it is not confident in these figures, which rest on assumptions the weights cannot confirm, and that they replace lower estimates in its first version.
In a post on X, DePue said senior researchers at OpenAI and Anthropic had checked his work. Storm Media points out that he did not identify them and that the report itself says nothing about such a review. His post also called the evidence conclusive for all six models, while the report reserves that word for DeepSeek-V4-Flash.
How have the companies responded?
None of the four labs, nor Nvidia, had responded by the time Storm Media published, and CASRAI said neither TechCrunch nor The Register reported a response. DeepSeek has denied an earlier claim, made by a senior US administration official in February, that it trained V4 on Blackwell chips, according to Storm Media. TechCrunch notes that White House technology adviser Michael Kratsios, citing US intelligence, said this year that Moonshot had used Nvidia chips barred from export, and according to Storm Media, the White House said in July that Moonshot got access to Blackwell chips via a facility in Thailand.
Questions readers are asking
Which models does the report name?
DeepSeek-V4-Flash and DeepSeek-V4-Pro, Moonshot's Kimi K3 and Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash from Ant Group.
Does the report prove export controls were broken?
No. The report says it identifies the hardware that created the starting weights, not the chips' location or owner, and that this alone does not prove smuggling or a breach of export rules.
How can weights reveal a chip?
Starting weights are generated from a seed and spread across the chip's processing units. Hopper, B200/B300 and GB200/GB300 chips have different numbers of them, so each leaves a different pattern, part of which can survive training.
Has the finding been independently checked?
The report is self-published and not peer reviewed, though it publishes seeds so others can try to replicate it. DePue said unnamed senior researchers checked it, which the report itself does not mention.
Sources
- Unemployment Labs: Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs
- TechCrunch: Export-banned Nvidia chips are behind China's most powerful AI models
- Tom's Hardware: Researchers allege DeepSeek, Kimi, and other Chinese AI models were trained using banned Nvidia Blackwell GPUs
- Storm Media: DeepSeek and Three Rivals Used Banned Nvidia Chips, Researcher Says
- CASRAI: Chinese AI Models Tied to Banned Nvidia Chips
- #nvidia
- #export-controls
- #china
- #open-weight-models
- #blackwell
Get the day's top AI & tech stories on WhatsApp: Follow the NextFish channel
Get stories like this every morning.
Related
AI2 min
HUMAIN sets Dec. 15 for its AI stack, promises a 3T-parameter model
The Saudi AI company's CEO calls the unnamed model the world's largest open-weight model and says it is far cheaper to run. No model card, licence or benchmark has been published.
Chips & Infra3 min
Megaspeed says Alice Huang left in 2025, hires counsel for internal probe
The Singapore-based Nvidia cloud partner says reports about its supply chain lack verifiable evidence, after a WSJ report on billions of dollars in Nvidia chip orders for China.
Chips & Infra3 min
Major Nvidia investors ask board to keep door open to Huang's return
Altimeter's Brad Gerstner and Atreides' Gavin Baker want a slow CEO search, while a petition backed by Elon Musk asks the board to reinstate Jensen Huang.
Data Centers & Energy3 min
Malaysia to require data centres to disclose power use and GPU imports
Trade minister Johari Abdul Ghani says the rule will be written into law so the government can monitor data centres. He also said some imported GPUs had been re-exported.