Six frontier AI models released by Chinese labs since April 2026 carry numerical fingerprints matching Nvidia's export-controlled Blackwell GPUs, according to a forensic analysis published November 12 by research group Unemployment Labs. The finding offers the first hardware-level technical evidence supporting allegations that leading Chinese AI developers have gained access to chips the United States restricts from export to China.
The analysis examined models from DeepSeek, Moonshot AI, MiniMax, and Ant Group, and found that DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3, and Ant Group's Ling-3.1-flash all match Blackwell-class hardware, while earlier models from the same companies match Nvidia's previous Hopper generation. For more context on this story, see our ongoing breaking AI news.
How a Model's Weights Reveal Its Hardware
The method is unusual because it relies entirely on publicly released model weights rather than leaks or insider accounts. Every trained model retains traces of the random initial values it started from, and those initialization patterns depend on the hardware configuration used to create them. The researchers recovered traces of the models' initial values and then searched for hardware configurations that reproduce them.
The report says the team went further for two models: for DeepSeek-V4-Flash and Kimi K3, the researchers confirmed the match on physical Blackwell chips. Blackwell-class B200 and B300 chips have 148 streaming multiprocessors, while the GB200 and GB300 configurations used in larger systems have 152 — and it is that kind of structural detail that leaves its mark in the initialization data the models carry.
The fingerprint identifies how a model's weights were created and, the report notes, supports — without proving — the inference that Blackwell hardware also performed the training.
What the Report Found, Model by Model
The results trace a clear timeline across four labs:
- DeepSeek: models read as Hopper-trained through DeepSeek-V3 (December 2024, with the lab having disclosed use of H800 chips), then as Blackwell from DeepSeek-V4-Flash and DeepSeek-V4-Pro in April 2026 — V4-Flash matching B200/B300 and V4-Pro matching GB200/GB300.
- Moonshot AI: Hopper fingerprints through Kimi K2.5 (January 2026), then Blackwell-class GB200/GB300 from Kimi K3 (July 2026) onward, including Kimi K3.1-Pro released this month.
- MiniMax: Hopper through MiniMax-M2.7 (April 2026), then B200/B300-class from MiniMax-M3 (June 2026).
- Ant Group: Ling-3.0-flash (August 2026) reads as Hopper, while Ling-3.1-flash (October 2026) reads as Blackwell GB200/GB300.
One model sits between eras: Kimi K2.7-Code, posted six weeks before K3, is built on Hopper-era weights but contains a small component matching the B200/B300 profile.
Blackwell chips require a US license for export to China, and the report states none have been reported. US officials and press reports had previously alleged that DeepSeek and Moonshot AI were using Blackwell hardware, but without publishing technical evidence. This analysis is the first to ground those allegations in the models' own published weights.
What It Does — and Does Not — Establish
The report is careful about scope. The findings provide technical evidence of access to advanced computing hardware, but do not establish where the chips were located, how access was obtained, or whether export controls were violated. The researchers also considered alternatives: specially configured AMD hardware, which is likewise restricted from export to China, is the only alternative they identified, and they excluded it for DeepSeek-V4-Flash while considering it very unlikely for the other five models.
The researchers acknowledge limits in their method: a Hopper fingerprint cannot distinguish an H800 from an H100 SXM or an H200, and the H800 identifications for older models come from lab disclosures rather than the forensic method itself.
Why It Matters for Export Policy
The timing is delicate. Nvidia's China business has been in flux all year, with H200 shipments resuming to Chinese customers under eased restrictions even as Washington scrutinizes enforcement, and the Commerce Department facing criticism over chip diversion. If Blackwell-class hardware has been reaching Chinese labs at scale since at least April, the gap between export policy and practice becomes a central question for the next round of policy debate.
The finding also has a verification angle for the AI industry itself. Model weights have typically been scrutinized for safety evaluations and benchmark claims; using them as forensic evidence of training infrastructure is a novel addition to the toolkit, and one that does not depend on cooperation from the labs involved.
For Nvidia, the report lands during an already turbulent period in which the company's board has been consumed by an unrelated governance crisis. For the Chinese labs named, none of which are accused of wrongdoing beyond the inference the fingerprints support, the analysis will likely prompt questions about their compute supply chains that their public communications have so far not addressed.
The full report, including lab histories, the complete model list, and its sources, is available from Unemployment Labs. The work drew significant attention on Hacker News, where it reached 277 points within hours of publication.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
