The honest view from inside technical leadership
Analysis

Engineering leadership

Independent reporting on the decisions, costs and contracts that shape how engineering organizations build with AI.

AI Governance · News

Six Chinese AI Models Carry Nvidia Blackwell Fingerprints in Their Published Weights

A random number generator spreads its work across a GPU’s processors, and the layout survives in weights that training never touches. Read back from public checkpoints, that layout points to export-controlled Blackwell hardware behind the newest Chinese open-weight models.

Line-art illustration of a GPU chip under a magnifying glass beside a grid of numbered weight cells, representing Blackwell GPU fingerprints found in Chinese AI models

Six Chinese AI models released since April 2026 carry Nvidia Blackwell GPU fingerprints in their published weights, according to a forensic report by Will DePue of the nonprofit Unemployment Labs. The models come from DeepSeek, Moonshot AI, MiniMax and Ant Group, and Blackwell chips have required a U.S. export license for China since before they shipped, with none reported granted.

Blackwell GPU Fingerprints Appear in Six Chinese AI Models From Four Labs

The report, published on 12 November 2026, identifies Blackwell-compatible patterns in DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash. DeepSeek-V4-Flash and MiniMax-M3 match the 148-processor layout of the B200 and B300. The other four match the 152-processor layout of the GB200 and GB300. Earlier models from the same four companies match Nvidia’s Hopper generation, which includes the H800 that DeepSeek and Moonshot have disclosed using before. DePue, a former OpenAI researcher, analyzed only openly published checkpoints and relied on no leaks or insider accounts, TechCrunch’s Tim Fernholz reported on 13 November.

PyTorch Leaves the Chip’s Processor Count Inside the Starting Weights

Every model starts from random values, and PyTorch generates them with a deterministic algorithm called Philox through Nvidia’s cuRAND library. The work is split across 2,048 threads for each streaming multiprocessor, so a chip with more multiprocessors places the same numbers in different positions. Hopper H100, H200 and H800 chips have 132 multiprocessors and use 270,336 threads. The B200 and B300 have 148 and use 303,104. The GB200 and GB300 have 152 and use 311,296. Some weights, such as rows for tokens the model never saw in training, keep their starting values almost exactly. DePue searched roughly six trillion combinations of seed and thread layout to find the pair that recreates those surviving rows, then tested each winning pair on weights held out of the search.

Physical Blackwell Chips Reproduced DeepSeek and Kimi Weights Bit for Bit

The team confirmed two results on real hardware with unmodified PyTorch. DeepSeek uses one seed, 953049583, across model generations. On a Blackwell B300, that seed reproduced 461 of 469 selected rows in DeepSeek-V4-Flash with correlations above 0.99, and 445 of those rows matched bit for bit. On a Hopper H100, the same seed reproduced all 57,344 preserved values in the older DeepSeek-V3. Neither chip could reproduce the other model, and changing the seed by one produced no matching rows. On a rented GB200, Moonshot’s seed reproduced all 13 preserved rows in Kimi K3. The method also agreed with all 24 cases where a developer has publicly named its training hardware, including Meta’s Llama 3.1 on H100s and Nvidia’s own Nemotron-3 Ultra on GB200s.

The Statistical Evidence Clears an 8-Sigma Bar by a Wide Margin

Because a search over trillions of candidates will produce some lucky matches, the report sets roughly 8 sigma as its threshold for ruling out chance. The strongest held-out evidence for the six models ranges from 49 to 1,385 sigma, and the weakest individual test reaches 20 sigma. The one alternative chip the team found is an AMD MI300X, which is also export-controlled and can reproduce the Blackwell layouts when deliberately limited to 148 or 152 compute units. For DeepSeek-V4-Flash, the team compared 377 values where AMD and Nvidia compute differently. The published weights matched Nvidia in 363 and AMD in none. DePue considers the AMD explanation very unlikely for the other five models.

The Fingerprint Shows Hardware Access and Leaves Location and Legality Open

The report is explicit that the weights cannot show where the chips sat, who owned them or how a lab got access. Blackwell systems can be rented from data centers outside China, which TechCrunch notes has become a common route for Chinese firms, and Epoch AI estimates 275,000 Nvidia AI accelerators were imported to Malaysia in 2024 and 2025. The fingerprint also records only the hardware that created the starting weights. A lab could initialize on one cluster and train on another. DePue still concludes that the evidence is strong for all six models and conclusive for DeepSeek-V4-Flash, and he writes that “this fact deserves urgent attention from the public.” The report lands days after Jensen Huang left Nvidia, having admitted to an affair with Alice Huang, an executive at Megaspeed, a company TechCrunch describes as implicated in smuggling Nvidia chips into China.

Why This Matters for AI Governance and Model Procurement

Provenance claims about open-weight models can now be tested by anyone with the checkpoint and a laptop. The code and preserved rows for five of the six models are public, and the report says the check takes seconds. Teams that approve open-weight models for internal use already ask about license terms and training data. Hardware provenance belongs on the same intake form, because it bears on sanctions and export exposure, and because it is evidence a lab’s public statements can be checked against. The method has hard limits that a policy should state plainly: closed models cannot be checked, CPU initialization and Baidu’s PaddlePaddle framework leave no trace, and a Blackwell fingerprint says nothing about location or legality. The same mechanism applies to models your own team publishes. If you release weights initialized with PyTorch on Nvidia GPUs, the seed and your chip class are recoverable from the file, so treat a weights release as a disclosure of your hardware and review it that way.

FAQ

Which Chinese AI models show Blackwell GPU fingerprints?

DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash. All six were posted since April 2026, and earlier models from the same four companies match Nvidia’s older Hopper chips.

How can model weights reveal which GPU created them?

PyTorch fills starting weights using a random number generator split across 2,048 threads per streaming multiprocessor, so the chip’s processor count decides where each number lands. Weights that training never updates keep that arrangement.

Does the study prove export controls were violated?

No. The report says the weights cannot show where the chips were, who owned them or how access was obtained, and Blackwell systems can be rented outside China.

Does a Blackwell fingerprint prove the whole model was trained on Blackwell?

It shows which hardware created the starting weights. The report says this strongly suggests training on the same hardware, but a lab could initialize on one system and train on another.

Could AMD chips produce the same pattern?

An AMD MI300X limited to 148 or 152 compute units reproduces the layouts. For DeepSeek-V4-Flash, 363 of 377 vendor-sensitive values matched Nvidia’s output and none matched AMD’s.

Can the results be checked independently?

Yes. The code and preserved weight rows for five of the models are public on GitHub, and the report says the check runs on a laptop in a few seconds.

Source: Export-banned Nvidia chips are behind China’s most powerful AI models by Tim Fernholz, TechCrunch, November 13, 2026. Primary source: Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs by Will DePue, Unemployment Labs, November 12, 2026, with reproduction code on GitHub.

Read more news & analysis