# Ex-OpenAI researcher finds Nvidia Blackwell fingerprints in six Chinese open models. What the weights prove, and what they don't

> Will DePue says the random-number layout left in published weights shows DeepSeek V4, Kimi K3, MiniMax-M3 and Ant's Ling-3.1-flash were initialised on export-controlled Blackwell GPUs. The method is clever and the code is public, but it can't show where the chips were or whether any law was broken.

Published Nov 15, 2026 by TechMeasured Staff · AI · Canonical: https://techmeasured.com/ai/blackwell-fingerprints-six-chinese-open-models-deepseek-kimi-what-it-proves

**Former OpenAI engineer Will DePue says he has found forensic traces of Nvidia's export-controlled Blackwell GPUs in the published weights of six Chinese AI models from DeepSeek, Moonshot AI, MiniMax and Ant Group.** His [report](https://unemploymentlabs.com/checkpoint-forensics), published November 12 by his research outfit Unemployment Labs, says earlier models from the same labs carry the signature of Nvidia's older Hopper chips. That is the pattern you would expect if the labs moved to newer hardware this year.

US officials and journalists have long suspected some Chinese labs had Blackwell access, but, as [The Decoder](https://the-decoder.com/former-openai-researcher-says-he-found-nvidia-blackwell-fingerprints-hidden-inside-six-chinese-ai-models/) notes, nobody had published technical proof. The report also lands in the middle of Nvidia's China crisis. [The Register](https://www.theregister.com/ai-and-ml/2026/11/13/chinese-ai-labs-are-training-models-on-nvidias-blackwell-despite-export-controls-says-ex-openai-engineer/5309891) ties it to Jensen Huang's resignation and to four pending bills that would tighten chip export rules. According to The Register, this is the first public link between MiniMax and Ant Group and Blackwell hardware.

## How the fingerprint works

Before training, PyTorch fills a model's weights with random numbers. It does this with the Philox generator through Nvidia's cuRAND, and it spreads the work across GPU threads. The thread count depends on how many streaming multiprocessors (SMs) the chip has, so the same seed lands values in different places on different chips:

| Hardware | SMs | Threads |
|---|---|---|
| Hopper (H100 SXM, H200, H800) | 132 | 270,336 |
| Blackwell B200/B300 | 148 | 303,104 |
| Grace Blackwell GB200/GB300 | 152 | 311,296 |

Some weights barely change during training, such as embeddings for unused vocabulary tokens, so they still hold their starting values. DePue searched about **six trillion** seed-and-layout combinations to find the ones that reproduce those values. He picked candidates on one set of weights and tested them on held-out weights. He sets about 8 sigma as the level chance matches could reach across a search that size.

## What he found

| Lab | Model | Layout | Strongest evidence (per report) |
|---|---|---|---|
| DeepSeek | V4-Flash (Apr 2026) | 148 SM | Exact copy, 1,385σ |
| DeepSeek | V4-Pro (Apr 2026) | 152 SM | Partial survival, 49σ |
| Moonshot AI | Kimi K3 (Jul 2026) | 152 SM | Exact copy, 305σ |
| Moonshot AI | Kimi K3.1-Pro (Nov 2026) | 152 SM | Same seed as K3 |
| MiniMax | M3 (Jun 2026) | 148 SM | Exact copy, 323σ |
| Ant Group | Ling-3.1-flash (Oct 2026) | 152 SM | Untrained rows, 244σ |

Qwen, Zhipu's GLM, Tencent, Xiaomi and StepFun models, plus earlier DeepSeek and Moonshot releases, read as Hopper. The method can't tell an H800 from an H100 or H200.

**Checks he ran.** Unmodified PyTorch on an H100 reproduced all 57,344 preserved values in DeepSeek-V3 bit for bit. On a B300 it reproduced 461 of 469 selected rows from V4-Flash, 445 of them bit for bit. Neither chip could reproduce the other model's rows. A rented GB200 reproduced 13 preserved rows from Kimi K3. DePue says the method also matched all 24 models whose training hardware is publicly known, including Meta's Llama 3.1 (H100) and Nvidia's Nemotron-3 Ultra (GB200), though he concedes some of those checks aren't independent. Code and preserved weight rows are on GitHub, and the report says the check runs in seconds on a laptop.

DePue says senior researchers at OpenAI and Anthropic confirmed the result. The Decoder notes it did not verify those confirmations, and The Register's report carried no comment from Nvidia, the four labs or the US Commerce Department.

## How many chips?

The report estimates five of the final training runs used about **12,000 Blackwell GPUs** in total (range 7,000–22,000). Kimi K3 alone accounts for about 4,500 and DeepSeek-V4-Pro for about 3,000. At about $70,000 per GPU, DePue puts the largest cluster at roughly $200–300 million and all five at more than $500 million. He revised these numbers upward on publication day and says he is **"not confident"** in them, because they rest on efficiency, token-count and run-length assumptions no lab has published.

## What the evidence does not show

- **Where the chips were.** Blackwell systems can be rented in data centres outside China. The fingerprint shows access to the hardware, not smuggling. The report says plainly that it does not find any law was broken.
- **Every chip used in training.** It shows the hardware that generated the starting weights. A lab could initialise on one system and train on another. DePue says the pattern "strongly suggests" the same hardware did the training.
- **AMD.** A specially configured AMD MI300X, also restricted, can mimic the 148- and 152-SM layouts. DePue rules it out for V4-Flash using rounding differences (363 of 377 differing values match Nvidia, none AMD) and calls it very unlikely for the rest.
- **Closed models.** Closed models can't be checked at all. Neither can models initialised on CPUs or built with frameworks that leave no pattern.

## Why it matters

This is the first public method anyone can rerun on published weights. For policymakers, it shifts the export-control debate from chip *sales* toward chip *access*, including cloud rental, which one of the pending bills, the Cloud Security Act, targets according to The Register. For labs that publish open weights, it is also a reminder that checkpoints leak more about how they were made than you might expect. Watch for independent replications of the GitHub code and any response from the labs, Nvidia or Commerce.


## Sources

1. [Unemployment Labs — Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs](https://unemploymentlabs.com/checkpoint-forensics)
2. [The Register — Chinese AI labs are training models on Nvidia's Blackwell despite export controls, says ex-OpenAI engineer](https://www.theregister.com/ai-and-ml/2026/11/13/chinese-ai-labs-are-training-models-on-nvidias-blackwell-despite-export-controls-says-ex-openai-engineer/5309891)
3. [The Decoder — Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models](https://the-decoder.com/former-openai-researcher-says-he-found-nvidia-blackwell-fingerprints-hidden-inside-six-chinese-ai-models/)
