Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
InfrastructureOpen Weights · ActionableTG: @data_secrets (Data Secrets)→ original

Former OpenAI Researcher Claims to Find Traces of Banned Blackwell Chips in Six Chinese AI Models' Weights

The US has banned exports of Nvidia Blackwell chips to China, but it is hard to verify what hardware Chinese labs train their models on. A former OpenAI employee claims that part of the GPU's fingerprint remains in the weights: the initial values depend on the seed and the chip. After checking about 6 trillion combinations, he says he found Blackwell traces in six models from four Chinese labs. So far, this is one author's claim, but if confirmed, the method would let regulators check sanctions compliance using open weights.

AI-processed from TG: @data_secrets (Data Secrets); edited by Hamidun News
Former OpenAI Researcher Claims to Find Traces of Banned Blackwell Chips in Six Chinese AI Models' Weights
Source: TG: @data_secrets (Data Secrets). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @data_secrets (Data Secrets) ↗
Editor-in-Chief
DeepSeek-V4-Flash and MiniMax-M3traces of B200 or B300
Kimi K3 and DeepSeek-V4-Pro2–3 thousand GPUs, a cluster of about $200 million
Key metric #3Earlier models from the same labs read as Hopper

Blackwell is Nvidia's newest generation of accelerators, and the US has banned shipping such chips to China. The problem is enforcement: purchases can be concealed, hardware can be rented through third countries, and a finished model looks the same no matter what hardware trained it. Now an independent researcher, a former OpenAI employee, claims he has found a way to identify the training chip directly from open weights. According to his data, six models from four Chinese labs carry Blackwell traces. This is so far one author's claim, but if it is confirmed, the implications for sanctions would be serious.

How a chip leaves a fingerprint in the weights

Before training, a model's weights are filled with random numbers. These initial values depend on the seed (the number that sets the randomness) and on the specific chip on which they were generated: different GPU architectures round floating-point numbers slightly differently. The conventional view is that the weeks of training completely erase this initial pattern. The author of the analysis argues the opposite: part of the pattern survives training and remains in the finished weights.

Think of it as a manufacturing defect in a printing press. Each print looks normal, but if you know how a specific machine "errs," you can identify the press from the print run. The researcher went through about six trillion combinations of chips and seeds, looking for the one that best matches the weights of the published models.

What he found

According to his results, matches exist for six models:

  • DeepSeek-V4-Flash and MiniMax-M3 were trained on B200 or B300;
  • DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash were trained on GB200 or GB300;
  • earlier models from the same labs read as trained on Hopper, that is, on previous-generation chips.

This means the fingerprint at these Chinese labs shifts from Hopper to Blackwell, and Blackwell is the chip family under export ban to China.

Next, the author estimated the scale. By his calculations, Kimi K3 and DeepSeek-V4-Pro each used between two and three thousand Blackwell GPUs. A cluster of that size costs about $200 million. That is full production infrastructure, not a small batch for experiments. The author adds that researchers from OpenAI and Anthropic checked the results. The source contains no public confirmation from those companies. The author has published details at unemploymentlabs.com.

What it means, and where the weak spots are

If the method holds up, regulators gain a tool that does not require access to a data center. Publicly released open weights are enough. That could become an argument in the sanctions debate. Publishing practices may also change: some teams may choose to add noise to or reinitialize their models.

But the story has caveats:

  • There is one source: the author's publication and its retelling on Telegram. The material contains no independent public verification.
  • A match with Blackwell alone does not prove a violation of the law. What matters is where the chips physically were: in China, at a foreign cloud provider, or somewhere else. Legal consequences depend on that.
  • The method rests on the assumption that the initial pattern survives all of training. It needs testing on models whose chips are known in advance, and the false-positive rate has to be measured.
  • The figures of 2–3 thousand GPUs and $200 million are the author's estimates, not statements from the labs.

What developers and product managers should do

First, open weights reveal more than their authors may assume: they can tell you something about the hardware and, possibly, about the model's origin. Second, if you build products on Chinese open models, watch how regulators respond: sanctions risk against the labs could also reach those who use their models. Third, wait for confirmation or refutation from major labs and independent groups. Until then, treat the story as a loud claim, not a settled finding.

Story

Key Questions Answered

How a chip leaves a fingerprint in the weights?

Before training, a model's weights are filled with random numbers. These initial values depend on the seed (the number that sets the randomness) and on the specific chip on which they were generated: different GPU architectures round floating-point numbers slightly differently. The conventional view is that the weeks of training completely erase this initial pattern. The author of the analysis argues the opposite: part of the pattern survives training and remains in the finished weights.

What he found?

According to his results, matches exist for six models: DeepSeek-V4-Flash and MiniMax-M3 were trained on B200 or B300; DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash were trained on GB200 or GB300; earlier models from the same labs read as trained on Hopper, that is, on previous-generation chips. This means the fingerprint at these Chinese labs shifts from Hopper to Blackwell, and Blackwell is the chip family under export ban to China. Next, the author estimated the scale.

What it means, and where the weak spots are?

If the method holds up, regulators gain a tool that does not require access to a data center. Publicly released open weights are enough. That could become an argument in the sanctions debate. Publishing practices may also change: some teams may choose to add noise to or reinitialize their models. But the story has caveats: There is one source: the author's publication and its retelling on Telegram. The material contains no independent public verification.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…