Will DePue, a former OpenAI researcher who now runs the independent group Unemployment Labs, published a report on November 12, 2026, titled "Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs." The analysis says six open-weight models released since April 2026 by DeepSeek, Moonshot AI, MiniMax, and Ant Group contain numerical patterns that match NVIDIA Blackwell GPUs. Those chips have required a U.S. license for export to China since before they shipped, and the report says no such licenses have been reported.

DePue writes that the evidence comes entirely from the models' published weights, not from leaks or insider accounts. For DeepSeek-V4-Flash and Kimi K3, the report says the match was also confirmed on physical Blackwell chips. Code and selected weight rows for reproducing several of the results are posted in the GitHub repository macrogrants/checkpoint-forensics.

The report is explicit about what the method can and cannot show. It says a match identifies the hardware that generated a model's starting weights and "supports, without proving, the inference that Blackwell also performed the training." It also says the results do not establish where the chips were located, who owned them, or how access was obtained, and that they "are not a finding that any law was broken." Blackwell systems, the report notes, can be rented from data centers outside China.

Key Facts

Item Detail
Publication November 12, 2026, by Will DePue through Unemployment Labs
Models matching B200 or B300 (148 streaming multiprocessors) DeepSeek-V4-Flash (April 2026) and MiniMax-M3 (June 2026)
Models matching GB200 or GB300 (152 streaming multiprocessors) DeepSeek-V4-Pro (April 2026), Kimi K3 (July 2026), Kimi K3.1-Pro (November 2026), and Ant Group's Ling-3.1-flash (October 2026)
Shared initialization The report says Kimi K3.1-Pro uses the same seed as Kimi K3 and may be a later checkpoint of that run with further post-training
Earlier models from the same labs Reported as matching NVIDIA Hopper (132 streaming multiprocessors), a class that includes the H800, H100 SXM, and H200, which the method cannot tell apart
Other Chinese labs checked The newest readable models from Alibaba, Zhipu, Tencent, Xiaomi, and StepFun are reported as Hopper, not Blackwell
Physical checks Unmodified PyTorch on a B300 reproduced 461 of 469 selected DeepSeek-V4-Flash rows; a rented GB200 reproduced all 13 preserved Kimi K3 rows. An H100 reproduced DeepSeek-V3 but none of the tested V4-Flash rows. Changing each recovered seed by one produced no matching rows.
Scale estimates A revised section of the report, if training used the same hardware as initialization, puts five final runs at about 12,000 Blackwell GPUs in total, with a stated range of 7,000 to 22,000. DePue writes that he is not confident in these estimates.

How the Analysis Works

The report examines the starting values assigned to a model's weights before training. In the PyTorch and cuRAND setup it studied, a Philox generator produces repeatable numbers from a seed, and those numbers are spread across GPU threads. On the NVIDIA chips in scope, a large enough initialization uses 2,048 threads per streaming multiprocessor, so a chip with more of those units places the same numbers in different positions. Traces of that layout can remain in published weights, especially in embedding rows for tokens that never appeared in training and therefore kept their initial values.

DePue says he searched roughly 6 trillion combinations of chip layout and seed. Under the report's statistical model, the strongest match expected by chance would reach about 8 standard deviations, or sigma. The strongest results for the six models range from 49 to 1,385 sigma, and the weakest individual test reported reaches 20 sigma. A seed and layout chosen on one set of weights were then tested on a held-out set.

The report also discusses an alternative. An AMD MI300X normally uses 304 processing units, but it can be divided into two groups of 152 or configured to use 148, which could reproduce the same layouts. For DeepSeek-V4-Flash, the report says 363 of 377 preserved values that differ between vendors matched NVIDIA and none matched AMD. That numerical test did not settle the other five models, though the report calls a specially configured AMD cluster very unlikely. The MI300X is also subject to export controls, according to the report.

Background

U.S. officials and earlier press reports had already alleged that some Chinese labs were using Blackwell chips, without publishing technical evidence drawn from model weights. Reuters reported on February 23, 2026, that a senior official in the Trump administration said DeepSeek's then-upcoming model had been trained on Blackwell.

In a June 12, 2025, comment on DeepSeek's DeepGEMM pull request 112, a DeepSeek engineer wrote, "we don't have SM100 GPUs." SM100 is NVIDIA's architecture name for Blackwell. The new report's readings on DeepSeek's 2026 models are, on their face, hard to square with that earlier statement, but DePue's own limits mean this is not by itself proof of an export violation.

The report also cites benchmark results in an Ant Group repository, dated March 31, 2026, and labeled as generated on a GB300. Successive releases in the report show a change over time: DeepSeek models are reported as Hopper through DeepSeek-V3 in December 2024 and as Blackwell from the V4 models in April 2026; Moonshot as Hopper through Kimi K2.5 in January 2026 and as Blackwell from K3 in July 2026; MiniMax as Hopper through M2.7 in April 2026 and as Blackwell from M3 in June 2026; and Ant Group as Hopper in Ling-3.0-flash in August 2026 and as Blackwell in Ling-3.1-flash in October 2026.

What We Don't Know Yet

  • As of this writing, we found no public comment on the report from NVIDIA, DeepSeek, Moonshot AI, MiniMax, or Ant Group.
  • The report does not establish where the chips were used, who owned them, or how access was obtained.
  • A match on starting weights does not inventory every chip used later in training. A lab could initialize a model on one system and train it on another, or use a mixed cluster.
  • On X, DePue said senior researchers at OpenAI and Anthropic had independently verified the result. The written report names no outside reviewers, and neither company has issued a public statement that we found.
  • The GPU-count and cost figures depend on assumptions the weights cannot confirm, including utilization, token counts, run length, and the premise that training stayed on the initializing hardware. The first version of the report used higher-efficiency assumptions and gave smaller cluster sizes; DePue revised those numbers on November 12, 2026, and still describes them as rough.

Sources