A former OpenAI researcher says he has discovered technical evidence suggesting that six artificial intelligence models developed by Chinese companies were initialized using Nvidia’s export-controlled Blackwell processors. The findings come from a forensic investigation by Will DePue, founder of research group Unemployment Labs, which examined numerical patterns hidden inside publicly available AI model weights. The investigation identified fingerprints associated with Nvidia’s Blackwell B200, B300, GB200 and GB300 hardware in models developed by DeepSeek, Moonshot AI, MiniMax and Ant Group.

The findings could have significant implications for US restrictions on advanced semiconductor exports to China. However, the research does not establish that the companies physically obtained restricted chips inside China or violated export-control laws. The processors could have been accessed through overseas data centres or cloud computing providers. The research also has limitations that independent investigators say still need to be examined, making it important to distinguish the technical evidence from conclusions about how the hardware was accessed.

Six Chinese AI Models Show Nvidia Blackwell Fingerprints

According to DePue’s November 12 report, six AI models released since April 2026 contain numerical patterns matching Nvidia’s Blackwell GPU architecture.

The models come from four Chinese technology companies:

CompanyAI modelReported hardware fingerprint
DeepSeekDeepSeek-V4-FlashNvidia B200/B300
DeepSeekDeepSeek-V4-ProNvidia GB200/GB300
Moonshot AIKimi K3Nvidia GB200/GB300
Moonshot AIKimi K3.1-ProNvidia GB200/GB300
MiniMaxMiniMax-M3Nvidia B200/B300
Ant GroupLing-3.1-flashNvidia GB200/GB300

The investigation suggests that these companies moved from older Nvidia Hopper-generation hardware to Blackwell for some of their newer AI models.

For comparison, earlier models such as DeepSeek-V3, Kimi K2.5, MiniMax-M2.7 and Ant Group’s Ling-3.0-flash showed patterns associated with Hopper-generation processors.

The findings do not suggest that every major Chinese AI developer has adopted Blackwell. The report found Hopper fingerprints in the latest examined models from Alibaba, Zhipu, Tencent, Xiaomi and StepFun. It also found no Nvidia fingerprint in Meituan’s LongCat-2.0, which the company says was trained on its own AI accelerator infrastructure.

How Did the Researcher Discover Hidden GPU Fingerprints?

The investigation uses an unusual property of how AI models are initially created.

An AI model contains billions of numerical values called weights. These values determine how the model processes information and generates responses.

Before training begins, the weights need initial values. Developers commonly use a pseudorandom number generator to create them. Although the numbers appear random, they are generated through a repeatable mathematical process controlled by a starting value known as a seed.

The important detail is that different GPU architectures distribute these calculations across different numbers of processing units.

Nvidia’s Hopper and Blackwell generations have different streaming multiprocessor configurations:

  • Hopper H100-class GPUs use 132 streaming multiprocessors.
  • Blackwell B200 and B300 GPUs use 148.
  • Blackwell GB200 and GB300 configurations use 152.

Because the initialization process distributes numbers differently depending on the hardware configuration, it can leave a recognizable pattern in the original model weights.

Training changes most of these values, but some unused or untouched portions of the model can preserve traces of their original initialization.

DePue used these surviving patterns to work backwards and identify the likely hardware configuration and seed used when the models were created.

He says the investigation examined more than six trillion combinations of hardware layouts and seed values.

DeepSeek-V4-Flash Provides the Strongest Evidence

One of the most detailed examples in the report involves DeepSeek-V4-Flash.

DePue says he recovered the seed used during the model’s initialization and reproduced the corresponding numerical patterns using an Nvidia B300 processor.

Out of 469 selected rows examined, 461 showed correlations above 0.99. Among them, 445 rows matched bit for bit.

The researcher also tested the same process using an older Nvidia H100 processor. It did not reproduce the corresponding patterns found in DeepSeek-V4-Flash.

A similar reproduction exercise involving Moonshot AI’s Kimi K3 reportedly matched preserved values using a rented GB200 processor.

The report also says its methodology correctly identified the publicly disclosed hardware for 24 control models.

These results form the central technical basis for the investigation’s claims.

However, the strength of the evidence differs between models. The physical GPU reproduction was particularly detailed for DeepSeek-V4-Flash and Kimi K3, while some of the other conclusions rely more heavily on statistical comparisons.

The Training Runs Could Have Required 12,000 Blackwell GPUs

The report also attempts to estimate how much computing infrastructure the five major training runs required.

Based on model sizes, estimated training workloads and assumptions about GPU utilisation, DePue estimates that approximately 12,000 Blackwell GPUs were used across five training runs.

The estimates exclude the Kimi K3.1-Pro model, which the report suggests may be a further post-training checkpoint of Kimi K3 rather than an entirely separate large training run.

AI modelEstimated Blackwell GPUsEstimated training duration
DeepSeek-V4-Flash1,50050 days
DeepSeek-V4-Pro3,00070 days
Kimi K34,50070 days
MiniMax-M31,50060 days
Ling-3.1-flash1,500100 days
Total12,000—

The researcher estimates that the combined infrastructure could have cost more than $500 million, assuming an approximate price of $70,000 per Blackwell GPU.

However, these figures are estimates rather than confirmed purchases or verified hardware inventories.

They depend on assumptions about training efficiency, model token counts, computing utilisation and the amount of time each training run required. The report itself acknowledges considerable uncertainty around the GPU estimates.

Why the Findings Matter for US Export Controls

The investigation arrives amid continuing US efforts to restrict China’s access to advanced semiconductor technology.

Washington has introduced restrictions on exports of high-performance AI processors to China, including Nvidia’s most advanced data-centre hardware.

Blackwell processors are particularly important because they are designed for large-scale AI training and inference workloads.

The report’s findings raise questions about whether Chinese AI developers can still access restricted computing infrastructure through alternative arrangements.

One possibility is overseas cloud computing.

A Chinese AI company does not necessarily need to physically possess a Blackwell processor to use its computing capacity. It could potentially rent access from a data centre located outside mainland China.

This distinction is crucial.

The investigation identifies evidence associated with the hardware used to initialize model weights. It does not establish where the processors were located, who owned them or how the companies accessed them.

Consequently, the research cannot independently establish that Nvidia chips were smuggled into China or that a particular company violated US export restrictions.

Independent Verification Still Has Important Gaps

The research has attracted attention, but independent analysis has identified areas that require further verification.

A November 14 review by New Runtime found that an outside developer could reproduce the report’s published results for five model checkpoints using its released code and data.

However, that exercise did not independently repeat the original hardware tests.

The review also found that the public repository did not include two of the six models highlighted in the report: DeepSeek-V4-Pro and Kimi K3.1-Pro. It also did not contain the full set of 24 control models used in DePue’s validation claims.

Another technical complication involves AMD processors.

The report acknowledges that specially configured AMD MI300X hardware could theoretically produce similar processing-unit layouts. DePue says he used numerical differences between AMD and Nvidia calculations to rule out this alternative for DeepSeek-V4-Flash, but could not conclusively exclude it for the other five models.

The distinction does not invalidate the investigation, but it means the six-model conclusion should not be treated as independently confirmed in every case.

The Chinese companies named in the report and Nvidia had not publicly provided detailed responses to the findings in the cited coverage.

The Bigger Picture

The investigation introduces a new dimension to the debate over semiconductor export controls. Until now, determining which processors were used to train a particular AI model has often depended on company disclosures, infrastructure reports, supply-chain investigations and government assessments. DePue’s approach suggests that publicly released model weights can themselves contain clues about the hardware used during development.

If the technique withstands independent verification, it could become a useful tool for examining AI supply chains. Researchers may be able to identify hardware patterns even when companies do not disclose their computing infrastructure. At the same time, the findings highlight the limitations of export restrictions when computing resources can potentially be accessed through international cloud providers.

Looking Ahead

The next important step will be independent reproduction of the original experiments using physical Nvidia Blackwell and AMD hardware. Researchers will also need to examine the models that are not included in the public verification repository and determine whether the reported fingerprints can be consistently reproduced. Responses from the Chinese AI companies, Nvidia and relevant US authorities could provide additional clarity about the findings and the infrastructure involved.

For the AI industry, the investigation raises a broader question about transparency in model development. As advanced processors become increasingly important to national technology strategies, the ability to identify computing hardware through published model weights could attract attention from researchers, regulators and semiconductor companies. However, identifying a hardware fingerprint is not the same as proving an export-control violation, and the distinction will remain central as the investigation receives further scrutiny.

Get the day’s top stories in your inbox

One concise email. No spam, unsubscribe anytime.