Six leading Chinese AI models may have been trained on export-banned US chips, weight analysis suggests

[Update, November 14, 2026, 4:44] Jensen Huang has already stepped down, so his title has been changed from "CEO" to "former CEO." We have also added that the House is seeking a hearing on December 1. The names of the six models, the analyst's background and the confidence figures were previously described as unknown. They have now been corrected to the values given in the analysis report.

An analysis has found that several of China's leading AI models were likely trained on Blackwell, NVIDIA's latest chip, which the United States bars from export to China. The analysis comes from Unemployment Labs, and TechCrunch reported on it on November 13. It covers six Chinese models with publicly released weights, including DeepSeek and Moonshot's Kimi. The clue came from traces of initialization, the step at the start of training, left in weight files that anyone can download. This is the first evidence to identify the hardware used for training from published weights alone. However, the analysis shows only which hardware the models were trained on. It does not show where the chips were or how they were obtained.

For Japanese companies: Most companies are not directly affected. Two groups are: companies that run DeepSeek or Kimi weights on their own servers, and trading firms and resellers that sell NVIDIA GPU servers in Asia. For the first group, IT and legal departments should check this month whether their contracts with the US government or large US companies contain clauses about where the models they use come from. The second group should review how they verify who ultimately uses the machines they sell, before the US Commerce Department issues new rules.

What the analysis found

The United States bans sales to China of NVIDIA's most powerful AI chips, including Blackwell. Chinese labs have often been described as training their models on lower-performance chips cleared for export, or on Chinese chips such as Huawei's Ascend.

The new analysis challenges that assumption with evidence that cannot be altered: the weights themselves. DeepSeek and Kimi are among the strongest of China's open models. Kimi's latest model, K3.1 Pro, had its weights released on November 4. With about 2.8 trillion total parameters, it is the largest open model from any Chinese lab. If export-banned hardware was used to train these models, US export controls have failed to stop China from building frontier models.

How the hardware was identified

When an AI model begins training, its billions of weights are filled with random numbers. The analysis rests on the idea that small quirks of the hardware, and of the software running on it, shape how this is done. Those quirks then remain as traces in the weights even after weeks of training. Because anyone can download the weight files, the findings can be checked without relying on company statements or internal documents.

Until now, the hardware used to train Chinese models could only be inferred from company papers, people familiar with the trade, or smuggling cases. What is new here is that the hardware was identified from the weights themselves.

What remains unknown: where the chips were and how they were obtained

The analysis goes only as far as showing that Blackwell was used for training. It does not show the following.

First, where the chips were. The weights cannot show whether the chips were brought into data centers in China or rented in data centers outside the country, such as in Southeast Asia or the Middle East. US rules focus mainly on banning chip shipments to China. Critics have long pointed to Chinese companies renting chips through cloud services abroad as a loophole.

Second, how the chips were obtained. Whether they were smuggled, resold through a third country, acquired before the rules took effect or rented legally determines whether any law was broken and who is responsible. The analysis says nothing about these questions. It has also not been confirmed whether DeepSeek, Moonshot or NVIDIA responded to the analysis.

Direct bearing on the US export-control debate

The analysis adds new material to a debate already under way in the United States. The Commerce Department is discussing a rule that would ban Chinese companies from renting chips located in other countries. It has also not decided whether to add Megaspeed, which is suspected of passing chips to Chinese companies, to its trade-restriction list (the Entity List). Testimony before Congress by NVIDIA's former CEO Jensen Huang has not been finalized either. So far, the House has requested a closed-door hearing on December 1.

In Congress, the main question is whether three export-control bills will be added to the National Defense Authorization Act (NDAA), the defense spending law due to pass by year-end. One of them, the Chip Security Act, would make it possible to verify where exported AI chips are located. The gap this analysis exposes, the inability to tell where the chips were, is exactly what the bill aims to close. At the APEC leaders' summit on November 18–19, the United States and China will also discuss whether to extend their trade truce.

Chinese labs are pursuing two paths at once

Chinese labs are preparing for both NVIDIA and Chinese chips. DeepSeek made the inference system it released in October run on both NVIDIA's Blackwell-generation GPUs and Huawei's Ascend 950. It has also kept releasing communication and compute components for Ascend. Meituan says it tuned a lightweight version of LongCat-2.5 for Chinese chips.

These moves show Chinese labs shifting toward domestic chips. The new analysis, however, suggests that for the most critical task, training their strongest models, they may still rely on US chips that are banned from export.

What we could not verify

We were unable to read the original analysis page or the full text of the TechCrunch article. The six models are DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ant's Ling-3.1-flash. The analysis was carried out by Will DePue, formerly of OpenAI. It puts its confidence at 49 to 1,385 sigma. This article does not say which Blackwell product the analysis points to. Reactions from NVIDIA, the US government and the Chinese companies named have not been confirmed. We have also found no reports so far of other researchers independently checking the analysis's method.