# Study says top Chinese AI models were trained on Nvidia Blackwell chips the US bars from export

Published 2026-11-15T07:06:47.000Z · AI Pulse · https://ai-pulse.news/policy-law/chinese-ai-models-nvidia-blackwell-export

*This story was written with the help of artificial intelligence, checked automatically, then reviewed by an editor before publication, under our AI policy : https://ai-pulse.news/ai-policy*

Unemployment Labs, in a report dated 12 November 2026, says six models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group show numerical traces of Nvidia Blackwell hardware, but it does not establish an export-control violation.

## Key points

- Unemployment Labs' report of 12 November 2026 says six Chinese models released since April 2026 match Blackwell hardware patterns, while earlier models match the Hopper generation.
- The group says its analysis proves access to Blackwell hardware but not where the chips were, how access was obtained, or whether export controls were broken.
- Its highly uncertain estimate is about 12,000 Blackwell GPUs across five training runs, with Kimi K3 the largest at about 4,500.
- Epoch AI data shows a gap between Chinese import values and Malaysian export values for servers, which it calls consistent with smuggling cases but not proof.
- TechCrunch's coverage also reports allegations around Megaspeed and Alice Huang, which remain unproven.

## What the report claims

According to [TechCrunch](https://techcrunch.com/2026/11/13/export-banned-nvidia-chips-are-behind-chinas-most-powerful-ai-models-uh-oh-jensen/), a new study asserts that leading open models from Chinese companies were trained on Nvidia chips that the United States has banned from export to China.

The work was carried out by Will DePue, a former OpenAI researcher who runs the non-profit research group Unemployment Labs. He analysed model weights from DeepSeek, Moonshot AI, MiniMax and Ant Group and found indications of training on Blackwell chips.

The [Unemployment Labs report](https://unemploymentlabs.com/checkpoint-forensics), dated 12 November 2026, says six models released by these companies since April 2026 carry numerical patterns matching Blackwell hardware. Their earlier models, it says, match the Hopper generation.

## What the analysis does not show

The authors themselves limit the conclusion. Unemployment Labs states that its analysis proves access to Blackwell hardware. It does not establish where the chips were located, how access was obtained, or whether export controls were violated.

That distinction matters for policy. The finding is a claim by a single research group, published in its own report and covered by TechCrunch. The available information includes no response from the companies named, from Nvidia or from regulators.

Unemployment Labs also offers a scale estimate, with what it describes as high uncertainty: about 12,000 Blackwell GPUs across the five training runs studied. The largest, Kimi K3, would have needed roughly 4,500 GPUs. How the six models relate to the five training runs is not explained.

## Context: smuggling routes and stockpiles

Separate data points on chip diversion sit alongside the study. Epoch AI reports that between April 2024 and June 2025, China recorded roughly $3.75 billion of servers imported from Malaysia, at about $106,000 per machine. Malaysia declared only about $17,000 per machine on the export side.

Epoch AI says this flow does not prove smuggling but is consistent with established cases. It estimates the flow could represent about 150,000 H100 equivalents. The flow collapsed after Malaysia required an export permit for AI chips in July 2025.

TechCrunch also reports that around 600,000 H200 chips are said to be stockpiled in Hong Kong awaiting shipment to China, with no consensus on why they are held up.

## Megaspeed, Alice Huang and Nvidia's leadership

TechCrunch reports that Jensen Huang left Nvidia in the week of its article after admitting an affair with Alice Huang, an executive at Megaspeed, a company suspected of smuggling Nvidia chips to China. TechCrunch also says Bloomberg reported that a company linked to Alice Huang was involved in smuggling banned Nvidia chips to a data centre in Shanghai.

TechCrunch additionally describes allegations that Alice Huang was an agent of Chinese intelligence services. These allegations are unproven.

None of these items, as presented, links Megaspeed to the specific models in the Unemployment Labs report. The connection between the corporate story and the forensic findings is not established.

## Open questions

- Where the Blackwell chips used for training were physically located.
- How the Chinese developers obtained access, whether by purchase, rental or another route.
- Whether any export-control rules were broken, which the report explicitly leaves open.
- How reliable the weight-based hardware attribution is, since it has so far been presented by one group.

**Epoch AI figures on servers imported by China from Malaysia, April 2024 to June 2025** (data as of 2025-07-01)

| Measure | Value |
| --- | --- |
| Servers recorded as imported by China | about $3.75 billion |
| Import value per machine (China's records) | about $106,000 |
| Export value per machine (Malaysia's declaration) | about $17,000 |
| Estimated H100 equivalents (Epoch AI) | about 150,000 |

## FAQ

### Does the study prove Chinese firms broke US export rules?

No. Unemployment Labs says its analysis shows access to Blackwell hardware but not where the chips were, how access was obtained, or whether export controls were violated.

### Which companies' models were analysed?

Models from DeepSeek, Moonshot AI, MiniMax and Ant Group, according to TechCrunch's account of Will DePue's analysis.

### How many chips does the report estimate were used?

About 12,000 Blackwell GPUs across five training runs, with high uncertainty. Kimi K3, the largest, needed about 4,500.

## Sources

1. [Export-banned Nvidia chips are behind China’s most powerful AI models-uh oh, Jensen!](https://techcrunch.com/2026/11/13/export-banned-nvidia-chips-are-behind-chinas-most-powerful-ai-models-uh-oh-jensen/), TechCrunch, 2026-11-13
2. [https://unemploymentlabs.com/checkpoint-forensics](https://unemploymentlabs.com/checkpoint-forensics), unemploymentlabs.com, 2026-11-12 (primary source)
3. [https://epoch.ai/data-insights/malaysia-china-chip-smuggling](https://epoch.ai/data-insights/malaysia-china-chip-smuggling), epoch.ai, 2025-07-01 (primary source)
