Are Export-Banned Nvidia Chips Powering China’s Top AI Models?

Introduction

U.S. export controls on advanced AI accelerators are designed to limit China’s access to the computing power needed to train frontier models. But a new research claim points to a difficult loophole: even if restricted chips are not shipped directly to mainland China, Chinese companies may still be able to use them through overseas facilities or indirect supply channels.

TechCrunch AI reports that some of China’s leading open-weight models may have been trained on Nvidia Blackwell GPUs, which are restricted from export to China and Chinese-owned entities. The claim is based on hardware-related patterns inferred from model weights rather than on complete procurement records, data-center logs, or training documentation. It should therefore be treated as an indication requiring independent verification, not as conclusive proof of a violation.

Key points

• Will DePue, a former OpenAI researcher who leads Unemployment Labs, analyzed the weights of models released by DeepSeek, Moonshot AI, MiniMax, and Ant Group.
• The analysis reportedly found features associated with Blackwell training in Moonshot’s Kimi K3 releases and DeepSeek’s V4 models.
• The models may not have been trained entirely on restricted Nvidia hardware, and the GPUs could have been installed in offshore data centers.
• Epoch AI estimates that roughly two million advanced chips entered China-related supply chains in 2024 and 2025 through schemes that concealed their final destination. That estimate is not, on its own, evidence of illegal conduct.
• Existing controls focus heavily on who owns a chip and where it is installed, while monitoring who actually consumes the resulting compute remains more difficult.

The offshore compute problem

The material cites an estimate that about 275,000 Nvidia AI accelerators were imported into Malaysia during 2024 and 2025. Overseas data centers have consequently become a potential route for Chinese companies seeking access to restricted computing capacity. The policy question is whether a company should be allowed to use high-end accelerators simply because the machines sit outside China, even when the resulting model development serves a Chinese customer.

The report also discusses Megaspeed and its subsidiary Speedmatrix, companies that have reportedly faced scrutiny from U.S. and Singaporean authorities. A separate Bloomberg report allegedly linked a company associated with the same network to the movement of restricted Nvidia chips to a data center in Shanghai. These remain reported allegations in the supplied material, not findings that have been established by a court or regulator.

Why it matters

If the hardware-signature research is independently replicated, export controls may need to evolve from a system centered on physical shipments toward one that also governs compute usage. Possible tools could include stronger customer identification by cloud providers, audits of remote accelerator use, disclosure of data-center beneficial owners, and records of cross-border compute transactions. For Nvidia, the risk would extend beyond product sales to distributor oversight, customer due diligence, and corporate compliance.

Tighter rules would also create costs. Global data-center operators and chipmakers could face more complex cross-border screening, while legitimate customers may encounter delays or higher compliance burdens. More fundamentally, a model’s technical fingerprint cannot by itself establish where a chip came from, who controlled it, or whether the law was broken. Weight analysis, trade records, and enforcement investigations will need to corroborate one another before the industry can determine whether the controls failed—and how.

Source: TechCrunch AI