Banned Nvidia Blackwell Chips Power China's Top AI Models: Investigation

Executive Overview

A former OpenAI researcher has uncovered evidence that China's most powerful open-weight AI models—including Moonshot's Kimi K3 and DeepSeek V4—were trained on Nvidia Blackwell chips, which are banned from export to China. The findings, led by Will DePue of the nonprofit Unemployment Labs, arrive amid a scandal involving Nvidia CEO Jensen Huang's resignation after admitting a relationship with Alice Huang, CEO of Megaspeed, a company accused of chip smuggling. Epoch AI estimates that nearly 2 million advanced chips entered China during 2024-2025 via schemes to conceal final destinations, while Nvidia shipped 275,000 AI accelerators to Malaysia in the same period. These developments place U.S. export control policies under existential scrutiny.

Investigation Details: How Weights Reveal Chips

Will DePue relied on analyzing model weights—the unique structure that reflects the configuration of chips used in training. He found compelling evidence that Moonshot's Kimi K3 and DeepSeek V4 models bear training fingerprints from Blackwell chips. Earlier this year, White House technology adviser Michael Kratsios stated that Moonshot used export-banned Nvidia chips, based on U.S. intelligence sources.

Epoch AI estimates that approximately 2,000,000 advanced chips entered China during 2024 and 2025 through schemes to conceal final destinations, while Malaysia alone imported 275,000 Nvidia AI accelerators in the same period. This does not necessarily mean the models were entirely trained on Nvidia chips or that export control laws were violated, as the chips could be in an overseas data center—a common pathway for Chinese companies to access computing.

Megaspeed, led by Alice Huang, has been at the center of this trade for years through its subsidiary Speedmatrix, and is under investigation by both the U.S. and Singapore governments. Bloomberg reported on Friday that a company linked to Alice Huang was involved in smuggling banned Nvidia chips to a data center in Shanghai. In May, short-seller Culper published a report arguing that Nvidia has a "China problem" because a large share of its revenue is driven by Chinese demand through remote data centers or smuggling.

Benchmark & Competitive Performance

Blackwell chips are the most powerful in Nvidia's portfolio, while the Trump administration allows export of the less powerful H200 chip on the condition of an indirect shipping route that enables a 25% tariff. Reports indicate that approximately 600,000 H200 chips are stored in Hong Kong awaiting release for import into China, though observers disagree on the reasons for the delay. Graham Webster of Stanford University says he is not convinced that holding back chips will drive domestic production, given the sheer scale of demand. Other theories include preparation for a potential trade deal or concerns over potential monitoring technologies in the chips.

Industry Impact & Enterprise Adoption

The revelations carry significant implications for developers and enterprises relying on open-weight models. If confirmed, the training of frontier models on banned hardware would undermine the effectiveness of U.S. export controls and raise questions about the integrity of the global AI supply chain. For enterprises, the availability of these models globally—without regional restrictions on weights—means that access to cutting-edge AI capabilities may not be curtailed by chip export bans, but the geopolitical and security risks are substantial. The scandal surrounding Jensen Huang and Megaspeed further complicates Nvidia's position, as U.S. and Singaporean investigations into alleged smuggling networks intensify. The AI industry now faces a critical test: whether export controls can be enforced effectively in an era of distributed computing and global data centers.

Conclusion

The evidence that China's leading open-weight models were trained on export-banned Nvidia Blackwell chips highlights the challenges of enforcing technology restrictions in a globalized AI landscape. With millions of advanced chips reportedly entering China through indirect routes and allegations of smuggling involving senior Nvidia leadership, U.S. export control policies are at a crossroads. The outcome will shape not only the future of AI development in China but also the integrity of international technology governance.

Media Source: TechCrunch AI | Fact Verification & Analysis: AI Tools Oasis