business 5 min read

How Banned Nvidia Chips Supercharged China's AI

Jensen Huang is gone, but the scandal he left behind exposes a deeper truth: U.S. chip bans have accidentally accelerated China's AI ambitions. Blackwell GPUs reached frontier labs through Malaysia, and now the country's most capable models bear the fingerprints of banned hardware.

  • Semiconductors
  • NVIDIA
  • US-China Tech Rivalry
  • China AI
  • Export Controls

The Ban That Built Chinese AI

Will DePue had been looking at model weights for months when the pattern became impossible to ignore. The distinctive neural network fingerprints left behind by Nvidia’s Blackwell GPUs were embedded deep inside the open-source releases from DeepSeek, Moonshot AI, MiniMax, and Ant Group — three of China’s most watched frontier labs.

It should not have been possible.

Blackwell chips are banned from export to China under U.S. restrictions designed to keep America’s leading rival from accessing the most powerful AI hardware on Earth. Yet there they were, baked into the architectures of Chinese models now competing on the global stage. The evidence, published this week by Unemployment Labs, suggests the most advanced Chinese AI models were trained on exactly the hardware Washington tried to deny them.

The finding arrives on a day marked by turmoil at Nvidia itself. CEO Jensen Huang stepped down after admitting to an affair with Alice Huang, an executive at Megaspeed — a company repeatedly flagged for smuggling Nvidia chips into China. Two people named Huang. One company. A personal scandal that opens a much larger corporate and strategic one.

Who Gains, Who Loses

For China’s AI sector, the result is unambiguous advantage. DeepSeek’s V4 models and Moonshot’s Kimi K3 releases were built with access to compute resources that include Blackwell clusters. These models are not theoretical exercises — they are live, publicly available systems that Western developers now have to compete against.

For the United States, the consequence is worse than a simple policy failure. Export controls were sold as a way to preserve American dominance in AI capability. Instead, they created a perverse incentive structure: Chinese firms paid premiums for smuggled chips, absorbed the cost, and kept building models that match or exceed what their U.S. counterparts produce.

Malaysia has become the critical node in this system. Approximately 275,000 Nvidia AI accelerators flowed into the country in 2024 and 2025 alone, according to Epoch AI. Some may have been legitimately used in offshore data centers — a popular workaround for Chinese companies seeking compute that is beyond reach domestically. But the distinction between legal overseas hosting and covert diversion is collapsing. Bloomberg reported this week that a firm linked to Alice Huang was involved in moving banned chips directly to a data center in Shanghai.

Michael Kratsios, the White House technology adviser, said earlier this year that intelligence sources confirmed Moonshot had used banned Nvidia chips. That assessment came from official channels. DePue’s research arrived independently, from the models themselves.

The Corporate Fallout

Nvidia sits in an untenable position. The company has publicly vouched for Megaspeed’s data centers, sending officials to visit and confirm legitimacy. It sold billions in chips to intermediaries that then redirected them to Chinese customers. Whether that constitutes a violation of export control law depends on whether investigators can prove Nvidia knew or should have known about the end use.

Jensen Huang’s departure raises the stakes. If government investigators determine that the CEO was aware of the smuggling activity — or benefited from it — the liability extends far beyond a personnel announcement. Corporate penalties for export control violations can run into hundreds of millions of dollars. Individual criminal charges are possible for those who facilitated or directed the shipments.

Howard Lutnick, the Secretary of Commerce who oversees the agencies licensing chip exports, may also face consequences. Reports suggest his investigators failed to uncover suspicious buyers and were too close to Nvidia officials. A regulatory framework that misses patterns as clear as those found in Chinese model weights raises questions about enforcement capacity that go beyond any single administration.

The short-selling firm Culper predicted this exact dynamic last May, arguing that Nvidia carried a structural “China problem.” The bet was not about U.S. crackdowns but about China’s response to restrictions: massive investment in domestic chip production that would eventually eliminate demand for Nvidia altogether. Six hundred thousand H200 chips are reportedly stockpiled in Hong Kong, held back from mainland shipment for reasons observers cannot agree on — surveillance concerns? Trade deal preparation? bureaucratic delay?

The hold-up makes no economic sense if you assume Chinese companies simply want the chips. Graham Webster at Stanford argues that the demand is so large, the supply chain so established, that holding inventory cannot meaningfully drive domestic production. But the longer the wait continues, the more room Chinese manufacturers fill the gap.

What Happens Next

The export control regime was built on a straightforward logic: restrict the tools, restrict the competitor. The reality, now documented in open-source model weights, is more complicated. Sanctions accelerated Chinese innovation, redirected capital toward domestic alternatives, and created black markets that enriched intermediaries while costing legitimate revenue.

Sam Bresnick at Georgetown captured the absurdity plainly this summer: allowing Chinese companies to access chips in mainland China and Hong Kong but not in Malaysia makes no strategic sense. Yet that is effectively the current posture. China hawks have called for controls based on who uses the chips, not just who owns them or where they are installed — a higher bar that would require tracking compute usage across multiple jurisdictions simultaneously.

The models are already built. DeepSeek and Moonshot are shipping products today that carry Blackwell fingerprints in their weight structures. Future chips may face tighter controls, but the knowledge transfer is complete. Chinese engineers know how to train on Nvidia hardware. They know what configurations work. They have benchmarks. All of this exists regardless of what happens at customs.

Nvidia will likely survive. It always does. But the episode confirms what critics have argued for years: containment through hardware restriction is not containment at all. It is a subsidy for domestic competitors and a catalyst for the very capability gap the policy was meant to prevent.

The question now is not whether the chips reached China. It is whether Washington realizes the ban achieved its opposite.