The Chip Smuggling War
The U.S. spent three years building an AI chip blockade around China. It didn’t work.
In late 2025, federal prosecutors in Texas unsealed details of Operation Gatekeeper: an investigation into a smuggling network that moved at least $160 million worth of Nvidia H100 and H200 GPUs to China between October 2024 and May 2025. The chips were restricted under U.S. export controls specifically designed to prevent China from building advanced AI systems.
The chips got there anyway. And they’re not the only ones.
The scale of the breach
The Center for a New American Security estimates that between 10,000 and several hundred thousand AI chips were smuggled to China in 2024 alone. The exact number is unknown because enforcement agencies can’t track inventory once chips leave authorized distribution channels.
The smuggling routes are diverse. In Singapore, authorities investigated Megaspeed — formerly 7Road International, a Chinese gaming company with Chinese government ties — as a potential conduit for restricted Nvidia GPUs. Megaspeed was Nvidia’s biggest Southeast Asian customer. The concern: it was buying more chips than a gaming company could plausibly use, and the surplus was allegedly flowing to China.
In February 2026, Chinese AI startup DeepSeek confirmed it had trained its latest model on Nvidia’s Blackwell chip — the company’s most advanced AI processor — despite U.S. export controls explicitly prohibiting such shipments to China. The company’s earlier breakthrough, DeepSeek-R1, was reportedly trained on a cluster of 10,000 Nvidia H800 chips, a less powerful variant designed specifically for the Chinese market under previous export rules.
The export control regime is leaking at every seam.
The policy contradiction
On December 8, 2025, the Trump administration announced a new AI chip export policy for China. Codified by the Department of Commerce on January 13, 2026, the policy set technical thresholds — chips with total processing performance below 21,000 or DRAM bandwidth below 6,500 GB/s could be exported — and capped volume at 50% of chips sold to U.S. customers per product.
The Council on Foreign Relations called the policy “strategically incoherent and unenforceable.”
Here’s why. The policy would allow approximately 1 million Nvidia H200 chips and 1 million H100 chips to be legally exported to China annually. CFR estimated this would increase China’s installed AI compute capacity by 250% in 2026. The policy that was supposed to constrain China’s AI development would, by its own terms, massively accelerate it.
The contradiction became sharper in February. Despite the licensing approval, Nvidia reported that zero H200 units had actually been sold to Chinese customers as of late February 2026. The reason: Chinese companies already had access to chips through other channels — smuggling, stockpiling, and domestic alternatives — and weren’t willing to pay a premium for chips that came with U.S. oversight strings attached.
How the chips move
The smuggling infrastructure is sophisticated. Operation Gatekeeper revealed a network that used shell companies, front organizations, and intermediary countries to move chips from authorized distributors to China. The basic method: a company in a non-restricted country buys chips legally from Nvidia’s distribution network, then re-exports them to China through a chain of entities that obscures the final destination.
Singapore, Malaysia, and the UAE have emerged as primary transit points. The chips are legal to buy in these countries. Tracking what happens after purchase is a different problem entirely. The Commerce Department’s enforcement capacity is limited: it can set rules, but verifying compliance across global supply chains with thousands of intermediaries is a fundamentally different challenge than writing regulations.
Nvidia itself has limited visibility. The company sells to authorized distributors, who sell to customers, who may resell or redirect inventory. By the time a chip reaches its final user, it may have passed through four or five entities across three countries. The paper trail exists in theory. In practice, it’s designed to be untraceable.
China’s domestic response
The export controls haven’t just failed to contain China’s AI capabilities. They’ve accelerated China’s investment in domestic chip production.
Huawei’s Ascend 910B processor, while less powerful than Nvidia’s top chips, is being manufactured at scale and deployed across Chinese AI infrastructure. Chinese cloud providers are optimizing software to run efficiently on less powerful hardware. DeepSeek’s breakthrough demonstrated that training-time efficiency can partially compensate for hardware limitations.
The strategic logic of export controls assumed that restricting access to the best chips would slow China’s AI development. The actual result is more complex: China’s AI labs are developing techniques to do more with less, while simultaneously acquiring restricted chips through smuggling. They’re getting both the efficiency and the hardware.
What happens next
The chip smuggling war has no obvious resolution. Tightening controls further would damage Nvidia’s revenue — China represented roughly 17% of Nvidia’s sales before restrictions began. Loosening them would contradict the national security rationale for the entire export control regime. And the current middle ground is producing the worst possible outcome: China gets the chips anyway, Nvidia loses the revenue, and the U.S. bears the diplomatic cost of a policy that doesn’t achieve its stated goal.
The deeper problem is structural. Semiconductors are small, valuable, and fungible. They can be shipped in a backpack. They can be relabeled. They can be routed through five countries in a week. Controlling their distribution with the precision required to enforce an export ban is like trying to control the flow of diamonds — theoretically possible, practically impossible at scale.
Three years into the AI chip export control experiment, the evidence suggests that the technology is moving faster than the policy can contain it. The chips are in China. They’re training models. And the smuggling networks that delivered them are more sophisticated now than when the controls began.
Originally published at https://noahaust2.github.io/strategist-dashboard/blog/the-chip-smuggling-war.html
Write a comment