678,000 Chips Later, US Export Controls on China Count as Failed

Nvidia shipped 678,000 AI accelerators to China in the first half of 2026, despite three years of US export controls designed to choke that exact flow. IDC data, cited by Reuters, also gives Nvidia a 34% share of China's data-center AI-chip market. The numbers attach a hard verdict to a policy debate that has moved from argument to evidence.

Nvidia CEO Jensen Huang said it plainly, in remarks covered by Reuters: "I think, all in all, the export control was a failure." He made the case at Computex in May 2025. Huang has an obvious commercial stake, so his view deserves scrutiny. But he is no longer alone. The Center for a New American Security (CNAS), a security think tank that wants tighter enforcement rather than looser rules, reaches a similar diagnosis in its analysis "The Nvidia Scandal Is a Wake-Up Call for Congress": the controls leak.

One caution first. No single, officially documented cause explains the entire failure. Claims that specific Chinese AI systems were trained on particular quantities of restricted chips remain unconfirmed. The Register relayed a claim from an ex-OpenAI engineer that Chinese labs trained models on Nvidia's Blackwell hardware despite the controls; no regulator has verified it.

Why did the controls leak?

The structural flaw is simple. US rules governed boxes shipped across borders, not who could log into a rack of GPUs (graphics processing units) once it sat in a legal jurisdiction. A Chinese lab that cannot import a Blackwell-class card can rent time on one in Kuala Lumpur or Singapore. Cloud computing turned a border control into a paperwork exercise.

CNAS identified Megaspeed as a company that hosted advanced AI chips for Chinese companies to access remotely and may have facilitated illicit transfers. Fortune reported a network in Singapore and Malaysia with a potential flow of about $7.5 billion in Nvidia chips. That figure comes from reporting alone. No court or regulator has confirmed it.

CNAS also documented a second hole. Overseas subsidiaries of Chinese companies obtained advanced Nvidia chips without licenses. A parent in China sets up a subsidiary abroad, buys hardware there, and runs workloads for the parent; unless regulators look through the corporate wrapper, the sale looks ordinary. Closing that gap requires ownership tracing and end-user checks, both of which cost time and people.

The rules had a technical escape hatch too. Export controls used interconnect bandwidth (the communication speed between processors) as a criterion. Vendors could trim the links while keeping much of the compute. The H20, custom-built for China and later banned, is the textbook case: regulators first accepted the design, then reversed it.

What do the numbers show?

• 678,000 Nvidia AI accelerators shipped to China, H1 2026 (IDC, cited by Reuters)
• Down 44% from H1 2025
• 34% Nvidia share of China's data-center AI-chip market, H1 2026
• 25% US share under the H200 reimportation arrangement (CNAS)
• About $7.5 billion in reported chip flow tied to Megaspeed (Fortune, unconfirmed)

A 44% drop looks like a policy win. A 34% share in a market the controls were meant to deny does not. Implied arithmetic suggests the year-earlier half carried roughly 1.2 million units; treat that as an estimate, not a published figure.

How did policy whiplash weaken deterrence?

A deterrent only works if buyers believe it will hold. Reports indicate the H20 was blocked in April 2025, then sales resumed in July 2025. The administration later permitted H200 sales to China under a reimportation arrangement giving the United States a 25% share. H200 licenses reportedly moved to case-by-case review in January 2026. The AI Diffusion Rule was suspended with no guidance on which provisions still applied.

The incentive effects are predictable. Buyers who expected a door to close reserved foreign capacity in advance. Huang, speaking to DW, argued the curbs backfired: "The local companies are very, very talented and very determined, and the export control gave them the spirit, the energy and the government support to accelerate their development."

Rakesh Kumar, a professor, told Fortune on October 16, 2026 that the policy had the opposite of its intended effect. His reading of Chinese industrial strategy is a judgment, not a measurement.

Did any controls work?

Yes. Restrictions on high-bandwidth memory and advanced lithography equipment held up better, because those products are specialized, slow to ship and made by few firms. A GPU ships in thousands of boxes. A leading-edge lithography tool fits no suitcase.

Enforcement remains the weak point. CNAS described the Bureau of Industry and Security (BIS), the agency that writes licenses and chases diversion cases, as under-resourced and understaffed. Tracing one chip through distributors, data centers and shell structures consumes analysts for months.

What comes next?

Reports say the House passed the Remote Access Security Act, targeting AI processors hosted outside China; the bill awaited Senate action in November 2026. A reported framework would require licenses for Blackwell-class or better chips in seven countries: Singapore, Malaysia, Thailand, Indonesia, Brazil, Norway and South Africa, with recipient disclosure, monitoring and an exemption for US-operated data centers that block access from embargoed countries. The plan is not final.

Beijing has already pushed back. Al Arabiya reported on November 13 that China warned against new controls. More agents and clearer authority over cloud access would raise the cost of evasion without a single new rule; that fix depends on Congress, not on chip specs.