Nvidia's most advanced AI chips are reaching China despite US bans

By Alex Mercer · 2026-08-19

A trickle, not a flood

Under Secretary of Commerce for Industry and Security Jeffrey I. Kessler put a number on the gap between what both governments have officially permitted and what has actually moved. Appearing before the House Foreign Affairs Committee on 14 July 2026, Kessler said "there have been minimal exports of any H200s to China so far," describing the volume as "very few" chips shipped to China or Hong Kong.^[1] That single admission tells you more about the state of the US-China chip standoff than any diplomatic statement issued in either capital.

Two separate regulatory systems have moved in the same direction at roughly the same time, with a wall of compliance requirements standing between policy and physical delivery. Understanding how those systems interlock is the only way to make sense of why a chip that both governments have nominally cleared is still barely crossing the border.

What Washington changed and what it kept

On 15 January 2026, the Bureau of Industry and Security revised its licence review policy for advanced semiconductors, including the H200, changing the default for exports to China from a presumption of denial to case-by-case review.^[3] A presumption of denial means an applicant starts from no and must argue their way to yes. Case-by-case review means each application is assessed on its own merits, with no thumb on the scale before the file is even opened.

Washington did not change the inspection architecture. Every shipment approved under the new framework must undergo third-party performance testing inside the United States before it is permitted to leave for China, verifying that each unit meets the performance specifications covered by the licence.^[2] A 25 per cent tariff applies on top of that, and the compliance burden on any exporter is substantial, which goes a long way toward explaining why Kessler's count remains so low months after the policy shifted.^[2]

By February 2026, the US government had granted a licence allowing Nvidia to ship small amounts of H200 chips to specific China-based customers under these conditions: named customers, inspection before departure, tariff on arrival.^[2]

Beijing's side of the equation

China approved its first batch of Nvidia H200 AI chip imports, covering several hundred thousand units, signalling Beijing's easing of restrictions on purchases.^[4] Beijing's clearance came around the same fortnight as Washington's policy change in January 2026, suggesting at minimum that both sides were aware the other was moving. Whether that convergence reflects back-channel coordination or parallel self-interest is not established by the available record.

China's AI developers, including cloud service providers and state-backed research institutes, have strong operational reasons to want the H200. The chip's Hopper architecture delivers over 4 terabytes per second of high-bandwidth memory and more than 6,000 gigabytes per second of interconnect bandwidth, a meaningful step up from the A800 and H100 variants already permitted under earlier US controls. For training large-scale language models and running inference workloads at scale, that bandwidth difference determines how large a model you can train in a given time at a given cost.

Chips permitted under previous control regimes have already been absorbed into the existing training infrastructure of China's leading developers. The H200 represents the next rung, the one that would allow training of foundation models in the same performance class as those built on unrestricted hardware in the United States.

Nvidia: licenced but unrevenued

Nvidia's own disclosure to US securities regulators is the clearest measure of how little has actually moved. The company said in its SEC filing: "To date, we have not generated any revenue under the H200 licensing program, and do not yet know whether any imports will be allowed into China."^[2] A licence exists. Revenue does not. Whether Chinese customs ultimately admit the tested and tariffed hardware remains, by Nvidia's own account, an open question.

The zero-revenue figure also constrains how analysts and investors should read any diplomatic signals around H200 exports. Both governments have moved from harder positions to softer ones, but the machinery that translates policy into shipments, and shipments into booked sales, is working very slowly. The testing pipeline, the named-customer restriction, and the tariff together create friction that a changed review standard alone cannot remove.

The broader contest

Washington has spent the better part of three years building an export control architecture designed to prevent China's AI sector from accessing chips above a certain performance threshold. Beijing has responded with its own approval and quota systems, while also investing in domestic alternatives through companies working on homegrown accelerator designs. The H200 situation is one data point in that longer and wider contest.

The move from presumption of denial to case-by-case review represents a real shift in US posture, acknowledging that a blanket ban on one chip model, when similar-capability alternatives exist or are in development, may impose costs on US exporters without delivering a durable strategic advantage. Case-by-case licensing lets officials tailor approvals to end-user risk, which is a more granular instrument than a blanket rule, though also a slower and more resource-intensive one.

For China's AI ecosystem, access to the H200 now depends on navigating two separate bureaucracies, paying a substantial tariff, and accepting that every shipment will be physically inspected before it travels. That is a workable path for large, well-resourced buyers. For smaller developers who lack the procurement infrastructure to manage a compliance process of that complexity, it is a prohibitive one.

Kessler told the House Foreign Affairs Committee on 14 July 2026 that exports to China remain minimal, and Nvidia's SEC filing confirmed the company had generated no revenue under the H200 licensing programme as of January 2026.

Alex Mercer

Alex Mercer writes about technology, energy and infrastructure. He likes the physical end of the story: the plants, the grids and the machines that everything else depends on.