How the U.S. could regulate Chinese access to Nvidia AI chips overseas

The Trump administration is reportedly preparing a new export-control approach aimed at a modern workaround: a Chinese AI company can obtain the computing power of restricted Nvidia hardware without importing the hardware into China. Instead, it can rent GPU capacity from a data center in another country and connect to the systems remotely. 5

The reported proposal would shift the focus of U.S. policy from where a chip is physically shipped to who can use its advanced computing capability. People familiar with the effort reportedly expected the rule could be circulated to technology companies and other stakeholders as early as September 2026, but its final scope, legal basis, and timing were not confirmed. 8

The loophole: controlling chips, not always compute access

U.S. export controls have restricted the export, reexport, transfer, and shipment of certain advanced-computing items. That framework is comparatively straightforward when a controlled GPU is physically sent to China or to a China-headquartered company.

The harder case is a foreign data center that already possesses the hardware. A Chinese company could contract for cloud or GPU time in a location such as Thailand, Malaysia, Japan, or Singapore, then use a network connection to run workloads on the machines. The processors remain overseas, while the customer receives their practical output: access to high-end computing capacity. 35

Reporting has linked this broader offshore-compute pattern to Chinese technology companies including Moonshot AI, Alibaba, ByteDance, and Tencent. The reports describe a regulatory gap rather than necessarily illegal physical smuggling: the customer may be accessing the capability of a controlled chip without taking possession of the chip itself. 2514

What the proposed rule could do

The reported rule would try to make remote use of covered advanced-computing systems an export-control event. Depending on its final design, it could require cloud and data-center providers to identify restricted end users, assess corporate ownership and control, and block or seek authorization for their access to designated systems abroad.

That would represent a significant change in compliance obligations. Providers could need to perform more detailed customer checks, trace ultimate parent companies, and monitor how cloud infrastructure is used—not simply verify where a physical server is located. A policy analysis of the proposed congressional framework similarly points to expanded know-your-customer requirements for cloud operators. 1

The proposal is still only reported policy development. It should not be treated as a final rule or as evidence that all offshore access by Chinese companies is currently prohibited.

Why enforcement would be difficult

Regulating remote access raises questions that do not arise in a conventional chip shipment. Existing export-control concepts are heavily tied to tangible items and defined transactions. Applying them to a foreign person operating a foreign server through a network could invite challenges over statutory authority and the limits of agency rulemaking.

Implementation would also be complicated. A provider may need to determine a customer’s real identity, corporate parent, affiliates, and beneficial relationships. Enforcement could be weakened by resellers, intermediaries, VPNs, shell companies, or accounts established in another customer’s name. U.S. authorities would also have to monitor providers and facilities outside the United States, where jurisdiction and cooperation may vary.

The likely result would be a more demanding compliance regime, not necessarily an airtight technical barrier. The legal and operational concerns are part of why Congress has been considering an explicit statutory solution for remote access. 1512

Why the Remote Access Security Act matters

The House-passed Remote Access Security Act, H.R. 2683, would amend the Export Control Reform Act framework to address remote access directly. Its definition covers access by a foreign person to a controlled item through a network connection—including the internet or a cloud-computing service—from a location other than the item’s physical location.

That language matters because it addresses the precise distinction behind the loophole: the customer may be far away from the GPU, but can still use the GPU’s capabilities. If enacted, the law would give the Commerce Department a clearer basis for regulating access itself rather than relying on rules designed primarily around shipping or transferring hardware.

H.R. 2683 had passed the House and been referred to the Senate Banking, Housing, and Urban Affairs Committee. It was not yet law, so it could not by itself authorize current enforcement.

The Moonshot AI example

The debate intensified after Michael Kratsios, a White House science and technology official, alleged that Moonshot AI had obtained servers equipped with Nvidia GB300 chips and accessed GB300 systems in Thailand, likely to train its Kimi K3 model. The allegation has not been presented here as an adjudicated finding, but it gave policymakers a concrete example of how frontier-model development might use advanced U.S.-controlled hardware located outside China. 13

The episode illustrates why physical export restrictions may not fully determine who benefits from a chip. If a model developer can send data and instructions to a foreign-hosted system, the location of the GPU may no longer be the same as the location of the company using its compute.

What the May 31 BIS guidance already changed

The Bureau of Industry and Security took an earlier step on May 31, 2026. Its guidance clarified that a license is required to export covered advanced-computing items to an entity headquartered in a Country Group D:5 jurisdiction, including China, or to an entity whose ultimate parent is headquartered there—even when the recipient itself is located outside that jurisdiction.

That guidance helps prevent a China-headquartered company from avoiding the licensing requirement simply by receiving chips through an overseas subsidiary or affiliate. It follows the ownership and corporate-parent relationship rather than relying only on the recipient’s postal address.

But the guidance did not require bona fide data-center operators to stop using or servicing chips they already held. That leaves a separate question: whether a Chinese customer may remotely rent computing time on hardware that is already installed abroad.

The bottom line

The reported Trump administration proposal is intended to close the gap between shipping a restricted Nvidia chip and renting access to that chip’s computing power. It would seek to extend export-control policy into cloud and remote-use relationships, where the hardware remains in a foreign data center but the computing capability is available to a restricted customer.

The policy faces two major tests: whether Commerce has enough authority to act without new legislation, and whether providers can reliably identify and control the ultimate users of globally distributed cloud infrastructure. The Remote Access Security Act could resolve much of the first problem if it becomes law, while the Moonshot case shows why officials are treating the second as an urgent practical issue.