White House Drafts Rule to Block China's Remote Access to US AI Chips

The Trump administration is drafting an AI rule aimed at cutting off China's ability to use advanced American chips remotely, The Information reports. If finalized, the measure would close a gap that existing export controls leave wide open: a chip can sit in Virginia while its compute serves Shanghai.

The reported rule addresses a structural weakness in the current control regime. US export restrictions govern where chips physically ship, not where their processing power ends up. Chinese firms can rent capacity on advanced GPUs housed overseas, or reach them through networked arrangements, without ever importing a restricted device. A remote-access rule would extend the control boundary from the wafer to the workload.

The report does not specify which agencies are leading the drafting, nor a timeline for publication. It also does not name the chip families the rule would cover. That absence matters, because the technical burden of compliance depends heavily on the definition: policing physical shipments at a port is one task; verifying the identity and location of every client tapping a GPU cluster is another entirely.

What does the rule change?

In principle, it converts export control from a customs question into a network question. NVIDIA's H100-class accelerators and their successors already face licensing requirements for shipment to China. A remote-access provision would reach the same silicon after lawful export, and would reach cloud operators and inference providers who never touch a restricted shipment at all.

The compliance machinery would be considerable. Providers would need to establish who is querying their hardware, from where, and on whose behalf. Enforcement would rest on audits rather than manifests. The draft's actual scope will determine whether this is a narrow fix or a broad licensing regime draped over the global cloud industry.

Who carries the cost?

The burden lands first on US chip vendors and the cloud platforms that buy their products. Each new restriction on where and how silicon may be used shrinks the addressable market for the most profitable product lines in semiconductors. NVIDIA, AMD and the hyperscalers have absorbed successive rounds of this since 2022, and each round has arrived with revenue adjustments attached.

There is also the substitution problem. Every restriction tightens the commercial case for domestic Chinese accelerators and for third-country supply. Washington has, in effect, been running a subsidy program for Huawei's Ascend line through the back door. A remote-access rule would deepen that effect even as it closes a loophole.

The geopolitical logic is straightforward: the frontier of AI capability is compute, and compute is now a service as much as a product. Controls written for crates of chips do not restrain a metered API. This reported draft is an attempt to write rules for the second model rather than the first.

What happens next?

Nothing yet, formally. A draft rule must clear interagency review and likely a public comment period before it binds anyone. Recent precedent suggests the final text will be narrower than the early reporting, after vendors and foreign governments press their cases. The open question is whether Washington can police remote access at all without the voluntary cooperation of the cloud industry it is regulating — and that industry has so far preferred lobbying to enforcement architecture. Watch the Federal Register; the scope of the published rule, whenever it arrives, will set the ceiling.