Forbes has reported that Washington's export controls, designed to keep advanced AI chips out of Chinese hands, are being undercut by a structural gap: the cloud.

The finding cuts to the core of US export policy. Regulators have focused their restrictions on physical hardware — the accelerators that train and run large AI models. But the same computing power is available as a rented service. A Chinese company that cannot legally buy a restricted chip can, in many cases, simply pay for time on that chip inside a data center abroad and access it over the internet.

That asymmetry defines the loophole. Hardware crosses borders and can be stopped at them. Compute does not. When a model trains on GPUs sitting in Virginia or Dublin, no controlled item ever enters China, yet the Chinese user captures most of the capability the rules were written to deny.

The policy implications are significant. If the objective of export controls is to slow China's progress in frontier AI, then restricting chip sales alone addresses only one delivery channel. The other channel — cloud inference and training delivered as a commercial service — operates largely through standard contracts and APIs, and enforcement at that layer is far harder than policing shipments of silicon.

The report also exposes a commercial tension inside the US industry itself. American cloud providers and AI platform operators sell access globally, and Chinese demand represents revenue. Tightening the cloud channel would mean turning away paying customers; leaving it open means the control regime leaks. Policymakers must weigh the competitiveness of US cloud firms against the strategic goal of the restrictions — a trade-off with no clean answer, since allies' cloud markets can absorb demand the US refuses to serve.

There is a second-order effect on Chinese domestic strategy. Every demonstration that compute can be rented abroad weakens the pressure on Beijing to build an indigenous alternative, and every tightening of that channel strengthens the case for domestic accelerators and foundry capacity. The loophole, in other words, shapes investment incentives on both sides of the Pacific, not just compliance behavior.

For semiconductor and cloud vendors, the practical question is how any future rulemaking would define a regulated transaction. Export law historically governs tangible items and specific technology transfers. Extending it to remote access — to bytes returned from a server rather than a chip shipped in a crate — would require new legal machinery, new monitoring obligations for providers, and new questions about how to verify the identity and location of customers at scale.

The Forbes report lands at a moment when US export policy toward China continues to tighten in iterative steps, with each round of rules chasing the workarounds exposed by the last. The cloud gap suggests the next iteration may have to reach beyond the shipment manifest entirely, and until it does, the effectiveness of the AI chip controls will depend less on what leaves the factory than on what anyone with a credit card can log into.