The Trump administration is working on rules that would bar Chinese customers from renting GPU compute out of countries such as Vietnam and Singapore, according to new reporting led by The Information. The paywalled piece included a diagram of the arrangement the rules are meant to break: data centers in those countries passing valuable data back to China. Two bills in Congress would push in the same direction by law. Together they amount to an admission that the chip export controls have a hole in them the size of a network connection.
The logic is easy to state and hard to execute. Washington has spent years trying to control where advanced silicon physically goes. Renting the same silicon by the hour, from a facility on foreign soil, produces the same trained model without a single restricted chip crossing into China. Whether the United States can police another country's access to remote services is the question the whole effort turns on.
The record on the simpler problem is not encouraging. Chinese companies were allowed to buy Nvidia's H20, then barred, then allowed again. As newer chips arrive, the restrictions on older ones get loosened. Sometimes the administration caps the number of GPUs that can be exported to a given company, Alibaba among them. And even under the bans that are in force, thousands of prohibited GPUs still reach China — smuggled in, relabelled, moved in deceptive packaging. Meetings between Nvidia chief executive Jensen Huang and Trump have shaped export decisions along the way. That is the enforcement record on objects that have serial numbers and have to fit in a crate.
The legislative track started earlier. The Remote Access Security Act, introduced last December, already targets Chinese access to remote services. Co-sponsor Dave McCormick frames it as closing a gap in current law: as things stand, bad actors can train AI models by accessing frontier chips that sit under US jurisdiction, and the Bureau of Industry and Security has no authority to require a licence for it. The bill would extend export controls to remote-access scenarios, so that physically holding sensitive technology and operating it over a network face the same scrutiny when national security is at stake.
The House is preparing a second bill along the same lines. Co-sponsor John Moolenaar (R-MI) says that in the AI race China buys everything it can and tries to steal the rest, and that it is seeking covert access to American data centers in order to train models on cloud compute. American cloud platforms, in his view, should be helping to contain a Chinese AI effort used for military and surveillance work; the bill would require them to verify the identity of their users to protect both their products and US national security. His co-sponsor Josh Gottheimer (D-NJ) puts it as a loophole: adversaries, China first among them, should not get around export controls by simply renting what they are forbidden to buy. He argues the law would give American companies legal certainty and oblige them to report attempts to turn US cloud infrastructure against national security.
Strip away the framing and what is being proposed is banking-style know-your-customer rules for compute. That is a workable mechanism and an expensive one, and it lands entirely on American providers. A US hyperscaler would have to identify and vet its customers; its competitors in jurisdictions that adopt no such rule would not. Every compliance regime creates a market for the non-compliant alternative, and this one is being written at exactly the moment other countries are building their own AI platforms in their own languages under their own control.
The more plausible reading is that nobody in Washington expects a wall. The goal looks like gradual attrition — many small restrictions that slow a competitor down rather than stop him, on the Cold War model, where the contest eventually forced one side to reconsider its imperial ambitions. That is consistent with how the Nvidia policy has actually behaved: China has not been cut off from the company's GPUs, but state bodies and private organisations cannot freely order the newest Grace Hopper and Vera Rubin architectures. Bottlenecks, not blockades.
The trouble with a friction strategy is that it has no failure condition. Nothing in what has been described sets a threshold, a review date, or a definition of how much delay would count as enough — which means the policy can be extended indefinitely without ever being judged. And the harder question sits underneath both bills and goes unaddressed: if Washington cannot reliably stop mislabelled crates of hardware from reaching China, what is the enforcement handle on a transaction where the provider is in Singapore, the customer is in China, and nothing but packets moves? The one real answer is that the chips in that Singapore facility were sold under American licence conditions — which makes this less an export control than a covenant running with the silicon, enforced forever, wherever it ends up.
That is a large claim to make about other countries' data centers, and it will be read as one. The rules are designed to teach China that renting American compute is not a reliable substitute for owning it. Everyone else renting American compute will learn the same lesson.