The US Bureau of Industry and Security has announced new restrictions on Chinese companies' remote access to advanced American AI chips through overseas data centres, targeting a loophole that has persisted since the first round of export controls.
The gap was structural. Controls governed the physical shipment of accelerators; they said much less about who could rent time on those accelerators once installed in a permitted country.
In practice, that allowed Chinese AI developers to train and serve models on restricted silicon by leasing capacity abroad - legally, under the old text.
The new rules shift the compliance object from hardware to access. Expect providers to require customer attestations, restrict console access by jurisdiction, and log workload metadata.
Enforcement will be the hard part. Compute is fungible and resold through layers of intermediaries; proving that a given training run used a given cluster requires cooperation from providers whose commercial interest runs the other way.
