The US is moving chip controls from the shipping address to the compute customer

The Commerce Department is finalising a rule intended to stop Chinese AI companies from remotely using advanced US chips installed in third-country data centres, according to The Information. The proposed regime would add Brazil, Indonesia, Malaysia, Norway, Singapore, South Africa and Thailand to licensing requirements covering Nvidia Blackwell-class chips and above. It has not yet been announced and could still change during inter-agency review.

The policy shifts the controlled object from a box crossing a border to access to computation. Customers seeking chip exports would have to identify recipients and intended uses, then accept continuing monitoring. US-run data centres could avoid a licence if they block remote use from arms-embargoed countries, but would have to report customers from those countries and the amount of compute they receive.

That architecture is more faithful to how AI infrastructure works. A Chinese laboratory does not need physical possession of a GPU if it can rent an equivalent cluster in Singapore. Existing Bureau of Industry and Security rules already require licence applicants to identify some remote infrastructure users; the reported change would widen both the geography and the operator’s duty.

The hard part is enforcement. Identity can be obscured through affiliates and resellers, workloads can move between regions, and utilisation is harder to audit than a customs declaration. The rule closes a legal loophole only if cloud providers can reliably attribute customers, constrain access and retain evidence without turning ordinary international cloud use into an indeterminate licensing process.