The US Bureau of Industry and Security published an interim final rule on November 16 titled Regulating Remote Access to Advanced Semiconductors, and it does something export control has not done before: it regulates a service rather than a shipment.

What the rule says

An overseas data centre that provides more than 12 million aggregated TPP of advanced compute to an entity headquartered in a D:5 country group nation or Macao must obtain a licence first. Existing arrangements get 90 days from the effective date to come into compliance.

The rule also adds Indonesia, Malaysia, Singapore, Thailand and Brazil to the list of countries requiring licences for advanced chips, tightens the presumption that advanced logic manufacturing and packaging abroad is controlled, and formally repeals the January 2025 AI Diffusion Rule. Compute used for commercial recommender systems is explicitly exempt.

Why the shift matters

Export control has always been about goods crossing a border. This rule is about who can call an API. The controlled object is the ability to run a workload, regardless of where the silicon sits.

That is a much harder thing to enforce, and a much easier thing to comply with badly. A cloud provider now has to know the ultimate headquarters of its customers, not just their billing address, and has to be able to prove it.

The compliance cost lands first

The immediate burden is not higher chip prices. It is customer vetting, ownership tracing, resource isolation, contract rewriting and continuous audit. The same GPU in two different isolation domains serves different customers with different legal exposure.

That changes what a data centre is. It has been a fairly neutral piece of infrastructure; it is becoming a compliance gateway with a power bill.

The second-order effects

Nomura estimated that if Nvidia chip demand from China disappeared entirely, high-bandwidth memory demand in 2028 would fall 1.7% to 2.5%, and commodity DRAM demand 0.4% to 0.5%. Those are small numbers — the point is that the memory supply chain is now exposed to a policy decision, not to a demand forecast.

Site selection logic shifts too. Cheap power and land used to decide where a data centre went. Now the questions are regulatory exposure, customer nationality mix, and whether an audit can actually be executed in that jurisdiction.

The response

China's commerce ministry and foreign ministry both responded on November 17, rejecting the rule and saying countermeasures would follow if it proceeds. That is the predictable part; the unpredictable part is whether enforcement is aimed at a handful of large providers or at the long tail.

What operators should do

Map every customer to an ultimate beneficial owner, not to a contracting entity. Identify which workloads cross the 12-million-TPP threshold and which do not, because the threshold creates a cliff that will drive architecture. Put the regulatory-change clause in contracts now. And assume the 90-day clock is real — retroactive licensing requests are slower than new ones.

The strategic conclusion is uncomfortable but clear: compute is no longer a commodity you route to wherever it is cheapest. It is a licensed service with a nationality.