US proposes new controls on remote access to advanced AI compute
The Bureau of Industry and Security wants new licences for advanced chips and overseas data centres that provide significant remote computing access to restricted destinations.
The US Department of Commerce is proposing new licence requirements for advanced semiconductors when overseas data centres provide significant remote computing access to companies headquartered in Macau or countries in Country Group D:5. The Bureau of Industry and Security (BIS) says the interim final rule is designed to stop advanced computing power based on US technology being used in ways that threaten US national security or foreign-policy interests.1
The proposal targets more than physical chip shipments. It would also cover exports, reexports and in-country transfers when the exporter knows that advanced “Super Intelligence” semiconductors will support a qualifying overseas data centre. Data-centre operators, foreign foundries and chip exporters would have to carry out due diligence on direct or indirect ownership and control connected to Macau or a D:5 destination.1
The rule would create a new licence requirement for advanced semiconductors sent to Indonesia, Malaysia, Singapore, Thailand and Brazil, which BIS describes as destinations carrying a higher-than-normal diversion risk. The proposed requirement would not apply to chips destined for companies headquartered in the US, or whose ultimate parent is headquartered there, provided the stated reporting conditions are met.1
BIS also wants to tighten some of the assumptions used for advanced-node logic-chip fabrication and packaging. The changes would reduce the routes available to overcome existing presumptions when chip designers are based in destinations subject to a licence requirement for advanced Super Intelligence semiconductors.1
What this means for remote AI compute
The important shift is that the proposal treats access to computing power as part of the export-control problem. A data centre does not necessarily need to ship a finished accelerator directly to a company in a restricted destination for the rules to matter. If the centre provides a significant volume of compute remotely, the operator or another party making the transfer may need a licence.
BIS says data centres already offering qualifying services before publication would have 90 days from the rule’s effective date to obtain a licence if they want to continue. The document also excludes commercial recommender systems, such as those used by online retailers and social-media companies, because BIS says those functions do not present the same national-security threat.1
The agency says the proposal extends existing controls on advanced Super Intelligence semiconductors. It would make remote or cloud-based access harder for companies headquartered in Macau or D:5 destinations, while adding destination-based controls for the five named countries. The text also warns that countries tolerating evasion could face further consequences, including import restrictions, new licence requirements, a lower standing in the US export-control regime, Entity List designations or temporary export prohibitions.1
The dates are not final yet
The document is labelled an interim final rule, but its publication and effective dates still contain placeholders. BIS also says the comment deadline will be 60 days after publication in the Federal Register, rather than giving a final calendar date in the document currently available.1
That matters for companies planning shipments or cloud-compute arrangements. The policy direction is clear, but the exact start date, comment deadline and final implementation timetable still need to be confirmed through the Federal Register. Businesses affected by the proposed controls should not treat the placeholder dates in the draft as a live compliance deadline.
The proposal also formally rescinds the regulations implementing the Biden-era AI Diffusion Rule, which BIS says was never enforced. The agency says that rescission formalises an announcement it made in May 2025, while the new controls establish a different framework focused on advanced semiconductor access and remote compute.1
Our opinion
This is a significant attempt to close the gap between controlling where a chip goes and controlling who can use the computing power that chip enables. Cloud access makes the physical location of a semiconductor a less useful guide to where its capability is being used, so the focus on overseas data centres is understandable.
The practical risk is ambiguity. Companies will need to distinguish ordinary international cloud workloads from access that crosses the rule’s thresholds, while the placeholder dates and detailed ownership tests leave important compliance questions unresolved. BIS has published a clear policy direction, but the final Federal Register notice and any guidance around enforcement will decide how disruptive these controls become in practice.
Photo by Sergei Starostin on Pexels.2