The rule in a nutshell
Washington has moved to close the remote-access loophole. On 16 November 2026, the US Bureau of Industry and Security (BIS) released an interim final rule, effective 19 November 2026, aimed at restricting companies in countries of concern from accessing significant volumes of advanced US computing power through offshore data centres. It does not prohibit that access outright. Instead, it imposes an end-use licence requirement on exports, reexports and in-country transfers of covered items, with licence applications reviewed case by case. Not every remote login or cloud service is caught.
- Worldwide licence trigger: a BIS licence is required to export, reexport or transfer (in-country) specified advanced-computing items if the relevant party knows they are destined for a data centre, anywhere, that provides or will provide more than 12,000,000 aggregate Total Processing Performance (TPP) to or for a covered company. TPP is the metric BIS uses under the Export Administration Regulations to measure the computing power of advanced chips. Roughly, it is a chip's peak operations per second multiplied by the bit length of each operation.
- Covered companies: companies headquartered, or with an ultimate parent headquartered, in Macau or Country Group D:5, including China and Russia. “For” includes compute supplied through an intermediary that trains, helps train or supplies models for a covered company.
- No fixed GPU count: the threshold is aggregate TPP, including across physically or digitally linked facilities; it is not a fixed number of GPUs.
- Later changes matter: a change from a permitted use to the prohibited data-centre use may be a transfer in-country. The control can apply to a possessor that was not involved in the original shipment.
- Other changes: new licence requirements for Indonesia, Malaysia, Singapore, Thailand and Brazil (with a limited US-parent exclusion); the AI Diffusion Rule (which was the earlier framework that capped exports of advanced AI chips by tiers of countries), has been rescinded. Foundry and packager controls (requiring chip foundries and packagers to take further steps to make sure they are not making advanced chips for restricted parties) have also been tightened.
Your business model decides your exposure
“Data centre operator” is not a legal category under the rule. What matters is who supplies, possesses and controls the hardware, and who allocates capacity.
- Neocloud / GPUaaS provider: typically owns or leases the GPUs, controls orchestration and allocates capacity. It is the most likely to face direct licensing, diligence and recordkeeping obligations.
- Facility-only colocation provider: supplies space, power, cooling and connectivity while the tenant owns and runs the GPUs. Hosting alone does not make it the compute provider, and it is unlikely to be the primary licence applicant. Its exposure turns on four things. First, possession or control: acting as consignee, taking delivery of or holding covered hardware can bring the provider into the chain of export, reexport or in-country transfer. Second, managed services: remote-hands, hardware management or orchestration support can blur the line between hosting and supplying compute. Third, any role in allocation or changed use: helping a tenant reallocate capacity, link facilities or move from a permitted use to the prohibited data-centre use may itself be a relevant transfer. Fourth, knowledge, which includes reason to know, not only actual knowledge: supporting covered equipment despite unresolved red flags about a tenant's ownership, ultimate parent or end use can create knowledge-based and contractual exposure, so ignorance is not a safe harbour. Upstream suppliers, lessors and financiers are also likely to expect colocation providers to give tenant representations and information rights.
- Hardware supplier, lessor or integrator: is squarely within the licensing and due-diligence framework when supplying covered items with relevant knowledge.
- Ordinary AI or cloud customer: generally needs no BIS licence merely to buy a service, unless it owns, supplies, controls or transfers the hardware or acts for a covered company. These customers should expect questionnaires and use restrictions.
Key distinction: a colocation provider hosts equipment; a neocloud sells compute. BIS will look at function, possession and the label on the door.
Where the risk sits
- Customer diligence: foreign data centres are expected to perform robust know-your-customer checks on their customers, including direct or indirect ownership or control. Exporters, reexporters and transferors must be able to obtain and provide sufficient information about those checks. The diligence extends to ownership or control by persons merely located in D:5 or Macau, which is wider than the headquarters-based licence trigger. A data centre declining or unable to provide that information is a red flag that presumptively creates "knowledge" unless resolved. For colocation data centre providers, this means hardware suppliers, lessors and tenants' financiers are likely to ask for evidence of the provider's own tenant screening. It also means existing onboarding checks, often built around creditworthiness and sanctions rather than export controls, may need to look through to the tenant's ultimate parent, controllers and end customers. A provider that cannot answer those questions may find that its tenants' supply chains treat the facility itself as a red flag.
- Downstream use: neoclouds should identify the customer, ultimate parent, intermediaries and beneficiaries; document end use; and monitor allocations across linked facilities. Contracts should support information requests, audits, suspension and termination.
- Colocation risk remains: a data centre facility-only provider may not be the primary licence applicant, but can face contractual and knowledge-based exposure if it supports covered equipment despite unresolved red flags.
Red flags may include a tenant that will not identify its ultimate parent, unexplained growth in GPU density or power draw, links to other facilities, or signs that capacity is being supplied to third parties. Because knowledge includes reason to know, continuing to host after such signs emerge may be enough. Exposure can also arise indirectly, through breach of warranties given to upstream suppliers or financiers, enforced suspension or removal of tenant equipment, and reputational damage, even where no licence obligation falls on the provider itself.
- Narrow recommender exception: the control does not apply where compute is used exclusively for a qualifying commercial recommender system, subject to the rule’s conditions. It excludes generative systems.
- Australia: the worldwide control can apply to covered US items in Australia. Australia is not among the five new destinations, but providers and users may still face end-use controls, questionnaires and contractual restrictions.
- Other regimes still apply: the BIS rule sits on top of, not instead of, Australian law.
Action items for Australian businesses
- Neoclouds and GPUaaS providers: classify the hardware; calculate aggregate TPP across linked facilities; map ownership, possession and control; screen customers, ultimate parents and beneficiaries; and assess existing services under the 90-day window.
- Colocation providers: document the facility-only boundary in contracts and operating procedures, so it is clear the provider does not own, control, allocate or manage the GPUs. Identify any consignee, possession, remote-hands, management or allocation role, and either remove it or assess it as a possible source of direct exposure. DC operators should obtain tenant representations on ownership, ultimate parent, end use and end users. Those representations should be backed with information and audit rights, notice of changes in use or control, and rights to suspend or terminate. DC operators should also create a red-flag escalation process, with clear internal owners, a record of how each flag was resolved, and a path to external advice. Review existing tenancies now, given the 90-day window for existing qualifying arrangements.
- Suppliers, lessors and financiers: verify recipient diligence, flow down notices on changes in end use or end user, and do not proceed on unresolved red flags.
- AI users: prepare ownership, ultimate-parent, end-use and beneficiary information, and review suspension, termination and capacity clauses in cloud and compute contracts.
Bottom line: labels will not protect you. The question is whether your business supplies or controls covered hardware or compute, makes a relevant transfer or changed use, or knowingly supports a prohibited arrangement. If you already serve covered companies, the 90-day clock is already running.
How G+T can help
Start building your compliance framework now: neoclouds, GPUaaS providers and DC operators should not wait for a supplier questionnaire or a tenant red flag. They should already be developing policies and procedures to manage the risk of breaching the EAR. These should cover customer screening, ultimate-parent checks, TPP calculations across linked facilities, red-flag escalation, recordkeeping and contract controls. A documented framework also helps providers answer the diligence requests that suppliers, lessors and financiers are likely to make.
Gilbert + Tobin can help DC operators and neoclouds design and put these policies and procedures in place, tailored to their business model and role in the AI compute chain.