The U.S. Commerce Department’s Bureau of Industry and Security announced an interim final rule on November 16, 2026, extending export-control protections to remote access to advanced American AI chips and servers. The target is an arrangement in which restricted entities obtain computing capacity from data centers in third countries, even though the hardware never enters the restricted destination. BIS’s announcement says the Treasury Department and National Security Council wrote and directed the rule.
The reported requirements reach chip exporters, overseas data center operators and arrangements involving compute brokers. They include a threshold above 12 million total processing performance, or TPP, and a 90-day licensing transition for existing operators providing covered services.
The practical change is that a server’s location becomes only part of the compliance question. Providers also need to examine who ultimately controls or accesses its computing capacity.
How the Rule Reaches Overseas Cloud Access
BIS says the rule is designed to prevent existing restrictions on sensitive American technology from being circumvented through remote computing arrangements. Its stated purpose is to align protections governing remote access with controls on physical hardware shipments. That is the government’s rationale, not evidence that the new requirements will successfully prevent every prohibited transaction. The announcement provides no enforcement results.
TechCrunch reports that licenses can be required to export, reexport or transfer covered chips and servers to data centers anywhere in the world if those facilities may provide covered compute to companies in restricted countries. The restrictions also reach indirect access through brokers or other intermediaries.
Consider an illustrative arrangement: a data center in a third country receives advanced chips, sells capacity to a broker, and the broker supplies that capacity to a restricted customer. Checking only the data center’s address or the broker’s incorporation documents would miss the ultimate user.
That is the consequential shift. The reported rule connects hardware transactions to downstream access, making the customer chain relevant even when the chips remain in an otherwise permitted location.
Who May Need a License?
The licensing obligations described in the reporting concern hardware transactions and existing operators supplying covered compute. They should not be read as saying every cloud customer or broker must automatically submit its own license application.
TechCrunch identifies Country Group D:5, the U.S. arms-embargoed destination group, as a central target. Examples include China, Russia, Iran and North Korea. Quartz’s account additionally identifies companies headquartered in Macau and says data centers, foreign foundries and chip exporters must perform due diligence on customers’ direct or indirect ownership or control.
| Participant | Reported relevance of the rule |
|---|---|
| Chip exporters and companies arranging reexports or transfers | Covered hardware transactions may require licenses because of the compute access the destination facility could provide. |
| Existing data center operators | Operators already providing covered services must obtain licenses within the reported transition period. |
| Foreign foundries | Quartz identifies them among businesses with customer due-diligence responsibilities. |
| Compute brokers and resellers | Indirect access through intermediaries is covered; an intermediary does not remove the ultimate customer from the analysis. |
| Customers in restricted destinations | Their headquarters, ownership, control and access can affect whether an arrangement requires authorization. |
The BIS press release confirms the policy’s purpose but does not spell out these operational details. The threshold, transition and exceptions discussed here come from TechCrunch’s and Quartz’s accounts of the interim rule. Businesses making an actual licensing determination need the operative regulatory text.
My view is that the intermediary provision deserves as much attention as the chip threshold. A provider can know its direct customer while still having incomplete information about the organizations receiving capacity through that customer. The reported ownership and control requirements make that information gap a compliance issue.
What the 12 Million TPP Threshold Means
The rule applies to a data center or network capable of supplying more than 12 million TPP, according to TechCrunch. TPP is an export-control measure of processing performance. It is not a measure of token output, cloud spending or customer accounts.
TechCrunch gives a rough hardware comparison: 12 million TPP corresponds to the maximum computing power of approximately 750 Nvidia H100 or H200 GPUs. That is the publication’s explanatory estimate, not an independent benchmark or a universal GPU-count rule.
Three details matter when interpreting it:
- The reported wording is “over” 12 million TPP. A summary should not silently turn that into “at least” 12 million.
- The scope includes a network. Looking only at the capacity of one rack or an individual customer’s allocation may miss the relevant regulated arrangement.
- The GPU comparison is approximate. A purchasing team should use the applicable performance calculation, not treat 750 GPUs as a substitute for the legal threshold.
The reference to a network creates an important planning question: what infrastructure counts together for the threshold? The opened summaries do not provide the aggregation methodology. It would be premature to claim that separate buildings, accounts or bookings necessarily create separate compliance boundaries.
I’d use the H100/H200 comparison to understand the scale of compute being targeted, then keep it out of any final licensing calculation. The evidence supports that distinction: the reported rule uses TPP, while the GPU count is an approximation offered to readers.
Five Countries Get an Additional Licensing Gate
The rule also adds licensing requirements for certain advanced semiconductor exports, reexports and transfers to Malaysia, Singapore, Indonesia, Thailand and Brazil. Commerce describes those destinations as presenting a higher diversion risk, according to Quartz.
Quartz reports that the five-country requirement applies in addition to existing licensing requirements. It also describes an exception for chips destined for companies headquartered in the United States, or companies with a U.S.-headquartered ultimate parent, provided reporting requirements are met. Applications will be reviewed case by case, according to its account.
These five destinations are a separate issue from the D:5 customer restrictions. Their inclusion should not be read as classifying them as arms-embargoed destinations.
The practical analysis therefore has two layers. One concerns where the hardware is going and whether that shipment needs a license. The other concerns who will ultimately receive covered remote compute.
A U.S.-headquartered company’s eligibility for the reported five-country exception should not be assumed to resolve every other obligation. Quartz describes a specific exception with reporting conditions, not a blanket authorization for all downstream uses.
The 90-Day Transition Starts With the Effective Date
Existing data center operators supplying services covered by the new rule must obtain licenses within , . Quartz says the rule takes effect upon publication in the Federal Register and provides a 60-day comment period.
Frequently Asked Questions
4 questions
1Does Every Overseas AI Cloud Customer Need a License?
No blanket requirement is reported. Licensing depends on covered hardware transactions and compute arrangements, including capacity, destination and the entities receiving access.