BIS announces remote advanced-compute rule with conditional licensing scope and timing

BIS announced an interim final rule on remote advanced-compute access on November 16, with publication scheduled for November 19. Its scope depends on chips, uses, end users and compute thresholds, with conditional exceptions and transition arrangements.
Editorial correction (November 18, 2026): Checked against the BIS announcement and Federal Register prepublication text as of November 18. Effective dates, applicability, recommendation-use exceptions and the ninety-day application arrangement have been corrected. The final official text governs specific cases.
Before the Nvidia controversy has settled, overseas GPU rentals face restrictions too.
Within the past twenty-four hours, the US Commerce Department’s Bureau of Industry and Security (BIS) formally announced a major interim final rule:
Where an overseas data center provides large-scale computing to entities with restricted headquarters or ultimate parent companies, the relevant chips’ export, reexport or in-country transfer requires a license.
How large is “large-scale”?
The rule uses a threshold exceeding twelve million aggregate total processing performance (TPP), alongside controlled-item classifications and specified end users. Roughly 750 chips is only the report’s computing-capacity analogy, not a substitute for legal classifications and calculations.

Document:
https://www.bis.gov/press-release/bis-issues-rule-addressing-remote-access-american-super-intelligence-technology
Restrictions previously focused on physical chip sales. They now address remote cloud access to computing.
Where, specifically, are the new boundaries?
Physically or digitally connected data centers operating as a unified resource may be aggregated under the rule. Chips spread across facilities are not automatically exempt, and calculation cannot be reduced to a single training task.
Requirements concern specified items, scale, end users and uses, including relevant headquarters or ultimate parents in Macau or Country Group D:5. Subleasing or intermediaries do not automatically remove applicability, but not every API call or overseas outsourced training task necessarily requires a license.
An exception covers certain items specifically designed and used for commercial recommendations, subject to reporting and other conditions. Generative models do not qualify for this recommendation-use exception. This does not mean all recommendation businesses need only report, or that every other legal authorization is canceled.
BIS announced the rule on November 16. As of November 18, the Federal Register prepublication page scheduled publication for November 19. The text takes effect upon publication; the announcement date is not the effective date.
For parties wishing to continue relevant existing services, the text requires a license application within ninety days after effectiveness; otherwise subsequent export eligibility may be affected. This does not mean connections must be unplugged after ninety days.
Beyond closing a cloud-access gap, the rule reshapes the global AI-infrastructure landscape.
Some advanced-chip transactions involving Singapore, Malaysia, Indonesia, Thailand and Brazil acquire licensing requirements, with conditional exceptions including certain US-headquartered end users. It is inaccurate to say all relevant transactions in those five countries require applications.
Over the past two years, Southeast Asia has become a major global computing-capacity intermediary.
Its location, inexpensive land and power and relatively accommodating trade environment have attracted an intense wave of data-center construction.
The original draft attributed the rule to the Megaspeed incident. The official announcement did not confirm that causal link, so this article discusses the rule and its publicly stated rationale instead.
Chip vendors and cloud providers face stringent know-your-customer due diligence under the rule—not just checking the contracting party, but tracing ownership and ultimate control.
Although announced by Commerce, the announcement specifically states that the Treasury Department and National Security Council helped write and guide the rule.
The announcement describes BIS collaborating with Treasury and the National Security Council. This provides no basis to claim those agencies were never involved in earlier controls.
The rule also repeals the January 2025 AI Diffusion Rule, replacing its complex country-tier and quota system with what the author describes as a blunt two-lock mechanism.
◈ A cat-and-mouse game
After the announcement, online commenters described the beginning of a cat-and-mouse game. That is because the company signing the contract, the one making the payment and the one ultimately using the computing power may not be the same company. Identifying the customers behind shell companies then becomes an ever-escalating contest between detection and evasion.
Although another layer of pressure clearly targets China, third countries also bear substantial pressure.
Singapore’s government has said it will not tolerate companies using the country to evade export controls, while protecting lawful companies’ access to advanced chips and compute. It is communicating with the US and affected businesses.

While the US closes gaps, Singapore still has its data-center business to run. Additional reviews, licenses and contract changes may raise operating costs.
The rule is expected to take effect November 19. Who receives licenses and how deeply overseas facilities can investigate ultimate users will depend on implementation.
This article was originally published by Xi Xiaoyao Tech Talk. Author: Luanma.
Related topics
After reading
Join the AINRK community
Read how to join the community and connect with people working in AI.




