A new US AI chip export rule requires licences for chips sent to overseas data centres that serve Chinese companies. The Bureau of Industry and Security filed the 132-page interim final rule on 16 November 2026, and the Federal Register has scheduled it for publication on 19 November, the day it takes effect.

When the AI Chip Export Rule Takes Effect

The AI chip export rule is not in force yet. It is on public inspection at the Federal Register as document 2026-23529, filed at 4:15pm US Eastern time on 16 November 2026 as a special filing, and the rule’s own text says it is effective on the date it is published.

That matters because every deadline in the rule counts from publication rather than from the day the announcement was made. As of 17 November 2026, the dates below follow from the scheduled publication date of 19 November 2026, which the Federal Register can still move.

MilestoneDate
Filed on public inspection16 November 2026
Scheduled publication and effective date19 November 2026
Deadline for shipments already in transit19 December 2026
Public comments due18 January 2027
Existing data centres must hold a licence17 February 2027

The rule carries a savings clause for goods caught mid-journey. Shipments that lost their licence exception because of this rule, and that were already en route to a port on the publication date under actual orders, may continue if they are delivered within 30 days.

What Triggers a Licence: 12 Million Aggregate TPP

A licence is required once a data centre crosses a single computing-power threshold. The rule bars exporting, reexporting or transferring controlled chips to a data centre “to or within any destination worldwide that will provide access to more than 12,000,000 ‘aggregate TPP'” to companies headquartered in, or ultimately owned from, Macau or Country Group D:5.

Total Processing Performance, or TPP, is the Export Administration Regulations’ own measure of a chip’s computing power, and “aggregate TPP” is a new definition this rule adds for the combined figure across a group of chips. The Bureau of Industry and Security publishes no chip-count equivalent. Reporters at TechCrunch and the trade outlet Beri each calculated the threshold at roughly 750 Nvidia H100 or H200 accelerators; those are journalists’ estimates, not government figures.

Two details widen the rule further than the headline number suggests:

  • Linked data centres count as one: capacity is added together across sites joined by networking infrastructure, and the rule says this includes digitally-linked infrastructure, not only physically-linked sites.
  • Due diligence falls on three parties: data centres, foreign foundries and chip exporters must all check direct or indirect ownership and control by persons in Country Group D:5 or Macau.

Country Group D:5 is the arms-embargoed list in the Export Administration Regulations. It covers more than 20 destinations, China among them, and the rule names Macau alongside it.

Five Countries Added to the Licence List

Separately from the data-centre control, advanced semiconductors now need a licence to reach five named countries. The Bureau of Industry and Security says each presents “a higher-than-normal risk of diversion”:

  • Brazil
  • Indonesia
  • Malaysia
  • Singapore
  • Thailand

The list is not closed. The rule lets the Secretary of Commerce add further countries after consulting the Secretary of State or the Assistant to the President for National Security Affairs, and it warns that countries which tolerate evasion could face import restrictions, a downgraded status in the US export-control regime, or Entity List designations for the people involved.

There is one carve-out. The five-country licence requirement does not apply to chips destined for an end user headquartered in the United States, or whose ultimate parent is, but only if the exporter files reports under a new section 743.10. Those reports name every party to the transaction, including ultimate parents, along with each product’s export classification, SKU, TPP and quantity.

The five-country list follows months of concern about chips reaching China indirectly. A federal indictment this year traced export-controlled Nvidia servers routed through Malaysia, and US officials have scrutinised Megaspeed, a Singapore-based cloud provider that denies sending chips to China, according to reporting by Fortune and The Next Web. The rule also lands days after Nvidia’s chief executive resigned, though the agency does not mention the company.

What the Rule Removes

The same document formally kills the Biden administration’s AI Diffusion Rule, which the Bureau of Industry and Security says was never enforced and which it announced it would rescind in May 2025. Gone with it are the country tiers, License Exception Artificial Intelligence Authorization, License Exception Low Processing Performance, and the Data Center Validated End User programme created in October 2024.

The rule also deletes the export control on AI model weights. ECCN 4E091 is removed from the Commerce Control List, along with the AI model weights Foreign Direct Product rule and the definition of “model weights” itself. License Exception Advanced Compute Manufacturing survives, because the agency says it supports chip supply chains.

What Is Still Unclear

The rule contradicts itself on recommendation systems. Its summary says the licence requirement “will not apply to commercial recommender systems, such as those used by online retailers or social media companies”, but a later passage says operators providing remote access for those systems “may continue providing these services for 90 days, but must obtain a license following expiration of this period”.

Two further questions have no answer in the text. Licence applications will be reviewed case by case, so the approval rate is unknown, and the rule does not spell out how an exporter is expected to establish that a customer will cross the 12 million aggregate TPP line in future rather than today. Comments on both points are open for 60 days after publication.

Because this is an interim final rule issued under the Export Control Reform Act of 2018, it binds before those comments are read. The Bureau of Industry and Security assessed the net paperwork change at about 970 extra submissions a year and 440 additional burden hours.

Frequently Asked Questions

When Does the AI Chip Export Rule Take Effect?

On the day it is published in the Federal Register, scheduled for 19 November 2026. It was filed on public inspection on 16 November 2026 and is not in force before publication.

Which Countries Now Need a Licence for Advanced Chips?

Brazil, Indonesia, Malaysia, Singapore and Thailand, in addition to destinations that already required one. The Secretary of Commerce can add more countries later.

What Is 12 Million Aggregate TPP?

Total Processing Performance is the US export-control measure of a chip’s computing power, and aggregate TPP is the combined total for a group of chips. A data centre supplying more than 12 million of it to Chinese-headquartered customers triggers the licence requirement.

Are US Companies Exempt?

Only partly. Chips going to end users headquartered in the United States escape the five-country licence requirement, but the exporter must file reports listing the parties, classifications, SKUs, TPP and quantities involved.

How Long Do Existing Data Centres Have?

Ninety days from the effective date, which falls on 17 February 2027 if the rule publishes on 19 November 2026. Operators that miss it face future export restrictions.