---
title: "US AI chip rule: a licence for Malaysia, Singapore, Thailand, Indonesia and Brazil"
summary: "The US Commerce Department’s interim rule of 16 November: the 12 million TPP line for data centres, licences for five countries, and the AI Diffusion Rule’s formal end."
url: https://www.thetechbag.com/blog/us-ai-chip-rule-licence-overseas-data-centres
format: Analysis
desk: "AI in the enterprise"
region: Global
published: 2026-11-16
facts_checked: 2026-11-16
author: TechBag (The TechBag Journal)
tags: ["US AI chip export rule","AI chip export controls","Nvidia chips Malaysia licence","AI Diffusion Rule rescinded","BIS interim final rule","remote access rule","Addressing Remote Access to Advanced Semiconductors","aggregate TPP","total processing performance","Country Group D:5","Bureau of Industry and Security","export licence","data centres","Malaysia","Singapore","Thailand","Indonesia","Brazil","Nvidia H100","commercial recommender system","know your customer"]
cover: https://www.thetechbag.com/blog/covers/us-ai-chip-rule-licence-overseas-data-centres.png
---

# US AI chip rule: a licence for Malaysia, Singapore, Thailand, Indonesia and Brazil

*The US Commerce Department’s interim rule of 16 November: the 12 million TPP line for data centres, licences for five countries, and the AI Diffusion Rule’s formal end.*

Analysis · AI in the enterprise · By TechBag · Published 2026-11-16 · Facts checked 2026-11-16 · Canonical: https://www.thetechbag.com/blog/us-ai-chip-rule-licence-overseas-data-centres

The short version

1. A US rule released on 16 November requires a licence to send advanced AI chips to any data centre, anywhere, that provides more than 12,000,000 “aggregate TPP” of computing to companies headquartered in Macau or in Country Group D:5, the US arms-embargo list. That is about 758 Nvidia H100s. Operators already serving such customers have 90 days to get a licence.[1]
2. Separately, every export of these chips to Indonesia, Malaysia, Singapore, Thailand and Brazil will need a licence, unless the buyer is headquartered in the US and the exporter reports each shipment to the Commerce Department monthly.[1]
3. The rule formally rescinds the AI Diffusion Rule of January 2025, which was never enforced. It takes effect on publication in the Federal Register, with comments due 60 days later. India is not among the destinations it names.[1][3]

The US Commerce Department’s Bureau of Industry and Security (BIS) says the interim final rule exists so that existing restrictions “cannot be circumvented through remote computing arrangements involving data centers in third countries”.[2] Its stated aim is to align the rules on remote access with “existing controls on the physical shipment of sensitive American technology”. In practice it looks past where a chip is shipped and who buys it, to who rents its computing, from which data centre, and how much.[2] For anyone who builds, runs or rents AI capacity outside the US, it answers three questions this piece walks through: which data centres need a licence, which countries now need one for every shipment, and what the rule takes away.

## The 12 million TPP line: what a data centre may rent, and to whom

The core of the rule is a new section, § 744.24. Without a licence, nobody may export, re-export or transfer within a country advanced AI chips, or servers built on them, if they know the chips are destined for a data centre “to or within any destination worldwide” that “is providing or will provide access to more than 12,000,000 ‘aggregate TPP’”, at one site or through “linked data centers”, to or for companies headquartered in, or whose ultimate parent is headquartered in, Macau or a Country Group D:5 destination. Applications are reviewed case by case.[1]

Two terms carry the rule. **TPP**, total processing performance, is the yardstick US chip controls use: two times a chip’s peak multiply-accumulate operations per second (in trillions), times the bit length of the operation, counted on dense rather than sparse maths. In practice it is a datasheet’s teraFLOPS multiplied by the bit length. A chip with a TPP of 4,800 or more is controlled, and so is one of 1,600 or more with a high performance density. “Aggregate TPP”, a definition the rule adds, is simply the sum across a group of chips.[1] **D:5** is a column in the export rules’ table of country groups headed “U.S. Arms Embargoed Countries”, drawn from the State Department’s arms-embargo notices.[4] The rule’s own text names the People’s Republic of China among the D:5 countries, and it lists Macau alongside the group.[1]

758

Nvidia H100 SXM chips: the count at which a data centre passes 12,000,000 aggregate TPP

About 95 eight-GPU servers. Nvidia lists the H100 SXM at 3,958 teraFLOPS of FP8 with sparsity, which is a TPP of 15,832 on the rule’s dense count.

The line is drawn wider than a single building. ‘Linked data centers’ include any sites joined by networking that lets them “share resources, balance workloads, and replicate data as a unified whole”, digitally linked as well as physically. And “for an entity” reaches a customer that is not itself headquartered in D:5 but is training models for, or supplying models to, a D:5 company.[1] BIS also restates how it reads its older headquarters rule of October 2023: it already covers chips sent anywhere on behalf of such a company, for instance where it is the buyer or the “exclusive or primary (greater than 50 percent of total compute) remote end user”.[1]

**Figure: Three questions decide whether a US AI chip shipment needs a licence**

*Licence tests for advanced AI chips and servers in the interim final rule, as released on 16 November 2026*

1. 01Is it a controlled AI chip?An ECCN 3A090 chip (TPP of 4,800 or more, or 1,600 with high density), or a server or board built on one.
2. 02Where is it going?Indonesia, Malaysia, Singapore, Thailand or Brazil: a licence, unless the buyer is headquartered in the US and the exporter reports monthly.
3. 03Who uses the compute?Anywhere: more than 12,000,000 aggregate TPP for firms headquartered in D:5 or Macau needs a licence, reviewed case by case.
4. 04The one carve-outCompute used only for a commercial recommender system needs no licence, if the customer, the TPP and the checks are reported to BIS.

Source: BIS, interim final rule RIN 0694-AK50, §§ 742.6, 743.10, 744.24 and ECCN 3A090 · as of 16 November 2026

The burden sits on both ends. A US exporter keeps its innocence only if it can show it lacked “knowledge”, and to do so it is expected to obtain the foreign data centre’s “know your customer” checks, including on direct or indirect ownership by D:5 or Macau persons. “A foreign data center declining or being unable to provide information sufficient to verify compliance” is a red flag that presumptively gives the exporter that knowledge.[1] The control also follows the chips after delivery: it applies whoever holds them, a data centre that later starts serving such customers needs authorisation for the change, and anyone who ships without a licence must tell the recipient that authorisation is needed if the chips are later used that way. A data centre already providing that much compute to such customers before the rule “must obtain a license with 90 days of the effective date”.[1]

The exception is narrow. A ‘commercial recommender system’ is a tool “exclusively designed and exclusively used to rank, filter, or suggest” products, services or content, and the definition excludes any system that generates text, images, audio, video or software, even if it also recommends. The provider still files the customer’s name, the TPP provided and the due diligence done with BIS.[1]

## Five countries, and an exemption only for US companies

The second change is a plain destination rule. A licence will be required to export, re-export or transfer these chips “to or within Indonesia, Malaysia, Singapore, Brazil, and Thailand”, which the rule describes as “countries which present a higher-than-normal risk of diversion”. The Secretary of Commerce may designate more.[1] The only exclusion is for chips going to an ultimate consignee or end user headquartered in the US, or whose ultimate parent is. Even then the exporter must report every shipment by the first business day of the next month: each party and its ultimate parent, the destination, the classification, product name and SKU, the TPP, the quantity and the date.[1]

**Figure: The new licence requirement covers four Southeast Asian countries and Brazil**

*Destinations added to the licence requirement for advanced AI chips, interim final rule of 16 November 2026*

|  | Value |
| --- | --- |
| Singapore | – |

Source: BIS, interim final rule RIN 0694-AK50, § 742.6(a)(6)(iii)(A)(1) · as of 16 November 2026 · Singapore is shown as a point

Licences for the five are reviewed case by case, including on “the effectiveness of a destination’s export control regimes and coordination with BIS”. The rule adds that Commerce will monitor for countries that “tolerate evasion” and may answer with import restrictions, new licence requirements, a downgraded export-control status or Entity List designations.[1] Shipments already on their way on the day of publication may still go under the old terms, if delivered within 30 days.[1]

The rest of the destination map is set out in the same paragraph. The full set of conditions, as the rule states them:

| Where the chips go, or who uses them | Licence needed? | Exception | How BIS reviews it |
| --- | --- | --- | --- |
| Indonesia, Malaysia, Singapore, Thailand, Brazil | Yes (new) | Buyer headquartered in the US, with monthly reports | Case by case |
| Country Groups D:1 and D:4 (not A:5 or A:6) | Yes | None | Presumption of approval |
| United Arab Emirates | Yes | Approved UAE entities | Presumption of approval |
| D:5 countries and Macau, or a firm headquartered there | Yes | None | Presumption of denial; case by case for China and Macau below 21,000 TPP and 6,500 GB/s |
| A data centre anywhere giving D:5 or Macau firms over 12,000,000 aggregate TPP | Yes (new) | Commercial recommender systems, reported | Case by case |
| Anywhere else, India included | No destination licence | The two end-use tests above still apply | Not applicable |

Malaysia already controls the other direction. Since 14 July 2025, exports, transshipments and transits of “high-performance AI chips of US origin” out of Malaysia need a Strategic Trade Permit under its Strategic Trade Act 2010.[9] The US rule now covers the chips going in and moving within the country. The projects it reaches are large: on 9 November, Singapore-based Grab and Japan’s SoftBank Group, with Sarawak’s PETROS, announced US$1.85bn of seed capital for AI data centres in Malaysia, with a long-term target of five gigawatts or more.[10]

Our viewOur reading: who in Malaysia is exempt

On the rule’s text, the exemption turns on where the buyer or end user is headquartered, not where the building is. Neither Grab nor SoftBank is headquartered in the US, so unless a US-headquartered company is the consignee or end user, US AI chips for such projects need a licence, reviewed case by case. A US cloud provider building in the same country can ship under the exclusion, with monthly reports.[1]

## What the rule removes: the AI Diffusion Rule’s tiers

The AI Diffusion Rule, published on 15 January 2025, would have sorted the world into tiers: 18 countries plus the US eligible for a licence exception, and a worldwide licence requirement behind it.[1][4] Its compliance date was 15 May 2025. Two days before, BIS announced it would rescind it and told its enforcement officers not to enforce it.[3] The new rule makes that formal. It removes the country tiers and the licence exceptions built on them, the Data Center Validated End-User programme of October 2024, and the controls on AI model weights (ECCN 4E091).[1] India, Indonesia, Malaysia, Singapore, Thailand and Brazil were all outside the 18-country list.[4]

**Figure: Four years of US AI chip rules end with a licence that follows the compute**

*US export-control rules on advanced AI chips, October 2022 to November 2026*

| Date | Event | Detail |
| --- | --- | --- |
| 7 Oct 2022 | First AI chip controls | The October 2022 rule on advanced computing chips |
| 25 Oct 2023 | Headquarters rule | Licence for chips to firms headquartered in D:5 or Macau, wherever located |
| 15 Jan 2025 | AI Diffusion Rule | Country tiers and a worldwide licence requirement |
| 13 May 2025 | Rescission announced | BIS says it will not enforce the Diffusion Rule |
| 15 Jan 2026 | H200 policy | Case-by-case review for H200-class chips to China |
| 31 May 2026 | Enforcement guidance | Licence enforced for D:5- and Macau-headquartered firms |
| 14 Jul 2026 | UAE rule | Approved UAE entities receive chips without a licence |
| 16 Nov 2026 | Remote-access rule | 12,000,000 aggregate TPP line; five countries; Diffusion Rule rescinded |

Source: BIS interim final rule RIN 0694-AK50; BIS releases; Federal Register · as of 16 November 2026

In place of the worldwide requirement, the rule keeps a licence for the D:1, D:4 and D:5 groups and the UAE, adds the five countries, and moves the headquarters test back into the end-use rules, § 744.23.[1] The January rule on H200-class chips survives as a case-by-case policy for China and Macau below 21,000 TPP and 6,500 GB/s of memory bandwidth, and the July rule’s approved UAE entities remain exempt.[1][6][7]

ReportedWho wrote it, and why it says SI

BIS says the rule was “written by and at the direction of the Department of Treasury and the National Security Council”.[2] It also says “Super Intelligence”, or SI, wherever earlier rules said artificial intelligence, following Executive Order 14434, “Inaugurating the Era of Super Intelligence”, published on 2 October.[1][8]

## What builders in Malaysia, Singapore and Thailand now have to do

For a data-centre operator in one of the five countries that is not US-headquartered, every new shipment of controlled chips will wait on a licence. For one serving a US cloud or AI company, the chips can flow, but its US suppliers file a monthly report naming each party and its ultimate parent.[1] Wherever the site is, an operator that wants to keep buying should expect suppliers to ask for its customer checks: who each large customer is, and who ultimately owns it.[1]

**Figure: The rule's clocks run 30, 60 and 90 days from its publication**

*Deadlines and BIS's own paperwork estimates in the interim final rule of 16 November 2026*

90 days

for data centres already serving D:5 or Macau firms to get a licence

30 days

for shipments already en route to arrive under the old terms

60 days

for public comments on the rule

~970

more filings a year to BIS, its own estimate, with about 100 more licences

Source: BIS, interim final rule RIN 0694-AK50 (dates, savings clause, § 743.10, Paperwork Reduction Act estimates) · as of 16 November 2026

Chip designers in the five countries face a further change. A foundry or packaging house making an advanced-node chip for a designer that is not an approved or authorised designer under BIS’s rules must presume the chip is controlled. Until now it could overcome that presumption with an attestation that the packaged chip stays under a transistor count; the rule withdraws those two routes for designers located in, or owned from, the D:1, D:4 and D:5 groups or the five countries.[1]

Our viewWhat to watch

The publication date in the Federal Register, which starts the 90-, 60- and 30-day clocks; whether the Secretary of Commerce designates further countries; what the governments of Malaysia, Singapore, Thailand, Indonesia and Brazil say; and how fast BIS turns licence applications for the five round, since the rule promises case-by-case review but no timetable.

India: not on the list, but its data centres will be asked about their customers

India is not one of the five countries, and it does not appear in the D:1, D:4 or D:5 groups, so the rule adds no destination licence for chips shipped to India.[1][4] The country’s name appears once in the rule’s 132 pages, in unchanged wording of the validated end-user programme. Under the rescinded Diffusion Rule, India was outside the 18-country list; that list is now gone.[1][4]

The two tests that follow the chip still apply in India, as everywhere: a licence for chips meant for a company headquartered in D:5 or Macau, and for an Indian data centre that would give such companies more than 12,000,000 aggregate TPP. In practice the rule reaches Indian operators through their suppliers, who are expected to collect their customer checks and to treat a refusal as a red flag. That includes the GPU clouds supplying the IndiaAI Mission, whose [subsidised GPUs we covered on 2 November](https://www.thetechbag.com/blog/india-ai-compute-pledges-gpu-use).[1]

How we did this

Figures as of 16 November 2026, 11:30pm IST

Every rule detail is from the interim final rule as BIS put it on display on 16 November (RIN 0694-AK50, 132 pages), read in full, and BIS’s press release of the same day. When we last checked the Federal Register’s API, no document for this rule had been published or placed on public inspection, so the rule’s own dates (“effective on the date of publication”, comments 60 days after, the 30-day savings clause) have no calendar date yet. The 90-day window runs from the effective date. Country groups and the Diffusion Rule’s 18-country list are from the eCFR, up to date as of 12 November 2026.

TechBag arithmetic: Nvidia’s H100 page lists the SXM version at 3,958 teraFLOPS of FP8 and 3,958 TOPS of INT8, both “with sparsity”. The rule counts dense performance, so we halve it to 1,979 and multiply by the 8-bit length: a TPP of 15,832 (FP16 gives the same: 989.5 × 16). 12,000,000 ÷ 15,832 = 757.96, so 758 chips is the first count above the line, and 95 eight-GPU servers (760 GPUs; 94 servers hold 752). The H100 NVL, at 3,341 teraFLOPS with sparsity, has a TPP of 13,364, so the line falls at 898 of them. Applying the rule to India and to the Grab and SoftBank projects is our reading of its text, not a BIS statement. We report the rule’s stated reasons only where they define its scope. Nothing published after 11:30pm IST on 16 November is included.

Spotted an error? Write to us; corrections are published at the foot of the piece.

Sources

1. [BIS, Department of Commerce: Addressing Remote Access to Advanced Semiconductors, interim final rule (RIN 0694-AK50), public display copy](https://www.bis.gov/media/documents/rin-0694-ak50.pdf) — 16 November 2026
2. [BIS: BIS Issues Rule Addressing Remote Access to American Super Intelligence Technology](https://www.bis.gov/press-release/bis-issues-rule-addressing-remote-access-american-super-intelligence-technology) — 16 November 2026
3. [BIS: Department of Commerce Announces Rescission of Biden-Era Artificial Intelligence Diffusion Rule, Strengthens Chip-Related Export Controls](https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens) — 13 May 2025
4. [eCFR: 15 CFR part 740, supplement no. 1 (Country Groups) and supplement no. 5 (Artificial Intelligence Authorization Countries)](https://www.ecfr.gov/current/title-15/subtitle-B/chapter-VII/subchapter-B/part-740) — up to date as of 12 November 2026
5. [Nvidia: H100 Tensor Core GPU, product specifications](https://www.nvidia.com/en-us/data-center/h100/) — undated page
6. [Federal Register: Revision to License Review Policy for Advanced Computing Commodities (91 FR 1684)](https://www.federalregister.gov/documents/2026/01/15/2026-00789/revision-to-license-review-policy-for-advanced-computing-commodities) — 15 January 2026
7. [Federal Register: Enhanced Favorable Treatment for the United Arab Emirates Under the Export Administration Regulations](https://www.federalregister.gov/documents/2026/07/14/2026-14132/enhanced-favorable-treatment-for-the-united-arab-emirates-under-the-export-administration) — 14 July 2026
8. [Federal Register: Executive Order 14434, Inaugurating the Era of Super Intelligence](https://www.federalregister.gov/documents/2026/10/02/2026-20321/inaugurating-the-era-of-super-intelligence) — 2 October 2026
9. [MITI Malaysia: Export, transshipment and transit of high-performance AI chips of US origin now subject to a Strategic Trade Permit](https://www.miti.gov.my/miti/resources/Media%20Release/%5BFINAL%5D_MITI_Press_Stmt_Malaysia_Regulates_Trade_of_US_AI_Chips_2025-07-14.pdf) — 14 July 2025
10. [Grab: SoftBank Group and Grab Sign Joint Development Agreements to Develop AI Infrastructure in Malaysia](https://www.grab.com/sg/press/others/softbank-group-and-grab-sign-joint-development-agreements-to-develop-ai-infrastructure-in-malaysia/) — 9 November 2026

**Disclosure.** TechBag is a software reseller: it earns when companies buy through it, including from vendors named here. The analysis is TechBag’s own; no vendor paid for or saw this piece before publication. Facts are labelled reported, TechBag count or our view.

**Read this next:** [Grab to put US$750m into Malaysian AI data centres with SoftBank](https://www.thetechbag.com/blog/grab-750m-malaysia-ai-data-centres-softbank) — Malaysia’s AI build-out in one deal: Grab and SoftBank’s US$1.85bn of seed capital for AI data centres in Sarawak and Peninsular Malaysia.

Go deeper on TechBag

- [US$300bn pledged to India's data centres; subsidised GPUs reportedly 20% used](https://www.thetechbag.com/blog/india-ai-compute-pledges-gpu-use) — the Indian side of the compute build-out the rule's customer checks now reach
- [GPU rent turned up: B200 hours rose 21% in a month, H100 hours 7%](https://www.thetechbag.com/blog/gpu-rental-prices-turn-up-b200-h100) — what the H100s and B200s the rule counts rent for by the hour
- [Nvidia’s Jensen Huang resigns as CEO after 33 years; CFO Colette Kress is interim chief](https://www.thetechbag.com/blog/nvidia-huang-resigns-kress-interim-ceo) — the largest maker of the chips the rule covers
