NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

The Rule Names Five Countries. The UAE Isn't One of Them.

132 pages. On November 16, the Bureau of Industry and Security posted an interim final rule on remote access to advanced semiconductors.

2026-11-17 · 4947 words · NeuroAI
The Rule Names Five Countries. The UAE Isn't One of Them.

132 pages. On November 16, the Bureau of Industry and Security posted an interim final rule on remote access to advanced semiconductors.

Five countries are on the list. The UAE is not.

It covers customers headquartered in China, Macau included, or in the D:5 group.

Five countries, no UAE

Malaysia, Singapore, Thailand, Indonesia, Brazil.

On November 11, The Information reported seven countries, adding Norway and South Africa. By November 16, five remained: four Southeast Asian hubs and one South American market.

According to a Caijing review, including Brazil points at the route Chinese internet companies use to rent compute abroad.

All five are regional data-centre hubs and priority monitoring points for chip transshipment. Malaysia and Singapore sit at either end of the Malacca Strait; Brazil is the region's main cloud market. The list follows transit routes, not economy size.

The UAE has G42 and its own AI campuses, but this rule governs remote access, not construction. Its absence says this round lands in Southeast Asia and South America, not the Gulf.

Which door is being closed

Advanced chips cannot enter China. Compute still has three routes.

Buy the chips; export controls closed that door years ago. Leave the cards in an overseas hall while a Chinese team logs in on bare metal, so only data crosses a border. Or call an inference API and pay per token.

That third route is the lightest.

Bare metal and API differ beyond latency. Bare metal gives card type, topology, long jobs, custody of the weights. An API gives inference output — enough for distillation, not pretraining.

So the rule aims at the second: the third yields outputs, the second whole machines.

It also pulls in third-country resale of the first. According to Caijing, calling a model through a US AI company's inference API falls outside this licence.

Not a door slammed shut but split in two: selling compute needs a licence, selling output does not.

How small 12 million TPP is

The line is 12 million TPP.

TPP is BIS's compute metric; it counts bit width, not just peak throughput. One H100 is roughly 16,000 TPP — an estimate by Wu Zihao of Ronghe Semiconductor Consulting, not an official figure.

Twelve million divided by 16,000 is 750. At eight cards a server, that is under 100 servers. A 10,000-card cluster runs 160 million TPP, 13.3 times the threshold.

The line is drawn very low.

Caijing cites industry estimates that a trillion-parameter pretraining run needs tens of thousands of cards. At 16,000 TPP each, 1,000 cards is 16 million — already over.

It counts aggregate compute: pooled halls combine, and inference counts alongside training. A low line plus aggregation targets subletting.

How licences are issued

BIS reviews case by case.

An application states end user, end use and compute scale, with KYC tracing ownership and checking direct and indirect control. Controlled chips are Nvidia Hopper and newer plus AMD equivalents, even inside foreign servers.

The hardest column is end use; get it wrong and the arithmetic is wasted.

A licence is not a long-term pass: changing use may mean reapplying. For an operator, compliance becomes ongoing customer management. For a customer, licence time is delivery time.

Existing halls get 90 days to apply. There is a 60-day comment period. The effective date is publication in the Federal Register.

Three dates: issued November 16, effective on publication, existing compliance at 90 days.

A window is not a buffer.

Who the exemptions are for

Two.

A US-headquartered end recipient or ultimate parent may keep receiving shipments with monthly reporting and no licence. And commercial recommendation systems on e-commerce and social platforms may apply — unless the recommender uses a generative model for copy, images or video.

The exemption looks at headquarters, not where the hall stands. Same data centre: recommendation exempt, generation not.

Generative copy, images or video defeat the exception even where recommendation is present. That boundary matters more than the country list: the rule is drawn by use, not geography.

Why now

The House passed the Remote Access Security Act, H.R.2683, in January 2026; it stalled in the Senate Banking Committee. BIS did not wait and used existing export-control authority.

The rule was drafted and guided by the Treasury and the National Security Council, issued by Commerce, and replaces the previous AI diffusion rule.

132 pages, 60 days of comment, 90 days for existing operations — an adjustment period, not a blanket ban. Unlicensed services face restrictions after 90 days.

Nvidia fell 2.9% intraday; the Philadelphia Semiconductor Index 5.5%.

Three groups feel it first

Chinese AI companies buying bare metal in the five countries, resellers moving compute through them, integrators building for those customers.

One must find compute again, one redo KYC, one rewrite contracts. The reseller has the worst of it: upstream needs a licence, downstream needs delivery.

A compute ledger is not an invoice: pooled halls combine, training and inference are recorded separately, a change of use is recalculated. Rack space used to be the product; now it is a customer list that can pass review.

Operators need new columns: headquarters, ultimate parent, compute scale, use, whether capacity is pooled.

Resellers cannot stop KYC at the contracting entity. Registered in Singapore with a Chinese parent still triggers; a US shell with unchanged ultimate control still triggers.

For a model company, API inference survives, but distillation cannot exceed the teacher, and the interaction data stays with the other side.

The rule governs where the compute sits, not the model's nationality.

The cards need not cross a border. The licence is still required.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

Related coverage

Cite this article

NeuroAI (脑机网), "The Rule Names Five Countries. The UAE Isn't One of Them," 17 November 2026, https://neuroai.site/a/na-policy-chip-remote-access-1117

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Policy & Governance” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories