What Washington issued on Monday

On Monday, 16 November, the US Commerce Department’s Bureau of Industry and Security (BIS) issued a rule titled “Addressing Remote Access to Advanced Semiconductors”. In its announcement, BIS says the aim is to make sure existing restrictions on sensitive American technology cannot be “circumvented” through remote computing.

It is an interim final rule. It takes effect as soon as it appears in the Federal Register, and comments are accepted for 60 days after that. BIS has already posted the full text on its own website. According to a client alert from the law firm Nixon Peabody, BIS expects Federal Register publication on 19 November.

The rule does two things. It adds a licence requirement for advanced AI chips sent to five countries. And it adds a worldwide check on data centres that sell large amounts of computing to companies tied to China.

The loophole: rent what you cannot buy

US export controls already stop the most powerful AI chips from being shipped to China. They did not stop a China-linked company from renting those chips in a data centre elsewhere and using them over the internet.

The rule’s text describes the worry as companies being able to “access significant volumes of advanced SI semiconductors remotely”, including through connected cloud computing. SI is the rule’s abbreviation for “super intelligence”, its term for advanced AI.

The Register reports that from 2021 to 2024, US cloud providers were required not to serve Chinese customers, while operators based outside the United States had no equivalent obligation. It adds that data-centre construction in Malaysia and Thailand has widely been read as partly aimed at Chinese customers who wanted to rent GPUs they could not import.

The chips never had to reach China. Only the computing did.

The new tripwire: 12 million TPP

The first part applies anywhere in the world. Under a new section of the export rules, § 744.24, a licence is needed to export, re-export or transfer covered chips when the seller knows that a data centre will provide more than 12,000,000 “aggregate TPP” to companies headquartered in China, Macau or another country in the US arms-embargo group D:5, or whose ultimate parent company is.

TPP, total processing performance, is the score export rules give a chip: twice its multiply-accumulate rate, times the bit length of the operation. The 12 million is counted across one data centre, or several “linked” ones that share resources and workloads over a network.

How big is that? GPU Smith, a site that explains export classifications, estimates an Nvidia H100 SXM, a common data-centre version, at 15,832 TPP. On that figure, 12 million is about 758 of those chips. That is SIGNAL’s arithmetic, not a BIS number, and only a manufacturer’s classification or a BIS ruling is official for a given product.

Data centres that already provide significant computing to such companies get 90 days from the effective date to obtain a licence.

Step 1Chips head to a data centreIn any country. The seller ships covered AI chips or servers.
Step 2Who rents the computing?Companies headquartered in China, Macau or another D:5 country, or owned by one.
Step 3More than 12 million TPP?Counted across the site and any linked data centres.
ResultLicence requiredNo licence, no shipment. Data centres already serving such customers get 90 days.
Schematic by SIGNAL, simplified from § 744.24 of the BIS rule. Not legal advice: the rule text governs.

Five countries move into the licence column

The second part is narrower. Exports, re-exports and in-country transfers of advanced AI chips to Indonesia, Malaysia, Singapore, Thailand and Brazil will need a licence. BIS will review applications “on a case-by-case basis”, and the rule says these countries were added because of concerns that advanced computing items are being diverted.

There is one exemption. Recipients headquartered in the United States, or whose ultimate parent is, do not need a licence if they report every transaction to BIS. The report lists the parties and their parent companies, the destination, each product’s export classification, SKU, TPP and quantity, and the date. It is due by the first business day of the following month.

That exemption is for US companies only. A European, Japanese or local operator in those five countries falls under the licence requirement, as the research publisher BERI also points out.

What the rule leaves alone

Not every AI workload is in scope. Systems “exclusively designed and exclusively used” to rank, filter or suggest commercial products, services or content, such as shop or feed recommendations, are exempt from the data-centre check, provided the operator sends BIS certain information. Systems that generate new text, images, audio, video or software are not covered by that exemption.

Shipments already on their way to a port on real orders when the rule is published can be completed within 30 days.

The rule also formally removes what remained of the Biden-era AI Diffusion Rule, including its licence exceptions AIA and LPP, the Data Center Validated End User programme and the export control on AI model weights. Its preamble is blunt: “The AI Diffusion Rule has never been enforced.”

Who feels it first

China-linked AI companies lose a route to computing. Counterpoint analyst Neil Shah told The Register the rule creates “a severe short-term compute bottleneck for Chinese tech giants”. That is an analyst’s forecast, not a measured effect.

Data-centre operators in the five countries gain paperwork, and so do chip sellers. According to Nixon Peabody, if a data centre refuses or fails to give an exporter information about its own customer checks, that counts as a red flag presuming the exporter’s knowledge. Sellers will have to ask those questions before they ship.

BIS itself expects a modest volume: about 100 extra licence applications and about 970 extra submissions a year, monthly reports included. Those are the agency’s paperwork estimates, not a forecast of how many deals will be blocked.

SIGNAL found no statement on the rule from Nvidia or AMD, and no official reaction from the five governments, in the sources it read.

What nobody knows yet

The clock depends on the Federal Register. The 30-day, 60-day and 90-day periods all run from publication, and the expected 19 November date comes from a law firm, not from BIS’s own announcement.

BIS has not said how quickly it will decide licences for the five countries, or how many it expects to approve.

Enforcement is the open question. Counting “linked” data centres and tracing ultimate parent companies needs information operators may not hold today. BERI also notes that the preamble and the rule text describe the recommender exemption in different terms. Expect questions, and possibly corrections, during the comment period.

Original sources