# US imposes licensing on chip rentals and tightens its security

> The BIS requires a license to send advanced chips to data centers that provide computing power to Chinese companies, including when access is remote or resold.
> By José Maldonado · Technology · 2026-11-18

The hole that had been open for two years in US export controls was not at a port, it was in a rental invoice. A Chinese company could never buy an H100 card and still train its model in a data center in Malaysia, Singapore or Brazil, paying for hours of computing power what it would previously have paid for hardware. The Commerce Department's Bureau of Industry and Security closed that path on November 16, 2026.

The rule is technically called the interim rule on remote access to advanced semiconductors and is published in the [Federal Register](https://www.federalregister.gov/public-inspection/2026-23529/addressing-remote-access-to-advanced-semiconductors) on November 19, the day it takes effect. [The analysis of the new license](https://www.nixonpeabody.com/insights/alerts/2026/11/16/bis-targets-remote-access-to-advanced-chips-through-overseas-data-centers) sums it up better than the official statement: if you know that your chip ends up providing computing power to a Chinese company, you need a license, even if the chip never crosses that border.

The figure that governs the rule is 12 million total processing performance aggregate, a unit equivalent to the maximum power of about 750 H100 or H200 accelerators. Above that threshold, the data center falls within the scope of the rule and the exporter is responsible for what happens inside.

## Tightening chip exports

The obligation falls on the exporter and on the reexporter, and also on anyone who transfers the equipment within the same country. It is triggered when that party has knowledge that the covered chips or servers are going to a data center anywhere in the world that provides access to more than 12 million TPP to companies headquartered, or with a parent company, in China, Macao or the D:5 group of countries.

That group includes Russia, Iran and North Korea along with another eighteen destinations under United Nations arms embargo, so the rule is not aimed only at Beijing. [The summary of the change](https://www.xataka.com/componentes/eeuu-no-dejaba-vender-chips-avanzados-nvidia-a-china-alquilarlos-remoto-eso-se-acabo) highlights what the industry finds most uncomfortable: distributed computing also counts.

The definition of a linked data center includes facilities connected digitally and not only physically, which closes the obvious shortcut of splitting a large cluster into four small buildings. Nor is there any de minimis: if a US-origin component ends up inside equipment manufactured abroad, the prohibition still applies.

The text adds a customer knowledge requirement that in practice works as a covert audit. The exporter must ask the foreign data center for information about who its customers are and who controls them, and a refusal to provide it is interpreted as a red flag that presumes knowledge.

## The threshold that decides everything

Twelve million TPP sounds like a technical figure, and it is the line that separates a small cloud provider from an industrial operator. A facility that stays below the threshold does not fall under the rule, which leaves the consumer services market intact and creates an obvious incentive to split capacity.

The rule also covers indirect access, and that is the point that will break the most contracts. A customer outside the list can fall within scope if it trains models for a D:5 group company, if it takes part in that training or if it supplies the resulting models.

Compliance does not end with the shipment. Changing the use of covered equipment from a permitted activity to a prohibited one counts as a transfer within the country and requires authorization, and anyone who has shipped material without a license since publication must notify the recipient in writing.

There is a narrow exception for commercial recommendation systems, understood as those designed and used only to classify, filter or suggest products or content. Any system that generates text, image, audio or video is excluded, even if it also recommends, and availing of it requires filing a declaration with the BIS.

Data centers that were already providing a significant volume of computing power to covered companies have 90 days from entry into force to obtain a license or expose themselves to future restrictions. Shipments already in transit can be completed without one if they arrive within 30 days.

That staggered timetable explains why the first consequences will not be seen on the publication date. Contracts signed before November remain valid during the quarter, and the real discussion will begin when an operator has to renew with a customer it cannot fully identify.

## Why are five countries on a separate list?

Indonesia, Malaysia, Singapore, Thailand and Brazil receive their own separate licensing requirement, due to a diversion risk the Department considers higher than normal. It is the part of the rule that has least to do with China and most to do with the geography of computing rentals.

That monthly paperwork is also a snapshot of the market. With each report, the US bureau will know which advanced chips entered those five countries and whose hands they actually ended up in.

The way out offered by the text is a paperwork regime. Transactions with recipients whose headquarters or ultimate parent company is in the United States are excluded from the requirement if the exporter submits monthly reports with parties, destinations, product classifications, quantities and dates, delivered on the first business day of the following month. Anyone who cannot substantiate that chain with documents simply has no exemption.

That reporting duty gives the bureau a transaction-by-transaction picture of where advanced chips go in five economies, material that could feed future cases or tighter controls. For the exporter, the exemption becomes a permanent administrative job.

The rule also repeals the previous stage's artificial intelligence diffusion rule, with its control over model weights and its country-tier system, and reorganizes regional stability licenses. Control over chips and servers, by contrast, stays and is strengthened.

## The control that remains physical

The move has an industrial logic that goes beyond security. If computing power is the product being sold, controlling where it is installed and who uses it is the only way an export control means anything, and cards have become too expensive to trust the buyer to declare its final destination in good faith.

The [investigation opened by Washington in 2024](https://observatorioblockchain.com/en/ai/eeuu-investiga-desvio-chips-nvidia-china-desde-2024/) into the diversion of Nvidia accelerators to China anticipated this ground, and the November rule puts in place the framework that was missing. What changes for a European operator is simple: it now has to know who is on the other side of the cloud contract.

It remains to be seen whether the threshold survives contact with commercial reality, because a rule that is triggered at 750 accelerators invites building 740 and splitting the rest among partners. The BIS has already warned that it will monitor countries that tolerate evasion, with sanctions that include import restrictions and entity lists.

For a European operator, the calculation is simpler than it seems. If its end customer is in the United States or has its parent company there, it qualifies for the exemption with monthly reports; if it trains for a D:5 group company, no exemption applies and the contract becomes a pending license.

The 60-day comment period will show whether the industry manages to soften the text before it becomes final. Meanwhile, the date of November 19 is the only one that matters for anyone with a pending shipment or a computing contract signed with a customer on the list.
