Isidora avatarIsidora & Cos Explain

The US Wants to Stop China Renting AI Chips Abroad. Can You Put Export Controls on the Cloud?

A large data centre campus among palm trees and green hills at dusk, with streams of blue light arcing from its buildings toward the sea
Summary

A new US rule, in force from 19 November, requires licences for advanced AI chips sent to data centres that rent large amounts of computing power to Chinese firms, and for chip shipments to Malaysia, Singapore, Thailand, Indonesia and Brazil. Why renting compute was the gap, how the rule tries to close it, why it is hard to enforce, what it costs Southeast Asia, and what Beijing says.

For three years Washington has tried to keep its best AI chips out of China. Chinese companies found a simple answer: if you cannot buy the chip, rent it. A data centre in Johor or Singapore buys the hardware legally, and a company in Beijing or Hangzhou trains its model on it over the internet.

On 16 November the US Commerce Department’s Bureau of Industry and Security (BIS) announced a rule to close that door. It takes effect on 19 November, when it is published in the Federal Register. The question behind it: can a country control computing power that is sold as a service, not shipped as a box?

What the rule actually does

The rule has two main parts, according to the text released for public inspection and a summary by the law firm Nixon Peabody.

First, a licence is now needed to send advanced AI chips and servers to any data centre, anywhere in the world, if the exporter knows it will give more than 12 million units of “aggregate total processing performance” to companies headquartered in China, Macau or the other arms-embargoed countries in the US “D:5” group. The count includes “linked” data centres, so splitting the work across several sites does not get around it. It also covers a middleman: a non-Chinese customer that trains models for a Chinese company.

Second, a licence is now needed for advanced chips shipped to five countries that BIS says carry an “unacceptable risk of diversion”: Indonesia, Malaysia, Singapore, Thailand and Brazil. Companies whose ultimate parent is in the United States are exempt, as long as they file monthly reports naming every party, the parent company and the final user. Data centres already serving large Chinese customers have 90 days to apply.

It also formally scraps the never-enforced Biden-era “AI Diffusion Rule”.

Why renting was a gap at all

Export controls were built for physical goods. A licence follows a crate from a factory to a port to a buyer. Cloud computing breaks that model: the chip stays in Malaysia, and only data and results travel. Until now, BIS did not read its own powers as covering that kind of remote access, a point made by the Washington think tank CNAS, which has pushed for tighter controls.

How big was the gap? CNAS cites one estimate that Chinese companies have remotely accessed about one million “H100 equivalents”, roughly a twentieth of the world’s AI computing power. That figure comes from a single estimate quoted by an advocacy-minded think tank, and I could not confirm it independently. The Institute for Progress, another Washington think tank, estimated in October that about 75% of China’s overseas AI computing would be in Southeast Asia by the end of 2026, according to CNBC. Firms such as Alibaba and ByteDance have rented capacity there.

The timing is not accidental. The rule came days after Nvidia’s chief executive resigned during a board investigation linked to the Singapore-based cloud company Megaspeed, and US Commerce and Justice officials confirmed a criminal investigation into Megaspeed over possible export-control violations. No court has ruled on that case. Beijing calls the related spying claims in US media “entirely groundless”.

Why it is hard to enforce

The rule does not police the cloud directly. It puts the duty on whoever sells or moves the chips. To show that they did not “know” a data centre serves Chinese firms, exporters will be expected to collect the data centre’s own know-your-customer checks, including on indirect ownership. A data centre that refuses to share that information is treated as a red flag.

That works against open arrangements, such as a Chinese tech giant renting a whole hall, and much less well against chains of small resellers or shell companies that rent compute and pass it on. The rule tries to catch the middleman case, but it relies on what exporters can find out, and on host governments’ willingness to check.

There is also a hole in the shape of the exemption. Chris McGuire, a former US National Security Council official now at the Council on Foreign Relations, warned of “a significant risk that Chinese companies will pivot from renting compute from data centers in Southeast Asia to renting compute from data centers in Virginia.” Analysts told CNBC that the exemption for US-headquartered companies leaves a narrow path for Chinese firms to keep renting from US cloud giants, so a large part of enforcement leans on those companies policing themselves. Pressure is already shaping that: the Financial Times reported, as cited by CNBC, that Oracle cancelled a chip-leasing deal with Tencent for about 100,000 chips after objections from US officials.

Who pays: the countries in the middle

Southeast Asia’s data-centre boom leaned heavily on Chinese demand. Malaysia has 55 operating data centres out of 117 approved projects, and data-centre construction contracts there hit a record RM18.2 billion (about US$4.4 billion) in 2026, The Business Times reported. Capital Economics estimates the rule could knock about 0.2 percentage points off Malaysia’s growth forecasts through 2028. Regional data-centre shares fell on 17 November.

None of the governments has said no. Malaysia’s trade minister, Johari Abdul Ghani, said: “If you want to be in the game, you have to follow U.S. regulations.” Thailand promised to “crack down” on data centres that re-export chips. Singapore said it does not condone companies using it to “circumvent or violate” other countries’ laws, and that it will seek “trusted access” for its firms. This is not servility. These countries cannot make the chips themselves, and losing US supply would cost more than losing Chinese tenants.

Brazil is the odd one on the list, far from the Chinese border. Its trade ministry (MDIC) stressed that this is a licence requirement, not a ban, and that the effect cannot yet be measured. The Brazilian Data Center Association expects limited impact, because much of the country’s planned AI capacity involves US companies. A Brazilian trade firm told Valor Econômico that delivery of such chips could stretch from about seven days to about 60, if the licence is granted at all. For a country that wants its own AI industry, being labelled a diversion risk adds delay and cost to projects that have nothing to do with China.

Beijing’s view, and the “backfire” argument

China’s Ministry of Commerce said Washington is stretching the idea of national security, abusing export controls and interfering in normal trade between Chinese companies and other countries’ companies. It said the rule goes against the consensus reached by the two presidents, and it promised countermeasures if the US goes ahead. The timing matters to Beijing: the rule landed days before Xi Jinping hosts Donald Trump in Shenzhen, and a tariff truce is due to expire in January.

There is also a serious argument, made from inside the United States, that controls defeat themselves. Nvidia’s former chief executive repeatedly said the curbs forced a US company to cede China’s $50 billion AI chip market to Huawei. The numbers point that way: Nvidia’s share of China’s data-centre AI chip market fell to 34% in the first half of 2026, while Huawei’s shipments rose to 572,000 units, a 28% share, according to IDC data reported by Reuters.

Research from the Asia Society adds a subtler point. Less computing power does not mean proportionally less capability. Chinese labs have squeezed more out of fewer chips through better algorithms and by designing models together with domestic chipmakers. Z.ai trained an image model entirely on Huawei hardware in January 2026. The same research says this does not prove the controls failed: the most advanced training still depends heavily on top-end chips, which is what renting abroad helped supply. CNAS, for its part, notes that China can make only around 8% of the advanced AI chips it needs at home, citing Goldman Sachs.

Scenarios

The rule mostly works. Large, visible Chinese training runs abroad stop, Southeast Asian operators shift to US and local clients, and China’s frontier training is slowed while its domestic chips catch up for everyday use.

The leak moves. Demand shifts to resellers, to Japan, Korea, the Gulf or Europe, or to US clouds. Each move forces another rule, and enforcement becomes a permanent chase.

Retaliation and bargaining. Beijing answers with its own controls, for example on critical minerals, and the rule becomes a chip in the wider trade talks, loosened or tightened by summit.

The likeliest outcome is some of all three.

Opinion: a real gap, closed in a lopsided way

Washington’s goal is coherent. If a chip is too sensitive to sell to a company, renting the same chip to the same company by the hour cannot be harmless.

But the rule is lopsided. It treats Brazil, Indonesia and Singapore as suspects and US cloud companies as trusted by definition, mostly on the strength of self-reporting. Countries that did nothing wrong pay with delays and lost investment, while the biggest loophole, US-owned clouds, stays open. A rule seen as protecting American market share as much as American security will be harder to sell to the governments whose cooperation it needs.

It also cannot stop the deeper trend. Controls buy time at the very top of AI training. They do not freeze China’s progress, and they push it to build a separate chip world that will one day compete with the American one. If time is what Washington is buying, the honest question is what it plans to do with it, beyond writing the next rule.

Sources

Comments