
The United States has formally closed the “cloud loophole” that let Chinese AI companies rent advanced American computing power from data centres abroad. The Commerce Department’s Bureau of Industry and Security issued an interim final rule on Monday, November 16, 2026, requiring a licence before advanced AI chips and servers can be supplied to any data centre that may provide more than 12 million total processing performance (TPP) of compute to firms in China or other embargoed countries. It was announced as President Donald Trump began a trip to Asia ahead of talks with Chinese President Xi Jinping.
The Bureau of Industry and Security, known as BIS, published the rule on Monday as Trump set off for Asia, days after reports that it was being finalised. That earlier reporting said it was coming, but no official text existed then. Now it does, and the rule is already in force as an interim measure. BIS said it was written at the direction of the Treasury Department and the National Security Council.
Reuters reported on November 12 that Chinese AI firms were already rushing to reserve overseas Nvidia-powered capacity before the rule landed. The rule does not ban every advanced chip sale. It targets a workaround: instead of buying restricted chips, a Chinese company could rent time on powerful servers in another country. Remote access is now treated like a physical export.
Washington has restricted direct sales of advanced AI chips to China since 2022, aiming to slow Beijing’s development of military-use AI. But the controls covered the chips, not the computing power they produced elsewhere — Chinese firms could rent Nvidia-powered capacity abroad without ever taking delivery of a chip.
The gap burst open with the Megaspeed scandal: a network tied to Alice Huang and her sister Huang Yu acquired about $7.5 billion of Nvidia chips and moved roughly 240,000 restricted chips to Chinese firms, while Speedmatrix bought at least $4.8 billion of AI equipment in 2024-26. Oracle cancelled a roughly $7 billion deal to lease capacity to Tencent, and Nvidia co-founder Jensen Huang resigned. Trump pushed for tighter controls while Commerce Secretary Howard Lutnick was reported to oppose them.
A licence is required to export, re-export or transfer advanced chips or servers to any data centre, anywhere, if it may provide more than 12 million aggregate TPP of compute to companies headquartered in, or ultimately parented by firms in, Macau or Country Group D:5. The D:5 group covers China, Russia, Iran, North Korea and 18 other embargoed countries. Twelve million TPP is roughly the maximum power of about 750 Nvidia H100 or H200 chips.
The rule also covers indirect access through go-betweens or brokers, and exporters and operators must monitor customers’ ownership structures and activities. Separately, it imposes licensing requirements for advanced-chip exports to Singapore, Malaysia, Indonesia, Thailand and Brazil, which the Commerce Department said present a higher-than-normal risk of diversion — though not for chips destined for US-based companies. Existing operators have 90 days to get a licence. Recommendation systems are exempt; generative systems are not.
The rule itself makes the administration’s case in blunt terms. “This new control is a logical extension of existing controls,” it reads. “These restrictions will undermine the ability of companies headquartered in countries of concern from using U.S. advanced Super Intelligence (SI) semiconductors to train advanced SI models.”
“Access to such computing power can be used to train and run advanced SI models capable of supporting military- and intelligence-related activities that could threaten U.S. national security or foreign policy interests,” the rule reads.
Computing power has become the choke point of the AI race: a frontier model cannot be trained without thousands of advanced chips, and the United States still makes the best of them. By governing remote access as well as physical sales, Washington is closing the last large gap in a control regime built up over four years — and telling the world’s data-centre operators they are now part of it.
For Nvidia and cloud providers, the rule means new paperwork, customer checks and licence applications. For Chinese AI firms, the cheapest route to frontier-scale computing — renting it abroad — is now gated by Washington. The timing matters too: it arrives days before Trump meets Xi, and in the same week as Nvidia’s quarterly earnings.
The 90-day clock for existing operators is the most immediate deadline, and the first licence applications will test how strictly BIS enforces the new line. As an interim final rule it takes effect quickly while public comment runs, so changes are still possible.
Attention now turns to diplomacy. Trump is due to meet Xi and Russian President Vladimir Putin in Shenzhen this week, and the chip rule is almost certain to come up. Beijing has answered past controls with export restrictions on critical minerals, and a response is likely.
It is the gap that let Chinese AI firms rent computing power on advanced chips in data centres outside China, rather than buying the chips. US controls covered physical sales, so remote access was largely unregulated.
TPP, or total processing performance, is the US metric for computing power. Twelve million TPP is roughly the power of about 750 Nvidia H100 or H200 chips — enough to train or run advanced AI models.
The remote-access controls target firms headquartered in, or ultimately parented by companies in, Macau or D:5 jurisdictions, including China, Russia, Iran and North Korea. Five diversion hubs — Singapore, Malaysia, Indonesia, Thailand and Brazil — face new licensing rules.
No. A licence is required for advanced-chip exports to those five countries, but not for chips destined for US-based companies. The aim is to stop diversion, not legitimate business.
Existing operators have 90 days from the rule taking effect to get a licence, and must monitor customers’ ownership structures to make sure they are not serving D:5 firms.