US closes the cloud loophole: new rule licenses remote access to advanced AI chips
For years, the workaround was simple: if you could not buy America’s most advanced AI chips, you rented them instead. A company barred from purchasing Nvidia’s top accelerators could still run its models on them — so long as the hardware sat in a data center in another country and the access came over the internet. On Monday, the US Commerce Department moved to close that gap, with a rule that treats renting the chips the same way as buying them.
- The rule: an interim final rule from the Bureau of Industry and Security (BIS), “Addressing Remote Access to Advanced Semiconductors” (RIN 0694-AK50), released November 16, 2026.
- The threshold: a license is required when a data center anywhere can supply more than 12 million aggregate TPP to companies headquartered in Macau or a Country Group D:5 country — roughly 750 Nvidia H100 or H200 chips.
- The targets: China, Russia, Iran, North Korea and 18 other embargoed destinations, plus Macau; indirect access through brokers and linked data centers is covered.
- Five more countries: Indonesia, Malaysia, Singapore, Thailand and Brazil now need licenses for advanced semiconductor exports, as higher-than-normal diversion risks.
- The deadline: existing operators have 90 days from the rule taking effect to get a license; the rule takes effect on Federal Register publication, with a 60-day comment period.
- The repeal: the Biden-era AI Diffusion Rule is formally rescinded, along with its model-weights control and country tiers.

Until now, export controls covered physical shipments: selling advanced chips to a restricted buyer needed a license, but renting time on those same chips from a server hall abroad did not. The new interim final rule from the Bureau of Industry and Security (BIS), part of the Commerce Department, erases that distinction.
The loophole, and how the rule closes it#
Under the rule, a license is now required to export, re-export or transfer advanced AI chips and servers to any data center, anywhere in the world, that may supply computing power to a company headquartered in Macau or in one of the Country Group D:5 embargoed destinations — China, Russia, Iran, North Korea and 18 others — or whose ultimate parent is based there.
The trigger is scale. The rule bites when a facility can deliver more than 12 million aggregate TPP, the export-control yardstick for computing power. That is roughly the maximum output of about 750 Nvidia H100 or H200 chips, according to TechCrunch’s reading of the rule.
The net is deliberately wide. Data centers that are linked — digitally or physically connected facilities that share resources and balance workloads — count together toward the threshold, so splitting capacity across sites does not dodge it. Indirect access through brokers or middlemen is covered too, and exporters and data center operators must check who their customers really are, including ultimate ownership. A customer that refuses to provide enough information to verify its ownership is treated as a red flag. Operators already selling that much compute to covered companies have 90 days from the rule taking effect to obtain a license or face further export restrictions.

Five countries face new license requirements#
Separately, the rule adds license requirements for exports, re-exports and transfers of advanced semiconductors to Indonesia, Malaysia, Singapore, Thailand and Brazil — five destinations Commerce says present a higher-than-normal risk of diversion. Applications will be reviewed case by case, taking each destination’s export-control regime and its enforcement coordination with BIS into account.
Shipments destined for companies headquartered in the US are excluded from those requirements, provided the exporter files new monthly reports identifying the parties, destinations, end users, chip classifications, quantities and dates. The carve-out that matters most to the tech industry is narrower than it first sounds: compute used exclusively for commercial recommender systems — the ranking engines behind online shops and social feeds — is exempt, but only if the operator tells BIS who the customer is and how much compute is involved, and only where the system does not generate new text, images, audio, video or code.

The end of the AI Diffusion Rule#
The rule also formally retires the Biden administration’s AI Diffusion Rule of January 2025, which sorted the world into tiers for chip access and would have restricted transfers of certain powerful AI models. That framework was never enforced, and the Trump administration said in May it would scrap it as an unnecessary burden on US companies and allies. The new rule removes its model-weights control, its country tiers and several license exceptions, replacing a geography-based regime with one aimed at the end user of the compute.
Why now#
The timing is pointed. The rule lands days after Nvidia co-founder Jensen Huang resigned amid a scandal involving an alleged Chinese spy linked to one of the chipmaker’s customers. The government has reportedly wanted to place that customer, Megaspeed, on a special entity list, believing it bought thousands of Nvidia chips to supply computing power to Chinese businesses. Reuters reported last week that Chinese AI companies are rushing to secure computing capacity overseas in anticipation of tighter restrictions.
BIS says the rule was written at the direction of the Treasury Department and the National Security Council, and frames the stakes in national-security terms: large-scale advanced computing can train models capable of supporting military and intelligence work, including weapons development, cyber-espionage and weaponized disinformation. The rule’s own text calls the new control “a logical extension of existing controls” that will “undermine the ability of companies headquartered in countries of concern from using US advanced Super Intelligence (SI) semiconductors to train advanced SI models.”
What to watch#
Three things matter next. First, how strictly BIS reviews license applications — case-by-case review leaves wide discretion, and the first decisions will set the pattern. Second, whether the 90-day window produces a rush of license applications or a quiet wind-down of existing arrangements. And third, how the five newly restricted destinations respond: Malaysia is already drafting its own data-center disclosure law, and Singapore is the region’s largest hub for the kind of compute the rule targets. The loophole is closed on paper; whether it stays closed depends on enforcement.
Sources: TechCrunch, “US tightens AI chip export controls to deter Chinese companies from accessing compute remotely”; Unite.AI, “BIS Releases Rule Targeting Remote Access to Advanced AI Chips”; CASRAI, “BIS issues rule on remote access to AI chips”; MadRobot, “US closes cloud loophole on China’s remote access to AI chips”.