Renting US AI chips from abroad now needs a license

The Commerce Department published an interim final rule on Monday that turns renting American compute from a data center in a third country into a licensable act whenever the customer is headquartered in China or Macau. It closes the route the Megaspeed affair put on every front page last week, and it landed the day before Nvidia reports earnings.
The mechanics are worth reading closely, because this is the first time Washington has regulated compute by the hour rather than by the box.
What the rule actually says
The new end-use control, section 744.24 of the export regulations, requires a license when a data center anywhere in the world gives a company headquartered in Country Group D:5 or Macau access to more than 12,000,000 "aggregate TPP" of advanced compute — roughly 760 H100-class accelerators at the rule's own performance unit — per the rule text. It counts a single site or "linked data centers," a definition broad enough to include facilities joined only by network infrastructure, not wires. Data centers, foundries and exporters all have to run due diligence on who ultimately owns and controls their customers. An existing provider already serving those customers has 90 days from the effective date to obtain a license or face future export restrictions.
There are three exemptions and they define the shape of the thing. Commercial recommender systems — the ranking models behind a shopping feed or a social timeline — are carved out entirely, and the definition explicitly does not cover systems that generate text, images, audio or code. Companies headquartered in the United States, or whose ultimate parent is, can still ship advanced chips to Indonesia, Malaysia, Singapore, Thailand and Brazil without a license, provided they file reports under section 743.10. Everyone else needs a license for those five countries, which the rule calls "higher-than-normal risk of diversion," and Commerce reserved the right to add more.
The rule also tightens the presumptions that chip fabricators and packagers face when producing advanced logic chips for designers in restricted destinations, and it formally rescinds the Biden-era AI Diffusion Rule, which was never enforced. Read the footnotes for the flavor of the year: the regulations no longer say artificial intelligence. Per Executive Order 14434, "Inaugurating the Era of Super Intelligence," the text now uses "SI" in place of "AI" throughout.
The signature does not match the draft
Commerce's own proposal, as the New York Times reported, covered seven countries — Singapore, Malaysia, Thailand, Indonesia, Brazil, South Africa and Norway — applied only to future shipments, exempted data centers already filled with American hardware, and asked companies to notify the government when Chinese customers trained models on their racks. Notification, not permission. Other parts of the administration wanted global coverage instead of a country list.
The published rule keeps five countries but reverses the concession that mattered most: incumbents now need licenses within 90 days. That is a materially harder rule than the one Commerce drafted, and the byline explains why. BIS's release says the interim final rule was "written by and at the direction of the Department of Treasury and the National Security Council." Treasury does not normally draft export controls. The enforcement bureau, meanwhile, is under pressure over exactly this failure — Semafor reports that Commerce Secretary Howard Lutnick faces growing criticism over weak enforcement, particularly around the Singapore-based company accused of smuggling chips to China, and the Times reported he had pushed to weaken rules that overlapped with his family's brokerage business. Whatever the intent, the outcome is a rule the enforcement agency did not get to write.
Who this actually costs
The Journal's reporting last week describes the business model the rule is aimed at: Megaspeed, the Singapore cloud provider at the center of the scandal, arranged nearly $12 billion of Nvidia chip orders for Chinese buyers between March 2024 and June 2025, covering several hundred thousand accelerators, according to documents the Journal reviewed. Those records do not establish how much hardware arrived, which is precisely the compliance problem — nobody downstream could tell either.
Now apply the threshold to the deals already signed. Tencent's five-year, roughly $7 billion arrangement with Oracle, reported by the Financial Times in October, is for access to about 100,000 advanced chips through Southeast Asian data centers — over a hundred times the rule's compute line. Under the new control, the operator of those racks needs a license within 90 days, and a license application naming a Chinese-headquartered end user runs into a presumption of denial. That is not a compliance chore. That is a contract that has to be renegotiated.
The carve-out for US-parented companies is where the competitive asymmetry sits. A US hyperscaler can keep selling into Singapore and Malaysia and file the paperwork; a Singapore- or Malaysia-headquartered operator serving the same Chinese tenant cannot. Malaysia has pulled in more than $34 billion of data center investment on the strength of that tenant base, according to the Journal's reporting. The regional builders who spent three years pitching themselves as the neutral option just discovered that neutrality needed a license.
Nvidia is exposed twice. It reported $12.4 billion of China revenue over six months, 7% of total, and the Southeast Asian billing route is what short sellers argue hides more than that. Kress now has to describe the exposure on a call where she would rather talk about why Nvidia stopped financing its own customers.
What the rule cannot do
The threshold is per data center, not per customer. A Chinese lab that spreads training across twenty providers, each supplying 11 million TPP, is under the line at every one of them, and the "linked data centers" clause only catches infrastructure that shares resources as a unified whole — separate accounts at separate companies are not obviously linked. The control also runs on "knowledge," which means a provider that declines to ask is only in trouble if BIS can prove it knew.
Then there is capacity. Every data center on earth serving a Chinese-headquartered customer above the line has 90 days to file, and applications are reviewed case by case with no presumption of approval. That is a queue, and queues are where policy gets softened in practice. The 60-day comment window is still open, so this is an interim rule that can move, and the political economy is already strange: Washington spent a year and a half dismantling a country-tiered compute regime and has now built a narrower one, with Treasury holding the pen. As we noted when the Pax Silica allies agreed their own protected-technology list, the direction of travel is coordination — the question was always whether the enforcement would follow the communiqué. This is the first test with a number attached.
What to watch: which licenses get granted in the first 90 days, and to whom — a single approval for a Chinese-linked tenant would tell you the rule is a tax, not a wall. Whether BIS publishes guidance on how it counts TPP across tenants. Oracle's next filing, for any language on the Tencent arrangement. And Tuesday, when Nvidia has to put a number on the region this rule targets.
If licensing compute works better than banning hardware, should Washington extend the same regime to inference, or does that break every cloud contract in Asia? Write to us via our contact page.







