The US extends AI chip controls to rented compute, adding licenses for five countries
Per a Nov 16 BIS rule, overseas data centers giving D:5 customers over 12M TPP of compute need licenses, and five countries face chip licenses.

The US Commerce Department’s Bureau of Industry and Security (BIS) published an interim final rule on November 16 that moves AI chip export control from who buys the hardware to who gets to use it. Under the new § 744.24, an overseas data center that gives a company headquartered in a Country Group D:5 destination such as China or Russia, or in Macau, access to more than 12 million “aggregate TPP” of compute triggers a license requirement for the covered chips behind it, reviewed case by case. The same rule adds destination licenses for Indonesia, Malaysia, Singapore, Thailand and Brazil, and formally rescinds the January 2025 AI Diffusion Rule. The text still carries a placeholder effective date of “publication in the Federal Register” and opens a 60-day comment period.
Key points
- The threshold is compute, not invoice value: a single data center, or linked data centers, serving D:5 or Macau-headquartered companies above 12,000,000 aggregate TPP needs a license for the exporter and for the operator.
- Indonesia, Malaysia, Singapore, Thailand and Brazil are treated as destinations with “a higher-than-normal risk of diversion” and need licenses for advanced chip shipments; chips bound for US-headquartered firms are exempt but must be reported monthly.
- The AI Diffusion Rule of January 15, 2025 (90 FR 4544) is rescinded, along with license exceptions AIA and LPP and ECCN 4E091, the entry that covered model weights.
Background
The rule being scrapped, the “Framework for Artificial Intelligence Diffusion,” was published on January 15, 2025. Its design was country quotas: wide allocations for allies, per-country caps elsewhere, plus limits on transferring the weights of the most capable models. It was never enforced. BIS said in May 2025 that it intended to rescind the framework, and this interim final rule writes that rescission into the regulations, taking the license exceptions built for it down with it.
Specific cases pushed the timing. On November 7 we covered the New Matrix story: a 17-month-old firm with an empty registered office was cleared to buy Nvidia H200s, and the due diligence behind the export license was described as a formality. Singapore’s investigation into the compute firm Megaspeed has been moving in parallel, and governments across Asia have already responded. Chips that never enter China but are rented remotely from a third-country data center are the path those cases keep tracing.
The shift in approach deserves its own sentence. Earlier controls worked through items and destinations: a chip class, a country, an entity list. Section 744.24 is an end-use control, and its test is how much compute a facility supplies to whom, which moves jurisdiction from the customs declaration to the lease and the rack.
The facts
Commerce’s press release titles the rule “Addressing Remote Access to American Super Intelligence Technology.” It is an interim final rule, RIN 0694-AK50, amending nine parts of the Export Administration Regulations: 15 CFR 732, 740, 742, 743, 744, 748, 762, 772 and 774.
| Item | Detail |
|---|---|
| Trigger | More than 12,000,000 aggregate TPP at one data center or across linked data centers, for D:5 or Macau-headquartered companies |
| Covered items | ECCNs including 3A001.z, 3A090 (except 3A090.c), 4A090 and 5A992.z, plus their software and technology |
| New destination licenses | Indonesia, Malaysia, Singapore, Thailand, Brazil (§ 742.6(a)(6)(iii)(A)(1)) |
| Existing operators | Must obtain a license within 90 days of the effective date to keep serving covered customers |
| Exception one | Compute used only for commercial recommender systems, with customer and TPP details filed with BIS |
| Exception two | Chips destined for US-headquartered companies, reported monthly under § 743.10 |
| Rescinded | AI Diffusion Rule (90 FR 4544), license exceptions AIA and LPP, ECCN 4E091 |
| Timing | Effective on Federal Register publication; 60-day comment period |
The arithmetic behind the threshold is worth spelling out. TechCrunch’s conversion is that 12 million TPP roughly equals the maximum compute of about 750 Nvidia H100 or H200 chips, which works back to roughly 16,000 TPP per card. TPP itself is not new: the technical notes to ECCN 3A090 have defined it for years. What is new is “aggregate TPP,” added to § 772.1, meaning the sum of TPP across a group of covered chips. The group can sit in one building or be assembled from several linked data centers, and the rule says linked includes digital links, not just physical ones.
Two kinds of object are being controlled, and they should not be blurred. The first is the item: exports, reexports and in-country transfers of chips, servers and their software and technology, which is about where goods move. The second is access to compute: who rents the machines and what they train on them, which is about a service. The second does not run on customs paperwork but on “knowledge.” An operator that knows, or should know, that the ultimate beneficiary is headquartered in D:5 has a license obligation; a customer whose ownership cannot be explained, or who refuses to document it, is itself a red flag. Exporters must also notify recipients when the end use later changes, which passes the compliance duty one more step down the chain.
Foundries and packagers are affected too. An IC designer headquartered in D:1, D:4, D:5 or one of the five countries that lacks “authorized” or “approved” status faces a presumption that its advanced logic chip is 3A090.a when it tapes out, which means the strictest treatment; authorized-designer status carries its own transition dates, December 31 and 180 days.
One more detail is easy to miss. The rule text uses “SI,” for super intelligence, instead of AI, citing Executive Order 14434, “Inaugurating the Era of Super Intelligence.” The vocabulary changed; the scope did not.
What others say
On how the rule was written, Chris McGuire of the Council on Foreign Relations called it “a major step forward” and added a line that several outlets picked up: “I have never seen a BIS rule that was not written by BIS.” This one was drafted by, and at the direction of, the Treasury Department and the National Security Council, then issued under BIS’s name. Joshua Kurlantzick, also at CFR, warned that controls on Malaysia “could be disastrous,” and Malaysia is where Chinese firms have concentrated much of their offshore compute.
Markets moved first. Per The Next Web, Nvidia shares fell as much as 2.9% and the Philadelphia Semiconductor Index as much as 5.5%. An analysis at 247wallst sizes the exposure differently: Nvidia’s direct revenue hit is limited, and the parties actually pushed to the front are the overseas data center operators and the customers renting capacity from them, who carry both the compliance cost and the risk of a terminated lease.
Chinese-language business coverage was not consistent before the rule landed. A Huxiu piece written before publication described planned licensing for seven countries, Norway and South Africa among them; the November 16 flash note and the BIS text both list five. The gap probably reflects an earlier draft, and it is a reminder that these lists can change until the Federal Register publishes. Huxiu’s industry figures are specific: ByteDance has planned about 36,000 B200 chips in Malaysia at an estimated hardware cost above $2.5 billion; Alibaba’s quarterly capex was 67.678 billion yuan, up 75% year on year; Tencent’s was 52.784 billion yuan, up 176%; Lambda lists H100 capacity at $5.54 per chip-hour, or about $4.08 million for 1,024 chips over 30 days.
Our take
For teams running overseas operations in Southeast Asia or the Middle East, or training on rented GPUs, this rule rewrites the diligence list before a contract, not after. Three questions decide most of it: where the ultimate parent is headquartered, whether the cluster’s aggregate TPP clears 12 million, and whether the termination, sublease and data-migration clauses survive a denied license. At roughly 750 H100/H200 equivalents, a thousand-GPU cluster is already near the line, so renting first and sorting out compliance later no longer works.
For the chip supply chain, licenses for five countries narrow the transshipment routes further. The US-headquartered exemption exists, but monthly reporting adds a legal and paperwork layer that smaller exporters will feel. For Chinese accelerator vendors, the higher risk attached to offshore rentals creates room for pilots, though migrating off CUDA still needs engineering validation that no policy can supply.
There are three honest gaps. First, the effective date is unset: the interim final rule still shows a placeholder where the Federal Register date belongs, and the 90-day clock runs from that date, so the only work available now is preparation. Second, deleting ECCN 4E091 is not liberalization of model weights; the control moves to destination and end-use tests, and reading the deletion as a loosening would be a mistake. Third, enforcement remains the open question. Shell companies and third-country relays have already been documented, including in the New Matrix case, and a tighter rule only writes the gap down in more detail.
What to watch
The Federal Register publication date is the next fixed point: it starts the 90-day window and tells existing operators when their filings are due. During the 60-day comment period, the fight will be over how aggregate TPP is counted, especially what qualifies as linked infrastructure across sites. BIS keeps the authority to add countries, so whether Norway and South Africa appear in a later version is worth tracking. The December 31 authorized-designer date sits in the same window, and design houses that tape out advanced logic need to plan both timelines together.