The keen-eyed readers of this blog will recall a story we told eighteen months ago about the time a young lawyer threw out his group-house’s struggling-but-still-functional toaster before he had purchased a new one. A housemate, whom we will continue to call Mike,[1] pointed out the folly of such action over his breakfast of disappointingly room-temperature bread.
The story seemed relevant at the time because BIS had just rescinded the AI Diffusion rule, but had not implemented any new controls on the export of AI chips or data center services.
A year and a half later, we are pleased to report (and to painfully overextend our metaphor) there is a new toaster on the counter. It has fewer slots, it came with a 132-page manual, and it requires you to know who will be eating the toast (ok, we will rest the toaster gag for a bit).
The new toaster (sorry, last one) is an interim final rule from the Bureau of Industry and Security (BIS) titled “Addressing Remote Access to Advanced Semiconductors” (the “Remote Access IFR”), posted on November 16, 2026. It takes effect on the day it publishes in the Federal Register, which had not happened as of this writing, and comments are due 60 days after that publication.
1. The Infrastructure-as-a-Service Problem
In the early 2020s, as BIS tightened controls on exports to China of the data center GPUs and related equipment and technology, an interesting phenomenon began. Suddenly, Southeast Asian countries vaulted up the list of top importers of the U.S.’s most advanced AI chips. Malaysia, Singapore, Indonesia, and Thailand were regularly cited as top-ten importers of AI chips, punching well above their economic weight class.
In October 2023, BIS issued a rule stating that advanced computing items could not be exported to third countries if they were “for an entity” that was headquartered in China or Macau. However, that rule did not expressly state that “for an entity” included delivering the compute from those chips to China as a service (an Infrastructure-as-a-Service arrangement, or IaaS). Further, in 2025, BIS rescinded the AI Diffusion Rule and stated that it would not enforce restrictions on AI models that would limit IaaS services into China.
So, the mid-2020s, companies headquartered in China could not lawfully buy the most advanced AI chips, but it appeared that they could rent time on those chips in data centers elsewhere. As demand plummeted in China, it rose in neighboring countries where data centers were built to provide the compute from those chips into China (well, yes, also probably to be diverted into China, but why do that when, ostensibly, you could provide compute services into China in compliance with the law, at least as it was enforced?)
The Remote Access IFR is a specific response to that phenomenon, targeting the practice of delivering Compute-as-a-Service into China from the region.[2] BIS’s press release says the IFR ensures existing restrictions “cannot be circumvented through remote computing arrangements involving data centers in third countries.”[3]
What may come as a surprise to many companies China and Southeast Asia,[4] is that the BIS takes the position that the prohibition on IaaS is retroactive(!)
That’s right, not only is the provision of IaaS to China a violation of U.S. regulations, it always was! It has been since 2023! The prohibition on the export of chips “for an entity” in China actually included the export of chips to third countries for (the provision of compute from those chips to) an entity in China!
Surprise! We got a new toaster years ago and it has always been there. We got it well before we threw away the old one, you just didn’t see it!
We will get into the deftness of that regulatory prestidigitation in Part II of this article. But first, the regs and their new restrictions.
2. Closing Three Routes to Advanced Compute[5]
Before the Remote Access IFR, a company in China that could not lawfully acquire advanced AI chips had three ways to obtain those chips’ computing power. It could rent time on chips sitting in someone else’s data center. It could have the chips diverted through a third country. It could design its own chips and have them made at a foundry outside China. The Remote
| Route to compute | What the IFR does | Where |
| Renting compute from data centers outside China | New data center end-use control with a 12,000,000 TPP threshold, plus BIS’s reading of an older control with a 50 percent test | §§ 744.24, 744.23(a)(3) |
| Diversion of chips through third countries | New license requirement for five countries, with an exclusion and monthly reports for U.S.-headquartered recipients | §§ 742.6(a)(6)(iii)(A)(1), 743.10 |
| Fabricating Chinese-designed chips at foreign foundries | Fewer ways for a foundry to treat a chip as uncontrolled, and a license requirement tied to the designer | Note 1 to ECCN 3A090.a, § 744.23(a)(3)(i)(C) |
3. Route One, Remote Access: Chip Sellers and Data Centers Must Know Who Is Renting the Compute
The IFR restricts remote access through customer due diligence, and two controls do the work. First, §744.24 is new and requires chip exporters to know how much compute their data center customer provides or will provide to covered companies. Second, §744.23(a)(3) dates to 2023 and, as BIS now interprets it, requires the covered company is the main user of a data center.
3.1. §744.24 - Chip sellers must know who their data center customers are serving
The new §744.24 requires a license to export, reexport, or transfer (in-country) advanced computing items if the person has “knowledge” at the time that the item is destined to a data center that “is providing or will provide access to more than 12,000,000 ‘aggregate TPP’” to or for companies headquartered in, or whose ultimate parent is headquartered in, Macau or a Country Group D:5 country.
3.1.1. The new control in § 744.24 starts at 12,000,000 TPP.
The elements are as follows:
| Element | What the IFR says |
| Covered items | ECCNs 3A090 (except 3A090.c), 4A090, the related “.z” items, and associated software and technology |
| Location of the data center | “Any destination worldwide” |
| Compute threshold | More than 12,000,000 aggregate total processing performance (TPP), at one data center or across “linked data centers,” which include “digitally-linked infrastructure, not just physically-linked infrastructure” |
| Covered customers | Companies headquartered in, or with an ultimate parent headquartered in, Macau or Country Group D:5 (China, Russia, Iran, and 19 other arms-embargoed countries) |
| Mental state | “Knowledge,” which under the EAR includes awareness of a high probability. The IFR explains what BIS expects an exporter to collect to show it lacks knowledge and when knowledge is presumed. See “Due diligence now runs two layers deep” below. |
| License exceptions | None |
| License review | Case-by-case |
3.1.2. Data centers inside the United States are outside the IFR’s text.
Section 744.24 restricts exports, reexports, and transfers (in-country), and the EAR defines a transfer (in-country) as a change within a “foreign country.” Note 3 speaks of the “foreign data center,” and the preamble refers to “data centers outside of the United States.” Chips installed in a U.S. data center have not been exported, so the IFR does not by its terms reach compute rented from a facility in the United States.
3.1.3. What 12,000,000 TPP looks like in hardware.
BIS does not translate the threshold into hardware configurations, so we ballparked some data center configurations below.[6] TPP is a per-chip figure derived from the manufacturer’s performance claims (twice the peak multiply-accumulate rate, times the bit length of the operation), and “aggregate TPP” is the sum across all the chips in a group.
BIS’s May 2025 guidance had pointed exporters toward scrutiny of data centers above 10 megawatts. The January 2025 AI Diffusion Rule allowed 26,900,000 TPP per customer per year without a license, an amount the White House then described as “roughly 1,700 advanced GPUs.” The new threshold is less than half of that, and it measures access provided to covered companies across linked facilities and not a single shipment.
| Hardware | Approximate TPP per unit | Units needed to exceed 12,000,000 aggregate TPP |
| NVIDIA H100 or H200 GPU | 15,832 | 758 GPUs, or 95 eight-GPU servers |
| NVIDIA B200 GPU | 36,000 | 334 GPUs |
| NVIDIA GB200 NVL72 rack (72 B200 GPUs) | 2,592,000 | 5 racks (4 racks total about 10.4 million) |
| NVIDIA B300 GPU | 54,000 | 223 GPUs |
| NVIDIA GB300 NVL72 rack (72 B300 GPUs) | 3,888,000 | 4 racks (3 racks total about 11.7 million) |
| AMD MI325X GPU | 20,919 | 574 GPUs |
3.2. §744.24(a)(1) - Chip Sellers and Data Center customers have diligence obligations.
3.2.1. Chip Exporters Due diligence now runs two layers deep.
The center of the IFR is Note 3 to §744.24(a)(1). To show it lacks knowledge, an exporter “will be expected to obtain and be able to provide sufficient information from the foreign data center regarding that data center’s ‘know your customer’ due diligence checks.” Three features stand out.
- The seller must know its customer’s customers. A chip or server supplier must collect the data center’s own customer screening and be able to hand it to BIS.
- The screening must reach ownership. The checks must cover “whether any data center customers are subject to direct or indirect ownership or control” by persons located or headquartered in Macau or a D:5 country. The preamble says this is to keep companies from avoiding the IFR “through corporate restructurings or spinoffs.”
- Silence counts against the seller. A data center “declining or being unable to provide information sufficient to verify compliance with this end use control shall be considered a red flag that presumptively provides you with ‘knowledge’ unless you are able to resolve that red flag before proceeding.”
For sellers of chips and servers, the obligations are the following
- License at the time of sale. A seller needs a license if it has knowledge that the data center is providing, or will provide, more than the threshold to covered companies. The diligence file described above is how the seller shows it lacked knowledge.
- Notice given to customer on every unlicensed sale. Where a sale proceeds without a license, the seller “must inform the recipient that a BIS authorization is required” if the items will later be used above the threshold for covered companies. The notice must also say that “the transferor, e.g., the data center operator, must obtain authorization for this change in end-use” and must follow BIS’s know-your-customer guidance in supplement no. 3 to part 732 (§ 744.24(e)).
- The same duties for resellers and integrators. The IFR speaks to every “exporter, reexporter, or transferor.” A reexporter cannot rely on the de minimis rules for U.S.-origin covered items built into foreign-made equipment (Note 2).
For data center operators, the obligations as follows:
- The operator is directly subject to the IFR. The prohibition applies “regardless of whether the person in possession of the EAR item was a party to the transaction to the original export” (Note 1). An operator that bought its servers from a local distributor, and never dealt with a U.S. exporter, is still covered.
- Signing a covered customer can be a transfer. The EAR has long defined a transfer (in-country) as “a change in end use or end user of an item within the same foreign country” (15 C.F.R. § 734.16), and the Export Control Reform Act uses the same definition (50 U.S.C. § 4801(6)). When an operator takes on covered customers and crosses the threshold, the IFR treats the operator as having transferred the chips already in its racks. The operator is the party that needs the license, and it needs the license before the change.
- The operator must screen its own customers. The preamble says that foreign data centers “will be expected to perform robust due diligence on their customers to identify direct or indirect ownership or control by Country Group D:5 or Macau persons.” An operator that cannot or will not share that screening becomes a red flag to every supplier, which puts its future supply of chips and servers at risk.
- Linked facilities are counted together. The threshold is measured at one data center or across “linked data centers,” so an operator cannot stay under it by spreading covered customers across sites that share workloads.
- Existing operators have 90 days. Data centers that were already providing a “significant volume of compute power” to covered companies before publication “must obtain a license with 90 days of the effective date of this IFR or face future export restrictions” (§744.24(b)(3)). The IFR does not define “significant volume” or say what the future restrictions would be.
3.2.2. §744.23(a)(3) has no size threshold and turns on a 50 percent test.
The new regulation described in 3.1 above (§744.24), is not the only rule that reaches rented compute. BIS takes the position that a second such rule has been in the regulations since October 2023, and the preamble to the IFR explains how BIS now interprets that rule to apply to data centers. BIS says the 50 percent test has applied since October 2023, with no grace period.
Section 744.23(a)(3) requires a license to send advanced computing items to most countries outside China when you know the items are “for an entity” headquartered in, or with an ultimate parent headquartered in, Macau or a D:5 country. The regulation’s own example is “a PRC-headquartered cloud or data server provider” located in a third country. The familiar case is a Chinese company’s overseas subsidiary buying servers for its own data center.
The preamble states that the control “is not, and has never been limited only to” shipments where the Chinese company is the named recipient or user of the hardware. It also covers shipments made “on behalf of” such a company, and BIS gives two examples.
- The covered company “was the purchaser in the transaction (directly or indirectly).” In plain terms, it paid for the chips, even through an intermediary.
- The covered company “was the exclusive or primary (greater than 50 percent of total compute) remote end user of the data center computing capacity.” In plain terms, it rents more than half of what the data center can do.
Suppose an independent operator in Malaysia buys servers, and a company headquartered in China rents 60 percent of the facility’s computing capacity. No Chinese company bought, received, or touched the servers. Under BIS’s reading, the servers were still shipped “for” the Chinese company, and anyone who supplied them knowing of that arrangement required a license. The size of the facility does not matter. A two-rack data center with one dominant Chinese tenant is covered.
3.2.3. Commercial recommender systems are excluded, with a report to BIS.
Section 744.24 does not apply “when the computing power access is only being provided for a ‘commercial recommender system.’” The definition is narrow. The system must be “exclusively designed and exclusively used to rank, filter, or suggest commercial products, services, or content to users.” It cannot be “designed or used to generate novel text, image, audio, video, software, or other content,” and a system that “incorporates such model” is disqualified “regardless of whether the system also performs recommendation functions.”
The exclusion is conditional. The exporter must submit to BIS the covered company’s name, the total TPP provided, and “the due diligence performed to ensure the computing power access is exclusively being provided” for that purpose. BIS will share the submission with the End-User Review Committee, the interagency body that decides Entity List additions.
4. Route Two, Diversion: Five Countries Now Require a License Unless the Recipient Is U.S.-Headquartered
Separately from the data center controls, the IFR adds a destination-based license requirement for advanced computing items going to Indonesia, Malaysia, Singapore, Brazil, and Thailand (§ 742.6(a)(6)(iii)(A)(1)). The summary calls them “countries which present a higher-than-normal risk of diversion.” Recent Justice Department cases allege transshipment of controlled chips and servers through Malaysia and Thailand, Singapore and Malaysia, and Thailand.
- The exclusion. No license is needed for shipments “to an ultimate consignee or end user headquartered in, or whose ultimate parent company is headquartered in, the United States,” if the exporter files reports under new § 743.10.
- The reports. Reports are monthly and due “no later than the first business day of the month following the transaction.” They must identify every party and its ultimate parent, the destination, each product’s ECCN, name, SKU, TPP and quantity, and the dates.
- Everyone else. A buyer in those countries that is not U.S.-headquartered needs a license, reviewed case-by-case, with attention to “the effectiveness of a destination’s export control regimes and coordination with BIS on the enforcement of export controls.”
- What remains available. License Exception ACA still covers chips that are not data-center 3A090.a items, so the practical weight falls on data center chips and the servers that contain them.
The Preamble warns that Commerce “will be actively monitoring for countries that tolerate evasion” and will consider “import restrictions, new license requirements on advanced semiconductor exports, downgrading the country’s status in the U.S. export control regime, and Entity List designations and temporary export prohibitions.” BIS may also impose a license requirement on any other destination by notice, including at the request of the Secretary of State or the National Security Advisor (§ 744.24(a)(2)).
5. Route Three, Fabrication: Foundries Lose Two Ways to Treat a Chip as Uncontrolled
The third route is manufacturing. A company that cannot buy a controlled chip can design one and send the design to a contract foundry. The IFR tightens the foundry’s side of that transaction in two ways. BIS describes the purpose as “critical oversight over companies headquartered in certain countries that seek to leverage advanced semiconductor production capabilities in third countries.”
5.1. The presumption is harder to overcome.
Under existing Note 1 to ECCN 3A090.a, a foundry or packaging house exporting a logic chip made at the 16/14 nanometer node or below must presume that it is a controlled data center chip. Until now the foundry could overcome that presumption in three ways. The first relies on a datasheet from an “approved” or “authorized” chip designer. The other two rely on the foundry’s or packager’s own attestation that the chip’s transistor count is below set limits. The Remote Access IFR removes the second and third where the chip’s designer is located in, or has an ultimate parent headquartered in, Country Group D:1, D:4, or D:5 or one of the five newly listed countries.
5.2. A designer in those places must be approved or authorized.
Approved designers are listed by BIS in supplement no. 6 to part 740. Authorized designers must be headquartered in Taiwan or a Country Group A:1 or A:5 country and cannot be located in, or have an ultimate parent in, Macau or a D:5 country. A designer headquartered in China therefore cannot be an authorized designer. Unless BIS has approved it, its foundry must treat every chip at those nodes as a controlled 3A090.a chip and comply with the license requirements that follow, whatever the chip is for.
5.3. A license requirement tied to the designer.
New § 744.23(a)(3)(i)(C) requires a license for a foundry or packaging house to export, reexport, or transfer a 3A090.a chip at those nodes “designed by a company that is neither an approved IC designer” nor an authorized one, “to or within any destination.” The preamble says this “ensures that a license requirement continues to apply” to foundries producing advanced chips “for non-approved, non-authorized IC designers wherever located.”
6. An Overview of the New Rule[7]—A Smaller Map and Heavier Diligence
The IFR formally removes the January 2025 AI Diffusion Rule from the Code of Federal Regulations. The AI Diffusion Rule and the Remote Access IFR were both crafted to address the same set of issues, but they do so in notably different ways. The old toaster had a slot for every country on earth. The new one has far fewer slots and a longer manual. Two differences deserve emphasis.
6.1. Under the Remote Access IFR, most of the world is license-free by destination.
For countries outside the listed groups, the EAR no longer imposes a list-based license requirement on advanced chips. The controls that remain are the end-use and end-user rules in part 744, which apply everywhere.
6.2. The model-weights controls are gone and nothing replaces them.
BIS states that “a more effective approach to address national security and foreign policy concerns that arise from the transfer of model weights is through the use of specified destination-based or end-use license requirements.”[8]
*Matthew Floyd is a law clerk in the firm's Washington, D.C. office.
FOOTNOTES
[1] Because that continues to be his name.
[2] Or even from outside the region, as we understand Brazil is setting up datacenters in partnership with some of China’s biggest technology firms. https://www.reuters.com/world/americas/brazil-launches-ai-supercomputer-push-splits-projects-between-chinese-us-firms-2026-08-20/
[3] The same press release states that the IFR was “written by and at the direction of the Department of Treasury and the National Security Council.” We cannot recall another BIS rule that arrived with a writing credit for two other agencies.
[4] And, frankly, to certain export control practitioners
[5] One housekeeping point before we brave these rules: The IFR uses “Super Intelligence” and “SI” where earlier rules said “artificial intelligence” and “AI.” That follows Executive Order 14434, which defines the new term by reference to the existing statutory definition of AI. The change is one of vocabulary and does not alter what is controlled. We will keep using AI. It took our spellchecker years to accept “AI,” and we are not putting it through that again.
[6] These are our own rough figures, built from published chip specifications. They are not classifications, and we would ask that nobody cite this table to BIS.
[7] . . . errr, toaster? Are we still doing that bit?
[8] We previously speculated that targeting models might be an easier way to address the IaaS gap because of the strong lobbying efforts of hardware developers. We stand corrected, though we did not predict the ousting of one of the biggest hardware lobbyists.