US Officials Defend New AI Chip Controls, Say Trump Could Make Changes

The controls, outlined in a Jan. 13 interim final rule, will cap the number of AI chips that can be sent to most countries while introducing an exception for a group of allies that the Biden administration said already have strong AI technology protection rules. It also creates a new license exception for certain supply chain activities involving powerful chips -- except for China, Russia and other U.S.-embargoed countries -- and updates the agency’s validated end-user (VEU) program to create a separate licensing track for certain data centers that meet stringent new security conditions.

Although exporters and others won’t need to comply with most changes until May 15 and the new VEU-related security requirements until Jan. 15, 2026, the Semiconductor Industry Association said the new rule was “rushed.”

John Neuffer, CEO of the Semiconductor Industry Association, said the group is “disappointed that a policy shift of this magnitude and impact” is being published days before a new administration takes over. He also said they lack “any meaningful input from industry.”

“The new rule risks causing unintended and lasting damage to America’s economy and global competitiveness in semiconductors and AI by ceding strategic markets to our competitors,” Neuffer said. “The stakes are high, and the timing is fraught.”

The Computer & Communications Industry Association said the rule likely will “hamper U.S. firms’ global competitiveness and could advantage the very rivals that the rule intends to constrain,” while the Information Technology Industry Council urged the incoming Trump administration to withdraw the restrictions. And although the National Foreign Trade Council said its members support U.S. government efforts to address national security concerns, they are also "deeply concerned that this rule appears to have been developed outside of standard rulemaking procedures, which unfortunately appears to be a pattern for regulatory activities in the waning days of the current Administration."

NVIDIA, which makes many of the advanced graphics processing units (GPUs) that will be controlled by this rule, said the restrictions put global AI progress in “jeopardy.” It called the restrictions “unprecedented and misguided,” adding that the rule “threatens to derail innovation and economic growth worldwide.” It also said they were “drafted in secret and without proper legislative review.”

Other industry groups and companies called the rule “underinformed” and poorly executed, saying it will damage U.S. technology leadership more than help (see 2501060015 and 2501080034).

Commerce Secretary Gina Raimondo defended the new rules during a Jan. 13 call with reporters, saying her agency had “taken pains” to consult both with U.S. industry and lawmakers before publishing them. National Security Adviser Jake Sullivan said U.S. agencies met multiple times over eight months to craft the restrictions, each time taking into account input from the private sector.

Raimondo also acknowledged that “this is very hard, and no rule is perfect.” It's "designed to safeguard the most advanced AI technology and ensure that it stays out of the hands of our foreign adversaries, but also enabling the broad diffusion and sharing of the benefits with partner countries," she said.

Raimondo added that she hopes the incoming Trump administration “takes full advantage” of the 120-day public comment period.

“I fully expect the next administration may make changes as a result of that input,” she said.

Public comments are due May 15.

The rule groups the world into roughly three categories of countries, each of which will face different license requirements for certain advanced chips to power data centers used to train AI models.

Most countries around the world will face new AI chip license requirements and volume caps for those advanced chips -- including those classified under Export Control Classification Numbers 3A090.a, 4A090.a and “corresponding .z items in new § 742.6(a)(6)(iii)(A)” -- along with a license requirement for new ECCN 4E091, which controls certain advanced closed-weight dual-use AI models.

U.S.-arms embargoed countries in Country Group D:5, plus Macau, will continue to face strict licensing rules for advanced AI chips in addition to new restrictions over exports of the most powerful closed-weight frontier AI models.

The third group includes 18 other countries that the Biden administration has said are “low-risk,” and “generally will be able to obtain the most advanced [integrated circuits] without a license,” along with AI models controlled under new ECCN 4E091, as long as they certify compliance with specific requirements outlined in New License Exception Artificial Intelligence Authorization (AIA). The new exception requires the ultimate consignee to certify to the exporter that they won’t use the items in a way that helps train controlled AI models for entities headquartered or whose parents are headquartered outside of the 18 AIA countries, among other things. The exporter must submit this certification to BIS.

The AIA countries are: Australia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, South Korea, Spain, Sweden, Taiwan and the U.K.

Another exception introduced by the rule, License Exception Low Processing Performance (LPP), places caps on the number of advanced chips that can be shipped to nations outside of Country Group D:5 -- including China -- and end-users that aren’t subject to specific BIS end-user controls. It authorizes exports of low volumes of controlled chips that BIS said don’t pose “significant national security risks,” including shipments of up to “26,900,000 Total Processing Performance (TPP) of advanced computing [integrated circuits] ICs per-calendar year to any individual ultimate consignee.”

The exception doesn’t place restrictions on the number of shipments -- only on the volume of controlled chips. But the agency stressed that the TPP limit applies to shipments to any individual ultimate consignee, even if those shipments are made by multiple exporters or through more than one intermediate consignee.

The exporter also must first obtain a certification from the ultimate consignee saying that they haven’t received more than 26,900,000 TPP during the calendar year under License Exception LPP from all exporters, and that the “requested TPP for the specific transaction will not result in the ultimate consignee exceeding the TPP limit.” The exporter must give this certification to BIS within 30 days of the export.

A third exception, License Exception Advanced Compute Manufacturing (ACM), authorizes exports of 3A090, 4A090 and related .z commodities, software and technology to certain commercial or non-government end users located outside of Country Group D:5 nations if the “ultimate end use” is the “‘development,’ ‘production,’ or storage (in a warehouse or other similar facility) of such eligible items,” BIS said.

The license can’t be used if the item will be used to train an AI model, the agency said. And although License Exception ACM shipments don’t count toward country caps, exporters “must maintain a system of distribution that allows them to account for the number of controlled items transferred to, and subsequently out, of the facility,” BIS said, “with records updated every six months or more frequently.”

The rule also updates the BIS Data Center Validated End User program -- which allows certain preapproved data facilities to more quickly obtain advanced semiconductors and other U.S.-controlled items with artificial intelligence uses (see 2409300029) -- by creating new categories of VEUs. Companies approved as VEUs can obtain a single authorization from BIS to build data centers around the world, except in arms-embargoed countries, without having to apply for additional licenses.

Companies headquartered in the 18 AIA nations can apply to become Universal VEUs (UVEUs), while all other companies headquartered outside arms-embargoed countries can apply to become National VEUs (NVEUs). They still will face limits “on where they can geographically allocate their AI computing power,” BIS said, which will be determined by “the aggregate TPP of chips that meet or exceed the scope of ECCN 3A090.a.”

For example: A UVEU headquartered in a country listed in one of the 18 AIA countries can’t transfer or install more than 25% of its total AI computing power to or in locations outside the AIA countries, and it can’t transfer or install more than 7% of its total AI computing power to or in “any single country” outside the AIA countries. BIS also said a UVEU headquartered in the U.S. can’t transfer or install more than 50% of its total AI computing power outside the U.S.

The rule also includes a table of the total “per-company, per-country installed base allocation of TPP” for all NVEUs. The table includes caps from the first quarter of 2025 through the start of 2027. These caps “were identified through an extensive analysis of the size of AI compute clusters necessary to train the largest AI models, and the rate at which those clusters are likely to grow over the next three years,” the agency said, adding that they “represent clusters approximately 12 months, or one generation, behind the cluster size BIS believes” will be used to train the most advanced AI models.

“By providing a three-year roadmap, BIS aims to give predictability to industry while reducing the risks posed by the unchecked proliferation of the most advanced AI models and largest clusters at any given time,” the rule said.

To become a UVEU or an NVEU, a company must go through an “intensive application process,” which will include proving it can comply with strict U.S. cybersecurity, transit security and export control standards. UVEUs must specifically notify BIS 60 days in advance before they transfer any chips in which they're using a “UVEU authorization for such transfer, as well as any planned construction or installations of data centers in countries not previously included in prior notifications to BIS.”

The rule outlines a host of other compliance requirements those companies must meet, including adherence to U.S. outbound investment restrictions, restrictions around training an AI model, semi-annual reporting to BIS about chip installations, recordkeeping rules, possible audits from BIS, “appropriate” sanitization and disposal procedures, and more.

Along with the new controls over advanced chips, BIS also is controlling exports of the most advanced AI models under new ECCN 4E091. The new restrictions will apply to model weights of any closed-weight AI model -- a model with weights that aren’t published -- that “has been trained on more than [10 to the 26th power] computational operations.” The agency will apply a presumption of denial license review policy “to every license application involving the model weights of those models,” although new License Exception AIA can be used to export these models to the narrow set of 18 countries that BIS said have sufficient AI-related technology protections.

BIS said it settled on this licensing policy because the risks to U.S. national security “from even one case of diversion are extreme.” It also said it didn’t want to control less powerful model weights because they are already stored around the world and available from foreign sources, so “imposing controls on such models would be ineffective.”

The rule also creates new foreign direct product rule controls -- which places export license restrictions on certain foreign-made items that are made with U.S.-origin software or technology -- for these AI models after finding that many foreign entities are training advanced AI models or plan to train them using advanced technology directly produced with U.S. technology. To fall under the FDPR product scope, the 4E091 item must be produced by a “complete plant or ‘major component’ of a plant that is located outside the United States, when the complete plant or ‘major component’ of a plant, whether made in the United States or a foreign country, is subject to the EAR” and controlled under ECCNs 3A001.z, 3A090, 4A003.z, 4A004.z, 4A005.z, 4A090, 5A002.z, 5A004.z or 5A992.z.

A new red flag outlined in the rule gives guidance to American cloud computing providers that may be subject to these new controls. The red flag describes a scenario in which a U.S. service provider provides infrastructure, “in the form of clusters of advanced ICs,” to train an AI model for a separate AI organization, and the model weights of the resulting AI model are transferred to the separate AI organization. But if the U.S. provider performs this service for a U.S. subsidiary of a foreign entity headquartered in a destination subject to an ECCN 4E091 licensing requirement, “the performance of the training run and the transfer of model weights creates a substantial risk that the model weights will be diverted to the entity’s ultimate parent in violation of the” Export Administration Regulations, BIS said. In this scenario, the agency said the U.S. provider “may have aided and abetted a violation of the EAR.”

BIS said it’s encouraging exporters and cloud service providers “to take additional steps as part of their compliance programs to determine whether the model weights in question will be exported, reexported, or transferred to a destination subject to a license requirement and, if so, either to apply for a license or inform their customers of the obligation to do so prior to export.”

Senior U.S. officials, including Raimondo and Sullivan, stressed that the new BIS rule takes a “balanced” approach to restricting the dangerous proliferation of AI while also preserving U.S. technology leadership. A White House fact sheet said that certain chip orders with collective computation power up to roughly 1,700 advanced GPUs don’t require a license and don’t count against the national chip caps outlined in the rule. This captures the “overwhelming majority” of chip-related orders.

“It ensures that the infrastructure for training frontier AI, the most exquisite AI systems at the frontier, happens either in America or in the jurisdictions of our closest allies,” Sullivan said, “and that that capacity does not get offshore, like chips and batteries and other industries that we've had to invest hundreds of billion dollars to bring back onshore.”

Sullivan also said the rule will make it “hard” for U.S. competitors to “use smuggling and remote access to evade our export controls” while creating incentives for close trading partners to use “trusted vendors,” including VEUs, that meet strong American security requirements.

He stressed that these controls will likely “evolve” as technology does. “The application of them, of course, will have to be adjusted as new realities come forward,” Sullivan said. “That's the kind of infrastructure we have built with this set of rules and regulations, and that is what we will pass off to the next team that's coming in.”

Officials declined to say whether they have spoken to the incoming Trump administration about the new rules. One official said the controls are “bipartisan,” noting that Republicans, including the chair of the House Select Committee on China, have called for increased restrictions on AI chip exports (see 2501070010).

“This is an area in which we've seen bipartisan consensus, because ultimately it's a question of U.S. national security,” the person said.

Another senior administration official said they didn’t want to “characterize the views of the incoming team” or how they will manage the new rule. Other officials noted that the rule has a lengthy 120-day comment period and delayed effective date, which they said will allow for feedback and possible changes, and said the Biden administration felt it couldn’t wait any longer to publish the rule.

“From our perspective, time is really of the essence,” the official said. “In this case, we believe we're in a critical window right now, particularly vis-a-vis China. If you think about where our models are today relative to [the People’s Republic of China] models, the estimates range from being six to 18 months ahead right now, and so every minute counts, from our perspective.”

If the Biden administration were to delay the rule, the U.S. likely would see “significant stockpiling” of advanced chips and other items controlled by the rule, including by China, the official said. The person also pointed to the case of Huawei, which was able to stockpile large amounts of chips before it became subject to strict U.S. export controls several years ago.

“Just in that period alone, Huawei stockpiled enough chips for its telecom base stations for years,” the official said, “and we certainly don't want to see something like that happen again in this space.”

Another official said BIS has enough resources to enforce this rule, although the agency could use more, including within its Office of Export Enforcement.

“I think that the way we structure this will actually help us monitor the flow of these chips around the globe,” the official said, “and give us better insight to where there are problems that we can then crack down.”