BIS Issues Interim Final Rule on Artificial Intelligence Diffusion as Biden Exits

On January 13, 2025, the Department of Commerce’s Bureau of Industry and Security (BIS) published an Interim Final Rule (IFR) that expands controls on advanced computing integrated circuits (ICs or chips) and model weights for artificial intelligence (AI) models, and established new security requirements for the storage of advanced chips related to AI. BIS’s announcement of the IFR indicated that the intent of this three-prong framework is to create a “trusted technology ecosystem” to prevent foreign adversaries from gaining access to advanced AI technologies. In its final days, the Biden administration moved forward with the IFR in the face of pushback from chip and AI industry executives, who say the rule will stymie development, and European Union officials, who fear the rule will limit access to needed chips.

The IFR was effective immediately, but compliance with export licensing requirements is not required until May 15, 2025. Items exported prior to that date are not controlled under the new rule so long as they arrive at their destination by June 16, 2025.

In this alert, we discuss each prong of the framework and detail the expected impact on the semiconductor, data center, AI, and related industries, as it moves to the Trump administration to continue with or perhaps adapt the IFR’s implementation.

IFR Imposes Further Restrictions on the Export of Chips, with Exceptions for Some Countries

The IFR imposes a global license requirement for the export of advanced computing ICs classified under Export Control Classification Numbers (ECCN) 3A090.a and 4A090.a and related items, but creates several license exceptions for transactions that pose a low risk of diversion or otherwise advance U.S. national security or foreign policy interests. The IFR also creates an authorization pathway for data center operators. It builds on previous rules from 2022, 2023, and 2024 that imposed license requirements on transactions exporting certain advanced computing ICs to the People’s Republic of China (PRC) and other countries of concern.

The IFR licensing scheme divides countries into three categories.

1. Artificial Intelligence Authorization (AIA) countries. The rules create a new and broad license exception for transactions with end users in low-risk destinations listed in Supplement 5 to Part 740 of the Export Administration Regulations (EAR) (Australia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, Republic of Korea, Spain, Sweden, Taiwan, the United Kingdom, and the United States), such that there are effectively no export restrictions to these destinations.
2. Restricted countries. The rule maintains the tight restrictions and licensing requirements for the PRC (including Hong Kong and Macau) and other countries subject to a U.S. arms embargo.
3. All others. For destinations and ultimate consignees located outside an AIA country but not in a restricted country, the IFR creates a “multi-part framework” for issuing licenses that considers (1) the compute power of the transaction and (2) security measures agreed to by the recipient, and (3) provides a limited license exception.

In an effort to ease the licensing burden, BIS added three new license exceptions applicable to advanced compute ICs.

1. The License Exception for AIA authorizes the export or transfer of some ICs and associated software and technologies to AIA countries unless the user or user’s parent company is headquartered outside of one of those destinations.
2. The License Exception for Low Processing Performance (LPP) authorizes the export (but not in-country transfer) of up to 26,900,000 total processing performance (TPP) of advanced chips per calendar year to any individual ultimate consignee or the entity that has ownership over the items. Eligible destinations are those outside restricted countries, provided that the ultimate consignee is not headquartered in or owned by an entity headquartered in a restricted country. The intent is for “all other” end users to be able to acquire low amounts of compute that do not present national security risks.
3. The License Exception for Advanced Compute Manufacturing (ACM) authorizes the export or transfer of certain ICs to “private sector end user[s]” where (1) the end user is located outside of a restricted country, (2) the end user is not headquartered in or owned by an entity headquartered in a restricted country, and (3) the ultimate end use is for the development, production, or storage of these items.

The IFR focuses primarily on total computing power, rather than number of chips, due to the difference in size of individual chips. In its press release, the White House estimated, for example, that the 26.9 million TPP threshold for license exception LPP equates to roughly 1,700 advanced chips.

Data Centers’ Authorization Program Will Expand

BIS is providing data center owners a pathway to import greater computing power by expanding its Data Center Validated End-User (VEU) Authorization and bifurcating it into two separate authorizations: Universal Validated End-User (UVEU) Authorization and National Validated End-User (NVEU) Authorization. The VEU program is designed to reduce or eliminate the need for specific licenses for export to trusted end-users that apply through an extensive and intensive application process and are cleared by BIS.

Entities that meet high trust and security standards and are headquartered in AIA countries can obtain a trusted status as a UVEU. These companies cannot transfer or install more than 25% of their AI computing power to or in locations outside of the countries not part of the AIA list and cannot transfer 7% or more of their total AI computing power, likely amounting to hundreds of thousands of chips, to any single country outside of those on the list. UVEU status allows users to build data centers around the world—except in restricted countries.

Companies with headquarters anywhere except restricted countries and not eligible for UVEU may apply for the NVEU Authorization, which enables them to purchase computational power equivalent to up to 320,000 advanced GPUs through 2027. For these companies, a per-company, per-country installed base allocation of TPP, measured by the collective computing power of items subject to certain export restrictions, will apply. The IFR includes the cumulative TPP per-company per-country allocation through Q4 of 2027. These figures represent one generation behind what BIS assesses is needed to train the most advanced dual-use AI models.

For companies that do not meet UVEU or NVEU requirements, the traditional license process under the EAR applies, subject to the license exceptions noted above. License applications must include a purchase order or contractual agreement reflecting the actual volume that the applicant plans to export, reexport, or transfer.

Model Weights Face Export Restrictions for the First Time

The IFR introduces Export Control Classification Number (ECCN) 4E091, a new control on closed-weight AI models that have been trained on more than 1026 computational operations. Licenses for the export or transfer of these model weights will be reviewed with a presumption of denial.

The IFR also creates a new Foreign Direct Product (FDP) rule for the model weights of certain foreign-produced closed-weight AI models. As with the new computing IC rules, BIS provides under license exception AIA for the export, reexport, or transfer (in-country) of model weights to end users in AIA countries, whose governments, according to BIS, (1) have implemented anti-diversion measures and (2) have an ecosystem that encourages firms to maximize their economic benefit by utilizing AI models. However, the license exception does not authorize the export or reexport to entities located in restricted countries.

The IFR also adds a new red flag to BIS’s “Know Your Customer” Guidance and Red Flags regarding the export of model weights. When a U.S. Infrastructure as a Service company provides products or services including model weights to an entity headquartered anywhere except AIA countries or a U.S. subsidiary thereof, and that customer uses these model weights to train an AI model falling under ECCN 4E091, a red flag will be raised. The red flag indicates that export of the model weights creates risk of diversion to a destination where a license is required, and the U.S. provider is expected to exercise additional diligence.

The IFR does not impose license requirements on open-weight models due to their smaller computational size, nor on model weights of closed models that are less powerful than the most powerful open-weight model.

Impact

The Trump administration has the power to implement the IFR as written or impose changes before compliance becomes mandatory in May of this year. We expect those in the industry to express their concerns about this rule directly to the Trump administration, which may impact the IFR’s next steps. We also expect the Trump administration to continue taking a tough stance on the PRC, as evidenced by a memorandum issued on President Trump’s first day in office that directs officials to conduct a review of all trade policies concerning the PRC. Regardless of the route this administration takes with respect to the IFR’s implementation, practitioners in the AI and semiconductor space should expect to see increasing restrictions moving forward. In just the past year, BIS has issued restrictions relating to cloud computing services, quantum technology, and semiconductor manufacturing equipment. We expect the fast pace of changing regulations to continue now that President Trump has taken office, but the contours of his plan to address the competition for semiconductor and AI technologies remain to be seen.

Associate Emilee Karr and paralegal Anabelle Lubin in the Washington, D.C. office also contributed to this article.