On January 13, 2025, the Department of Commerce Bureau of Industry and Security (BIS) announced new rules restricting the export of advanced artificial intelligence (AI) chips and certain closed AI model weights in an expected move that was preemptively criticized by giants in the tech and semiconductor industries. The 168-page "Framework for Artificial Intelligence Diffusion" interim final rule (the "Rule") adds a global licensing requirement for the export of advanced AI chips and closed AI model weights but with certain exclusions for some allied countries. Compliance with most portions of the new rule is required by May 15, 2025, and interested persons may submit public comments on the rule until May 15, 2025.
Previous related controls involved restrictions on exports of advanced semiconductors to countries of concern. For example, see our previous articles, New Rules Further Restrict China's Access to Semiconductor Technology and BIS Releases New Rules Updating Restrictions on Advanced Computing Chips, Manufacturing Equipment, and Supercomputing Items to Countries of Concern.) The new Rule, however, is much broader in its application to non-adversarial destinations. BIS explained the reasoning behind the new Rule: "A global licensing requirement is the most effective way to ensure that BIS can address the various risks associated with exporting large quantities of advanced computing integrated circuits (ICs) and the model weights of the most advanced AI models."
Items Covered by the Rule
The new export controls described in the Rule expand controls on advanced computing ICs controlled under Export Control Classification Numbers (ECCN) 3A090.a and 4A090.a and items in the corresponding .z ECCNs.
ECCN 3A090.a controls: integrated circuits having one or more digital processing units having either: 1) "total processing performance"1 of 4800 or more, or 2) A "total processing performance" of 1600 or more and a "performance density"2 of 5.92 or more. ECCN 4A090.a controls "Computers, electronic assemblies, and components containing integrated circuits, any of which meets or exceeds the limits in 3A090.a." The "corresponding .z ECCNs include any ECCNs ending in .z, which reference the controls in ECCNs 3A090.a or 4A090.a (e.g., 3A001.z.1.a and 5A002.z.1.a).
BIS also created new ECCN 4E091 to control model weights of certain advanced closed-weight dual-use AI models. Model weights are "numerical parameters within an AI model that help determine the model's outputs in response to inputs." Because model weights are important to the functioning of advanced AI models, BIS is imposing licensing requirements on the export, reexport, or transfer (in-country) of model weights of any closed-weight AI model – i.e., a model with weights that are not published – that has been trained on more than 1026 computational operations.
Three-Tier Country Classification System
The Rule creates three different tiers of countries, each having different levels of restrictions for the advanced AI chips.
Generally unrestricted countries: A list including 18 U.S. allies that will generally not be subject to the restrictions on advanced AI chips in the Rule. The complete list of these countries includes:
- Australia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, South Korea, Spain, Sweden, Taiwan, and the United Kingdom.
Completely restricted countries: Those countries identified in Country Group D:5 of Supplement No. 1 to Part 740 of the Export Administration Regulations (EAR), which are countries subject to a U.S. arms embargo, and Macau. These countries are excluded from receiving the advanced AI chips subject to the Rule. The full list of these countries includes:
- China (including Macau), Russia, Afghanistan, Belarus, Burma, Cambodia, Central African Republic, Cuba, Eritrea, Democratic Republic of the Congo, Haiti, Iran, Iraq, North Korea, Lebanon, Libya, Nicaragua, Somalia, South Sudan, Sudan, Syria, Venezuela, and Zimbabwe.
Countries subject to specified "Allocation Amount": All other countries fall under this category. BIS has created a license exception for exports of limited quantities of advanced AI chips to these destinations. For transactions involving larger quantities of advanced AI chips, BIS will review under a license policy "that enables the export of ICs up to a specified country allocation." From 2025 to 2027, these countries will be subject to a cumulative maximum installed base allocation of 790,000,000 total processing performance (TPP). This TPP allocation "represents permitted cumulative installed base for the entire period to and inclusive of 2027."
Foreign Direct Product Rules
The Rule creates a new Foreign Direct Product (FDP) rule related to the AI model weights controlled under new ECCN 4E091. The product scope of the new "AI Model Weights FDP Rule" applies to foreign-produced items specified in ECCN 4E091 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 is subject to the EAR and specified in ECCN 3A001.z, 3A090, 4A003.z, 4A004.z, 4A005.z, 4A090, 5A002.z, 5A004.z, or 5A992.z. A foreign-produced 4E091 item meets the destination scope of the AI Model Weights FDP Rule if the foreign-produced item is destined to any location worldwide.
The Rule also expands the scope of the Advanced Computing FDP Rule. Whereas the destination scope of the Advanced Computing FDP Rule currently applies to only certain Country Group D countries (see Supplement No. 1 to Part 740 of the EAR), the destination scope will expand to include any destination worldwide.
License Exceptions
The Rule creates three new license exceptions and expands the previously created Validated End User (VEU) program. BIS's press release announcing the Rule describes the new license exceptions as follows:
Exceptions for certain allies and partners: New License Exception Artificial Intelligence Authorization (AIA) allows for the export, reexport, or transfer (in-country) of advanced computing chips, without an authorization, to a set of allies and partners.
Exceptions for supply chains: New License Exception Advanced Compute Manufacturing (ACM) allows for the export, reexport, or transfer (in-country) of advanced computing chips, without an authorization, for the purposes of development, production, and storage of these chips, except to arms-embargoed countries.
Low volume exception: New License Exception Low Processing Performance (LPP) allows limited amounts of compute to flow globally, except to arms-embargoed countries.
The "Framework for Artificial Intelligence Diffusion" interim final rule is complex and highly technical, and it has already come under fire from large corporations and industry groups claiming the Rule will stifle competitiveness and innovation. The above brief description contains only a portion of the revisions described in the lengthy Rule. For more detailed analysis of specific issues related to the new Rule or how it may impact your business, or if you are interested in submitting a public comment, please contact the attorneys at Torres Trade Law.
Footnotes
1. "Total processing performance" is defined in the Export Administration Regulations (EAR) as "2 × 'MacTOPS' × 'bit length of the operation', aggregated over all processing units on the integrated circuit. 'MacTOPS' is the theoretical peak number of Tera (1012) operations per second for multiply-accumulate computation (D = A × B + C). The 2 in the 'TPP' formula is based on industry convention of counting one multiply-accumulate computation, D = A × B + C, as 2 operations for purpose of datasheets. Therefore, 2 × MacTOPS may correspond to the reported TOPS or FLOPS on a datasheet. 'Bit length of the operation' for a multiply-accumulate computation is the largest bit-length of the inputs to the multiply operation. Aggregate the TPPs for each processing unit on the integrated circuit to arrive at a total. 'TPP' = TPP1 + TPP2 + . . . . + TPPn (where n is the number or processing units on the integrated circuit)."
2. "Performance density" is defined in the EAR as 'TPP' divided by 'applicable die area," and the "applicable die area" is measured in millimeters squared and includes all die area of logic dies manufactured with a process node that uses a non-planar transistor architecture."
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.