The US Department of Commerce’s Bureau of Industry and Security (BIS) introduced its long-anticipated (some would say dreaded) Framework for Artificial Intelligence Diffusion (the AI Diffusion Rule) on January 15. Highlights include the first-ever export controls on artificial intelligence (AI) models and worldwide license requirements for advanced chips.

Unfazed by warnings from many in the US tech industry that the regulation would disrupt and harm US companies, BIS published an interim final rule that amends the Export Administration Regulations (EAR) and aims to strengthen controls on the global diffusion of advanced AI models and large clusters of advanced computing integrated circuits.

These changes are effective as of January 13, but the compliance date for the new license requirements is not until May 13. Public comments on the AI Diffusion Rule should also be submitted by May 13.

Overview

Here is a roadmap of some of the biggest changes in the AI Diffusion Rule:

• Worldwide licensing requirement: advanced integrated circuits (ICs) under Export Control Classification Number (ECCN) 3A090.a, 4A090.a, and the corresponding .z paragraphs will require an export license for all destinations starting May 13. To go with the worldwide requirements, BIS is introducing a new licensing policy that determines who can receive chips using country-based total processing performance (TPP) allocations.
• New ECCN 4E091: BIS is imposing a worldwide licensing requirement on model weights, or parameters, of certain advanced closed-weight dual-use AI models. And why not have yet another one? Yes — there is a new Foreign Direct Product Rule (FDPR) that covers 4E091.
• AI Authorization (AIA) Countries : These lower-risk destinations, identified in new Supplement No. 5 to Part 740, are eligible for much more favorable treatment under the new export controls: Australia, Belgium, Canada, Denmark, Finland, Germany, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, South Korea, Spain, Sweden, Taiwan, the United Kingdom, and the United States.
• New license exceptions: introducing License Exceptions Artificial Intelligence Authorization (License Exception AIA), Advanced Computing Manufacturing (License Exception ACM), and Low Processing Performance (License Exception LPP).
• Other changes include:
  • License Exception Notified Advanced Computing (NAC) was amended to give clearer notification procedures.
  • Data center validated end users (DC VEUs) divided into universal validated end users (UVEUs) and national validated end users (NVEUs).
  • New red flag guidance for companies involved with AI, including cloud providers with customers training large AI models.

Worldwide Licensing Requirements

The AI Diffusion Rule establishes a worldwide licensing requirement for ECCNs 3A090.a, 4A090.a, and the corresponding .z subparagraphs. These items are:

• ECCN 3A090.a: ICs having one or more digital processing units having either of the following:
  • a.1. TPP of 4800 or more.
  • a.2. TPP of 1600 ore more and a ‘performance density’ of 5.92 or more.
• ECCN 4A090.a: Computers, “electronic assemblies,” and “components” containing integrated circuits, any of which meets or exceeds the limits in 3A090.a.

License applications for an AIA country, or for a recipient with its headquarters or ultimate parent company in an AIA country, will be reviewed under a presumption of approval. Applications for Country Group D:5 or Macau, or for a recipient with its headquarters or ultimate parent company in one of those locations, will be subject to a presumption of denial.

For all other destinations, BIS will apply a presumption of approval up to a per-country allocation of 790,000,000 TPP from 2025 to 2027, after which a presumption of denial will apply. Having one across-the-board allocation for every country, be it India or the Solomon Islands, seems like it will be a compliance nightmare for everyone, including BIS. To implement the TPP scheme, license applications for items subject to the worldwide requirement must include documentation (e.g. purchase order or equivalent firm contractual statement) reflecting the actual volume of items to be sold. Exporters can check whether their destination country’s allocation has been met here.

The worldwide licensing requirement will be reflected in the expanded destination scopes of both the Advanced Computing FDPR and License Exception Advanced Computing Authorized (ACA).

A friendly reminder that the other subparagraphs of 3A090 and 4A090 are still subject to another “worldwide” licensing requirement under the end use controls at EAR Section 744.23, although only for recipients with headquarters or ultimate parents in Macau or Country Group D:5.

Introducing ECCN 4E091 — AI Model Controls

The new ECCN 4E091 covers “‘parameters’ for an AI model trained utilizing 10^26 or more ‘operations.’ “Parameters” refers to any value learned during training (in other words, model weights), while “operations” includes mathematical operations used in pre-training as well as subsequent training.

A license is required to export, re-export, or transfer (in-country) 4E091 parameters to all destinations worldwide, with a presumption of approval for AIA countries and a presumption of denial everywhere else. No license is required for deemed exports and deemed re-exports to “permanent regular employees” of companies headquartered in or with an ultimate parent headquartered in AIA countries.

Open-Weight AI Models That Are Published Are Not Controlled

Because “published” software and technology are not subject to the EAR, ECCN 4E091 does not control parameters of or derived from open-weight models, except where a downstream model was derived from an open-weight model using additional training ‘operations’ greater than whichever is higher of: 2.5 x 10^25 ‘operations,’ or over 25% of number of ‘operations’ used to train the most advanced open-weight model. ECCN 4E091 also does not control the parameters of any AI model trained using fewer operations needed to train the most advanced AI model that has been published, thereby allowing the scale to float upwards based on published open-weight models.

To see if your AI model falls under this exclusion, you can self-classify relying on BIS guidance, or technical opinions issued by the US AI Safety Institute and the US Department of Energy. If the idea of self-classifying your closed model against an open-weight published model to see if it falls under 4E091 gives you a headache (or nightmares), you can also submit a classification request to BIS.

AI Model Weights FDPR

Do not worry, foreign-produced AI model parameters, BIS did not forget about you!

The new AI Model Weights FDPR makes non-US technology described by 4E091 subject to the EAR if it is:

“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 specified in ECCN 3A001.z, 3A090, 4A003.z, 4A004.z, 4A005.z, 4A090, 5A002.z, 5A004.z, or 5A992.z”

The destination scope of the AI Model Weights FDPR is — surprise, surprise — everywhere in the world. The rule also clarifies that further training of foreign-produced 4E091 items does not change their status under the EAR through post-training techniques like fine-tuning or quantization.

New License Exceptions

With all the worldwide export licensing requirements popping up left and right, we should be thankful that BIS issued the following license exceptions, but others are pining for the days when the main license exceptions were GBS, RPL, TMP, and ENC (which seems practically easy in comparison!). Below is a summary of the eligible items and recipients, but we have not included all the conditions because, well, space constraints. Do not attempt these on your own: Always consult export controls counsel before relying one of these license exceptions!

License Exception AIA

• Eligible items: 3A090.a, 4A090.a, corresponding .z paragraphs, and 4E091.
• Eligible recipients:
  • For 3A090.a, 4A090.a, and corresponding .z paragraphs (as well as corresponding software and technology for .z paragraphs) — entities located in and headquartered in or whose ultimate parent company is headquartered in an AIA country.
  • For 4E091 — entities outside of Macau or Country Group D:5, if headquartered in or has an ultimate parent company in an AIA country.

License Exception ACM

• Eligible items:3A090.a, 4A090.a, and corresponding .z paragraphs (as well as corresponding software and technology for .z paragraphs).
• Eligible recipients: ‘private sector end user’ located outside Macau and Country Group D:5, provided that its headquarters and ultimate parent company are also outside of those locations. End use is limited to development, production, or storage of the eligible items.

License Exception LPP — exports and re-exports only!

• Eligible items:3A090.a, 4A090.a, and corresponding .z paragraphs.
• Eligible recipients: entities located outside Macau and Country Group D:5, whose headquarters and ultimate parent company are also outside Macau and Country Group D:5. Each ultimate consignee may receive up to 26,900,000 TPP in a calendar year from all exporters and re-exporters combined.

Other Changes

Revisions to NAC Procedures

BIS is amending the notification procedures for License Exception Notified Advanced Computing (NAC), which applies to certain lower-end chips (not 3A090.a chips) going to Macau or Country Group D:5 if the exporter gives BIS notification 25 days beforehand and obtains the agency’s approval.

The AI Diffusion Rule clarifies notification procedures by establishing criteria for interagency review of potential security risks when evaluating NAC requests. NAC notifications will also now require new information, such as, NAC and license approvals to the end-user over the past 12 months, memory bandwidth of items requested, and information about whether the items eventually will be aggregated into a datacenter or computing cluster.

UVEUs and NVEUs

The AI Diffusion Rule splits DC VEUs into the two new categories of UVEUs and NVEUs. DC VEU authorization was introduced last September, making it possible to export certain ICs under ECCNs 3A090, 4A090, and .z paragraphs for the purpose of building data centers.

UVEU authorization is available to applicants headquartered in or whose ultimate parents are headquartered in AIA countries, and would potentially allow approved companies to more easily export ICs anywhere besides Macau and Country Group D:5 for the purpose of building data centers.

NVEU authorization, which would apply only for a specific destination country, is available for applicants headquartered in or whose ultimate parents are headquartered in locations outside of Macau and Country Group D:5. NVEUs will be subject to a per-company, per-country installed base allocation of TPP, as measured by the collective computing power of items subject to ECCNs 3A090.a, 4A090.a, or corresponding .z paragraphs, as follows:

Red Flag Guidance

BIS is updating its red flag guidance at EAR Part 732, Supplement No. 3 to help exporters, re-exporters, and transferors of AI model weights, as well as cloud service providers whose customers are training AI models, identify possible diversion. Specifically, companies should be on alert if:

“You will be providing Infrastructure-as-a-Service (IaaS) products or services, or other computing products or services, to assist in training an AI model with model weights captured by ECCN 4E091 for an entity headquartered, or whose ultimate parent is headquartered, in any destination other than [the AIA countries]. Such assistance creates a substantial risk that such AI model weights, due to their digital nature, will be exported or reexported to a destination for which a license is required and, if a license is not obtained, that the IaaS provider will have aided and abetted in a violation of the EAR. In such cases, the IaaS provider should inquire if the customer intends to export the model and if so, apply for a license as required or inform the customer of their obligation to do so prior to export.”

Put another way, IaaS providers who provide AI model training assistance to companies headquartered or with parent companies headquartered outside of the AIA countries — Australia, Belgium Canada, Denmark, Finland, Germany, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, South Korea, Spain Sweden, Taiwan, the United Kingdom, and the United States — are expected to ask their customers if they intend to export the AI models (or their parameters) and if so, get a license or make sure their customer does so. The AI Diffusion Rule lays out some especially scary scenarios for cloud service companies, warning that IaaS providers could run afoul of the EAR if a US subsidiary of a foreign corporation headquartered in a high-risk country uses the provider’s services to train AI models and divert 4E091 model weights to the parent company without a license.

Depending on the incoming Administration’s position on this red flag (and the rule in general), this could be a vast expansion of the EAR to cover AI training services offered in the United States to US subsidiaries of foreign parent companies.

Conclusion

The Biden Administration is refusing to go quietly, leaving behind a pile of compliance headaches for companies operating in the AI (and automotive, health care, and other industries) spaces, and potentially for the third parties who provide cloud services to these companies. Even for those that qualify for one of the new license exceptions or whose license applications will be favorably reviewed, just figuring out where you fit within the AI Diffusion Rule’s framework could well prove challenging and costly. And with an incoming Trump Administration absolutely packed with China hawks, we do not expect BIS to reverse course on AI-related export controls anytime soon.