Interim Final Rule on AI Diffusion
Introduction
On Monday morning, U.S. tech stocks tumbled after an announcement from China: R1, an open-source AI model developed by DeepSeek, a relatively unknown AI startup, had overtaken U.S. rival ChatGPT in Apple's app store. According to DeepSeek, the company spent less than $6 million on computing power for its base model–by comparison, spending hundreds of millions of dollars less than U.S.-based companies such as OpenAI or Google. For years now, the United States has implemented measures to restrict the supply of advanced AI chips to other countries, particularly China. (It's not clear whether DeepSeek's claims about its costs are accurate or whether R1 was developed without U.S.-made technology.) Newly announced regulations also seek to restrict the export of model weights. But have they come too late?
On January 13, 2025, the Bureau of Industry and Security (“BIS”) under the Biden-Harris Administration released an Interim Final Rule on Artificial Intelligence Diffusion. The rule:
i) revises controls on advanced AI chips (a specialized type of integrated circuit optimized for handling AI algorithms’ high computational demands) and creates a new licensing framework, and
ii) implements a licensing regime for exporting “model weights” (the parameters that encode an AI system’s core). While model architectures are often public information, the weights assigned during the model training process to certain operations’ results are valuable intellectual property.
Although the measures apply primarily to advanced models, the Interim Final Rule also implicates general-purpose AI (“GPAI”), depending on the deployment and use case. The restrictions set forth in the rule are effective immediately but subject to a 120-day public comment period and may be modified. Impacted parties will have to meet most compliance requirements on May 15, 2025 (with certain requirements effective in January 2026). We discuss the new requirements below.
Context
As GPAI software is not an effective target for controls, governments are increasingly looking to the computer hardware necessary to develop more powerful AI systems. Chips are critical for AI development, because they accelerate the complex calculations necessary for machine learning. Moreover, unlike general purpose integrated circuits (“ICs”), AI chips are specially designed to handle the unique workloads required for AI computation tasks, such as training and inference. For example, they are power-efficient and capable of parallel processing. Historically, they have also been costly. According to AI Index estimates, OpenAI’s GPT-4 used an estimated $78 million worth of compute to train, while Google’s Gemini Ultra cost $191 million. (Technological advancement, however, is making advanced integrated circuits more accessible. For example, NVIDIA’s project DIGITS offers an AI supercomputer capable of running and training a large language model (“LLM”) at an astonishing cost of $3,000.) Therefore, the central idea behind the Interim Final Rule is to maintain AI chips within the United States and its allied countries, while also creating new regulations on exporting model weights.
New Global Licensing Requirement for Advanced Computing Integrated Circuits
While the BIS’s previous regulations already controlled exports of advanced chips—as well as Semiconductor Manufacturing Equipment (SME) and parts and components—even in smaller quantities, those controls were more targeted, applying to certain destinations, entities, or end-users/uses. The BIS’s Interim Final Rule, however, is meant to address the sales of large quantities of advanced computing integrated circuits sufficient to train advanced AI models. To accomplish this goal, the Interim Final Rule creates a global license requirement, with several exceptions for allied countries and low diversion risk use cases.
For large clusters of integrated circuits sufficient to train AI models, the rules establish a three-tiered system for data centers:
Entities in certain allied country destinations may gain validated status to receive exports;
End-users in other countries—except for Macau and destinations in Country Group D:5, such as China, Iran, and Russia—may obtain authorizations to be treated as entities in allied countries; and
Other end-users in such other countries—except for Macau and destinations in Country Group D:5, such as China, Iran, and Russia—would be subject to uniform default country allocations of advanced ICs.
Finally, there is an exception to the licensing requirement, for transactions involving certain types of end users in certain low-risk destinations. Companies will be able to export advanced integrated circuits to the United States’ allies, with virtually no restrictions. Most other countries will face importation limits (or caps) on total computing power, unless able to verify a trusted and secure environment prior to import. Companies that meet high security and trust standards and are headquartered in close allies and partners will be able to obtain highly trusted “Universal Verified End User” (universal VEU) status. With this status, they may place up to 7% of their global AI computational capacity in countries around the world.
New Export Regulations for Frontier Model Weights
In addition, the Interim Final Rule amends the Export Administration Regulations (EAR) to create a new classification EECN (4E091) to cover “frontier” AI model weights. Frontier AI models—such as OpenAI’s GPT-4 and Google’s Gemini 1.5—are advanced models, capable of performing more than a single type of task. They include both foundational AI models and GPAI models. The Interim Final Rule also implements additional due diligence measures for advanced computing integrated circuits, lays out new reporting requirements for “front-end fabricators” producing advancing computing ICs for authorized IC, new “Know Your Customer” (KYC) vetting form, and revisions to clarify the scope of ECCN 3A090 (to include possible shell companies), as well as other changes and amendments.
In addition to being costly to develop, BIS determined that the unrestricted export of such model weights presents a threat to national security. To mitigate the threat, the Interim Final Rule imposes a mandatory licensing schema under EAR to export, reexport, or perform in-country transfers of AI model weights from closed-source code that have been trained on more than 10^26 computational operations. According to BIS, no more than five AI models exceed this computing power at present. But, BIS intends to adjust the threshold for computing power to meet the licensing requirement as needed—in response to advances in computing and AI software and hardware. This might mean updates to this requirement within the calendar year.
The BIS’s Interim Final Rule also established a new Foreign Direct Product. Any export of the qualifying frontier AI model weights are now subject to EAR jurisdiction, even if the model weights were produced outside of the United States, if the model weights were created using ICs, servers, or other technology with U.S. origins. (U.S. allies may qualify for license exceptions under the new Artificial intelligence Authorization exception to the Interim Final Rule. Applications for licenses to export qualifying frontier AI model weights to any country other than those on the close allies list will be “presumptively denied.”)
Takeaways
The Interim Final Rule will require companies to make many changes. Significantly, companies must take measures to obtain approval before exporting information about model weights, including advanced close-weight AI models, to certain countries. These measures apply to foundational and some GPAI models (depending on use case and export location). The BIS did not impose restrictions for weights that are publicly available (“open weight” models), or for closed models that are less powerful than the most advanced open-weight models.
We are already seeing AI-powered service providers update their Terms of Service to account for these new rules, including requiring that users comply with “all domestic and international export laws”—a new item in AI software agreements. But, this is arguably not enough to absolve themselves of liability, given that various requirements of the EAR depend on a person’s “knowledge of the end-use, end-user, ultimate destination, or other facts relating to a transaction or activity.”
For example, Infrastructure as a Service (“IaaS”) cloud computing providers that train AI models will have to ask more questions of their clients, when working with companies located abroad. Indeed, BIS-introduced a new “red flag” (28), as part of its “Know Your Customer” guidance, that states IaaS providers should exercise heightened due diligence when their client is headquartered in a country other than one on the list of Artificial Intelligence Authorization Countries. The IaaS provider may be liable for aiding and abetting a violation of EAR, if it provides cloud computing services to train advanced AI models and transfers the resulting model weights (because of their digital nature) to such clients without a license.
