The Trump administration recently announced its intent to scrap the so-called AI diffusion rule, a complex regulation the Biden administration published in January to govern the global export of AI technology and maintain US AI leadership. The rule was designed to limit adversary access to the inputs needed to build and train advanced AI systems, as well as to keep most AI data centers in the United States and allied countries. In a statement denouncing the rule, the US Department of Commerce’s Bureau of Industry and Security called it “overly complex, overly bureaucratic,” and likely to “stymie American innovation,” promising to replace it “with a much simpler rule that unleashes American innovation and ensures American AI dominance.” The Department plans to issue a Federal Register notice formalizing the rescission shortly.
It is not yet clear what such a replacement might look like, but the policy concerns that motivated the rule remain as salient as ever: In the words of Jeremy Kessler, the Under Secretary of Commerce for Industry and Security, the Trump administration intends to “pursue a bold, inclusive strategy to [share] American AI technology with trusted foreign countries around the world, while keeping the technology out of the hands of our adversaries.” The administration may use new AI export restrictions to address issues such as chip smuggling and give the United States leverage in bilateral trade and other negotiations. Whatever their ultimate form, such restrictions will have significant implications for model developers, hyperscalers, chip manufacturers, and investors across the industry.
Background
The diffusion rule—formally titled the “Framework for Artificial Intelligence Diffusion”—established a new global licensing regime for the export of advanced AI chips and related technologies under the Export Control Reform Act of 2018. Restrictions varied geographically based on perceived risk levels. “Tier 1” countries, a group of 18 US allies including Australia, Canada, New Zealand, and the United Kingdom, faced virtually no restriction on their ability to obtain advanced chips. “Tier 2” countries, a category comprising most of the rest of the world, were subject to chip caps that could be increased if providers met robust security standards and agreed to stringent reporting rules or if a country reached a bilateral government-to-government understanding with the United States. Exports to “Tier 3” countries, arms-embargoed nations including China and Russia, remained subject to the strictest controls: prohibitions on any export of advanced chips.
The rule also imposed first-of-its-kind restrictions on the export of model weights, the numerical parameters that encode an AI model’s intelligence. Weights of models exceeding a certain compute threshold—a proxy for a model’s capabilities also adopted in the Biden administration’s now-repealed October 2023 executive order on artificial intelligence—could not be exported to Tier 3 countries; in the rest of the world, they could be stored only by companies headquartered in Tier 1 in data centers that met strict standards for cybersecurity, physical security, personnel security, and supply chain independence. Publicly available models, known as “open-weight models,” were exempted from the regulation.
Published amid intensifying US-China military and technology competition, the rule was partly an attempt to clamp down on Chinese efforts to evade US export controls. It was also a response to intensifying interest in AI exports from countries, such as Saudi Arabia and the United Arab Emirates, with their own foreign policy and technology interests and ambitions. The rule also sought to extend US and allied dominance in AI infrastructure, including by requiring Tier 1 companies authorized to build above the caps in Tier 2 to keep 75 percent of their AI computing power in Tier 1 countries.
What’s next?
Since January, the rule has been the subject of increasingly pitched debate within the executive branch, Congress, and industry. Some argue that stringent controls are necessary to maintain America’s compute advantage and prevent adversaries from developing and deploying advanced AI models with critical national security applications, while others maintain that the regulation stifles global economic growth and undermines US technology leadership abroad. In April, seven Republican US senators wrote to US Commerce Secretary Howard Lutnick urging him to withdraw the rule, arguing that it placed “burdensome constraints on US companies” and would incentivize countries to turn to “unregulated, cheap” Chinese substitutes or create their own AI technology stack.
In scrapping the rule, the Trump administration lent support to its critics. But the policy concerns that animated the rule have not gone away: China is reportedly still smuggling chips and seeking to develop advanced AI models, and the United States still needs to figure out the appropriate balance between spreading a transformative technology worldwide and controlling innovations that may have critical national security implications, including for intelligence collection and analysis, weapons development, and cyber capabilities. At the same time, given its stated desire to strengthen US AI leadership, the administration will need to decide how best to ensure that the most advanced AI data centers are built in the United States.
The Trump administration has therefore promised a “replacement rule,” though it is not yet clear what its approach to AI diffusion will be, what form such an approach might take, or when it might arrive. Among other possibilities, the new administration could implement a stricter global licensing regime without any tiers; remove license exceptions for low volumes of chips; eschew a global system in favor of new targeted controls on Southeast Asia and/or other hotspots for chip smuggling; and/or focus on a series of direct bilateral government-to-government negotiations over access to US computing power.
In the meantime, companies should prepare for new controls of one kind or another and develop a strategic approach to manage risks that may arise when locating these technologies overseas. They should not assume that a repeal means a comprehensive loosening of US AI chip exports. In addition, they should take care to follow newly issued Department of Commerce guidance—largely a restatement of existing policy—for the export of AI chips abroad. These include (1) a warning that the use, transfer, purchase, or transport of specific advanced Chinese chips, among other actions, could result in substantial criminal and administrative penalties; (2) a policy statement that a license may be required for AI training chips when an exporter has knowledge that the AI model has the potential to enable end uses related to military intelligence and weapons of mass destruction in arms-embargoed countries, including China and Macau; and (3) red flag guidance and due diligence actions companies should take to prevent the diversion of advanced chips to and within China.
[View source.]
