.jpg)
One of the final actions of president Joe Biden’s administration was to publish the “AI diffusion rule,” a complex regulation governing the export of advanced computing chips to most countries in the world. President Donald Trump’s administration has signaled that it intends to repeal the rule, which is set to come into force on May 15, and replace it with a simpler framework to regulate the export of AI technology, although the administration has reportedly not reached a final decision.
The existing rule seeks to regulate the sale of the world’s most powerful AI chips and the storage of advanced AI model weights (the parameters that encode a system’s core intelligence) in every country in the world. It does so by splitting the world into three groups: a small set of close U.S. allies to which almost no restrictions apply; a group of arms-embargoed countries, including China and Russia, that were already banned from receiving U.S. chips; and a large middle category where most shipments can proceed but those necessary to build very large computing clusters are subjected to additional scrutiny or, at the largest levels, banned entirely.
The rule’s designers were attempting to balance the need to preserve U.S. control over powerful AI systems against the importance of promoting exports of U.S. AI products and services abroad. Their solution was to try to ensure that the largest computing clusters and the most powerful AI systems remained in the United States and a small group of close allies, while allowing the vast majority of commercial activity to carry on largely unimpeded. At the same time, they created a framework for using AI exports as leverage over geopolitical swing states, establishing incentives for other governments to adopt U.S. technology standards and protections in exchange for U.S. chips. In other words, the rule was a compromise between what one of this article’s authors has elsewhere described as three possible goals of international AI policy: control, promotion, and leverage.
Public reporting suggests that the administration is considering a range of options to replace the rule. Officials could, like the Biden administration, seek to balance control, promotion, and leverage by continuing to rely on the use of country tiers. Or they could lean more aggressively into one of those three goals. To maximize control, they could impose significantly tighter restrictions, including on U.S. allies, in an effort to lock in a U.S. monopoly on advanced AI technology. To maximize promotion, they could dramatically loosen controls, likely by repealing without a replacement. Finally, in an effort to strengthen their bargaining position, they could double down on bespoke government-to-government deals as part of a strategy of bilateral horse-trading that uses chip exports as leverage in broader negotiations.
It’s not yet clear which of these routes the Trump administration will take. There are multiple competing factions within the administration, and policymakers so far have sent few signals of their intentions beyond a desire to rewrite the existing rule. This article thus considers a broad range of options to help frame the debate and highlight the trade-offs involved. Ultimately, it argues that however the new administration designs its replacement for the diffusion rule, policymakers’ best bet is to pursue some version of a country tier framework, although with a series of changes to Biden’s proposal, including giving more favorable treatment to major U.S. partners like India, Israel, and Poland; doing more to crack down on smuggling to China; and using chips as leverage to cut smart deals with geopolitical swing states.
Each of the other options would come with serious risks. If the administration leans too far toward control, it will cut U.S. tech companies off from major foreign markets and likely slow the pace of AI development, worsening the U.S. trade balance and jeopardizing America’s AI lead. Meanwhile, if policymakers largely remove controls on exports to countries other than China, they will enable continued large-scale chip smuggling and accelerate the offshoring of a core strategic technology to a set of countries, including the Gulf states, that do not share the United States’ interests. Finally, if the administration attempts to avoid an overarching global framework and instead win leverage in a series of bilateral deals, it will find itself bogged down in dozens of negotiations and thousands of licensing decisions that it lacks the bureaucratic capacity to handle. The result will be either lopsided deals that give away valuable technology in exchange for illusory trade concessions or a logjam in which, by accident rather than design, the U.S. government turns off exports of one of its most important technologies.
Option One: Return to a Country Tier Framework
The administration appears to be planning to repeal the diffusion rule in full, while keeping the existing controls on arms-embargoed countries that were first imposed in 2022 and on two additional groups of countries in Central Asia and the Middle East that were first imposed in 2023. It also seems poised to impose additional controls on some countries in Southeast Asia, including Malaysia and Thailand, that have emerged as smuggling hotspots.
This approach is a reasonable short-term response to the pressures the administration faces, especially from U.S. companies eager to open up foreign markets. But sooner or later, officials may find that they have moved, rather than solved, the problems the diffusion rule was designed to address. The Biden administration imposed controls on Central Asia and the Middle East in 2023 in part because of concerns over diversion to China. Smuggling has since cropped up in Southeast Asia. If Washington imposes controls on a new set of countries, smuggling is likely to shift once again, perhaps to India, elsewhere in Asia, or Latin America, forcing yet more rounds of controls.
If officials eventually tire of playing whack-a-mole, they may return to something like the logic of the diffusion rule: strict controls on arms-embargoed countries, looser controls on a broad middle group of states that pose diversion and misuse risks, and few or no controls on an inner ring of U.S. allies.
In that scenario, the Trump administration would have the chance to create a simpler, stronger version of the framework. Among other things, that would involve correcting Biden-era policy errors and bringing things more into line with the Trump administration’s priorities and worldview. To understand this option, it helps to recall the underlying policy problem the Biden administration was grappling with and the logic behind the rule it created.
The core idea behind the rule was simple—the need to balance control of a technology that could have critical national security implications with the promotion of U.S. exports—but the rules the administration created were complex. Premised on the idea that scale—in computing power, data, and consumer use—was likely to unlock major breakthroughs in AI technology in the next decade, the diffusion rule set caps on the number of advanced chips that could be exported to most countries. Data center providers could exceed these caps only if they met stringent security standards, followed extensive reporting rules, and, in the case of U.S. companies, committed to keeping the majority of their advanced chips in the United States. Even then, they were still blocked from building the very largest data centers—a level that the Biden administration set to rise each year—in much of the world. The weights of models trained on those very large clusters were also controlled: Companies were not allowed to store them in arms-embargoed countries, and they could be exported only to entities headquartered in the favored group of close allies under strict cybersecurity standards.
The basic structure made sense as a way to balance protection with promotion and economic partnerships with swing states. But the Biden administration made several mistakes in its rule, which the Trump administration could fix in any replacement. A new rule would also let the administration pursue its own geopolitical priorities, such as closer relations with countries like India and Israel. Along with other reforms, these changes would help expand U.S. market opportunities, create a powerful incentive for countries to stay in the U.S. orbit, reward those who act in U.S. interests, and crack down on smuggling to China. In any replacement rule that adopts a tiered structure, the administration should consider:
- A larger group of favored countries. The diffusion rule’s list of close U.S. allies, where no license was required to export advanced chips and model weights, did not include major U.S. partners such as Israel, Poland, Portugal, Switzerland, and much of Eastern Europe. A new such group could include these countries, plus Iceland and the entire EU. These countries currently present low smuggling and governance risks, and they represent important markets for U.S. cloud service providers. There are also high diplomatic costs to the perception that Washington is shutting its partners out of their AI ambitions.
- A clearer pathway to favored status. Although the diffusion rule contemplated government-to-government deals to raise chip caps, it did not mention the possibility of countries moving from the middle group to the favored ring of close U.S. allies and partners. However the new administration chooses to shake up the country groups, it could say more explicitly that moving from one group to another is on the table in any potential government deal. It doesn’t have to specify exactly what reforms it would take to move up a level—keeping bargaining leverage is a good idea—but it could let foreign governments know that if they can develop an attractive offer, significantly improved access to chips is available. In addition to improving U.S. leverage, this would also reduce the perception that any policy is designed to permanently exclude the Global South.
- A lower threshold for license-free chip sales to middle-group countries. Countries in the diffusion rule’s middle group could receive up to the equivalent of about 1,700 Nvidia H100s per company, per year, without a license. This exception was intended to enable research and small-scale commercial use, but the threshold was set higher than these applications typically need, and by setting up multiple shell organizations, a company could accumulate large numbers of chips without going through the approval process. For countries in any future middle group, the administration could reduce this number to around 300 H100-equivalents, and then double the threshold each year for the next three years to account for chip performance improvements.
- Higher export limits for major U.S. partners like India. The rule allowed up to 50,000 H100-equivalents to each country in the middle group from 2025 to 2027, a cap that could be doubled if the country signed a government-to-government agreement with the United States. For most countries, this is enough to encompass much commercial use: Most data centers contain far fewer than 50,000 such chips. But India, due to its size and strategic importance, was likely to hit the cap quickly and should receive a higher allocation in a future rule, with further increases on the table in a government-to-government agreement.
- Allowing trusted companies to put more of their chips in each middle-group country. The rule prevented major cloud providers that gained trusted status (known as Universal Validated End Users) from installing more than 7 percent of their AI computing power in any single middle-group country (or more than 25 percent of their AI chips in the middle group overall). The 7 percent limit would have done little to protect national security given the other security requirements and the 25 percent cap. Any future framework should require that major U.S. cloud providers build most of their AI computing power in the United States, and more in close allies, but should not impose country-level requirements like the 7 percent rule, instead allowing businesses to decide independently what geographic distribution best suits their operations, within overall limits.
Option Two: American Monopoly
The second option is for the administration to double down on a strategy of control—an attempt to create a U.S. monopoly of AI by drastically limiting the export of advanced AI chips to all other countries. To do so, the administration would apply licensing requirements, likely above a low chip threshold, to the entire world.
The logic behind this approach is straightforward. Influential voices in the national security community and the Republican Party believe that the United States is locked in an “AI arms race with China.” Keeping as many AI chips in the United States as possible, and preventing companies from storing frontier model weights in risky locations, might give Washington control over the development and use of powerful AI and reduce the risk that another country could divert U.S. technology to China, leave it vulnerable to cyber espionage, or even seize physical control of data centers owned by U.S. companies abroad. Last year, the U.S.-China Economic and Security Review Commission recommended that Congress launch a “Manhattan Project-like program” for AI. Any such program would, like the original Manhattan Project, require amassing (or even monopolizing) the raw materials—then, uranium; today, GPUs—in the United States.
Although advocates of such a strategy may believe it would help the United States push the frontier faster than other countries, placing restrictions on close allies would be more likely to hurt the U.S. AI industry than to help it. For one thing, the United States will likely struggle to bring sufficient power online fast enough to support the AI industry’s plans. Although many worries about power and permitting are overblown—and the Trump administration is already working hard to solve the issue—without the large-scale power available in U.S. allies like Australia, Canada, and Norway, U.S. companies will have to slow their data center build-outs, holding the United States back in the very race Trump is determined to win.
Draconian restrictions on chip sales will also hurt U.S. AI developers by shutting them out of most foreign markets. Cloud service providers like Google, Microsoft, and AWS have built out global data center networks, driven in part by data localization laws and the need to reduce latency; if they are unable to build corresponding AI data centers, they will be at a disadvantage in almost every global market. Although U.S. firms currently dominate the cloud service market, other countries—especially China—represent a latent but real source of competition that could surge in the coming years if American companies face excessive restrictions on foreign build-outs. The diffusion rule attempted to thread this needle by allowing development (except of the very largest projects) to continue around the world through the Validated End User program, while keeping the most significant training data centers in the United States and close allies. Clamping down further would risk upsetting that balance.
Undermining relationships with close allies will bring other costs, too. The original diffusion rule prompted complaints from several U.S. partners—including natural allies of the Trump administration such as India and Israel—that they were excluded from the most favored group; shutting out the whole world would cause far greater diplomatic upset. U.S. allies might restrict joint research until they regained access to chips or retaliate on other issues. As the administration has discovered in the tariff context, a policy that applies high trade barriers to the entire world is not politically or economically sustainable.
Critical partners on U.S. export controls to China, including Japan and the Netherlands, might refuse to cooperate on additional bans on the sale of chip-making equipment or limits on servicing machines already in China. These countries have many options to slow-walk cooperation, and the administration has few tools to force them to comply. In a worst-case scenario, they might even roll back existing restrictions, a move that would allow Chinese chip makers to rapidly catch up to their foreign counterparts. Other countries would have an especially strong incentive to boost Chinese chip-making if doing so became their only path to reliable sources of advanced AI chips. In that scenario, U.S. officials might fall back on extraterritorial controls on foreign chips and equipment produced with U.S. technology through a legal mechanism known as the foreign direct product rule, but a broad application of this rule would be diplomatically and strategically risky.
Option Three: Repeal Without Replacement
Rather than attempting to keep all advanced AI development in the United States, the administration might instead want to simply abandon anything that resembles the diffusion policy, hoping to make life easier for industry and for itself. It could pursue a strategy of promotion, seeking to “flood the zone,” as the software giant Oracle put it, by allowing U.S. companies to export their products with few restrictions. To do so, the administration could roll things back to the state of the world before January 2025, when restrictions applied to arms-embargoed countries (a group that includes China, Iran, and Russia) and another set of countries that presented a high risk of technology diversion, including Saudi Arabia, the United Arab Emirates, and other countries in the Persian Gulf and Central Asia. Reporting suggests they may do this in the short term, and that would be far from the worst outcome. Alternatively, the administration could abolish all restrictions on countries outside the arms-embargoed group, greenlighting the export of U.S. chips to much of the world.
Pursuing that approach would make the administration’s job easier, and industry happier, in the short term, but the hard questions the rule sought to address won’t go away. Smuggling to China through third countries, especially those in Southeast Asia, will continue. (Even if the administration cracks down on current hotspots, smugglers will likely shift their operations rather than quit altogether.) If the administration keeps restrictions on the Gulf and Central Asia, it will need to come up with a policy for license applications to those countries. And if it scraps all restrictions outside arms-embargoed countries—or cuts overly generous deals with a few partners in the Gulf—Washington will lose its leverage over geopolitical swing states, and it will watch as yet another industry gets offshored to countries that offer generous subsidies, flexible regulations, and lower wages—but that do not share U.S. interests or values. And in this case, the technology that the United States would offshore is one that could reshape the global balance of power by driving rapid economic growth, transforming intelligence and surveillance capabilities, enabling strategically significant weapons development, and creating new cyber threats.
Option Four: Bilateral Horse-Trading
Finally, the administration might be tempted to lean into a “strategy of leverage.” Beyond the technology-focused agreements envisioned by the diffusion framework, it could seek to use AI exports to negotiate country-by-country deals that could advance a wide range of U.S. political and economic objectives. For example, the administration could condition AI exports on trade concessions, such as tariff reductions. It could tie chip sales to increases in military spending, divestment from China, or looser regulatory constraints on U.S. tech companies. The list of potential demands is long; the common theme is linking access to U.S. computing power to the administration’s broader set of foreign policy priorities.
The diffusion rule already contemplated unlocking greater chip exports through government-to-government agreements, but the administration may believe it can give itself greater leverage by imposing export limits on all countries as a default, and promising relaxed requirements in exchange for concessions. A blanket license requirement is the same legal move the administration might make if it wanted to pursue an American monopoly over AI. But whereas that approach would involve presumptively denying all licenses, using chips for leverage would explicitly promise that other countries could get out of these requirements through individual deals. (Of course, as in the tariff context, the administration might try to signal both strategies at the same time.)
Using chips as bargaining tools would give the administration additional leverage in international negotiations, and it would align with both Trump’s broader transactionalism and his trade strategy of making deals with individual countries. The United States has legitimate interests on a broad range of issues and has good reasons to pursue leverage over other states if they want cooperation. In the case of the Gulf states, for example, Washington should use access to computing power to drive a hard bargain: Among other things, it should push the Gulf to cut off investments in China’s AI and semiconductor sectors in exchange for U.S. chips. When it comes to India, Washington could push for a range of concessions, including lower trade barriers, a shift from Russian to U.S. oil and gas, and crackdowns on chip smuggling and arms control violations.
But the administration should be wary of overplaying its hand. If it locks down too hard, it will struggle to negotiate dozens of bespoke compute-for-concessions deals, alongside a vast number of broader trade deals. As in the tariff context, if Washington takes a hard line with every country simultaneously, it will realize that deals take time, and will likely have to walk back its initial restrictions in the interim. It took the Biden administration months to help broker a deal between Microsoft and the Emirati tech company G42, and officials failed to reach a broader government-to-government agreement with the UAE—delays that frustrated Microsoft and the Emiratis as they waited for the relevant licenses. At the best of times, the Departments of State and Commerce have limited bandwidth. Expecting them to accurately assess each side’s leverage and then design and negotiate numerous technology deals that respond to developments in a fast-moving field seems overly optimistic.
If the administration miscalculates, the costs could be significant. If Washington underplays its hand, then the United States might give away a major strategic technology in exchange for what will ultimately seem like minor concessions: Trading increased foreign purchases of U.S. soybeans for access to cutting-edge AI chips might help a politically important domestic constituency but is unlikely to be the best use of U.S. leverage. Conversely, if Washington overestimates the value of its AI offering to countries that may not share Silicon Valley’s views about the strategic importance of this technology, it may fail to reach deals, alienate potential partners, and ultimately lose market share to China.
In that scenario, the administration may want to approve some exports even without deals in place, but unless it walks back any license requirements, it will find doing so difficult. Four agencies—Commerce, Defense, Energy, and State—all have the authority to review license applications and can vote to deny or impose additional conditions. A large or complex license request can involve many rounds of back-and-forth with the companies involved as agencies request more information, ask about potential conditions, and debate the merits among themselves. If agencies can’t agree at the staff level, they can escalate to progressively more senior officials until, in theory at least, a dispute reaches the president’s desk. Without overarching policies to guide officials in making these decisions, following this procedure for thousands of export requests to every country in the world is a recipe for exporting nothing at all, with all the resulting downsides from shutting U.S. companies off from global markets. In the end, a strategy of pure transactionalism will have broader diplomatic costs. Maximal use of U.S. leverage on every economic and political front simultaneously is unlikely to provide a stable foundation for long-term U.S. technological primacy.
Conclusion
Confronted with a lengthy, complex rule, the Trump administration may naturally wish to adopt a radically simpler approach. Some simplifications would represent real improvements. There can be a trade-off between technocratic nuance and ease of implementation, and keeping the rule as it’s currently written would generate plenty of work for lawyers and compliance experts.
But at bottom, the rule’s complexity was a response to the complex nature of the problems the United States confronts. The rule attempted to maintain the U.S. lead in AI, bring partner countries closer to the U.S. technology ecosystem, safeguard advanced U.S. technology from China, promote the U.S. AI industry around the world, and limit the proliferation of potentially dangerous future AI capabilities. As the Trump administration grapples with what comes next, it too should seek to protect national security while minimizing the economic cost. If it picks an apparently more straightforward approach, it is likely to find that each alternative has its own downsides, and the same difficult problems rear their heads time and again.
