US AI Diffusion Framework Needs Tweaks, Not Sledgehammer, Researchers Say

The Biden-era AI diffusion rule was a logical way to balance the need for protection against China while setting some guardrails around nations that could be used to smuggle chips to China, said Alasdair Phillips-Robins and Sam Winter-Levy, fellows with the Carnegie Endowment for International Peace. But the Biden administration also made “several mistakes” with the rule, including by not including enough countries in Tier 1 -- the group of nations subject to the fewest chip-related export license requirements. It also didn’t give countries in Tier 2 -- the mid-tier group of countries subject to caps on AI chips and other restrictions -- a clear pathway to move into Tier 1, the researchers said.

They outlined four options the Trump administration could take to regulate exports of advanced AI chips, including one that would tweak the current AI diffusion rule to a slightly different version, which they said would be “policymakers’ best bet.” The U.S. could do this by expanding the group of “favored countries” that need no license to buy advanced chips and AI model weights, they said, specifically pointing to Israel, Poland, Portugal, Switzerland, Iceland, the entire EU and much of Eastern Europe.

Those nations “present low smuggling and governance risks” and are “important markets” for U.S. cloud service providers, the researchers said. “There are also high diplomatic costs to the perception that Washington is shutting its partners out of their AI ambitions.”

Although it’s not yet clear how the administration will change the rule, David Sacks, the president's AI policy adviser, acknowledged that the rule “strained relationships” with close allies and risked pushing them toward “non-American technology alternatives” (see 2505090011).

The Carnegie researchers also said the U.S. could give allies a more explicit path to benefit from looser chip restrictions. The administration wouldn’t need to spell out “exactly what reforms it would take to move up a level,” they said, adding that withholding that information could give the U.S. leverage in potential trade negotiations. But it could at least let foreign governments know that the U.S. is open to hearing an “attractive offer” in exchange for “significantly improved access to chips.”

For the middle group of countries that would still be subject to some restrictions but not an embargo, the administration could tighten the country caps outlined in the AI diffusion rule. The rule said those nations could buy the equivalent of about 1,700 Nvidia H100 chips per company, per year, without a license, but the Carnegie researchers said that threshold “was set higher than these applications typically need.” They also said businesses would be able to evade the cap by setting up shell companies.

The administration could instead set the threshold at around 300 H100-equivalents, “and then double the threshold each year for the next three years to account for chip performance improvements,” the researchers said.

They also said exceptions could be made for certain countries, such as India. Although the rule allowed up to 50,000 H100-equivalents to each country in the middle group from 2025 to 2027, the researchers said India was likely to hit that cap quickly because of “its size and strategic importance.” The country “should receive a higher allocation in a future rule, with further increases on the table in a government-to-government agreement.”

The researchers also said the Trump administration could take another path if it chooses to double down on restrictions by “drastically limiting” exports of advanced AI chips to all countries. But they said that strategy would “be more likely to hurt the U.S. AI industry than to help it,” including by shutting American AI developers from foreign markets.

A third option would be to repeal the AI diffusion rule without a replacement, and instead seek to promote advanced chip exports, except to China and other adversaries or countries that pose a high diversion risk, such as Saudi Arabia and the United Arab Emirates. Carnegie said it appears the administration “may do this in the short term, and that would be far from the worst outcome.”

Although this approach would “make the administration’s job easier, and industry happier,” they said, “the hard questions the rule sought to address won’t go away.” Smuggling to China will likely continue, the U.S. would lose leverage over “geopolitical swing states,” and it could “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.”

The researchers floated a final option that would see the U.S. negotiate AI chip deals on a country-by-country basis, tying chip access to lower tariffs, divestments from China or fewer regulations on American technology companies. This would give the U.S. more leverage as it seeks to negotiate trade deals and would align with Trump’s “broader transactionalism,” they said. But they said the U.S. could also struggle to negotiate “dozens of bespoke compute-for-concessions deals, alongside a vast number of broader trade deals.”

“At the best of times, the Departments of State and Commerce have limited bandwidth,” the researchers said. “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.”

The researchers said the administration “may naturally wish to adopt a radically simpler approach,” which could improve the rule’s implementation and enforcement. “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,” they said.

But they also said any rule to restrict advanced AI chips may need to be complex, because the “nature of the problems the United States confronts” are also complex.

“As the Trump administration grapples with what comes next, it too should seek to protect national security while minimizing the economic cost,” they said. “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.”