Overview
On Monday, the Trump administration revoked the Biden-era Framework for Artificial Intelligence (AI) Diffusion, a week before its expected enforcement on May 15, stating that it will replace it with “much simpler” guidance than the complex tier-based system. While the reprieve carries upside risks for the tech industry in the short term, AI’s potential to disrupt economic and military power means that some form of strict US export controls will persist in a bid to maintain dominance. The Trump administration has signaled an appetite for government-to-government AI deals, which carry upside risks for AI growth and innovation but downside risks in the fragmentation of the export control regime, making chip flow diversions to China easier.
The Scrapped AI Diffusion Rule, Explained
The Framework for Artificial Intelligence, issued at the close of the Biden administration on January 13, 2025, was created to streamline a case-by-case export control regime on advanced chips into a generalized tier-based system. AI chips provide the requisite computing power (or “compute”) that enables the development of AI applications, providing an effective chokepoint as a dimension of US industrial policy. The Framework succeeds a patchwork of AI chip export controls that began in October 2022 and was successively expanded, including applying the foreign direct product rule to the export of semiconductor manufacturing equipment (SME) in December 2024—subjecting even semiconductor chips made using US-made SME to export controls.
The Framework divided the world into three groups with different export licensing requirements. The first group, which mostly comprises allies in Western Europe and East Asia, was largely exempt from the rules. The third group corresponded to those with US-imposed arms embargos, such as China and Russia, who were the targets of previous export controls. The Framework also introduced new authorization rules on validated end-users (VEUs) from the Bureau of Industry and Security’s validated end-user program to control exports to specific data centers.
The middle group, comprising the majority of US trading partners, was the most complex. Per the rule, Tier Two countries could order up to 50,000 H-100-equivalent chips until 2027 without a national validated end-user (NVEU) licensing requirement, as well as up to 1,700 H100-equivalent chips without a license per middle group entity, which would not count toward the 50,000 country-wide cap. Companies from middle group countries could double their volume of chips by securing government-to-government agreements.
Entities exclusively from the first group could have applied for universal validated end-user (UVEUs) authorization, establishing the entity as a major trusted cloud provider that implements leading practices to prevent technological diffusion or theft. A UVEU headquartered in a country in Tier One must have kept at least 75% of their total AI computing power in a first group country, but they could not install more than 7% of their AI computing power in any single middle-group country and no more than 25% of their total AI compute power in middle category countries. US UVEUs in particular were required to maintain at least a 50% share of compute in the US. Meanwhile, companies headquartered in the middle group must have obtained NVEU authorization on a country-to-country basis and could only purchase up to 100,000 controlled H100 or equivalent chips by end of 2025, 270,000 chips in 2026, and 320,000 chips in 2027—a limit designed to keep middle category countries’ data infrastructure at least twelve months, or one generation, behind the capacity to support the most cutting-edge AI models. To illustrate, the 2027 chip limit would use a predicted maximum of only 0.2 GW of power. Meanwhile, Goldman Sachs predicts that AI data centers will create an extra 68 GW in additional global power demand by 2027—340 times the amount allotted to non-US-based NVEUs.
In addition to these chip regulations, the Framework localized cloud access to US-headquartered companies and imposed license requirements on frontier closed model weights—the parameters controlling an AI model’s learned capabilities, sort of like an instruction manual—on models trained on more than 1026 computational operations (about $70 million in compute), which no system currently reaches. Through these restrictions, the Diffusion Rule targeted the whole “AI access triad”—the physical hardware, remote access of compute power, and resultant products (i.e., model weights).
Most critics of the Framework argue that rules on “Tier Two” countries created an overly complex and restrictive export control regime that takes the wrong approach to geopolitical “swing states”—pushing them toward alternative partners like China and undermining US hyperscalers and chip producers as they planned for large investments to fulfill global demand. Moreover, the determination of who was in which tier aggravated several US allies in “Tier Two,” including Poland and Israel. Some corporate players argue that the rule, by isolating the AI market in the US, would have stymied the buildout of compute power to meet global demand.
The Enduring Geopolitics of Compute
Despite market jubilation over scrapping the Diffusion Rule, the underlying challenge it sought to address—China’s creative circumvention of the US export control regime—remains a pressing concern. The US has treated chips, the building blocks of compute, as its chokepoint of choice to defend its dominance in AI, but Chinese firms have found creative ways to circumvent restrictions. These include creating subsidiaries to bypass licensing limits, procuring advanced chips from third countries or the informal economy, rerouting orders destined for external data centers, and simply frontloading orders to stockpile in anticipation of export controls. Although an exact estimate of diverted chips is difficult to predict, the Center for a New American Security, through a simulation-based analysis, previously estimated in 2023 that up to 12,500 chips could be smuggled into China annually, but the true number may be far higher.
Although these smuggled chips comprise a small portion of China’s compute, they have underpinned its near-parity in AI engineering—as demonstrated by DeepSeek’s V3 and R1 models—and policymakers worry that these smuggled chips could accelerate breakthroughs to allow China’s chip industry to catch up. Currently, China’s most advanced chip is four years behind the US, and its total compute capacity is about ten times smaller. If export controls are successful and American chips are cut off, it will take time for China to upscale the intelligence of its AI capabilities (which require a logistical scale of additional compute). The US will remain keen to protect this compute advantage, even as it adjusts its AI Diffusion policy.
The Trump administration has stated that it will issue “much simpler” guidance in lieu of Biden’s AI Diffusion rule. Although it is impossible to speculate what a replacement will look like, the administration has given clear signals that it will maintain a strong export control regime, as exemplified by last month’s new licensing requirements on Nvidia’s H20 chip—originally designed to comply with export controls to China. In terms of replacing the diffusion rule, current administration officials have suggested refining rules as a product of government-to-government diplomacy—which could be interlinked with broader trade negotiations. Other former officials suggested either implementing stronger “know your customer” regulations or reducing non-licensed orders from 1,700 H100 or equivalent chips to something smaller (say 500) to improve oversight of more chip shipments. Separate from the administration, Sen. Tom Cotton (R-AR) and Congressman Bill Foster (D-IL) plan to propose a bill requiring on-chip verification technology to improve traceability of advanced US-made chips, prevent smuggling, and improve export control enforcement, potentially eliminating the need for complicated diffusion rules.
Government-to-government negotiations may be the preferred method for the dealmaker-in-chief. During President Trump’s visit to the Middle East this week, he is expected to announce large chip deals with UAE-based G42 and Saudi Arabia-based Humain, two companies that would have otherwise been barred from such orders under the AI Diffusion rule. Such deals may follow the example made by the Biden administration, which negotiated chip deals with G42 in 2024 under the condition that it implement new security measures and reduce its partnerships with Chinese technology firms.
The Risks
By rewriting the Diffusion Rule, the Trump administration will reopen old policy tradeoffs. US tech dominance has tried to balance multiple priorities at once, ranging from monopolization of the industry, unleashing growth and innovation, and seeking to shape the global AI ecosystem, either unilaterally or in tandem with US allies. Nonetheless, while the reprieve carries upside risks for tech players and underscores the administration’s receptiveness to their concerns, more information on a new Diffusion Rule is needed for long-term planning.
If government-to-government partnerships are embraced, they could take many forms, ranging from transactional trade agreements to boost US chip exports, building consortia with allies with VEUs, or coordinating export controls on additional chokepoints in the AI value chain (such as in the case of coordinating on semiconductor manufacturing equipment controls alongside the Netherlands and Japan). If a government-to-government strategy prefers transactionalism, the administration could create negotiating leverage by ratcheting up expanded licensing or VEU requirements, creating upside risks for willing US partners who can strike deals but downside risks for industry as they navigate a patchwork of export controls. Other AI aspirants, such as the EU, may accelerate their efforts to establish AI capabilities that are either autonomous or indispensable to the US to hedge for dealmaking, ironically accelerating risks of AI diffusion. On the other hand, a government-to-government strategy could embrace minilateral initiatives or the creation of small clubs of like-minded states that coordinate AI policy and sharing. However, minilateral initiatives would diminish the administration’s leverage in shaping the AI ecosystem or negotiating bilateral trade agreements.
A potential downside risk in delayed or case-by-case diffusion rules is a fragmented enforcement regime that allows diversion of chip flows to blacklisted firms, accelerating China’s efforts to establish bipolarity in the AI ecosystem. If China closes the gap in its compute or chipmaking capacity, it could go on a commercial offensive to displace US firms in foreign markets and shape tech norms. AI, augmenting economic and military power, could become central to a modern arms race and carry unpredictable ramifications for global stability.
