The Diffusion Dilemma

Protecting vs. Leading in AI

Over the past several years, the United States has taken steps to restrict Chinese access to high-end chips and advanced AI-related technologies. From Huawei’s placement on the Entity List to sweeping rules on semiconductor manufacturing, each regulatory update has sharpened the government’s tools for controlling dual-use tech. Now, with the Framework for Artificial Intelligence Diffusion, The Department of Commerce’s Bureau of Industry and Security (BIS) is taking an even bolder step: moving beyond hardware constraints and into regulating large-scale AI capabilities themselves. Below we trace the policy continuum from Trump 45-era blacklists to this latest system of “tiered” country designations and new export rules for AI models. We also highlight a key dilemma: while these measures aim to safeguard national security interests by denying frontier AI to strategic rivals, they may also undercut the global collaboration and foreign market access that have sustained America’s leadership in emerging technologies.

Incremental Additions to Advanced Computing Export Controls

On May 21st, 2019, the Bureau of Industry and Security (BIS) issued a rule to add Huawei to the Entity List of the Export Administration Regulations (EAR). This imposes license requirements on exports, reexports, and in-country transfers to Huawei. BIS justified Huawei’s addition, arguing that they had reasonable cause to believe that Huawei has been involved in activities contrary to the national security or foreign policy interests of the United States. Adding Huawei to the EAR’s Entity List restricted the company from developing products using certain US-designed manufacturing equipment and software.

On May 18, 2020, BIS, under the Foreign Direct Product Rule, issued a interim final rule (IFR) extending the export controls targeting Huawei by restricting its access to semiconductors and products manufactured with U.S. technology. BISaimed to curtail Huawei's ability to design and manufacture advanced chips critical for its operation.

On December 18, 2020, BIS amended the EAR to add 69 entities to the Entity List. The added entities included various Huawei subsidiaries, limiting the access to U.S. technology.

On January 20th, 2021, BIS issued an interim final rule to secure the Information and Communications Technology and Services (ICTS) supply chain from foreign adversaries. It established procedures for BIS to assess and address transactions involving ICTS that pose risks to U.S. national security. The rule provides a framework for identifying, mitigating, or prohibiting transactions involving foreign-owned or controlled entities, with a focus on addressing vulnerabilities that could be exploited to undermine critical infrastructure, data security, and economic stability.

On December 17, 2021, BIS added 34 more entities to the Entity List, including China’s Academy of Military Medical Sciences (AMMS), which develops military biotechnology, including purported “brain-control weaponry”.

On October 11th, 2022, BIS implemented additional export controls targeting advanced semiconductor chips, instrumental transactions for the development of supercomputers, and transactions for certain entities on the Entity List. Additionally, BIS implemented controls on semiconductor manufacturing equipment and transactions for integrated circuits used for military end-uses in China.

BIS also amended the “Foreign-Direct Product” rules to expand the Export Administration Regulations (EAR), covering advanced computing items and those for supercomputer uses in China, if they are made with U.S. technology.

On October 25, 2023, BIS issued a new rule on the export of advanced computing and semiconductor manufacturing items to China as well as transactions related to supercomputer end-uses in China. New rules:

1. Amend controls for chips to prevent workarounds and to add license requirements for non-China regions such as the Middle East.

2. Increase export reporting, including of chips with near-threshold performance

3. Increase the scope of semiconductor manufacturing equipment covered by arms embargoed countries

4. Asserted that any foreign-made lithography equipment with U.S. components was also subject to export controls.

5. Imposed end-use ownership restrictions to mitigate circumvention risks for entities with ownership ties to China.

BIS issued another interim final rule on December 5th, 2024 revising the Export Administration Regulations (EAR) to address China’s ongoing efforts to develop independent and controllable semiconductor manufacturing capacities. The IFR expands controls on advanced computing items, supercomputers, and semiconductor manufacturing equipment (“SME”), including newly created Foreign Direct Product (FDP) rules. These revisions target commodities that are the products of U.S.-origin technology or software and are used to produce “advanced-node integrated circuits” for end users of concern. Although effective December 2, 2024, certain provisions have later compliance dates, and BIS has invited public comments on the IFR no later than January 31, 2025.

Framework for Artificial Intelligence Diffusion

Having surveyed some of the intricate patchwork of export controls under both the first Trump administration and the Biden administration, it’s clear that the United States has been ratcheting up oversight of cutting-edge computing technology. Yet the new AI Diffusion Framework isn’t incremental–it is a paradigm shift. This framework leaps far beyond hardware restrictions and regulates the software that underpins artificial intelligence, drawing new global lines on who can access, build, and deploy the most powerful AI chips and models.

The new global framework sets export controls on both AI chips (hardware) and advanced AI models (software). The rule clusters the world into three “tiers” of countries. Tier 1 (the U.S. plus 18 key allies) face minimal restrictions; Tier 3 (e.g., China, Russia, and 21 other arms-embargoed nations) are essentially barred from obtaining high-performance AI chips; and Tier 2 (approximately 153 other countries) can access chips only through strict quotas or via “Validated End User” (VEU) licenses. These licenses divide further into “Universal” VEU status (for large U.S. cloud providers or similar) and “National” VEU status (for major non-U.S. players). Each VEU must meet robust security requirements and maintain detailed inventory and usage reporting on advanced AI hardware. Tighter rules also apply to the model weights of extremely large AI systems (those trained with roughly 10^26 or more mathematical operations), which in effect will prohibit unauthorized exports of closed-weight frontier models. BIS does not require licenses for the export of models with open-weights.

To define what counts as a “high-performance” AI chip, BIS employs a Total Processing Performance (TPP) threshold: chips above certain TPP and “performance density” levels are subject to these controls. Countries in Tier 2 have an aggregate TPP “cap” that limits how many such chips they can install before 2027. For lower-power (sub-threshold) chips, limited annual shipments may proceed without a formal license, but still require reporting. Overall, these measures are intended to significantly reduce loopholes that previously allowed re-exports or stealth shipments of high-end chips to Tier 3 destinations, encourage Tier 2 countries to seek formal VEU approvals, and tighten control over frontier AI training for entities outside Tier 1.

Competing National Security Interests Introduce a Key Contradiction

The Framework for AI Diffusion places the United States in a position of attempting to achieve two fundamentally competing goals. On one side, there is a clear national security imperative: preventing nation-state adversaries from misusing dangerous AI capabilities, especially for weapons of mass destruction or surveillance. On the other side, U.S. national security depends on America remaining the undisputed leader in foundational science and technology, including cutting-edge AI research and development. (The Export Control Reform Act of 2018, 15 U.S.C. 4801 explicitly states this.) By restricting the diffusion of AI model weights, BIS aims to minimize misuse–yet these controls inherently risk the global collaboration that is a major contributor to U.S. leadership in frontier technologies.

Restricting large-scale AI models under a dual-use rationale is a potentially impossible balancing act between shielding frontier technology from harmful applications and fostering the innovation of beneficial applications. Dual-use AI models can indeed have harmful applications, such as aiding in chemical or biological weapons design, but they can also yield groundbreaking medical advances and scientific discoveries when shared responsibly. By imposing complex licensing requirements and total processing performance (TPP) thresholds on exports, BIS may deter transfers for illicit purposes, but will almost certainly also discourage legitimate uses overseas and cap demand of U.S. chips and AI models in Tier 2 countries. Tier 2 countries may look elsewhere for this technology, including China's growing chip industry. This tradeoff risks eroding the economic reach, research feedback loops, and data partnerships that reinforce America’s AI leadership.

Moreover, the intangible nature of software-based model weights raises practical concerns about enforcement and potential unintended consequences. Determined adversaries could steal proprietary technology through hacking or espionage – or simply invest in implementing open-source alternatives. Meanwhile, legitimate partners and U.S. developers bear compliance costs for negligible security gains. The result is a central contradiction: measures designed to protect U.S. AI leadership might instead undermine it if they push global AI ecosystems to develop or adopt non-U.S. AI products.

Capping U.S. Chip Demand is not a Recipe for U.S. AI Leadership

The chip caps that apply to most countries will hurt American interests globally. Each Tier 2 country is allowed a maximum quota of 49,901 H100 equivalents shipped from 2025 until 2027 (credit to SemiAnalysis for these calculations). This quota does not apply to chips that are transacted to a Validated End User (VEU) in a Tier 2 country, but the VEU process also adds significant friction and expense. For example, an Israeli AI startup training a frontier-level foundation model not only has to purchase a large number of GPUs, but it also has to become a VEU, which requires a licensing process and legal expenses. In practice, the quota plus the VEU licensing process adds friction for foreign technology companies and places a ceiling on international demand for U.S. chips.

Export controls will also harm U.S. companies. Recent empirical research by staff from the Federal Reserve Bank of New York suggests that affected U.S. suppliers will lose Chinese sales and will not replace it with more domestic or other international customers, at least in the three years following export controls. Their data show a net demand decline for the relevant U.S. products once export controls take effect, with no compensating uptick in new markets. They also document immediate stock market losses for impacted U.S. firms–an aggregate $130 billion drop in market capitalization–along with a 20% decrease in cash flow and 8.6% decline in total employees. However, these firms did not significantly reduce capital expenditures.

Additionally, the same research finds that Chinese firms often reorganize their supply chains by turning to domestic suppliers–an adaptation that appears relatively swift, partly due to state-owned enterprises benefiting from top-down coordination. This indicates that Chinese corporations might more effectively “reshore” capacity than U.S. suppliers can “friend-shore” or pivot to other global customers. Such pivots for American companies will be even more difficult given the Framework’s imposition of additional compliance costs for non-VEU firms in Tier 2 countries.

Another recent empirical study conducted at Harvard Business School analyzes the economic consequences of a 2007 U.S. policy called the “China Military Catch All Rule.” This BIS rule expanded licensing requirements for commercial products destined for military end-uses by the People’s Republic of China. In line with its goals, the policy depressed the performance of Chinese firms that relied on controlled technologies. Over the longer term, however, these same Chinese firms experienced a notable uptick in R&D expenditures and total patents, suggesting they adapted by innovating around restrictions. While patents and R&D are only rough proxies for invention, the study underscores that export controls can spur targeted countries to develop their own capabilities.

Taken together, these findings reinforce an intuitive point about export controls: although they damage the businesses of both U.S. exporters and their foreign customers in the short run, they also tend to incentivize the targeted country–in this case China–to accelerate domestic innovation. As a result, overuse of export controls may yield diminishing returns for the United States’ long-term tech leadership if competitors are effectively pressured into scaling their own alternatives.

Export Controls Are Too Much - and Not Enough

For the reasons above, the Biden diffusion rules are likely to create significant unintended consequences. At the very least, the number of countries in Tier  2 should be narrowed; for instance, grouping India, Yemen, Israel, Poland, Singapore, and Portugal all under Tier 2 ignores the stark differences in their relationships with the U.S. and U.S. adversaries, effectively forcing fundamentally dissimilar partners to follow the same licensing requirements.

None of this implies that export controls should be entirely eliminated. Indeed, they remain an important tool in preventing truly hostile states or malicious actors from acquiring militarily sensitive or dual-use AI technology. However, in the long-term, export controls cannot guarantee America’s leadership in advanced AI. The evidence suggests that if export controls are used in isolation, they may unintentionally harm U.S. firms while accelerating innovation in rival nations.

To truly secure U.S. global competitiveness in AI, policymakers may need to pair export controls with complementary measures–such as favorable federal and state regulatory environments, public-private R&D partnerships, deeper collaborations with closely allied nations,enhanced cybersecurity measures for domestic data centers, and workforce development. In short, export controls remain a critical policy lever for safeguarding U.S. national security, but they must be part of a broader strategy that supports continued domestic and targeted international innovation under agreeable terms.