Export Controls on AI Chips: Biden’s Overreach Risks U.S. Leadership in Tech

As the Biden administration draws to a close, one of its final acts has been to propose an overdesigned, yet underinformed, Interim Final Rule (IFR) for a new “Export Control Framework for Artificial Intelligence (AI) Diffusion.” This rule would impose country-specific caps, along with a license regime on U.S. exports of semiconductors that would most notably impact GPU chips that underpin key AI applications such as the development of large language models (LLMs). While the Biden administration’s interest to use export controls to limit the ability of geostrategic competitors such as China and Russia to use advanced, U.S.-developed technologies for military applications is certainly warranted, and while the challenge of such advanced chips reaching U.S. competitors through third-party countries is quite real, the proposed framework fails to address the core challenge in a targeted way, and would have potentially catastrophic consequences for U.S. digital industry leadership. Moreover, the Biden administration developed the proposed rules almost wholly without industry consultation and has ignored far better approaches to address the central challenge.

At the core of the over 200-page proposed framework is the establishment of a mandatory global license requirement for AI technology and GPUs that is connected to country-specific caps on U.S. exports of these technologies. Twenty trusted countries—Australia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, the Republic of Korea, Poland, Spain, Sweden, Switzerland, Taiwan, and the United Kingdom—would be exempted from the export caps. But, for the more than 140 countries not exempted—a list which includes some of America’s most important allies such as Israel and Singapore; key trading partners such as Brazil, India, Indonesia, Malaysia, and Mexico; and other allies such as Saudi Arabia and UAE—the regulations create a Low Processing Performance (LPP) license exemption, essentially permitting certain levels of GPUs to be shipped to these nations on a country-specific capped basis. To manage the challenge of U.S. companies sending AI systems to their own branches operating in foreign countries, or to the data centers they manage, the regulations would establish a Universal Validated End User (UVEU) regime that would allow trusted partners (such as U.S. hyperscalers) to receive speedier access to the GPUs, but only so long as they comply with federal FedRAMP High requirements, a technical certification designed to protect the federal government’s most sensitive non-classified data.

The problems with the proposed framework are manifold; indeed, an approach that seeks to limit and control all global sales of U.S. advanced chips through aggregate country caps would introduce significant economic competitiveness, security, and even foreign policy concerns. Most fundamentally, placing caps on U.S. exports of AI GPUs will limit market opportunities for U.S. companies while providing an open door for foreign suppliers of AI chips—and Chinese AI chipmakers such as Biren are increasingly competitive in this field—to swoop in and take market share. By artificially restricting AI chip exports to the vast majority of the world, the U.S. government essentially forces open the door for foreign competition across the globe, redirecting revenue that could be flowing to U.S. chipmakers to invest in next-generation chips with competitors abroad.

In particular, the proposed regulations misunderstand a key element of how AI chips are being used to develop LLMs and work on other computing challenges: GPUs scale both up and out, meaning that their capability comes from operating many GPUs simultaneously to attack a computing challenge. Therefore, controlling GPUs makes little sense when a competitor can achieve parity by simply adding more, if less-powerful, GPUs to solve the computing challenge. In other words, even if Chinese-made GPUs aren’t as capable as U.S.-made GPUs in the short term, Chinese firms (or those from other competitors) will happily supply the computing power needed to meet customers’ challenges, vitiating the administration’s goal of limiting overall AI computing while simultaneously harming U.S. firms’ leadership in global AI compute technology and market share.

Moreover, the framework would introduce considerable—likely intractable—compliance challenges. First, it would be extraordinarily difficult (if not well-nigh impossible) for companies to know if a product export would bring a country over the compute cap unless there were a burdensome process whereby some U.S. government entity reviewed all sales and aggregated sales information across all industry (including from foreign firms). Even the existing AI chip export license process, which applies to only a few countries (and China already), is unpredictable, expensive, and time consuming; so expanding that to a global process would be highly impracticable, and that would hold true for any country-level cap, regardless of how high.

Beyond this, the policy could introduce significant market distortions, especially by putting the U.S. government in the difficult position of picking winners and losers, both in specifying what are the country-specific caps and potentially which firms would get to fulfill these caps. For instance, consider that the U.S. government assigned a given country an annual cap of 25,000 GPUs and three U.S. companies submitted license requests to ship 10,000 GPUs to the nation; how would the U.S. government arbitrate that request? Similarly, if a company informed the U.S. government that it had contracted to sell 25,000 GPUs to that nation, would other companies be precluded from selling their product into that nation, even if the order had not been fulfilled yet? How often would the country-specific caps be updated? In short, the U.S. government would effectively have to develop a regime to oversee and manage U.S. companies’ GPU sales to non-exempted foreign nations.

Further, as noted, if the U.S. government wildly missed in its estimate—say a friendly ally like Singapore could have readily consumed twice as much GPU compute power—that market demand wouldn’t go unfulfilled. Instead, it would simply be filled by a foreign competitor, who would then have a foot in the door for future sales. This risks significantly damaging a high-value-added industry pioneered by U.S. innovators that lead the world in their technology and markets. Indeed, analysts estimate that U.S. firms currently account for 70 to 90 percent of the global market share for generative AI solutions. U.S. chipmaker NVIDIA itself accounts for 70 to 95 percent of the global market for AI accelerators. Yet U.S. competitiveness in the AI/cloud/hyperscaler market goes well beyond well-known firms like AWS, Google, Oracle, and NVIDIA but includes other leaders such as Cerebras, Tenstorrent, Groq, and D-Matrix and dozens of more startups beyond. These companies employ tens of thousands in high-earning jobs and generate significant exports for the U.S. economy. In short, it makes little sense that the Biden administration would expend so much effort and resources to revitalize the U.S. semiconductor industry—notably spearheading passage of the $52.7 billion CHIPS Act—only now to potentially compromise that leadership by tying one hand behind companies’ backs as they seek to sell their world-leading chips into global markets.

Moreover, with U.S. AI GPU solutions still the most sought after globally, that raises the specter that countries would likely jockey and campaign for higher caps, introducing heightened foreign policy considerations into the equation. The United States would potentially risk injuring diplomatic relations with countries; for instance, U.S. policy could unwittingly drive developing nations that might not enjoy high cap levels, or don’t see their caps increased over time, directly into the arms of the Chinese Communist Party and Chinese firms.

Another conceptual problem with the framework’s design is that it applies a highly blunt instrument instead of addressing what should be the core goal: limiting AI compute capacity for activities opposed to core U.S. national security interests from core competitors—notably China, Iran, North Korea, and Russia. U.S. export control policy has long been linked to targeting specific end users and end uses of U.S.-developed technologies that the United States wishes to deny adversaries the ability to use. The proposed framework turns decades of U.S. export control policy directly on its head, moving it from a tailored, narrow approach to imposing a blanket license regime on all countries (even if a subset is exempted). This represents a quite significant expansion of the U.S. export control regime, both in conceptualization and operation, and as such points to how overwrought these proposed regulations truly are. Indeed, their aim is nothing less than to control the global speed and diffusion of advanced AI technologies.

Beyond the conceptual and operational challenges with the proposed rules lies the reality that they preclude approaches that could achieve key objectives while being far more effective, less burdensome, and less damaging to U.S. industry competitiveness. In particular, controlling so-called frontier AI system installs (or “clusters”) instead of the total number of GPUs in a country would be a better approach to achieve U.S. national security objectives in a more effective and targeted manner. (Frontier AI models refer to the most advanced, cutting-edge models pushing the boundaries of what AI can accomplish.) These frontier systems are massive—requiring 100,000 to 400,000 GPUs—operating in hyperscale data centers so large that in some cases they can be detected from satellite imagery. These frontier systems require significant technical expertise to build and support (i.e., networking, cooling, etc.) and require ongoing maintenance operations and support. There exist only a few companies and countries with the resources requisite to buy and install these systems, and concentrating on these would enable the U.S. government to focus on a smaller set of higher-risk systems.

In other words, as ITIF has previously argued to the U.S. Bureau of Industry and Security (BIS), regulators should move from focusing on compute-based to performance-based evaluations of AI systems. Indeed, compute power represents a poor measure of AI model risk. Some high-compute models could pose minimal threats, while lower-compute models with superior performance could present far greater risks. A shift to performance-based thresholds would provide a more accurate assessment of capability and risks, better identifying the most advanced AI models. While ITIF made this point with regard to BIS developing reporting requirements for development of advanced AI models and computing clusters, the point that regulators in general should be focused on is overall system capabilities and performance as opposed to raw compute (e.g., in this case of country-capped exports of GPUs) which still stands and provides a better framework for developing policy in this area.

The Biden administration is trying to force other countries to pick a side—the United States or China—and it is likely going to discover that if it issues this ultimatum, many will pick China. After all, only one of these countries is actively threatening to cut them off from valuable AI chips that they will need to compete in the digital economy. The next administration should focus on improving U.S. competitiveness in AI by expanding market access for U.S. chips and AI technologies and limiting the influence of its geostrategic competitors in the sector. The vast majority of uses for these chips will be for legitimate, lawful purposes, and the goal should be to capture and retain as much of this market as possible. Moreover, the United States should be countering efforts by China and Russia, which recently launched an AI Alliance Network among BRICS countries, to offer its own allies and partners access to AI data, models, and compute resources.

As articulated, many concerns pertain to the Biden administration’s “Export Control Framework for Artificial Intelligence (AI) Diffusion.” However, perhaps the most significant concern is that they’re already being submitted as an Interim Final Rule (which despite the obfuscatory bureaucratic jargon effectively means “final”) despite their promulgation taking place with virtually no industry consultation. Moreover, the rules would be set to go into effect just 60 days after their publication in the Federal Register, scarcely enough time to erect such a massive regulatory edifice, let alone give industry a chance to adapt.

In conclusion, establishing a globalized country cap system for AI compute is every bit as bad an idea as the U.S. government putting price controls on groceries (or pharmaceuticals). It’s also transparently clear that the Biden administration is trying to shoehorn these rules under the wire in the dying days of its term and hamstring the following administration from making its own assessment of the central challenge and how to respond to it. The Export Control Framework for Artificial Intelligence Diffusion represents a flawed policy that should be immediately withdrawn and replaced with a better approach in the next administration based upon more-extensive stakeholder input and consultation.