FLAWS OF THE COMPUTE-DENIAL REGIME: Evaluating the United States’ Strategy to Counter the People’s Republic of China’s Access to Advanced Artificial Intelligence Capabilities

Abstract

Over the last decade, the United States has instated significant restrictions on advanced semiconductor chips to achieve the policy goal of restricting PRC access to the computational power (compute) required for advanced artificial intelligence (AI) development. The United States has a vested interest in preventing the PRC’s access to artificial intelligence technology, which may be used in military applications to bolster the People’s Liberation Army’s (PLA) weapons and surveillance systems. Further, the U.S. has an interest in preserving its technological superiority for the purposes of economic competition. Though policymakers have operated under the assumption that advanced semiconductor restrictions are effective, emerging evidence suggests that the PRC engages in several activities that not only allow access to controlled graphics processing units (GPU) but also undermine the reliance on these hardware technologies entirely. This thesis seeks to answer the following research question: is the current U.S. export control strategy effective at achieving the policy goal of preventing the PRC from achieving advanced AI capabilities? More specifically, it will investigate how semiconductor smuggling operations, remote access to compute, and recent innovations in AI development have complicated the United States’ aspirations of compromising the PRC’s AI industry.

The thesis hypothesizes that U.S. compute-denial strategy is much less effective at preventing Chinese access to artificial intelligence than policymakers believe. However, due to the uncertainty and rapid evolution of the artificial intelligence industry, the thesis hypothesizes that semiconductors may continue to play an important role in facilitating the United States’ advantage as the leader of AI technology. To investigate the research question, the thesis employs a qualitative research design underpinned by rigorous secondary research, a series of anonymous interviews with experts around the world, and a comprehensive survey of current events and news related to the rapidly evolving AI industry.

Ultimately, the research findings conclude that the U.S. compute-denial regime is ineffective at preventing the PRC’s access to advanced AI models. Rather, export controls on advanced semiconductors only contribute to solving a small part of the larger issue of PRC AI development. Though the media and scholars alike have hyper fixated on the issues associated with current export controls, namely semiconductor smuggling operations, the PRC has pursued remote compute access options through cloud service providers and decentralized GPUs, effectively gaining access to export-controlled chips. Further, the PRC has participated in a global initiative to create more efficient AI models, which has lowered the amount of compute and hardware infrastructure necessary to develop AI models of a particular caliber.

The research findings support the cohort of scholars who agree that expanded U.S. restrictions on legacy and consumer-level GPUs are not a viable solution to current policy issues, siding with those who have decentralized export controls from the scholarly discussion more generally. Further, whereas the relevant U.S. foreign policy is often discussed in disjointed terms of export controls and restrictions on remote access to hardware, the present thesis reframes U.S. strategy in terms of a ‘compute-denial regime,’ a term that encapsulates both physical and remote access methods.

Based on the findings, U.S. policy makers should pursue broad strategies of shifting their focus away from export controls and supplementing negative foreign pressures with positive, domestics ones. Likewise, U.S. officials should not expand export controls to encapsulate legacy or consumer-level GPUs already available to the PRC.