The Geopolitics of Compute in the Aftermath of the AI Diffusion Framework

Abstract

Recent U.S. government policy decisions demonstrate a willingness to curtail AI development within strategic competitors such as China. Compute has served as a reliable proxy for AI system capabilities and is an easily targetable input for AI development. Thus, it is an ideal target to constrain the diffusion of AI capabilities. The release of the 'Framework for Artificial Intelligence Diffusion' by the Biden administration in its last months was the most expansive of these efforts. Though the policy is no longer in place, lessons can be learned from its implementation. This paper acts as a strategic and governance risk analysis of the Biden administration's most significant AI policy, examining its motivations and drawbacks. It does not argue that compute-governance or export controls should be abandoned or that it is unimportant. Rather, it discusses the regulatory tiering of the European Union, that the frontier models of today will be replicated by models using less compute in the future, and the potential for China to adapt to a constrained compute environment and indigenise computing capabilities in five to fifteen years. The paper concludes with a brief discussion of the U.S.'s evolving compute-based policies since the Trump administration began.