Computing Power and the Governance of AI

This comprehensive paper presents compute governance as a promising approach to managing AI development. The authors argue that computing power offers a distinctive opportunity for intervention due to its detectability, excludability, quantifiability, and concentrated supply chain. Unlike other AI inputs like data and algorithms, compute is a tangible resource that can be monitored, controlled, and regulated. The paper systematically explores how compute governance can enhance three key governance capacities: increasing visibility into AI capabilities, steering AI progress through resource allocation, and enforcing prohibitions against reckless AI development. The authors propose numerous policy mechanisms while maintaining a balanced perspective, acknowledging potential risks such as privacy concerns, centralization of power, and unintended economic consequences. They emphasize that the design and implementation of compute governance strategies are crucial, and recommend implementing safeguards to mitigate potential negative impacts.