Compute Governance and International Security

Summary

This research paper from the Centre for the Governance of AI examines how computing power (compute) serves as a critical lever for AI governance, analyzing how controlling access to compute infrastructure can enable oversight, safety enforcement, and international coordination on AI development. It explores practical mechanisms through which policymakers can use compute as a governance tool to manage AI risks.

Key Points

• •Compute is a uniquely governable bottleneck in AI development due to its physical, traceable, and concentrated supply chain properties.
• •Hardware-level controls, export restrictions, and compute monitoring could enable states to enforce AI safety standards and limit dangerous capabilities.
• •International agreements around compute allocation could facilitate coordination on AI development norms, analogous to nuclear nonproliferation regimes.
• •Compute governance faces challenges including enforcement difficulties, potential to entrench existing power asymmetries, and risks of compute arms races.
• •The paper proposes a tiered governance framework linking compute access to safety evaluations and compliance with international AI standards.