Compute governance is the idea that policymakers can oversee AI most effectively by focusing on the computing power, or “compute,” used to develop and operate advanced models, rather than trying to regulate the algorithms or data directly. The argument is set out in the 2024 paper “Computing Power and the Governance of Artificial Intelligence” by Girish Sastry, Lennart Heim, and a large group of co-authors including Yoshua Bengio and Diane Coyle, submitted on 13 February 2024.
The paper contends that compute has four properties that make it an unusually practical lever for governance. It is detectable, because large training runs require identifiable hardware and facilities; excludable, because access to chips and data centers can be restricted; quantifiable, because computing power can be measured and reported; and produced through a highly concentrated supply chain dominated by a small number of firms. These features let governments gain visibility into who is building the most capable systems, steer resources toward beneficial work, and enforce restrictions against irresponsible or malicious development in ways that would be far harder if they tried to monitor software or datasets.
Compute governance underlies real policy: the US executive order on AI used training-compute thresholds to decide which models trigger reporting, and export controls on advanced chips reflect the same logic. The approach has critics who warn it could entrench incumbents or grow ineffective as algorithms become more efficient. For the public, it explains why semiconductors and data centers, not just AI models, have become central to national AI policy.