Compute Governance

Compute governance is the use of control over computing resources, especially the specialized chips and data centers used to train and run large AI models, as a lever for shaping who can build advanced AI and under what safeguards. The idea treats compute as a governable input in the way policymakers have long treated other inputs to dangerous or strategic technologies: it is physical, measurable, and passes through a small number of chokepoints, whereas algorithms and data are easy to copy and hard to track. It is one tool within the wider field of AI policy and is distinct from governing model behavior or deployment directly.

The scholarly case was set out in Computing Power and the Governance of Artificial Intelligence (Sastry et al., 2024), a paper from authors including Girish Sastry, Lennart Heim, Haydn Belfield, Markus Anderljung, and Miles Brundage, with many co-authors from academia and industry. It argues that compute has properties that make it a good policy target: it is detectable, because large training runs need large concentrated facilities; excludable, because chips are physical goods that can be withheld; quantifiable, because compute can be counted in floating-point operations; and produced through an unusually concentrated supply chain, with a few firms dominating chip design, fabrication tools, and manufacturing. The paper organizes uses into increasing visibility (tracking and reporting), allocating compute (for example subsidizing safety research or restricting access), and enforcement (such as hardware-level limits), and it also discusses the risks, including privacy, concentration of power, and the chance of ineffective or counterproductive measures.

Export controls are the most prominent real-world instrument. In October 2022 the US Bureau of Industry and Security issued rules restricting exports to China of advanced AI chips and of equipment for making advanced semiconductors, and it tightened and extended them in October 2023 and later updates, after companies introduced chip variants designed to fall just beneath earlier performance thresholds. Allied governments, notably the Netherlands and Japan, adopted related restrictions on advanced chipmaking equipment. Controls work by exploiting supply-chain concentration, but they also carry costs: they spur domestic substitute efforts, create smuggling and cloud-access workarounds, and can affect the revenue of the firms they target. Their effectiveness at slowing capability is debated, and evidence is still being collected.

Compute thresholds turn the lever into rules, and the numbers are tied to dated instruments. The EU AI Act (Regulation 2024/1689) presumes that a general-purpose AI model has high-impact capabilities, and therefore systemic risk, when the cumulative compute used for its training exceeds 10 to the power of 25 floating-point operations, with the Commission empowered to amend the threshold and to designate models on other grounds; providers of such models face additional duties on evaluation, adversarial testing, incident reporting, and cybersecurity. US Executive Order 14110, signed in October 2023, set a reporting requirement for models trained above 10 to the 26 operations, with a lower threshold for models trained mainly on biological sequence data, but it was rescinded in January 2025. California's SB 1047, which used a similar 10 to the 26 figure, was vetoed in September 2024. The thresholds were chosen to capture only a handful of the largest runs of the time, and they are explicitly proxies for capability and risk, not measures of either.

The proxy is the central weakness, and the main misconception is that compute equals capability. Algorithmic progress, better data, and distillation mean smaller models trained with less compute can reach capabilities that once required more, so a fixed threshold gradually captures fewer of the concerning systems, and rising compute budgets can make it capture too many. Newer methods that spend large compute at inference time, such as long reasoning chains, are not counted by training-compute thresholds. Cloud rental, distributed or pooled training, and chips smuggled through intermediaries complicate monitoring, and reporting regimes raise questions of confidentiality and who verifies the numbers. Proposals therefore favor combining compute thresholds with capability evaluations and with regular updating, rather than relying on any single number.

Compute governance also has legitimate critics and trade-offs. Concentrating control over chips can entrench incumbents, restrict academic and small-company research, and create geopolitical friction; hardware-level monitoring raises security and privacy risks; and rules aimed at distant frontier risks may add burden for ordinary developers. Its supporters respond that visibility into large training runs is a minimal, targeted measure compared with regulating all AI use. Either way, the debate shows why compute figures now appear in statutes and trade rules and why they should be read as policy proxies, with a date attached.

Read compute governance as the attempt to regulate AI through the one input that is physical and countable, accepting that the count is only a rough stand-in for what the models can do.