Chips, compute and data centres
Compute governance
In short
Compute governance is the use of rules on advanced computing power to oversee who can develop and run advanced AI. A 2024 research paper treats that hardware as easier to detect, exclude and measure than data or algorithms.
- Also called
- Governance of compute, Computing-power governance
- Topic
- Chips, compute and data centres
- Level
- Technical
- First described
- Sastry, Heim, Belfield and colleagues, in a paper submitted on 13 February 2024, say Tim Hwang's 2018 paper was among the first to highlight the role of computing power in governing AI1.
In this literature, compute means computing power used to develop and run AI. Sastry, Heim, Belfield and colleagues argue that, next to data and algorithms, AI-relevant compute is detectable, excludable and quantifiable, and that it comes from a highly concentrated supply chain. They say large-scale work often needs thousands of specialised chips in a data centre that uses a large amount of power1.
The paper says governments had already invested in domestic compute, placed export controls on sales of computing hardware to competing countries, and subsidised access outside the largest technology companies. Beyond those steps it describes visibility, allocation and enforcement, and says compute rules are not the whole of AI governance, because risky uses that need little compute call for other tools1.
The ITU's 2025 AI governance report uses the phrase for rules on access to, and use of, advanced computing resources. Its examples are monitoring who has the most capable hardware, setting oversight thresholds for large training runs, and requiring safety measures. It also notes proposals for licensing large training runs, registries of high-risk systems, and shared compute programmes2.
Drawn badly, the same tools can do harm. The paper says naive or poorly scoped approaches carry risks for privacy, economic impact and concentration of power. Its guardrails include leaving small-scale AI compute and non-AI compute outside the rules, using privacy-preserving methods, and adding procedural safeguards1.
Example
Writing in February 2024, the paper says US Executive Order 14110 required reporting of large compute clusters, including cloud use by a foreign customer above a training threshold, and it calls that cloud rule a know-your-customer provision1.
Why it matters
Advanced chips are physical and pass through a concentrated supply chain, so governments can see and withhold them in ways they cannot with data or code. The same controls can expose private information or concentrate power if they are drawn too widely1.
Questions people ask
Why govern compute instead of only governing AI models?
The paper says AI hardware is tangible and the supply chain is concentrated, while data, algorithms and trained models are easy to copy. It also says compute rules do not replace other rules, including for risky systems that need little compute1.
Is chip export control a form of compute governance?
The paper treats export controls on computing hardware as something governments had already done. The ITU's description is wider: it also includes monitoring who holds the most capable hardware, thresholds for large training runs, and safety requirements12.
Could compute governance affect ordinary computers?
The paper says the risk is higher if rules cover small-scale AI computing or computing that is not for AI. It recommends keeping both of those outside the regime, and using privacy-preserving methods where personal data is involved1.