AppsGuidesAI DictionarySmall FixesOn RecordAboutFollow the lab

Compute governance

Compute governance means using the chips and computing power that AI needs as a handle for oversight: tracking who trains big models, sharing access, and restricting misuse.

Safety and societyTechnical

In plain words

Software is hard to police. It can be copied in seconds and run anywhere. But training a cutting-edge AI model needs something very physical: thousands of specialised chips running for weeks in huge data centres.

Those chips come from a handful of companies, and the amount of calculation used can be counted. That makes compute a practical place for rules to grip. Compute governance is the idea of using it that way.

In practice it takes three forms. Seeing: requiring companies to report or test models trained with more than a set amount of compute. Sharing: governments funding or subsidising compute so more people can build AI, as India does. Restricting: controlling who may buy the most powerful chips, as US export rules do.

Everyday example

Tracking big building projects by cement

A city cannot inspect every renovation, but it can notice who orders thousands of tonnes of cement. Only large projects need that much, so the order itself flags the project for a building permit and a safety inspection. Compute governance does the same with chips and computing time: the scale of the order points to the projects worth checking.

Where the example breaks: Cement use maps neatly onto building size. Compute is a rougher signal: smaller, efficient models can be very capable, and a fixed compute line can miss them or catch harmless ones as chips get cheaper.

Go deeper

How it actually works

A February 2024 paper by Girish Sastry, Lennart Heim and 17 co-authors set out the case, arguing that compute is detectable, excludable and quantifiable.

  1. Why compute works as a handle

    Advanced AI chips have an extremely concentrated supply chain, from chip design to the machines that make them. Training compute is measured in floating-point operations, so a law can name a number.

  2. Visibility: compute thresholds

    The EU AI Act presumes a general-purpose model has systemic risk if training used more than 10^25 operations, bringing extra duties. California's SB 53 (from January 2026) and New York's RAISE Act (from January 2027) use more than 10^26 operations to define frontier models. A 2023 US executive order also used 10^26, but was revoked on 20 January 2025.

  3. Allocation: sharing compute

    Governments can steer who gets to build. Under the IndiaAI Mission, the government said in March 2026 that 38,231 GPUs were available through a common compute facility at subsidised rates for startups and researchers.

  4. Enforcement: chip controls

    Export controls limit sales of top chips and chipmaking tools to some countries, mainly China. A US plan to cap AI chip sales to most countries, including India, was rescinded in May 2025 before it took effect.

  5. The trade-offs

    The same paper warns that poorly designed compute rules carry risks to privacy, the economy and concentration of power. Thresholds also age as training gets cheaper and efficiency improves.

Key paper
"Computing Power and the Governance of Artificial Intelligence", February 2024
EU threshold
10^25 training operations: presumed systemic risk under the AI Act
US state threshold
10^26 operations, plus over US$500 million revenue for the main duties (California SB 53, New York RAISE Act)
India
No compute thresholds in law as of November 2026; compute policy focuses on subsidised access through the IndiaAI Mission

Compute governance is one tool among many. It works best for the largest models and says little about how smaller ones are used.

Is compute a good way to regulate AI?

One view

Yes, it is measurable

Compute is one of the few inputs to AI that is physical, countable and made by few suppliers. Thresholds give clear, objective lines and catch the biggest training runs early, before a model is released.

The other view

It is a blunt proxy

Capability does not track compute closely: efficient methods and distillation make strong models cheaper. Fixed lines go stale, and chip controls can hurt allies, concentrate power and push rivals to build their own chips.

Most likely: Most policy experts see compute rules as a useful first filter for the very largest models, not a full solution. Laws now pair compute thresholds with other tests such as revenue, capability evaluations and how a model is used.

Questions people ask

What does 10^26 FLOPs mean?

It is 1 followed by 26 zeros floating-point operations, the total calculation used to train a model. California and New York use it as the line for frontier models. The EU uses 10^25, ten times smaller, to presume systemic risk.

Is compute governance the same as chip export controls?

No. Export controls are one tool within compute governance, the restricting side. Compute governance also covers reporting duties tied to training size and public programmes that share compute.

Does India have compute rules for AI?

Not thresholds in law, as of November 2026. India's approach is mainly about access: the IndiaAI Mission offers subsidised GPU time to startups, researchers and students through a common compute portal.

Why do governments care about AI chips?

Because the most capable AI needs vast numbers of them, and only a few firms can make them. Whoever controls chips influences who can build frontier AI, which makes chips a matter of economics and national security.

Was this clear?

1 is confusing, 5 is crystal clear. One tap, no account needed.

Spotted something wrong, or know a better example? Suggest a change

Further reading

Prepared by Mundhra Labs. Reviewed by Aditya Mundhra. Checked November 2026.