Governing the Energy Footprint of Artificial Intelligence for Sustainable Development

Artificial intelligence can accelerate progress on the Sustainable Development Goals, yet the electricity infrastructure supporting AI can also increase emissions, strain power systems, and compete with other development priorities. This paper reframes AI energy demand as a sustainabledevelopment governance problem rather than a narrow efficiency problem. Building on the thermodynamics of computation and updated energy-system evidence, it combines physical accounting with a scenario analysis of a representative twenty megawatt data center.

The analysis distinguishes four levers that are often conflated in claims of green AI: computational efficiency, facility efficiency, electricity carbon intensity, and system-level energy additionality. It then proposes an SDG-Compatible Compute framework for evaluating whether AI infrastructure contributes to, rather than merely consumes from, sustainable energy transitions. The framework requires efficiency disclosure, time-and locationaware carbon accounting, credible clean-energy additionality, and assessment of local grid and heat-reuse effects.

The results show that efficiency gains alone cannot ensure sustainability when AI demand grows faster than energy intensity falls. Conversely, low-carbon power can substantially reduce operational emissions, but procurement that lacks temporal matching or additional generation may shift burdens elsewhere on the grid. Sustainable AI therefore requires coordinated design of models, data centers, electricity supply, and public policy.

The paper links these choices directly to Sustainable Development Goals 7, 9, 12, and 13 and offers a compact governance approach suitable for developers, utilities, regulators, and public purchasers.