Can renewable energy meet the surging power demand of artificial intelligence? A systematic review

Whether renewable energy can scale fast enough to meet the surging power demand of artificial intelligence (AI) without undermining broader decarbonization goals is the central question of this review. A PRISMA 2020-compliant systematic review of 88 peer-reviewed studies (2019–2025; Scopus, Web of Science, IEEE Xplore) yields four findings. First, peer-reviewed per-task inference energy spans more than three orders of magnitude, with image generation at approximately 1450 times a text-classification baseline and video generation substantially higher per frame; reasoning and agentic AI are an emerging modality with no peer-reviewed end-to-end measurements as of the search date. Second, corporate procurement instruments resolve into four tiers—unbundled RECs, physical PPAs, project-anchoring VPPAs, and 24/7 hourly-matched carbon-free energy—and the binding “additionality gap” lies between the lowest and highest tier. Third, four actionable pathways are identified: carbon-aware scheduling, long-duration energy storage, AI-optimized grid management, and optimal data center siting under joint grid-carbon, renewable-resource, cooling-climate, and network-latency constraints. Fourth, institutional differences across the United States, China, the European Union, and the Republic of Korea make one-size-fits-all disclosure and matching mandates unlikely to be effective. The IEA's April 2025 Energy and AI report projects a Base Case global data center demand of approximately 945 TWh by 2030—comparable to Japan's annual electricity consumption—rising to approximately 1200 TWh by 2035. Renewable energy can meet AI-driven demand in principle, but only conditional on accelerated deployment, a procurement shift from annual to hourly matching, and jurisdiction-matched policy frameworks for transparent energy reporting.