Scenario-based forecasting of the global energy demand and carbon footprint of artificial intelligence

Abstract

Artificial intelligence (AI) is advancing rapidly and is emerging as a significant driver of global electricity consumption, yet its long-term energy and emissions implications remain poorly quantified. This study develops a scenario-based, simulation-driven modeling framework that links mathematical representations of AI computational demand with life-cycle carbon accounting for global AI-related energy use and emissions through 2050. We evaluate alternative development pathways that differ in model scale, deployment structure, and electricity mix assumptions. Across all scenarios, improvements in hardware and algorithmic efficiency substantially reduce energy use per operation

however, aggregate AI electricity demand still increases by roughly an order of magnitude due to rapid growth in training and inference workloads. Under the continuation of current trends, AI electricity consumption could reach up to 30% of global demand by 2050, corresponding to more than 8 gigatons of annual CO2-equivalent emissions. Even under optimistic efficiency trajectories, total AI-related electricity demand remains more than six times higher than 2024 levels. In contrast, scenarios that combine consolidation toward fewer, larger models with transitions to low-carbon electricity sources reduce total emissions by up to 40% relative to business-as-usual pathways, exceeding the reductions achievable through efficiency gains alone by more than 20 percentage points. These results highlight widening regional disparities and indicate that policy choices affecting AI deployment patterns and electricity system decarbonization play a central role in shaping the carbon intensity of computation.