The electricity behind artificial intelligence is both concentrated and diffuse: concentrated in the high-performance machines that train large models, and diffuse in the millions of devices and servers that run those models for everyday tasks. That duality helps explain why simple headline numbers about “A.I. energy use” rarely tell the full story.

The global context is modest but rising. The International Energy Agency and academic studies have put data centers — the broader infrastructure that hosts A.I. workloads — at roughly 1 percent of global electricity demand in recent years. Within that slice, A.I. work represents a growing but uncertain fraction because it can be buried inside broader cloud, storage and networking operations. “The big picture is that data centers are not the dominant consumer of electricity, but A.I. workloads are spiking demand for the most energy-intensive parts of the cloud,” said Jonathan Koomey, an energy analyst who has tracked computing’s power trends.

Energy consumption breaks into two distinct phases: training and inference. Training a state-of-the-art language model requires thousands of high-end accelerators — GPUs or TPUs — running for days or weeks. Estimates vary widely depending on model size and hardware. Studies and industry disclosures indicate training can consume anywhere from tens to thousands of megawatt-hours, with accompanying carbon emissions that depend heavily on where and how the electricity is generated. Emma Strubell, a researcher who has studied model carbon footprints, cautioned that early headline numbers exaggerated some worst-case scenarios, but still underscored a key point: “Large-scale training is carbon-intensive if powered by fossil-rich grids.”

By contrast, inference — the day-to-day computation that answers users’ questions — can, over time, consume more energy than training because models are queried millions or billions of times. For widely deployed systems, inference-related electricity and the cooling that supports it often dominate the lifecycle footprint. That has pushed companies to invest in model compression, distillation and more efficient chips to cut per-query energy costs.

Hardware and data-center design have improved efficiency markedly. Hyperscale facilities today operate with power usage effectiveness ratios typically near 1.1 to 1.2, meaning relatively little overhead beyond IT load. New accelerators also deliver more compute per watt than previous generations. Cloud providers report that software optimizations, model sparsity techniques and shifting workloads to low-carbon grids have reduced relative emissions per unit of compute.

Still, the societal stakes are not only climate-related. The geographic concentration of A.I. infrastructure raises water, land-use and equity concerns, and rapidly expanding demand can strain local grids. Firms are responding with renewable power purchases, on-site solar and long-term clean-energy contracts, while some regulators examine disclosure rules for A.I. energy and emissions.

“Transparency is the next frontier,” said Karen Weise, a technology reporter who has covered data centers. “Without granular reporting on compute, location and grid mix, we can’t meaningfully compare claims of efficiency or sustainability.” As A.I. becomes embedded in health care, transportation and government services, the choices companies make about where and how to run models will shape not just climate outcomes, but the resilience and fairness of critical infrastructure.