The Real Cost of AI: Energy, Water, and Infrastructure in 2025

Training a large language model like GPT-4 reportedly consumed approximately 50 gigawatt-hours of electricity — equivalent to the annual energy use of roughly 4,500 U.S. homes. As AI inference at scale becomes ubiquitous, cumulative energy demand is projected to grow dramatically. The International Energy Agency estimates that data centers could account for 4% of global electricity consumption by 2026, up from 1.5% today.

Water use is equally striking: Microsoft disclosed that its global data centers consumed 6.4 million cubic meters of water in 2022 for cooling, and AI-intensive workloads are intensifying this demand. The geographic concentration of data centers in water-stressed regions — Arizona, Nevada, the Netherlands — is prompting local regulatory scrutiny.

The response from hyperscalers involves nuclear power (Microsoft's Three Mile Island deal), direct liquid cooling, and AI-optimized chip architectures (Google's TPU v5, AWS Trainium 2) that deliver more compute per watt. But the fundamental tension between AI's growth trajectory and resource sustainability remains one of the defining challenges of the decade.