Generative AI's Environmental Impact

Generative AI’s rapid growth brings environmental costs that extend beyond the electricity used to train large models. Training clusters can consume seven or eight times as much energy as typical computing workloads, while data centers’ global electricity use reached 460 terawatt-hours in 2022 and is projected to approach 1,050 terawatt-hours by 2026. One estimate puts GPT-3’s training at 1,287 megawatt-hours and 552 tons of carbon dioxide. Expanding data-center demand also strains power grids and can lock in fossil-fuel generation. The impacts continue after training: every user prompt requires energy, and a ChatGPT query has been estimated to use about five times the electricity of a standard web search. As AI tools spread and models grow more complex, inference - the act of using a trained model - is expected to account for an increasing share of energy demand.

Data centers also consume water to cool computing equipment, with an estimate of two liters per kilowatt-hour of electricity used. Their hardware has further costs: manufacturing powerful GPUs requires energy and materials, while mining, chemical processing, and transportation add pollution and emissions. Frequent releases of larger models can make earlier training investment obsolete and drive repeated resource use. The central argument is that AI’s environmental footprint must be assessed across its full lifecycle and broader effects on energy systems, water supplies, ecosystems, and materials - not just the power used by a computer. Responsible development requires weighing those costs against AI’s claimed benefits, such as productivity gains and scientific advances.