Artificial Intelligence
The Carbon Footprint of AI: Energy per Query, Data Center Costs, and Comparisons
A single query to ChatGPT uses 0.5–2 kWh of energy, equivalent to running a 100-watt light bulb for 2–8 hours. Training a large language model like GPT-4 consumes 1,200–1,900 MWh of electricity, roughly the annual energy consumption of 300 US households. The data centers powering AI consume 4–6% of global electricity as of 2026, consuming vast amounts of water for cooling. AI is not uniquely energy-intensive compared to data centers generally, but the rapid scaling of AI training is pushing energy demand upward at 15–20% annually. This guide covers the real energy costs of AI (training vs. inference), per-query estimates, data center impacts, how AI compares to other activities, and what is being done to improve efficiency.