Definition
The electricity, water, and carbon that data centers use to train and run AI models, weighed against the efficiency gains AI can unlock elsewhere.
At a glance
- One query is tiny: a typical Gemini prompt uses ~0.24 watt-hours, like a microwave running for one second[3]. The concern is scale, not your chat.
- Data centers are the real footprint: their power use jumped ~17% in 2025, and AI-focused centers grew ~50%[1].
- The IEA projects data-center power to more than double by 2030 to ~945 TWh (near Japan’s total demand), with emissions near 1% of global CO2[2].
- Water counts too: U.S. data centers used ~66 billion liters in 2023, triple their 2014 level[4].
Where the impact comes from
The footprint is in physical data centers, not the app on your screen. They draw power to train models (a huge one-time cost) and to answer everyday requests (which adds up across billions of users). They also use water to cool servers and at the power plants feeding them[4]. Because much of that power is still gas and coal, the result is carbon.
What it means for you
Two trends partly offset the growth: per-query efficiency is improving fast (~33x in a year for Gemini)[3], and AI can cut emissions elsewhere, such as optimizing power grids and renewables[5]. But the rebound effect — cheaper AI simply used far more — can erase those savings. Your direct footprint is modest; the real lever is choosing vendors who run on clean power and publish their numbers.
Bottom line
Your own AI use barely registers, but data-center electricity, water, and carbon are climbing fast, so favor vendors who run on clean power and disclose their footprint.