An energy-first method for estimating the water and energy footprint of AI use (methodology 2026-10)

Abstract

Published estimates of the water used by a single AI chatbot prompt differ by roughly a hundredfold, from about 0.3 mL to 50 mL. Most of that spread comes from four choices that are rarely stated together: when the measurement was made, where the accounting boundary is drawn, what unit of work is measured, and whose data centre is assumed. This note sets out the method behind a public calculator and an open-source server that estimate the water and energy of AI use for individuals, coding and agent sessions, exported chat histories and organisations. The method is energy-first. Energy per prompt comes from 2025 vendor disclosures (0.34 Wh for a typical text prompt), and water is derived from energy under one of two stated boundaries: onsite cooling only (1.0 mL per Wh) or comprehensive, which adds water consumed in generating the electricity (4.1 mL per Wh). A token-based extension, calibrated to the same 0.34 Wh prompt, covers sessions in which the model re-reads a growing context. Every result is reported as a central estimate with a likely range of 0.3× to 3×, a judgement-based band that summarises the spread in published energy figures and in provider and grid water intensities. A typical text prompt comes to about 0.34 mL onsite and 1.4 mL comprehensive (range 0.42–4.2 mL). The note gives every constant, worked examples, a reconciliation with widely quoted figures, and the limitations. Results carry the methodology version, and changes are logged publicly.