Published October 22, 2026 | Version 2026.10.22.2

Data Center Water Benchmarks

Authors/Creators

  • 1. GreenCalculus

Description

86 publicly disclosed data center and AI water figures, each traced to its publisher: Google's 2025 water withdrawal, discharge and consumption for 40 named data center locations, other sites and the fleet total (10,523 million US gallons consumed); water usage effectiveness (WUE) for Google (1.15 L/kWh, 2023 and 2024), Amazon Web Services (global 0.12 L/kWh and 17 cloud regions, 2025), Microsoft (global and three regions, FY24 and FY25), Equinix (0.91) and Digital Realty (0.59) portfolios in 2025, Meta (0.19 L/kWh in 2024, and 2020 to 2023; 2,974 megalitres consumed by its data centers in 2024), and the European Commission's 2024 average of 0.58 L/kWh for EU reporting data centres that use water; and water per AI prompt and per model: Google Gemini (0.26 mL per median text prompt, May 2025), OpenAI ChatGPT (0.32 mL per average query, 2025), Microsoft (0.0 to 0.067 mL per typical query, modelled by Microsoft researchers, 2026), Mistral AI (45 mL per 400-token Le Chat response, life cycle; 281 million litres for Mistral Large 2 training plus 18 months of use) and Li et al. (16.904 mL per GPT-3 response and 5.439 million litres for GPT-3 training, US average estimate). Values are copied exactly as published, with source URL, period, scope and retrieval date (11, 12 and 22 October 2026).

Access. Interactive versions with calculators: data center water use at greencalculus.com/data/data-center-water-usage/ and water per AI prompt at greencalculus.com/data/ai-water-usage/. Source repository and build script: github.com/greencalculus/data-center-water-benchmarks. This deposit is a frozen edition; cite it for reproducibility, or cite the concept DOI for the latest version.

Read before comparing. WUE definitions differ slightly between publishers (water withdrawn, water for cooling and humidification, or total water input), and the EU average excludes dry-cooled sites. The AI per-prompt figures count different things (column water_scope): Google's is on-site cooling water only, Li et al. add the water used to generate the electricity, Mistral's is a full life cycle including server manufacturing, and OpenAI has not stated its scope. Correction in this version: earlier versions said Meta publishes no data center water figures, which was wrong; Meta's 2025 Environmental Data Index gives annual WUE and water withdrawal and consumption, now included. Companion dataset: Data Centre PUE Benchmarks, doi:10.5281/zenodo.23280657.

Get started in one minute: emission factors and calculations with full provenance from the GreenCalculus API.

v2026.10.22.2: rows 87 to 104 add water on record for AI and hyperscale campuses (xAI Colossus utility bills, Meta Hyperion projection and state registration, Project Blue, OpenAI and Oracle Stargate Abilene, Microsoft Fairwater), each with a figure_type (bills, projection, permit, statement or model), plus US totals and projections from Lawrence Berkeley National Laboratory (2024) and a Texas estimate from HARC.

Files

data-center-water-benchmarks-2026.10.22.2.zip

Files (165.6 kB)

Name Size
md5:e6ab7249e00835647195b1f5c66bb546
39.8 kB Preview Download
md5:6bffaf2e353a6c31efe0a58fc0b8232c
45.0 kB Preview Download
md5:fde4f4f6a0e05358b95ea272afc6b448
74.8 kB Preview Download
md5:932d12eea9ceced9e8f9add75d1b3e55
6.1 kB Preview Download

Additional details