Who Pays for AI?
All resource ledgers

Local resources · resource ledger

How much water do AI data centers use?

Contested>10,000×observed workload-level water-use variation

Measured workload-level water use varies by more than 10,000-fold (LBNL, 2025), because cooling design, climate, utilization and the local power source dominate the result. Operator averages for on-site cooling are 0.26 mL per Gemini prompt (Google, 2025) and 0.32 mL per ChatGPT query (OpenAI, 2025). A top-down estimate puts AI’s total 2025 water footprint at 312.5 to 764.6 billion litres (de Vries-Gao, 2026). Grade: Contested.

ContestedChecked 2026-07-17 · next review 2026-11-06 · reviewed by David Veksler

Key numbers
ValueMeasuresSourceGrade
>10,000×variation in measured workload-level data-center water use, per LBNLLawrence Berkeley National Laboratory (2025) · U.S. Government Accountability Office (2025)Background
0.26 mLwater for a median Gemini text prompt, comprehensive methodGoogle Cloud (2025)Background
0.32 mLwater for an average ChatGPT query, per OpenAISam Altman (2025)Background
312.5 to 764.6 billion LAI’s total 2025 water footprint, top-down estimate incl. power-plant waterAlex de Vries-Gao (Patterns) (2026)Background

How does AI demand affect this market?

Water can be consumed directly by cooling and indirectly by electricity generation. Climate, cooling design, utilization and grid mix dominate the result.

Who pays, and who benefits?

In water-stressed communities, new demand leaves less water available for other users. Siting, cooling systems and reclaimed water determine the local effect.

What limits supply?

The main limits are local water availability and the lack of transparent reporting at the facility level.

What the evidence cannot show

A global average does not measure a particular project’s pressure on its watershed. Three kinds of number answer different questions: LBNL’s workload-level range, operator fleet averages for on-site cooling (Google, OpenAI), and top-down annual totals such as de Vries-Gao’s 312.5 to 764.6 billion litres for AI in 2025, which count cooling plus power-plant water and are not per-prompt or per-facility figures.

What would change the grade

Individual campuses can be graded when their permits, water sources, consumptive use and seasonal operations are public.

Sources

Public data, agency work and company reports

  1. The water use of data center workloadsLawrence Berkeley National LaboratoryPublished 2025-06 · checked here 2026-07-17 ↗
  2. Generative AI’s Environmental and Human EffectsU.S. Government Accountability OfficePublished 2025-04-22 · checked here 2026-07-17 ↗
  3. The carbon and water footprints of data centers and what this could mean for artificial intelligenceAlex de Vries-Gao (Patterns)Published 2026-01 · checked here 2026-07-17 ↗
  4. United States Data Center Energy Usage Report: 2025 UpdateLawrence Berkeley National LaboratoryPublished 2026-06 · checked here 2026-07-17 ↗

How much water does one AI prompt use?

It depends on the model, workload, facility and accounting boundary. Google reports 0.26 mL of on-site cooling water for a median Gemini text prompt, using a fleet-wide average water-use efficiency (2025); OpenAI reports about 0.32 mL per ChatGPT query without publishing its method (2025). These are operator-level averages that compress site and workload variation, so their similarity neither contradicts nor illustrates LBNL’s range. The 2023 bottle estimate modeled one GPT-3-era scenario: about 500 mL per 10 to 50 responses.

Why does cooling design matter for data-center water use?

LBNL’s review found that cooling design, local climate, server utilization and the electricity source used to generate power all shape how much water a workload consumes, producing more than 10,000-fold variation across measured workloads. A per-prompt figure can hold for a clearly specified model, workload and boundary; none applies everywhere.

Related claim report cards

Background reading

Companion references from the editor

These reference pages on cheatsheets.davidveksler.com are maintained by the same editor. They provide background context; they are not graded sources for this ledger.