How much water do data centers actually use?
Follow the source, the season, the cooling mode, and the point where heat finally leaves the system.
Primary sources ↓ · Limitations ↓
SHORT ANSWER
Start here.
There is no universal data-center water number. A facility can withdraw water and return part of it, consume water through evaporation, substitute reclaimed water for potable water, or use little direct cooling water while still carrying water use in the electricity that powers it. Annual averages can also hide hot-day demand and drought constraints. The useful local questions are the water source, monthly and peak-day withdrawal and consumption, what returns after treatment, how often wet cooling operates, and what the outdoor heat-rejection equipment actually does. “Liquid cooled” describes only part of that system.
- Bright publication
- Source review
- Reading question
- Which water crossed which boundary, when, and what heat-rejection system required it?
INTERACTIVE EXPLANATION
Follow the heat all the way outdoors.
Rack cooling and outdoor heat rejection are separate choices. Select both to see why “liquid cooled” does not settle the water question.
Liquid at chip or rack → Evaporative heat rejection
A contained liquid loop collects heat closer to the processors. It still must hand that heat to an outdoor rejection system.
A cooling tower rejects heat by evaporating water. Recirculation reduces repeated withdrawal, but make-up replaces evaporation, blowdown and small drift losses.
Tradeoff: Often energy-efficient in hot conditions; direct water consumption and summer demand remain material.
CLAIMS EXAMINED
Test the exact claim.
“One ChatGPT query uses a 500-milliliter bottle of water.”
The original paper estimated 500 milliliters for roughly 10–50 medium-length GPT-3-style responses, depending on location. It assumed 0.004 kilowatt-hours per request and combined onsite cooling with electricity-generation water. It did not meter today’s ChatGPT, a named facility, or every prompt.
What would change this assessment: Model-, hardware-, token-, site-, hour- and cooling-specific metered energy and water data, with the electricity source and accounting boundary disclosed.
Sources: Making AI Less “Thirsty” · The water use of data center workloads: A review and assessment of key determinants
“Liquid cooling or a closed loop means the facility consumes no water.”
Liquid can move heat from the chip or rack without being consumed, but that heat must still leave the site. The final equipment may be a wet tower, dry cooler, chiller, hybrid system, or heat-reuse network. A wet final stage can evaporate water even when the internal loop is closed.
What would change this assessment: A complete site design and operating record showing the rack loop, heat exchanger, final outdoor rejection, make-up water, blowdown, all-site water and electricity source.
Sources: DOE cooling water efficiency opportunities for federal data centers · 2024 United States Data Center Energy Usage Report
“A zero-water-for-cooling design means zero site or lifecycle water use.”
Microsoft says its post-August-2024 design avoids water evaporation for cooling and still uses water for kitchens and restrooms. That is narrower than zero total site, construction, electricity-associated, or lifecycle water. The announcement did not provide named operating-facility meter records.
What would change this assessment: Public meter data from named operating facilities covering cooling, total site water, seasonal peaks, maintenance losses and the electricity-associated boundary.
Sources: Microsoft’s next-generation data center cooling design
“Data centers are draining towns dry.”
Some projects can be material users in a constrained watershed; others use reclaimed water, dry rejection, or comparatively little direct cooling water. A national total or annual company average cannot establish local scarcity. Source, season, permit limits, drought rules, competing uses and cumulative projects determine the local risk.
What would change this assessment: Facility-specific monthly and peak-day withdrawal and consumption by source, drought-stage rules, watershed supply and cumulative demand from existing and approved projects.
Sources: Virginia JLARC: Data Centers in Virginia · 2024 United States Data Center Energy Usage Report
DOCUMENTED SOLUTIONS
What helps—and under which conditions.
OPERATING · WASHINGTON · SINCE JUNE 2021
Quincy’s city-owned reuse loop
A city utility treats Microsoft cooling-water blowdown and returns it to the campus. EPA estimates the system offsets 138 million gallons a year of potable-groundwater demand.
Conditions and limits: The system still uses canal and groundwater make-up water, principally in summer, consumes water through evaporation, and sends a concentrated residual stream to lined ponds. Potable demand offset is not total water saved.
Sources: EPA Quincy water-reuse case study
OPERATING RESULT · GOLDEN, COLORADO · DOE 2019 CASE
NREL’s thermosyphon hybrid cooling
NREL added a dry thermosyphon cooler to a warm-water liquid-cooled HPC system, using campus heat reuse first, dry rejection when weather permits, and the existing cooling tower at peak. DOE reports 2.10 million gallons avoided over two years and data-center water use cut in half.
Conditions and limits: This is a 10,000-square-foot national-laboratory HPC facility, not a hyperscale benchmark. The system retains full tower capacity for the hottest hours and reports reduced—not zero—direct water use.
Sources: Using Thermosyphon Hybrid Cooling System to Optimize Data Center Water Efficiency
COMPANY-REPORTED DESIGN · ANNOUNCED 2024
Microsoft’s zero-evaporation cooling design
Microsoft says a newer direct-to-chip, closed-loop design avoids ongoing water evaporation for cooling and began entering new designs in August 2024.
Conditions and limits: No named operating-facility meter series was published with the announcement. The claim excludes other site water, construction, maintenance and electricity-associated water and should not be applied to the existing fleet.
Sources: Microsoft’s next-generation data center cooling design
Withdrawal, consumption, reuse and potable demand avoided are different
Withdrawal is water taken from a surface or groundwater source. Consumption is the portion not returned to the immediate water environment because it evaporates, becomes incorporated or is otherwise removed. Reuse means water is treated and used again. Potable demand avoided compares against drinking-quality water that otherwise might have been supplied; it does not automatically equal total withdrawal or consumption.
Electricity-associated water sits outside the site boundary. It depends on where and how electricity is generated, which can change hourly. A dry-cooled site can lower direct water use while using more electricity; a wet system may use water onsite while reducing compressor energy. Report both boundaries instead of declaring one design universally best.
Sources: USGS water-use concepts and definitions · 2024 United States Data Center Energy Usage Report
A wet tower recirculates water—and still consumes it
In an evaporative system, warm water releases heat as a fraction evaporates into outdoor air. The remaining water recirculates, but make-up water must replace evaporation, mineral-rich blowdown and small drift losses. Treatment can increase cycles of concentration and reduce blowdown, yet evaporation remains the main heat-rejection mechanism.
Quincy shows why the complete balance matters. EPA reports both an estimated potable-groundwater offset and 260 million gallons a year of canal make-up water, principally in summer. When the canal became unavailable during the hot, dry 2021 season, the system shifted to groundwater. An annual positive result therefore coexists with a real peak and drought condition.
Sources: DOE cooling tower management · EPA Quincy water-reuse case study
Dry, hybrid and liquid systems solve different parts of the problem
Dry coolers use coils and fans to move heat to outdoor air with no routine cooling-water evaporation. Hybrid systems use dry rejection when conditions allow and wet or adiabatic assistance under defined hot conditions. Their annual water use depends on climate, loop temperatures, equipment size and control strategy—especially how often wet assist runs during the hottest hours.
Direct-to-chip, rear-door and immersion systems move heat efficiently from electronics into a liquid loop. That loop can feed a dry cooler, a cooling tower, a chiller or heat reuse. “Liquid cooled” is therefore not a complete answer to either direct water use or total energy use.
Sources: DOE cooling water efficiency opportunities for federal data centers · 2024 United States Data Center Energy Usage Report
Per-query estimates are scenarios, not universal measurements
The revised “Making AI Less Thirsty” paper is useful because it exposes a method: assumed energy for a medium GPT-3-style request, site PUE and WUE, and a location-specific electricity-water factor. Its table spans materially different results by location. The popular “one bottle per query” retelling removes the paper’s 10–50 response range and most of the assumptions.
Later LBNL research found workload water use can vary by more than four orders of magnitude across model, server, utilization, cooling, climate and grid conditions. The responsible conclusion is not a replacement universal number. It is a request for measured, time- and place-specific disclosure.
Sources: Making AI Less “Thirsty” · The water use of data center workloads: A review and assessment of key determinants
FOR A LOCAL PROJECT
What a resident should ask for.
- What is the water source—potable, reclaimed, canal or surface water, groundwater, or another non-potable supply?
- What are monthly, peak-day and drought-stage withdrawal and consumptive-use limits?
- Which equipment rejects heat outdoors, and how many hours does it run in wet, dry or hybrid mode?
- How much make-up water, evaporation, blowdown and return flow is metered, and where does the residual go?
- Does a “closed loop” claim cover only the rack loop, the full cooling system, total site water, or lifecycle water?
- Does the reported metric include electricity-associated water, and which grid mix and time period does it assume?
- Are operating records public at monthly and peak-season resolution, including drought contingencies?
What this guide does not establish
- WUE values are sensitive to climate, load, technology and reporting boundary.
- Annual totals can hide summer stress.
- Modeled national totals do not predict a local permit or watershed impact.
- Company “zero water” and fleet-average claims need named-facility meter records and explicit boundaries.
- “Water positive” claims require a separate accounting method and local hydrology.
Primary sources
USGS water-use concepts and definitions ↗
Government definitions for withdrawal and consumptive use.
2024 United States Data Center Energy Usage Report ↗
DOE-funded national modeling of direct and electricity-associated water, cooling configurations and reporting limits.
DOE cooling tower management ↗
Federal guidance for make-up water, evaporation, blowdown, drift and cycles of concentration.
DOE cooling water efficiency opportunities for federal data centers ↗
Functional diagrams and tradeoffs among cooling towers, direct liquid cooling and hybrid heat rejection.
EPA Quincy water-reuse case study ↗
Operating reuse system, potable-groundwater offset, canal make-up water, seasonal limits and financing terms.
Using Thermosyphon Hybrid Cooling System to Optimize Data Center Water Efficiency ↗
DOE operating case for a national-laboratory HPC facility, including two years of reported water reduction and design conditions.
Making AI Less “Thirsty” ↗
Academic estimation method and the original 500-milliliter-for-10–50-responses scenario, revised March 26, 2025.
The water use of data center workloads: A review and assessment of key determinants ↗
June 2025 peer-reviewed review of model, hardware, utilization, cooling, climate and grid determinants.
Microsoft’s next-generation data center cooling design ↗
Company-reported design claim for zero ongoing cooling-water evaporation, with a narrower boundary than total site or lifecycle water.
Virginia JLARC: Data Centers in Virginia ↗
Virginia-specific utility and water-resource findings, including locality and oversight limits.
Source-reviewed 2026-09-19. Source review means Bright compared this explanation with the linked records; it is not independent replication or expert peer review by Bright.