Global AI cooling + power
Evidence dashboard · Global view
How thirsty is AI,
really?
An interactive, source-led view of AI and data-center water use. Explore what is reported, what is estimated, and what is still unknown.
WithdrawalWater taken from a source
ConsumptionWater not returned locally
Indirect useWater used to make electricity
Global outlook
AI’s modeled annual water footprint
Trillion liters per year · Direct cooling + electricity generation
Modeled · medium confidence
≈ 2.1× 2025
At 2030 base, using ~4,000 km³/yr benchmark
Perspective: globally, agriculture accounts for about 70% of freshwater withdrawals. AI’s global share is much smaller, while its local impact can still be severe in water-stressed basins.
Scale comparison
AI beside the world's water users
Approximate annual withdrawals · Trillion liters · Logarithmic scale
Like-for-like withdrawal basis
Why a log scale? On a linear chart, AI's bar would be nearly invisible. Bar length shows order of magnitude; trust the printed values. FAO sector shares are approximate, and the AI figure is Li et al.'s modeled 2027 range. AI is a zoom-in within Industry, not a fourth additive sector.
Per-prompt reality
One prompt has no universal footprint
Published figures use different models, workloads and accounting boundaries
Do not rank unlike boundaries
Measured0.26 mL
Gemini Apps
Median text prompt
Google production measurement; direct cooling water only.
Primary source ↗Company claim0.32 mL
ChatGPT
Average query
OpenAI CEO figure; methodology and system boundary were not published.
Primary source ↗Audited LCA45 mL
Le Chat
400-token response
Mistral and Carbone 4 marginal lifecycle assessment; a broader boundary than direct cooling.
Primary source ↗Modeled10–50 mL
GPT-3-era estimate
One response
Derived from 500 mL for roughly 10–50 responses; cooling plus electricity and highly location-dependent.
Primary source ↗Task intensity
More compute usually means more water
Illustrative cross-study estimates apply a 4.97 L/kWh U.S. data-center water factor to published energy ranges. These are not direct measurements of any named model.
Short text0.2–0.5 Wh≈1–2.5 mL
One image0.06–2.9 Wh≈0.3–14 mL
Long reasoning>29–33 Wh≈144–164 mL
5-sec AI video≈0.94 kWh≈4.7 L
Facility lens
Specific data centers, disclosed
Million U.S. gallons per year · Logarithmic scale · Mixed metrics clearly labeled
Consumption ≠ withdrawal ≠ cap
Temple is intentionally shown as a contractual ceiling, not measured use. Facility totals cannot be cleanly ranked unless metric, boundary, year and water source match.
Company disclosures
What companies actually report
Latest public direct operational totals · Billions of liters per year
Not an AI-only ranking
The honest answer: No major AI lab currently publishes a complete, independently comparable “AI-only water use” total. Google and Meta provide strong fleet-level data; AWS reports withdrawal; Microsoft’s scope is broader; AI labs using shared clouds are especially hard to attribute.
Fact check
Four claims worth slowing down for
01
Misleading“Every AI prompt uses a bottle of water.”
The often-repeated estimate was roughly 500 mL for 10–50 GPT-3 responses, not one prompt. It varies sharply with model, location, weather, cooling and the electric grid.
02
False“All water taken by a data center is lost.”
Withdrawal is water taken from a source. Consumption is the portion not returned locally, often because it evaporates. Mixing the two can exaggerate or understate impact.
03
False“A global total tells us where the harm is.”
A liter used in a water-stressed basin is not equivalent to a liter in a water-abundant region. Location, season and water source are essential context.
04
Not yet“We can rank AI companies precisely.”
Most disclosures cover entire companies or data-center fleets, not AI workloads. Cloud customers share infrastructure, so attribution can double count the same water.
Methodology
Built to expose uncertainty
This tool separates three evidence levels so a forecast never masquerades as an audited figure.
- 1Reported
Direct company or government disclosure, with its original scope preserved.
- 2Modeled
Research estimate using electricity, cooling and grid-water intensity.
- 3Scenario
Transparent extrapolation used to explore plausible futures, not predict one certain outcome.
What could work
Plans to reduce water stress
Current initiatives and proposed interventions · Likelihood of material benefit by 2031
Analyst synthesis · not a measured probability
How to read the score
Our estimate that a materially scaled version of the initiative will reduce freshwater stress versus business-as-usual by 2031. Scores weigh technical maturity, deployment evidence, scalability, cost, local constraints and rebound risk. The range shows uncertainty.
Bottom line: The strongest near-term path is a portfolio—not one silver bullet: waterless or closed-loop cooling for new builds, reclaimed water where infrastructure allows, better efficiency, and enforceable basin-level disclosure. Even successful intensity reductions may not lower total industry water use if AI demand keeps growing faster.
The recursive question
How much water did this website take?
Estimated water consumed to research, generate, revise, test and deploy AI Water Reality
Estimated · not directly metered
Midpoint estimate: about 9 liters, or roughly 2.4 U.S. gallons.
1
Estimated compute≈ 0.6–4.0 kWh across AI inference, research calls, revisions, testing and deployment
2
Water intensity≈ 4.97 L/kWh, combining the cited U.S. average direct cooling and electricity-related water factor
3
Rounded result≈ 3–20 liters consumed, with a midpoint near 9 liters
Why the range is wide: the exact models, chips, utilization, token counts, data-center locations, cooling systems and grid mix were not metered for this project. This estimate covers the full multi-step build process, not a single prompt, and should be treated as an order-of-magnitude estimate.
Adversarial review
Where this report can be challenged
The strongest fair criticisms, how the dashboard responds, and what remains unresolved
Limitations are part of the finding
A credible report should make its own weak points easy to find.
These objections do not erase the evidence. They define how far the evidence can responsibly be taken. Open each challenge for our response and the remaining gap.
01The global forecast looks more precise than the evidenceMajor limitation
The criticism
The 4.5T-to-9.3T-liter trajectory depends on assumptions about AI electricity demand, cooling, grid water intensity and growth. Small changes compound quickly, and no global meter separates AI from ordinary data-center workloads.
How we address it
We call the curve a scenario, expose its growth rate, provide low and high paths, and avoid presenting it as an audited total. It is useful for scale and sensitivity—not as a point prediction.
What remains unresolved
Supporting evidence · Nature Sustainability — deep uncertainty in AI server projections ↗The baseline itself could be revised substantially as better workload-level and geographic data become available.
02Most company numbers are not AI-onlyMajor limitation
The criticism
Fleet totals can include search, storage, video, office operations and cloud customers. Shared infrastructure makes attribution to OpenAI, Anthropic or another lab vulnerable to double counting.
How we address it
The dashboard does not sum company totals or claim a precise AI-company ranking. Every figure preserves its reported scope, year and metric, and undisclosed AI-only values remain blank.
What remains unresolved
Supporting evidence · Uptime Institute — sustainability reporting remains incomplete ↗A defensible ranking requires standardized, workload-level disclosure that companies generally do not publish.
03Withdrawal, consumption, caps and WUE are not interchangeableComparison risk
The criticism
A facility that reports withdrawal cannot be cleanly ranked against one reporting consumption. A permit ceiling is not actual use, and WUE is an intensity ratio rather than an annual total.
How we address it
The interface labels each metric, separates withdrawal from consumption, flags the Temple figure as a contractual ceiling and warns against unlike rankings.
What remains unresolved
Supporting evidence · Uptime Institute — water figures vary by source, return and quality ↗Readers can still carry a number out of context. Uniform facility reporting would be stronger than interface warnings.
04Per-prompt numbers are apples and orangesComparison risk
The criticism
Published estimates use different models, token counts, hardware, locations and boundaries. Some count direct cooling only; others include electricity or lifecycle impacts. They also age quickly as models and chips change.
How we address it
We show the boundary and evidence type beside each number and explicitly say not to rank unlike estimates. The range demonstrates variability rather than declaring a universal prompt footprint.
What remains unresolved
Supporting evidence · Berkeley Lab — water use varies at the workload level ↗Without reproducible provider telemetry, no public comparison can fully normalize model quality, task length and infrastructure.
05Global liters can hide local harmMajor limitation
The criticism
A liter consumed during drought in a stressed basin is not equivalent to a liter in a water-abundant place. Annual global totals erase season, source, watershed condition and community utility capacity.
How we address it
We repeatedly distinguish global share from local impact and include facility-specific context instead of treating every liter as equally damaging.
What remains unresolved
Supporting evidence · Talukder et al. — stress-adjusted water accounting ↗The dashboard does not yet apply a basin- and season-adjusted stress score to every facility or company total.
06The comparison with agriculture can minimize AI’s impactFraming risk
The criticism
Comparing a new, concentrated industrial load with global agriculture can make AI look trivial even where a single project strains a local system. AI is also nested inside the broader industrial category, not an additive fourth sector.
How we address it
The chart states that AI is a zoom-in within Industry, uses a log scale with printed values and pairs the global comparison with a warning about local basin pressure.
What remains unresolved
Supporting evidence · Water and AI Feedback Loop — local burden varies by orders of magnitude ↗The most decision-useful comparison is often local: facility demand versus utility capacity, drought-stage supply and competing users.
07The footprint omits parts of the lifecycleLikely underestimate
The criticism
Operational cooling and electricity do not capture all water used to manufacture chips, build servers and data centers, transport equipment or retire hardware. Semiconductor fabrication depends heavily on ultrapure water.
How we address it
Our global forecast and most prompt figures are labeled according to their stated operational boundaries; we do not describe them as complete cradle-to-grave totals.
What remains unresolved
Supporting evidence · ThirstyFLOPS — embodied and operational water are distinct ↗The report likely understates total lifecycle water because public, model-specific embodied-water data remain sparse.
08Company disclosures are interested-party evidenceEvidence risk
The criticism
Corporate reports may be selective, use favorable boundaries or emphasize replenishment without showing whether benefits occur in the same place and season as withdrawals.
How we address it
We link directly to disclosures, retain their scope and add government or academic evidence where available. A company claim is labeled as such rather than upgraded to independent measurement.
What remains unresolved
Supporting evidence · Uptime Institute — gaps persist in operator reporting ↗Independent facility-level audits and standardized public reporting would materially improve confidence.
09Our solution scores and site estimate are judgment callsAnalyst judgment
The criticism
The initiative percentages are not statistically calibrated probabilities, and the 3–20 L website estimate was not metered. Both could create false confidence if quoted without their caveats.
How we address it
The interface labels the scores as analyst synthesis, shows uncertainty ranges and publishes the website estimate’s arithmetic and assumptions.
What remains unresolved
Supporting evidence · Berkeley Lab — regional water intensity varies materially ↗These should be read as structured judgment and order-of-magnitude estimation, not empirical measurements.
Bottom line: The dashboard is strongest as a map of reported evidence, incompatible boundaries and plausible scale. It is weakest when any modeled number is treated as a precise global measurement, a company ranking or a prediction. New disclosure should change the dashboard—not be forced to fit it.
Frequently asked questions
Clear answers about AI and water
Concise answers grounded in the evidence and limitations shown above
How much water does AI use per prompt?
There is no universal number. Published text-prompt figures range from about 0.26 mL for Google's measured median Gemini Apps prompt to 45 mL for Mistral's broader lifecycle estimate of a 400-token Le Chat response. Model size, response length, cooling, weather, location and the electricity mix all matter.
Does one ChatGPT prompt use a bottle of water?
No. The bottle claim usually misstates a study that estimated roughly 500 mL for about 10 to 50 GPT-3-era responses. OpenAI's CEO later stated an average ChatGPT query uses about 0.32 mL, but the company did not publish the full methodology or accounting boundary for that figure.
Why do AI data centers use water?
Most direct water use removes heat from servers, often through evaporative cooling. Data centers also have an indirect water footprint because many power plants use water to generate electricity. Direct consumption, withdrawal and electricity-related water are different measurements.
How does AI water use compare with agriculture?
Agriculture accounts for roughly 70% of global freshwater withdrawals, while the modeled 2027 AI midpoint is about 0.14%. AI is a small global share, but a single facility can still create meaningful pressure in a dry or water-stressed local basin.
Which AI company uses the most water?
A defensible AI-only ranking is not currently possible. Google, Microsoft, Meta and AWS disclose different metrics, scopes and reporting years, while OpenAI, Anthropic and xAI do not publish comparable company-wide AI-only totals. Shared cloud infrastructure also creates double-counting risk.
What is the difference between water withdrawal and water consumption?
Withdrawal is all water taken from a source. Consumption is the portion not returned to the same local water system, commonly because it evaporates. A facility can withdraw a large volume while returning some of it, so the terms should never be used interchangeably.
Can a data center operate without consuming water for cooling?
Yes, some closed-loop liquid-cooling and dry heat-rejection designs can nearly eliminate routine evaporative cooling water. The tradeoff is that dry cooling can require more energy or backup systems during very hot weather, and retrofitting older facilities can be difficult.
How much water will Meta's Temple, Texas data center use?
Actual annual use has not yet been published. Temple's agreement caps domestic water use at 4,000 gallons per day, or about 1.46 million gallons per year. Meta says its closed-loop design should use zero cooling water for most of the year, so the cap is not a forecast of actual consumption.
How much water could AI use over the next five years?
This dashboard's base scenario rises from 4.5 trillion liters per year in 2025 to 9.3 trillion liters in 2030, following the cited UN University trajectory. The interactive low and high scenarios are transparent extrapolations, not meter readings or certain predictions.
What can reduce data-center water use?
The strongest approach combines closed-loop or dry cooling, reclaimed water, efficient chips and models, water-aware siting, workload shifting, transparent reporting and enforceable local conservation plans. No single intervention solves every geography or reduces both direct and indirect water at once.
How much water did it take to create AI Water Reality?
The site was not directly metered. A transparent estimate is about 3 to 20 liters of water consumed, with a midpoint near 9 liters, for AI inference plus coding, testing and deployment compute. The wide range reflects unknown model energy, hardware, cooling systems, electricity sources and the number of internal reasoning steps.
Sources
Primary sources & key references
Company disclosures, government records and research used throughout this dashboard
2026 · Global forecastUN University — Environmental Cost of AI2024 · U.S. data centersLawrence Berkeley National Laboratory2023/2025 · AI methodologyLi et al. — Making AI Less “Thirsty”2025 · Company disclosureGoogle Environmental Report2025 · Company disclosureMicrosoft Environmental Data Fact Sheet2025 · Company disclosureMeta Environmental Data Index2026 · Company disclosureAWS Water StewardshipGlobal comparisonFAO — Global Water Withdrawals2025 · Prompt measurementGoogle — Gemini inference footprint2025 · Lifecycle assessmentMistral AI — Le Chat footprint2026 · Facility disclosureCity of Temple — Meta data center2023 · Task energyLuccioni et al. — AI task energyCooling technologyBerkeley Lab — Liquid CoolingRecycled waterAWS — Recycled-Water Data CentersDisclosure policyEuropean Commission — Data Center Performance2026 · Water-aware computingTalukder et al. — Balancing Bits and Drops2026 · Rainwater captureMicrosoft — Community-First AI Infrastructure2025 · Reporting gapsUptime Institute — Global Data Center Survey2025 · Workload variabilityBerkeley Lab — Water Use of Data Center Workloads2025 · Projection uncertaintyNature Sustainability — U.S. AI Server Impacts2025 · Lifecycle accountingThirstyFLOPS — Embodied and Operational Water
Conversions: 1 U.S. gallon = 3.78541 liters; 1 megaliter = 1 million liters. Forecast paths and the site-creation footprint are calculated from stated assumptions. Corporate totals are not summed because their scopes overlap.