# ThirstTrap.ai

> Evidence-first intelligence for the water, power, permits, and local infrastructure behind the AI boom. Every number carries its source, date, unit, boundary, and evidence class.

Research cutoff: 2026-07-15. Latest verification: 2026-08-28. Every figure below states its source, date, unit, boundary, and evidence class. Reported means a first-party or government record. Modeled means calculated from explicit assumptions. Do not present modeled values as measurements, do not derive per-model water use from provider fleet totals, and do not add figures with different boundaries.

## Citation guidance

- Cite as: ThirstTrap.ai, "<question or figure>", <permalink>, verified <date>. Each answer below has a permalink.
- There is no universal water-per-prompt figure. Quote the specific source, service, date, and boundary.
- Withdrawal, consumption, and discharge are different metrics. Consumption = withdrawal minus discharge.
- WUE is a ratio (liters per kWh), not a total.
- Provider fleet totals (Google, Amazon, Meta, Microsoft) mix every workload and cannot be assigned to ChatGPT, Claude, or Gemini.
- Evidence classes and the corrections policy are at https://thirsttrap.ai/methods.

## Headline figures

- 66B L: direct site consumption. Modeled 2023 · all U.S. data centers · not AI alone. Source: LBNL, https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf
- ~800B L: electricity-related consumption. Modeled 2023 · all U.S. data centers · not AI alone · separate boundary. Source: LBNL, https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf
- 521–843 TWh: 2030 U.S. electricity range. Scenario · all U.S. data centers · not a water forecast. Source: LBNL 2025 update, https://eta-publications.lbl.gov/publications/united-states-data-center-energy-2025

## Company disclosures (not comparable with each other)

- Google: 10.523B gal, 2025 global data-center consumption. Gap: Owned and operated fleet, withdrawal minus discharge, limited assurance. Not AI-specific and not a site-level peak view. Source: https://sustainability.google/files/google-2026-environmental-report.pdf
- Amazon / AWS: 9.4B L, 2025 global data-center withdrawal. Gap: Leased, owned, and shared data centers; withdrawal is not comparable with consumption. Source: https://sustainability.aboutamazon.com/2025-amazon-sustainability-report.pdf
- Meta: 2.974B L, 2024 global data-center consumption. Gap: Fleet-wide and not generative-AI-specific; construction is separate. Source: https://sustainability.atmeta.com/wp-content/uploads/2025/10/Meta_2025-Environmental-Data-Index.pdf

## Per-interaction scenarios (different boundaries, not a ranking)

- Google · Gemini Apps: 0.26 mL per prompt, 0.24 Wh. Measured May 2025 · first-party. Google’s point-in-time median included active and idle compute, CPU and RAM, and facility overhead. The water figure is on-site cooling water only, not water used to generate the electricity. It was not independently verified and should not be read as every prompt or a full lifecycle footprint. Source: https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference/
- Sam Altman · ChatGPT: 0.32 mL per query, 0.34 Wh. 2025 public average claim. A public claim by OpenAI’s CEO, not an independently assured measurement. The post did not publish the system boundary, sample distribution, or methodology needed to reproduce it. Source: https://blog.samaltman.com/the-gentle-singularity
- Li et al. · GPT-3-based scenario: 10 to 50 mL per response. 2023 modeled scenario · preprint. A 2023 modeled operational scenario covering direct and electricity-related water, so it is a wider boundary than Google’s on-site figure, not a meter reading from today’s services. It predates major changes in models, hardware, serving, and cooling. Source: https://arxiv.org/abs/2304.03271

## Recent signals

- 24 JUN 2026: Microsoft reports fleet WUE of 0.27 L/kWh for FY2025. Down from 2.3 L/kWh in Microsoft's telling. Its 2024 AI-optimized design uses zero water for cooling during operations, a material reduction within that boundary. Electricity, construction, and chip manufacturing remain outside it. Evidence: Reported · primary. Source: https://blogs.microsoft.com/blog/2026/06/24/inside-microsofts-two-decade-push-to-cut-water-intensity-while-scaling-for-growth/
- JUN 2026: Berkeley Lab models a 649 TWh reference case for 2030. Published 18 Jun 2026. The reference case is 11.8% of forecast 2030 U.S. electricity; the compounded range is 521–843 TWh, about 9.5%–15.3%. This is an electricity scenario, not a direct water projection. Evidence: Modeled. Source: https://eta-publications.lbl.gov/publications/united-states-data-center-energy-2025
- 08 JAN 2026: Data Center Transparency Act introduced. H.R. 6984 proposed recurring reporting on water, reuse, utilities, pollution, and electricity. As of the research cutoff, it was introduced, not enacted. Evidence: Reported · primary. Source: https://www.congress.gov/bill/119th-congress/house-bill/6984/text/ih

## Straight answers

### How much water does one ChatGPT prompt use?

There is no single verified number. OpenAI's CEO claimed about 0.32 mL per average query in June 2025, but published no method. The only measured first-party figure is Google's 0.26 mL median for a Gemini Apps text prompt (May 2025), which counts on-site cooling water only.

Published figures range from 0.26 mL (Google, measured, on-site water only) to 10 to 50 mL per response (Li et al., 2023 modeled GPT-3 scenario that also counts water used to generate electricity). They cover different services, dates, and boundaries, so they are not comparable and cannot be used to rank AI models.

Boundary: Water consumption per prompt, query, or response; Three different services and system boundaries; 2023 to 2025, by source.
Does not say: That every AI prompt uses a bottle of water. That ChatGPT uses more or less water than Gemini or Claude. That the 0.32 mL claim has been independently measured.
Evidence: Reported · primary, Modeled. Source: Google inference study; Altman, The Gentle Singularity; Li et al. 2023, https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference/. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/water-per-chatgpt-prompt

### Does every AI prompt use a bottle of water?

No. The claim traces to a 2023 modeled estimate that GPT-3 consumed about one 500 mL bottle across roughly 10 to 50 medium-length responses (the first April 2023 draft said 20 to 50 questions and answers). That is per dozens of responses, modeled, and not a current meter reading.

The Li et al. preprint 'Making AI Less Thirsty' modeled on-site and electricity-related water for GPT-3 in 2023. Later measured figures from Google (0.26 mL per median prompt, May 2025) are roughly 40 to 200 times smaller, partly because of a narrower boundary and partly because hardware, serving, and cooling changed.

Boundary: Modeled water consumption, direct plus electricity-related; GPT-3 served from Microsoft U.S. data centers, as modeled; 2023 scenario.
Does not say: That the study was wrong. It was a modeled scenario with a stated boundary. That today's services use zero water per prompt. That the bottle figure applies to any model other than the one modeled.
Evidence: Modeled. Source: Li et al., arXiv 2304.03271, https://arxiv.org/abs/2304.03271. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/bottle-of-water-per-prompt

### How much water does Claude use?

No public figure exists. As of the July 15, 2026 research cutoff, Anthropic had published no standalone water disclosure for Claude, and the providers that host it (AWS as primary, plus Google Cloud) report fleet-wide totals that mix Claude with every other workload. Provider totals cannot be assigned to one model.

Anthropic describes AWS as its primary training and cloud provider and has announced large additional Google Cloud TPU capacity, so Claude runs across several companies' facilities. GAO reached the same conclusion for generative AI generally: company details are sparse and separating AI from other data-center workloads is not currently possible with public data.

Boundary: Any water metric attributable to Claude specifically; Anthropic disclosures plus AWS and Google fleet reports; Located by July 15, 2026.
Does not say: That Claude uses no water. Every served request has a water footprint somewhere. That Claude uses more or less water than ChatGPT. That Amazon's 9.4 billion liters or Google's 10.523 billion gallons belong to Claude.
Evidence: Analysis, Reported · primary. Source: GAO-25-107172; Anthropic and Amazon compute announcement, https://www.gao.gov/products/gao-25-107172. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/how-much-water-does-claude-use

### How much water do U.S. data centers use?

Berkeley Lab modeled about 66 billion liters of direct on-site water consumption by all U.S. data centers in 2023, plus nearly 800 billion liters consumed indirectly through electricity generation. Both figures cover every data center, not AI alone.

The 2024 LBNL report estimates hyperscale and colocation facilities account for 84 percent of the direct total. Indirect water uses a national average of 4.52 liters per kWh across 176 TWh of 2023 data center electricity. For scale, U.S. data centers used about 4.4 percent of national electricity in 2023.

Boundary: Modeled consumption, direct (site) and indirect (electricity) shown separately; All U.S. data centers, every workload; Calendar 2023.
Does not say: That AI alone consumed 66 billion liters. That the two figures should be added without noting they are different boundaries. That any single site or state matches the national average.
Evidence: Modeled. Source: LBNL 2024 United States Data Center Energy Usage Report, https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/us-data-center-water-use

### How much water do data centers use worldwide?

The IEA models roughly 560 billion liters of annual water consumption by data centers today, rising to about 1.2 trillion liters by 2030. In its 2023 breakdown, about two-thirds came from electricity supply, about one-quarter from on-site cooling, and the remainder from chip manufacturing.

This is a global model with its own system boundary and assumptions, so it is not directly comparable with Berkeley Lab's U.S. figures. The IEA also notes that historical data-center electricity is itself estimated rather than directly measured, which carries into the water estimate.

Boundary: Modeled consumption, electricity supply plus cooling plus chips; Global data centers, every workload; 2023 baseline and 2030 scenario.
Does not say: That the 2030 figure is a forecast of what will happen; it is a scenario. That this can be added to or compared directly with the LBNL U.S. numbers. That AI is the whole footprint.
Evidence: Modeled. Source: IEA, Energy and AI (April 2025), https://www.iea.org/reports/energy-and-ai. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/global-data-center-water-use

### What is the difference between water withdrawal and water consumption?

Withdrawal is the total water taken from a source. Consumption is the part that does not return, usually because it evaporated in cooling. Discharge is what goes back. Consumption equals withdrawal minus discharge. Mixing the two makes companies look better or worse than the data supports.

Amazon reports data center withdrawal (9.4 billion liters in 2025). Google and Meta report consumption (10.523 billion gallons in 2025 and 2.974 billion liters in 2024). A withdrawal number cannot be ranked against a consumption number.

Boundary: Definitions (USGS) applied to company reporting; Any facility or fleet; Definitions are stable; examples are 2024 and 2025.
Does not say: That withdrawal is harmless because most of it returns; timing, temperature, and quality of the return flow matter. That consumption means the water is destroyed. It re-enters the hydrologic cycle, but not locally or on time.
Evidence: Definition, Reported · primary. Source: USGS water-use definitions; Google 2026 Environmental Report data table, https://www.usgs.gov/mission-areas/water-resources/science/water-use-united-states. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/withdrawal-vs-consumption

### What is WUE (water usage effectiveness) and why is it not the whole story?

WUE is liters of site water used per kilowatt-hour of IT energy. It is an efficiency ratio, not a total. A large campus with a low WUE can still consume more water than a small campus with a high one, and the numerator may mean withdrawal or consumption depending on the company.

Recent fleet WUE disclosures: Microsoft 0.27 L/kWh (FY2025, down from 2.3), Amazon 0.12 L/kWh (2025), Meta 0.19 L/kWh (2024). These use different boundaries and years and are not a ranking.

Boundary: Liters of site water per kWh of IT energy (ratio); Fleet averages as reported by each company; FY2024 and FY2025, by company.
Does not say: That the company with the lowest WUE uses the least water. That WUE includes water used to generate electricity. It does not unless a source WUE is added.
Evidence: Reported · primary. Source: Microsoft, June 24 2026, https://blogs.microsoft.com/blog/2026/06/24/inside-microsofts-two-decade-push-to-cut-water-intensity-while-scaling-for-growth/. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/what-is-wue

### Which AI company uses the most water?

Nobody can say honestly from public disclosures. Google reported 10.523 billion gallons of data center water consumption in 2025, Amazon 9.4 billion liters of withdrawal in 2025, and Meta 2.974 billion liters of consumption in 2024. Different metrics, years, and boundaries; none is AI-specific.

Google's figure is owned and operated facilities, withdrawal minus discharge, under limited third-party assurance. Amazon's covers leased, owned, and shared data centers. Meta's excludes construction water. A tidy leaderboard would outrun the evidence.

Boundary: Mixed: consumption (Google, Meta), withdrawal (Amazon), intensity (Microsoft); Global fleets, every workload; 2024 and 2025, by company.
Does not say: That Google is the thirstiest company. Gallons and liters, consumption and withdrawal, and 2024 and 2025 are not comparable. That any of these totals belong to one AI product.
Evidence: Reported · primary. Source: Google, Amazon, Meta, and Microsoft 2025 and 2026 reports, https://sustainability.aboutamazon.com/2025-amazon-sustainability-report.pdf. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/which-ai-company-uses-most-water

### How much water does a single data center use per day?

It depends on cooling design, climate, and load, so there is no standard figure. The IEA's illustrative 100 MW U.S. hyperscale facility consumes about 2 million liters per day in total, with more than 60 percent of that indirect (from electricity generation). A dry-cooled site can be far lower on site; a water-cooled chiller plant can be higher.

'This data center uses as much water as X homes' is a scenario, not a facility fact, unless the source discloses metered data and the household denominator it used. The same nameplate capacity can also run at very different load factors, and evaporative cooling demand peaks on the hottest days.

Boundary: Modeled total consumption, direct plus indirect; One illustrative 100 MW U.S. hyperscale facility; IEA 2025 illustration.
Does not say: That every data center uses 2 million liters per day. That nameplate megawatts equal operating load. That annual averages describe the maximum-day demand a local utility has to plan for.
Evidence: Modeled. Source: IEA, Energy and AI (April 2025), https://www.iea.org/reports/energy-and-ai. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/data-center-water-per-day

### How much electricity will U.S. data centers use by 2030?

Berkeley Lab's June 2026 update models a 649 TWh reference case for 2030, about 11.8 percent of forecast U.S. electricity, with a compounded uncertainty range of 521 to 843 TWh (9.5 to 15.3 percent). It is a scenario, not a forecast of water use.

The range reflects uncertainty in AI adoption, hardware efficiency, and grid constraints. Electricity drives indirect water consumption, but the ratio depends on the generation mix in each region, so the electricity number cannot be converted to a national water number without that mix.

Boundary: Modeled electricity consumption (TWh) and share of U.S. total; All U.S. data centers, every workload; 2030 scenario, published June 2026.
Does not say: That water use will grow by the same ratio. That 649 TWh is a prediction rather than a reference case with a wide range.
Evidence: Modeled. Source: LBNL United States Data Center Energy Usage Report: 2025 Update, https://eta-publications.lbl.gov/publications/united-states-data-center-energy-2025. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/data-center-electricity-2030

### Do closed-loop or 'zero water' data centers really use no water?

They can eliminate evaporative cooling water on site, which is a real reduction. They do not eliminate water used to generate their electricity, water used in construction, or water used to manufacture chips. 'Zero water' is a cooling claim, not a footprint claim.

Microsoft says its AI-optimized design introduced in 2024 uses zero water for cooling during operations. Dry cooling usually trades water for more electricity, which carries its own indirect water, so the net effect depends on the local grid.

Boundary: On-site evaporative cooling water; Specific new designs, as described by the operator; Designs introduced from August 2024.
Does not say: That the site's total water footprint is zero. That 'liquid cooled' means waterless. The internal loop is closed; heat rejection outside it may still be evaporative. That dry cooling is always the better environmental choice; it can raise electricity use.
Evidence: Reported · primary. Source: Microsoft, June 24 2026, https://blogs.microsoft.com/blog/2026/06/24/inside-microsofts-two-decade-push-to-cut-water-intensity-while-scaling-for-growth/. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/zero-water-cooling

### Is the water AI data centers use destroyed?

No. Water consumed by evaporative cooling re-enters the hydrologic cycle. But USGS defines consumptive use as water not returned to the immediate water environment, which is the point: it is unavailable locally, at the time and quality it was taken, and that is what a stressed basin or a hot-week utility peak actually cares about.

This is why 'the water comes back' and 'the water is gone' are both incomplete. Return flow only offsets local impact if its quantity, timing, temperature, and quality do. Replenishment projects are a separate, modeled benefit that may occur elsewhere or later, not a one-for-one cancel of peak withdrawal.

Boundary: Consumptive use as defined by USGS; Any evaporatively cooled facility; Definition; accessed July 15, 2026.
Does not say: That evaporation is locally harmless. That replenishment commitments cancel consumption one-for-one.
Evidence: Definition. Source: USGS, Water Use in the United States, https://www.usgs.gov/mission-areas/water-resources/science/water-use-united-states. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/is-ai-water-destroyed

### How much water do Virginia data centers use?

Virginia's legislative audit agency (JLARC) reported that the state's data centers used about 2.1 billion gallons of water in 2023, with more than one-third of it reclaimed water. It judged statewide supply manageable while flagging that local planning and growth still warranted attention.

Virginia hosts the largest data-center cluster in the world, so it is the most-audited state. The reclaimed-water share and local utility conditions there are not representative of every region, and a statewide total does not describe any single county's peak-day demand.

Boundary: Reported water use (withdrawal basis as audited), potable and reclaimed shown separately; Data centers in Virginia; Calendar 2023, published December 2024.
Does not say: That other states have a similar reclaimed-water share. That a manageable statewide total means every local system is fine.
Evidence: Reported · primary. Source: Virginia JLARC, Data Centers in Virginia (Report 598), https://jlarc.virginia.gov/pdfs/reports/Rpt598.pdf. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/virginia-data-center-water-use

### Is there a federal law requiring data centers to report water use?

Not yet. H.R. 6984, the Data Center Transparency Act, was introduced in the House on January 8, 2026 and referred to the Energy and Commerce Committee. It would require recurring reports on water, electricity, air quality, and local utility effects. As of the research cutoff it had not been enacted.

Today most public water figures come from voluntary corporate sustainability reports, state utility commission dockets, and local permits, which is why boundaries and years differ from company to company.

Boundary: Legislative status; United States, federal; As of July 15, 2026.
Does not say: That the bill has passed or that reporting is currently required. That no reporting exists anywhere; the EU has harmonized data-center reporting rules under Delegated Regulation 2024/1364.
Evidence: Reported · primary. Source: Congress.gov, H.R. 6984 (119th Congress), https://www.congress.gov/bill/119th-congress/house-bill/6984. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/data-center-transparency-act

### If data centers are a small share of national water use, does it matter?

Two things are true. Nationally, data centers are a minority water user next to agriculture and municipal demand. Locally, a large campus can be material when it draws potable water in a stressed basin, during summer peaks, alongside other projects. The national share does not settle the local question.

The useful questions are local: what source, what season, what peak demand, what watershed condition, and what the permit or docket actually says. That is what ThirstTrap.ai tracks.

Boundary: Framing, not a single figure; National totals versus local systems; Ongoing.
Does not say: That data centers are a top national water user. That a small national share makes a local project irrelevant.
Evidence: Analysis. Source: WRI Aqueduct; LBNL 2024, https://www.wri.org/aqueduct. Verified 2026-08-28. Permalink: https://thirsttrap.ai/answers/does-ai-water-use-matter

## Pages

- https://thirsttrap.ai/: homepage with receipts, signal feed, impact lens calculator, and disclosures
- https://thirsttrap.ai/answers: every direct answer, each with its own permalink page
- https://thirsttrap.ai/methods: evidence classes, metric definitions, boundary rules, verification, and the corrections log
- https://thirsttrap.ai/about: who runs the site, how it is funded, and what exists today
- https://thirsttrap.ai/advertise: advertising placements and sponsor rules
- https://thirsttrap.ai/merch: merch that keeps the research free
- https://thirsttrap.ai/privacy: privacy policy
- https://thirsttrap.ai/feed.xml: RSS feed of new signals and answers

Contact: thirsttrapai@pm.me