66B L
DATA-CENTER INFRASTRUCTURE INTELLIGENCE
The AI boomhas a water address.
ThirstTrap.ai is being built to turn chaotic permits, utility dockets, water records, and company claims into source-linked signal you can actually use.
No doomscrolling. No greenwashing. No fake precision.
TT / BASIN SIGNAL 001 MODELED
MODELED U.S. DIRECT CONSUMPTION COUNTER
Modeled illustration. We divide Berkeley Lab's 66 billion liter estimate of direct water consumption by all United States data centers in 2023 evenly across a year, about 2,100 liters per second. This is a derived average, not a live meter, a peak, or an AI-only figure.
≈2.1K L / SECOND · 2023 AVERAGE
66B L/year spread evenly over time · all U.S. data centers · not AI alone · not a live meter or measured peak
View the LBNL receipt~800B L
electricity-related consumption
Modeled 2023 · all U.S. data centers · not AI alone · separate boundaryLBNL521–843TWh
2030 U.S. electricity range
Scenario · all U.S. data centers · not a water forecastLBNL 2025 updateTHE TRUST PROTOCOL
Every number gets a receipt.
Boundary
Withdrawal, consumption, or what?
Source
Who actually said it?
Date
When was it true?
Confidence
Reported, modeled, or inferred?
01 / RECEIPTS
The discourse needs a denominator.
Personal prompt math is the hook. The evidence is conditional. Data centers remain a minority water user at national and global scales, but aggregate totals do not settle local questions.
THE VIRAL CLAIM
“A bottle of water for every AI prompt.”
NOT A CONSTANT
The famous estimate was roughly one 500 mL bottle for 10–50 medium-length GPT-3 responses in a 2023 modeled scenario (the first April 2023 draft said 20–50 questions and answers), not one bottle per prompt and not a current meter reading.
Read the modeled study Our full answerTHE EFFICIENCY CHECK
A low WUE is useful, not a verdict.
WUE is an efficiency ratio. A large efficient campus can still consume more total water than a smaller inefficient one. The numerator may also mean withdrawal or consumption.
RATIO ≠ TOTALRead the DOE guidanceTHE LOCAL CONTEXT
A modest national share can still be locally material.
Basin stress, potable versus reclaimed supply, summer peaks, cumulative development, and utility costs decide whether a project fits the place and the moment.
PLACE + PEAK + SOURCEOpen WRI AqueductTHIRSTTRAP.AI / SOURCE RECEIPTREPORTED · PRIMARY
Every claim keeps its chain of custody.
- Metric
- Withdrawal, consumption, discharge, or replenishment. Never undefined “water use.”
- Boundary
- Site, electricity, construction, or lifecycle
- Status
- Measured, proposed, modeled, or inferred
- Revision
- Corrections stay visible; history does not vanish
THE MOAT
A pin is easy. A defensible record is hard.
What we're building is not another generic map. It is an evidence graph: projects, owners, permits, utility filings, water claims, confidence, and change history, all connected to the original record.
- Exact sources, not circular citations
- Proposed demand never styled as actual use
- Reported and modeled values never quietly mixed
02 / SIGNAL
Know before the headline lands.
ThirstTrap.ai is designed to watch the boring documents where the important stuff happens, then make the change legible.
REAL SOURCES · PRODUCT PREVIEW01
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.
Reported · primaryOpen source
02
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.
ModeledOpen source
03
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.
Reported · primaryOpen source
WHAT WE WATCH
The boring paperwork is where the signal lives.
- Permit and utility docket changes
- Water agreements and cooling claims
- Original documents and visible corrections
- Dates, boundaries, and evidence classes
03 / IMPACT LENS
Three non-comparable published examples.
Move the slider to see what selected published figures would imply, not what your own activity actually consumed. Different answers can reveal different boundaries, not a winner.
Same number, every day, for 365 days.
Reality check.
This is a scenario lens, not a meter. These sources describe different services, dates, workloads, locations, and accounting boundaries. The spread reflects real system differences and uncertainty. It cannot support a brand ranking or estimate your own use.
04 / DISCLOSURE
Disclosure is improving. Comparison is limited.
Fleet totals mix workloads, places, and accounting methods. We show the number and the gap beside it because confidence without caveats is incomplete.
2025 global data-center consumption
WHAT'S MISSING
Open disclosure Owned and operated fleet, withdrawal minus discharge, limited assurance. Not AI-specific and not a site-level peak view.
Amazon / AWS
9.4B L2025 global data-center withdrawal
WHAT'S MISSING
Open disclosure Leased, owned, and shared data centers; withdrawal is not comparable with consumption.
Meta
2.974B L2024 global data-center consumption
WHAT'S MISSING
Open disclosure Fleet-wide and not generative-AI-specific; construction is separate.
No fake leaderboard: fleet disclosures use different boundaries, so a tidy company ranking would outrun the evidence.
05 / STRAIGHT ANSWERS
The questions everyone asks. With receipts.
Number first, boundary second, source and verification date always. Each one has its own page you can quote, link, and argue with. The rules behind them are on the methods page.
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.
Full answer, boundary, and citationDoes 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.
Full answer, boundary, and citationHow 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.
Full answer, boundary, and citationHow 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.
Full answer, boundary, and citation06 / KEEP IT FREE
Receipts stay free. Here is how that gets paid for.
No paywall, no subscription, no paid corrections. The site runs on clearly labeled contextual ads and merch that says what the research says.
ADVERTISE
Reach people who read the filing.
Residents, reporters, planners, and industry folks who want the number and the boundary. Labeled placements, low density, zero say over coverage.
See placementsMERCH
Wear the denominator.
Tees, hats, mugs, and stickers with the lines this site is built on. Every design is a boundary lesson you can put on a laptop.
Shop the merch