How AQUAMIND estimates, and what it does not claim

The calculator is deterministic: the same inputs always give the same result, and no figure is generated by a chatbot.

The model in one picture

Same inputs, same answer
Your usageprompts, images, GPU hours
×
Energy per unitWh per item
=
IT energy EkWh = usage × Wh ÷ 1000
Direct water
E × WUE
evaporated on site for cooling
Indirect water
E × PUE × grid intensity
consumed at power plants
Direct Indirect= Your water footprint, as a low, central and high range

WUE

Water usage effectiveness

Liters of on site water consumed per kWh of IT energy, mostly from evaporative cooling.

PUE

Power usage effectiveness

Total facility energy divided by IT energy. It scales server energy up to what the grid supplies.

Grid intensity

Water per kWh generated

Liters consumed per kWh at power plants, mainly thermal plants cooling with water.

Withdrawal versus consumption

Withdrawal

Water taken from a source. Much of it may be returned.

Consumption

Water that is not returned, mainly because it evaporates.

What AQUAMIND estimates

Uncertainty and confidence

Every input is a range. Unknown inputs use wide ranges and provider data narrows them. Each of the four inputs (energy, WUE, PUE, grid) scores 2 points for provider data, 1 for a chosen category and 0 when unknown.

Low
0 to 3 points: mostly unknown
Medium
4 to 6 points: categories chosen
High
7 to 8 points: provider data

The ranges behind every estimate

Illustrative modelling ranges for the MVP, deliberately wide. Google reports 0.24 Wh and 0.26 mL for its median Gemini text prompt (May 2025), which sits inside our text generation range.

Energy per unitWh per item · log scale
Text generation
0.1–3
Image generation
1–10
Document processing
0.5–15
Cloud computing
300–1000
Organizational AI
0.2–5
Cooling water (WUE)L per kWh
Unknown
0–2.5
Air or closed loop (low water)
0–0.2
Seasonal evaporative
0.2–1
Evaporative, hot climate
1–2.5
Grid water intensityL per kWh
Unknown
0.2–5
Mostly wind and solar
0.2–1
Mixed grid
1.5–3.5
Mostly thermal (coal, gas, nuclear)
3–5
Facility efficiency (PUE)ratio, 1.0 is no overhead
Unknown
1.1–1.8
Efficient, modern hyperscale
1.08–1.2
Typical facility
1.4–1.6
Legacy or on premises
1.6–2

What AQUAMIND does not claim

  • Results are estimates, not measurements of any user's actual water consumption.
  • Energy per AI request varies with model, hardware, utilization and output length, and is rarely published.
  • Water used to manufacture chips and build facilities is not included in the MVP.
  • Water stress is shown as context; it does not change the liters calculated.
  • Proving real reductions requires data from providers and data centers. That is the next stage of AQUAMIND.

Sources behind the platform

International Energy Agency

Energy and AI (2025)

Data center electricity demand and the water implications of cooling, electricity generation and semiconductor manufacturing.

Google Cloud

Measuring the environmental impact of AI inference

Energy, carbon and water per median Gemini text prompt, with a discussion of why methodology changes the answer.

Google Sustainability

2026 Environmental Report

Corporate reporting on data center energy, water and replenishment.

World Resources Institute

25 countries face extremely high water stress (Aqueduct 4.0)

Country and regional water stress, including the Middle East and North Africa finding used on our dashboard.

Why AQUAMIND exists

Mission

To make the hidden water footprint of artificial intelligence visible, understandable, and actionable.

Vision

A future where technological innovation and environmental sustainability advance together.

Measuring our impact

In the MVP stage we track users, completed calculations, Academy modules completed, gains in understanding and institutions expressing interest. Actual water savings will be verified through future implementation projects.