AI Impact Calculator

How Much Water Does Your AI Use?

Estimate the hidden water and energy footprint of ChatGPT, Gemini, Claude and other AI tools in under 30 seconds – and discover practical ways to reduce it.

✓ Reviewed September 2026 · figures from 2025–26 vendor disclosures

Your AI Usage

Adjust the options – results update instantly

Select Your Usage Profile

What Do You Mostly Use AI For?

These tiers model chat-style use. Agentic and coding workflows (tool-calling agents, coding assistants) can consume far more per task – see the methodology note below.

Estimate Basis

Comprehensive includes water consumed generating the electricity that powers data centres – the wider boundary most independent researchers use.

8.5 Wh
Daily Energy Use
≈ 5 min of TV
35 mL
Daily Water Use
≈ 7 teaspoons
Drag to adjust query volume
25 queries/day
Annual Impact
2.1 kWh energy ≈ 21 hrs of TV
8.8 L water ≈ 1 min of showering

🔒 Runs entirely in your browser – nothing you enter is stored or sent anywhere.

How We Calculate This

Last reviewed: September 2026. The per-query figures are unchanged. We re-checked them against the most recent public claim – Sam Altman's September 2026 comparison of 38,000 ChatGPT queries to a single almond – which is addressed in the FAQ below. Earlier versions of this calculator used 2023-era research estimates (roughly 40 mL of water per query). Vendor disclosures published since report far lower figures for typical text prompts, so we have rebuilt the model – energy first, with water derived from it.

Energy per query: Google reports a median Gemini text prompt at 0.24 Wh; OpenAI cites ~0.34 Wh per ChatGPT query. Independent estimates put image generation and long reasoning tasks at roughly 3 Wh or more. Our usage-mix options reflect this range.

Water per query: "Vendor-reported" uses ~1 mL per Wh – the onsite cooling figure implied by Google's disclosure (0.26 mL per 0.24 Wh prompt). "Comprehensive" adds ~3.1 mL per Wh for water consumed generating the electricity (US grid average, per Li et al.). Vendor figures are not peer-reviewed, exclude model training, and describe median prompts – treat all results as indicative estimates. Onsite intensity also varies sharply by provider: Amazon reports a fleet water-usage effectiveness of 0.12 L/kWh and Microsoft 0.27 L/kWh, against roughly 1.08 L/kWh implied by Google's per-prompt figures. This calculator uses the Google-derived value — the most conservative of the three.

What this calculator doesn't cover: These figures model chat-style use – one prompt, one response. Agentic workflows (coding assistants like Claude Code, tool-calling agents, long multi-step sessions) work differently: the full conversation history, attached files and tool outputs are typically reprocessed on every turn, so a single task can involve dozens of model calls over an ever-growing context. Per-task consumption in these workflows can exceed the per-query figures here by orders of magnitude.

Location matters too: Water intensity varies enormously with where a data centre sits and how its electricity is generated. As Andrew Shepherd puts it, a geothermal-powered data centre in Iceland has a rather different footprint to an evaporatively cooled one in the middle of a desert. Our figures use US grid averages; your actual footprint depends on which region serves your queries.

Sources: Google (2025), "Measuring the environmental impact of AI inference"; OpenAI per-query figures (2025); Li et al., "Making AI Less Thirsty"; MIT Technology Review analysis. Several refinements (agentic workloads, cooling methods, prompting habits) draw on Andrew Shepherd, "AI resource use and sustainability" (2026).

Why AI Water Estimates Vary by 100×

Published figures run from 0.3 mL to 50 mL per prompt. Most are defensible — they are measuring different things.

2023 research estimate
~40 mL
Comprehensive, 2026 — used here
~1.2 mL
Vendor-reported, onsite only
~0.3 mL

Logarithmic scale — each step is 10×. Figures are per typical text prompt.

When it was measured

Efficiency has improved fast — Google reports a 33× drop in energy per prompt in twelve months. Estimates built on 2023 hardware describe a system that no longer exists, yet are still widely quoted.

Where the line is drawn

Counting only data centre cooling gives one figure. Adding the water used to generate that electricity gives one several times higher. Neither is wrong — but a calculator that does not say which cannot be checked.

What is being measured

These are chat-style prompts. An agentic task — a coding assistant or tool-calling agent — reprocesses a growing context over dozens of calls, and can exceed any per-prompt figure by orders of magnitude.

Whose data centre

Onsite water intensity differs about ninefold between providers — 0.12 L/kWh at Amazon, 0.27 at Microsoft, roughly 1.08 implied by Google. Same prompt, same energy, very different water. We use the highest of the three.

What Uses More Water?

Everyday context for your AI footprint

Bar lengths use a logarithmic scale – each step to the right represents roughly ten times more water. On a linear scale, the AI bar would be thinner than a hair.

The AI figure is operational water use (including electricity generation). Paper, tea and burger figures are full lifecycle "virtual water" footprints from the Water Footprint Network – a wider boundary, shown here for everyday scale. The point isn't that AI is free; it's that context matters.

💧 Why AI Uses Water

AI models run on powerful servers that generate heat, and data centres can consume water two ways: directly, where evaporative cooling is used, and indirectly, through the water used to generate their electricity. Not every facility is thirsty – many newer data centres use closed-loop or free-air cooling, which dramatically reduces onsite water consumption – but evaporative systems remain common, especially in hot regions. Per-query figures are far smaller than early estimates suggested – Google and OpenAI now report well under a millilitre per typical text prompt – but with billions of queries every day, and data centres often sited in water-stressed regions, the totals still matter. As the pharmaceutical industry scales up AI adoption, understanding your share is the first step to reducing it.

Visualization of data center resource management systems
Every AI query requires data centre cooling

Frequently Asked Questions

Is it true that 38,000 ChatGPT queries use as much water as one almond?
Sam Altman made this claim in September 2026. The arithmetic is self-consistent, but it pairs two different accounting boundaries: the total water footprint of a California almond (~12 L, including rainwater and dilution water) against the onsite cooling only figure for a query (~0.32 mL). Match the boundaries and the number falls – roughly 13,000 queries against an almond's irrigation water alone, or about 10,000 using the comprehensive per-query figure shown above. "Thousands of queries per almond" is fair; 38,000 is the most flattering of the available pairings. Volume is not the whole story either: almond irrigation and data centre cooling draw on different supplies, and data centres typically use treated potable water from municipal systems.
How much water does a single AI query use?
A typical text prompt consumes about 0.3 mL of water onsite for data centre cooling – around five drops. Including the water used to generate the electricity, it's roughly 1–1.5 mL. Image generation and long reasoning tasks can use ten times more. Early 2023 estimates of 40–50 mL per query referred to GPT-3-era infrastructure and wider system boundaries; efficiency has improved dramatically since.
Why do data centres use water at all?
Servers generate heat, and evaporative cooling – which consumes freshwater – is one of the most energy-efficient ways to remove it. It isn't the only approach, though: many facilities now use closed-loop or free-air cooling, which consumes far less water onsite. Data centres also consume water indirectly: most electricity generation (thermal and hydro) evaporates water in the process.
If per-query numbers are tiny, why does it matter?
Scale. Billions of queries run every day, AI usage is growing rapidly, and data centres are often located in water-stressed regions where even modest consumption competes with local needs. Small per-query savings multiplied across an organisation – or an industry – add up.
How is this calculator different from others?
Most calculators still use 2023-era estimates. Ours is energy-first (water is physically a function of energy, not the other way round), uses 2025–2026 vendor disclosures, and lets you toggle between vendor-reported onsite figures and a comprehensive boundary that includes electricity-generation water.
How much water does AI use per day?
That depends almost entirely on how much you use it. On the comprehensive basis used above, 50 typical text prompts come to roughly 70 mL – so someone running 50 prompts every working day uses around 2 litres over a month, less than a large bottle of water. Long documents, image generation and larger models push that up considerably. Use the calculator to model your own pattern rather than relying on an average.
How much water does generating an AI image use?
Substantially more than a short text reply – roughly ten times, on the estimates we use here. Image generation keeps accelerator hardware busy far longer than answering a text prompt, so its energy footprint, and therefore its water footprint, is much higher. Published figures for images are also less settled than for text, because no vendor has disclosed per-image numbers the way Google has for text prompts. If images are a large part of your usage, select the intensive mix above to see the effect.
Does ChatGPT, Claude or Gemini use more water?
There is no reliable basis for ranking them. Only Google and OpenAI have published first-party per-prompt figures, and those sit in the same range – fractions of a millilitre of onsite water for a typical text prompt. Anthropic and Meta have published no per-prompt water figures at all, so any number you see for Claude or Llama is a third-party estimate, not a disclosure. What genuinely moves your footprint is what you ask for (text or images, short or long), the model size you choose, and where the data centre sits and how it is cooled – not which assistant brand is on the tab.

Smart Strategies to Reduce Your AI Footprint

Practical ways for medical affairs, regulatory and commercial teams to cut water and energy use

Batch Your Queries

Group related work – literature reviews, safety signal monitoring, competitor intelligence – into single sessions rather than scattering queries through the day. Batching reduces redundant server spin-up and keeps your own context together too.

Choose Purpose-Built Tools

Specialised tools for compliance checking, poster analysis, or reference verification use far fewer resources per task than prompting a general-purpose frontier model – and typically produce more reliable output for regulated work.

Prioritise Text Over Images

Image generation uses roughly ten times the resources of a text query. For congress materials and internal documents, generate visuals only when needed – and reuse approved assets rather than regenerating variations.

Cache Common Responses

Standard response documents, template summaries, and recurring analyses don't need regenerating every time. Build a knowledge base of approved AI-generated content your team can reuse.

Select the Right Model Size

Larger models aren't always better. Use smaller, faster models for grammar checks and simple summaries; reserve heavyweight reasoning models for complex regulatory analysis and strategy work.

Monitor and Optimise Usage

Track AI consumption patterns across medical affairs, MLR review, and commercial projects. Identify redundant queries, consolidate similar requests, and set team-level best practices for efficient AI use.

Skip "Just in Case" Attachments

Every document you attach is reprocessed on each turn of the conversation, not just once. Attach only what the task actually needs – and start a fresh chat when you move to an unrelated question.

Be Precise Up Front

A specific first prompt beats a vague one refined through five follow-ups – each follow-up reprocesses the whole conversation. Showing one example of the output you want is often worth paragraphs of description.

From the PharmaTools.AI lab

This free tool is built in the open by Nick Lamb, PhD – an independent AI researcher and engineer working on evaluation, verification and oversight of AI systems. Methodology, sources and the reasoning behind every figure are published above.

Explore the research →  ·  More free tools →

Build More Responsible AI Workflows

Explore PharmaTools.AI solutions designed for regulated healthcare communication and content excellence.

Patiently AI

Turn complex medical notes into clear, plain-language patient communication.

Learn more →

RefCheckr

Verify medical claims against their cited references – built for medical writers.

Learn more →

MedCheckr

AI compliance checking against 11 national and international pharma codes of practice.

Learn more →
Explore AI Solutions →