Estimate the hidden water and energy footprint of ChatGPT, Gemini, Claude and other AI tools in under 30 seconds – using AI companies' own 2025–26 figures.
per typical text prompt – about a quarter of a teaspoon, including the water used to generate the electricity.
✓ Reviewed October 2026 · figures from 2025–26 vendor disclosures · Methodology 2026-10 · cite
Water per typical text prompt · log scale, each step is 10×
Adjust anything – the result updates as you go.
These model chat-style use. Agentic and coding workflows (tool-calling agents, coding assistants) can consume far more per task – see the methodology.
Comprehensive includes water consumed generating the electricity that powers data centres – the wider boundary most independent researchers use.
The white mark is our central estimate. Real use could be roughly 0.3× to 3× that – energy 2.6–26 Wh – depending on prompt length, the model and where the data centre is. Published figures for a similar prompt, and the water intensity of different providers and power grids, vary by about this much. How we set the range.
On these settings: ≈ 1.4 mL water / query
🔒 Runs entirely in your browser – nothing you enter is stored or sent anywhere.
Published figures run from 0.3 mL to 50 mL per prompt. Most are defensible – they are measuring different things. A figure that doesn't say which can't be checked.
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.
33× drop in twelve monthsCounting 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.
~4× between boundariesThese 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.
orders of magnitude per taskOnsite 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.
~9× between providersBar lengths use a logarithmic scale – each step to the right is roughly ten times more water. On a linear scale, the AI bar would be thinner than a hair.
Paste a claim about AI's water or energy use. We'll show what it assumes and where it sits against today's figures.
Well-known claims and simple figures are read in your browser. If we can't read a figure, you can choose to send the text to a small AI model that only pulls out the number and what it's per – the explanation is always worked out by fixed rules, not by the AI.
Estimate a year of AI use across your teams and download a report with the figures, methodology version and sources. Everything is calculated in your browser – nothing you enter is sent anywhere.
| Team | People | % using AI | Prompts/day | Mostly |
|---|
Prompts per user per day is an average across working days. Licence dashboards and usage exports are the best source; a quick staff survey works too.
Methodology version 2026-10 · central estimates with a 0.3×–3× likely range · indicative, not measured supplier data.
Practical ways to cut the water and energy behind your AI use – starting with getting a usable answer first time.
Group related questions into one conversation rather than starting fresh each time. It keeps the context together and saves you repeating background the model already has.
A calculator, spell-checker or search engine uses a tiny fraction of the energy of a chatbot. Save AI for the tasks that actually need it.
Image generation uses roughly ten times the resources of a text reply. Generate images only when you need them, and reuse ones you already have rather than regenerating variations.
If you ask for the same kind of thing regularly – a cover letter, a meal plan, a summary format – save the answer that worked and adapt it rather than asking again.
Larger models aren't always better. Fast, lightweight models handle everyday questions and simple rewrites well; save the heavyweight reasoning modes for genuinely hard problems.
Use the calculator above to see where your prompts go. If you're running AI across a team, the same habits applied at scale make the biggest difference.
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.
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.
Paste a draft prompt and a small AI model rewrites it so you get a usable answer first time. We also show what the rewrite itself cost – because it isn't free either.
Opt-in: unlike the calculator above, this sends your text to Anthropic's Claude API via our server. It isn't stored. Don't paste personal or confidential information.
This rewrite cost
Follow-ups avoided
Net water
How we count: every prompt, including the rewrite, is costed at the page's comprehensive figure for one text prompt (~0.34 Wh, ~1.4 mL). That is an upper bound for the small model used here. Follow-ups are counted as single prompts, ignoring the extra cost of reprocessing the conversation – so real savings are likely larger. The follow-up count is the model's own estimate.
Last reviewed: October 2026 (methodology 2026-10). The per-query figures are unchanged since the September 2026 review, which re-checked them against Sam Altman's comparison of 38,000 ChatGPT queries to a single almond (see the FAQ below). Version 2026-10 adds a likely range to every estimate and is published as a citable method note – see the changelog and citation 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. Benchmarks put reasoning-model queries at roughly 1–12 Wh depending on length (o3: 1.2 Wh short, 12 Wh long), so we use 3 Wh for images and long reasoning. 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.
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.
The same figures, sources and versioned methodology as four read-only tools and a JSON API. No model in any tool. footprint.pharmatools.ai
AI scientists live in a simulated universe whose laws secretly change. They gathered the evidence. Almost none of them noticed.
Read the research → ToyA goblin that walks down your LinkedIn feed eating corporate jargon and spitting out plain English. No AI model involved.
Feed him → ExperimentPlay chess engines trained on the games of another era: the gambit-happy attackers of 1850, the Classical masters or the Soviet school.
Play →Add your ChatGPT or Claude data export and see the water and energy behind every conversation you've had. This page opens the file on your own device. It is never uploaded.
Usually arrives within minutes; OpenAI says it can take longer. Not available on Business or Enterprise workspaces. OpenAI's guide
Not available from the mobile apps. On Team and Enterprise plans, only the Primary Owner can export. Anthropic's guide
Network requests this page has made since you added your file, counted by your browser's own network log.
Your browser has to open the file to count it. The text is let go as soon as it's counted and nothing is saved – refresh and it's gone.
Turn off your Wi-Fi before you add the file. It still works, because everything happens here.
Or read the exact code that opens your file – it runs in this page and you can check it line by line.
This is the code running on this page right now (shown with Function.toString(), so it can't differ from what reads your file). There are no network calls in it.
The method behind every figure on this page is published as a versioned, open-access note with a DOI, so you can cite the exact version you used.
Lamb, N. (2026). An energy-first method for estimating the water and energy footprint of AI use (methodology 2026-10) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23417340
Citing a figure? Say which boundary you used (cooling only or comprehensive) and quote the likely range alongside the central estimate. To always point to the latest version, use the concept DOI 10.5281/zenodo.23375360. The code is open too: github.com/nickjlamb/ai-footprint.
Every change to the figures, sources or method gets a new version. Results everywhere – this page, reports, the MCP server – carry the version they used.
Free to embed with one line of code. It uses the same figures and methodology as this page, works out everything in your readers' browsers and sets no cookies.
Free for any site, including commercial ones. Please keep the small credit link. Results are labelled with the methodology version, and the figures update automatically when the method changes.
