Live
What It Costs
Calculator estimating the freshwater cost of an AI session, from peer-reviewed data-centre research.
Next.js / TypeScript / Tailwind CSS / Vercel
Visit siteWhat the calculator estimates
What It Costs is a public calculator that estimates how much freshwater an AI session uses, for ChatGPT (GPT-4o or GPT-3.5), Claude and Gemini. A visitor sets the model, the number of queries (1 to 200), the query type and the data-centre region, and the estimate is recalculated on every change.
The total is translated into 250 ml drinking glasses, a share of a 60 litre shower, one person's 2 litre daily water supply in a water-stressed region and a 30-day projection. A global panel scales the session to an estimated 100 million daily AI users and gives the total in Olympic swimming pools.
How the estimate is calculated
Each estimate adds two terms, both multiplied by the number of queries and a query-type factor: the model's water per query, and its energy per query times the region's Water Usage Effectiveness (1.8 litres per kWh for the US, 1.2 for Europe, 2.1 for Asia, 2.5 for India). The factor is 1 for simple Q&A, 1.8 for code generation, 2.5 for long-form writing and 6 for image generation.
The ChatGPT figures come from Li et al. (2023), 'Making AI Less Thirsty', and Gemini's from Google's 2023 Environmental Report; the Claude figure is labelled an estimate. The page states that no AI company publicly discloses per-query water data, presents results as directional, and measures inference only, not training.
Shareable results and search metadata
Every change rewrites the URL with the four inputs, so each result is a link, shared through the Web Share API or, failing that, copied to the clipboard. For a shared link, the server validates the parameters (unknown values fall back to defaults), pre-fills the calculator, and sets a title, description and preview image for that result; the image is rendered for those inputs on the edge runtime. One calculation function serves the browser, the metadata and the image, so a link preview shows the same figure as the page.
Search metadata covers a sitemap, robots rules, a canonical URL and JSON-LD for WebSite, WebApplication and an FAQPage generated from the questions on the page. Built with Next.js 16, React 19, TypeScript and Tailwind CSS 4.