Independent · Nonprofit · Open methodology
You asked an AI a question today. What did it cost the Earth?
Every prompt you send runs on electricity, is cooled with water, and leaves a trace of carbon. OpenButterfly exists to make that footprint visible, both yours and the industry's, with honest numbers and honest uncertainty.
Your AI year
Set the sliders to your own habits, from casual chatting to heavier coding-agent work, and see what a year of AI adds up to, in things you can picture.
Claude Code, Codex, and similar tools work very differently from chat: one session churns through hundreds of thousands of tokens, so a single task costs about as much as 140 chat prompts.
Electricity / year
24 kWh
about 1,604 full phone charges
Water / year
24 L
about 48 half-litre bottles
Carbon / year
8.8 kg CO₂e
like driving 46 km in a petrol car
Text prompt ≈ 0.3 Wh (Google, OpenAI, Epoch AI, 2025); median coding session ≈ 41 Wh across ~600k tokens (Couch, 2026). Both already account for prompt caching: re-reading cached context costs about a tenth of processing it fresh; without caching, agent sessions would run roughly ten times higher.
Water assumes typical evaporative cooling (~1 mL per Wh); air-cooled and closed-loop data centers evaporate far less on site but draw more electricity, and the water behind power generation is often larger still. Carbon uses US average grid intensity. Excludes model training; image, video, and long “reasoning” requests cost far more.
Chatting barely registers: a year of it costs less electricity than a refrigerator uses in a week. Heavy agentic coding is real but still household-scale: a full-time developer's year of AI runs about half a refrigerator. If the story ended at your keyboard, there would be no story. It doesn't.
Small for you. Enormous in total.
Your prompt is one drop in a very large river: billions of daily requests, the training runs behind every model, the chip factories, and the data centers being built years ahead of demand. Across the latest reported year for each organization, the 27 companies this observatory tracks consumed roughly 384 TWh of electricity, the annual use of 36.6 million US homes. They withdrew around 507 GL of water, some 203,000 Olympic swimming pools, and their reported and estimated emissions total roughly 365 Mt CO₂e, comparable to 79.3 million passenger cars.
And not one of them will tell you how much of that is AI. Not one discloses AI-specific figures. These totals mix company-disclosed figures with our documented estimates; every underlying number carries its confidence tier and source on the entity profiles.
Plain answers to fair questions
I read that one prompt uses a bottle of water. True?
Not for a typical text prompt; that's closer to a few drops. The viral bottle figure came from early estimates that spread training costs and older assumptions across single queries. But water is a real issue at the facility level: a large data center can draw millions of litres a day, and some are sited in places already short of water. The right place to worry is the map, not your chat window.
So my AI use is fine and I can stop feeling guilty?
Mostly, yes: guilt over individual text prompts is misplaced. But not all use is equal: generating images and video costs far more per request than text, and running AI agents that fire thousands of requests multiplies everything. And smallness per prompt is exactly how enormous totals hide. The industry's footprint is growing faster than almost any other electricity demand on the planet.
Why doesn't anyone just publish the real numbers?
Because almost no law requires it. Sustainability reports blend AI into company-wide totals, some omit whole categories of emissions, and water often goes unreported entirely. That's why we treat silence as a data point: when a company doesn't publish its water use, we record that refusal and score it.
What can I actually do?
The lever isn't using AI less; it's demanding that its costs be visible. Prefer providers that publish real environmental figures. If a data center is proposed near you, ask about its water source and grid deal; those hearings are where the footprint gets decided. And when a number seems off, check it against sourced data. That's what this site is for.
What this observatory watches
27 organizations and 152 infrastructure sites in 45 countries, drawing on 219 cited sources: government records, company reports, academic research, satellite imagery, and local journalism.
Look up a company
Energy, carbon, water, and honesty scores for the AI labs, cloud providers, and chipmakers behind the tools you use. Every figure carries a confidence tier and its source.
Browse the dataSee where it happens
A map of AI data center and power sites: water stress, grid carbon, drought risk, and documented community concerns, each claim backed by cited evidence.
Explore the map