Rebuilding an AI water calculator, energy first

What I changed

The old model started from water. The new one starts from energy, because the water figure follows from the electricity used. OpenAI puts an average ChatGPT query at about 0.34 Wh, and Google reports 0.24 Wh for a median Gemini text prompt, along with 0.26 mL of cooling water. That works out at roughly 1 mL of water per Wh at the data centre.

Data centres also use water indirectly, through the power stations that supply them. Adding that (about 3.1 mL per Wh on the US grid average) gives 4.1 mL per Wh. So a typical text prompt comes to about 0.3 mL counting cooling only, or about 1.4 mL counting the electricity as well. The calculator shows both, labelled, and every estimate now carries a likely range of 0.3× to 3×.

Why published figures differ by 100×

Published figures for one prompt run from about 0.3 mL to 50 mL. Most of them are defensible: they measure different things. Four choices account for most of the spread.

Choice	Effect
When it was measured: Google reports a 33× drop in energy per prompt in twelve months	~33×
Where the boundary is drawn: cooling only, or cooling plus electricity generation	~4×
Whose data centre: Amazon 0.12, Microsoft 0.27 and Google about 1.08 L per kWh	~9×
What is measured: a chat prompt, or a coding agent re-reading its context on every turn	orders of magnitude

That is also why Sam Altman’s comparison of 38,000 ChatGPT queries to one almond is self-consistent but flattering. It sets the almond’s total water footprint (about 12 L, rain included) against the cooling-only figure for a query. Match the boundaries and it is roughly 13,000 queries against the almond’s irrigation water, or about 8,600 counting electricity too. Thousands of queries per almond is fair.

What else turned up

Re-checking every source turned up one I had cited wrongly. The 3 Wh figure I use for images and long reasoning was attributed to a news analysis that, on a re-read, didn’t support it. Benchmark data does: Jegham et al. (2025) put reasoning queries at roughly 1 to 12 Wh depending on length. The figure stayed the same; the source changed, and the changelog says so.

I also added a token-based model for coding assistants. They re-read the conversation and your files on every turn, so an hour building a feature comes to roughly 90 Wh and 370 mL – around 270 text prompts’ worth.

What I took from it

• A water figure without a boundary can’t be checked. Most disagreements about AI water use are two people counting different things. Stating it matters more than adding decimal places.
• Re-read your sources, not just your arithmetic. The maths was right. One citation wasn’t, and it took a full re-check to see it.
• Version the method, not just the code. The method is published as a citable note with a DOI, every result carries its methodology version, and a changelog records what changed and why. Corrections are then part of the record rather than something to hide.

The same figures drive the calculator, a JSON API and MCP server and an embeddable widget, and tests pin the server to the page. I built and checked this with Claude as a research and engineering partner. The method, judgement calls and conclusions are mine. Method note: doi:10.5281/zenodo.23375361.