An honest accounting
The real cost of a
writing session.
People are right to worry about AI’s footprint. So here is the math for what this tool spends — text only, every number sourced, the work shown so you can check it.
A full session — your 50 messages and 50 replies
costs about one second
of your morning shower.
That’s the water. The electricity runs near a few phone charges — about what the lamp on your desk burns while you write.
Two estimates below. The conservative figure pads three to four times above what Google and OpenAI measured in production, so the case holds even at the high end. The measured proxy uses their reported numbers direct.
The energy
One session, against things you already plug in.
Measured in watt-hours. A standard iPhone holds about 14 Wh and pulls near 16 Wh from the wall for a full charge.7
Even padded to 1 Wh a reply, a full session draws about what your desk lamp does while you sit and write it.
The water
Now the part the headlines shout about.
This counts both layers: the water that cools the data center, and the larger amount used at the power plant to make the electricity.45 Then set it beside a shower.
No asterisks hidden
What these numbers leave out.
A fair page names its own gaps before a critic does. Three of them.
One. This is the running cost, not the training cost.
Building a model takes a large, one-time burst of energy. Spread across the billions of conversations that follow, it rounds down to a sliver per session — real, not zero, but small once shared out. The figures above measure the day-to-day act of using the tool.
Two. The frightening totals are about industrial scale.
The shocking water and power figures in the news describe entire data-center build-outs, every user at once, and image and video generation combined. Those numbers are true. They are not one person writing. That gap is the whole point of this page.
Three. Text is the light end. This tool is text only.
Generating images and video costs many times more than generating words. This tool makes no images and no video. It takes your pages and hands them back sharper. That keeps it near the floor of what any screen task spends.
Show your work
The math, open for inspection.
No company has published a per-message figure for the model behind this tool. So these estimates lean on the two firms that did publish production data, padded upward for safety.
The assumptions, in plain numbers
A “session” here means 100 messages — 50 from you, 50 replies — or 50 response-cycles.
| Input | Conservative |
|---|---|
| Energy per reply | 1 Wh |
| Energy, full session (×50) | 50 Wh |
| Phone charge from the wall | ~16 Wh |
| Water per kWh generated | ~2 L |
| Data-center cooling water | ~1 mL / Wh |
| Session water (cooling + generation) | ~150 mL |
| Shower flow | 2.1 gal/min (~8 L) |
For reference, Google measured a median text prompt at 0.24 Wh and 0.26 mL of water;1 OpenAI put a query near 0.34 Wh.2 The 1 Wh figure used here sits three to four times higher on purpose. Swap in the measured numbers with the toggle above and every total shrinks.
Sources
- Google Cloud — Measuring the environmental impact of AI inference (median Gemini text prompt: 0.24 Wh, 0.26 mL water, under 9 seconds of TV).
- Data Center Dynamics — reporting OpenAI’s figure of ~0.34 Wh and ~one-fifteenth of a teaspoon of water per query.
- MIT Technology Review — analysis of Google’s per-prompt energy disclosure.
- University of Illinois CEE — the two water layers: data-center cooling and electricity generation.
- NREL — Consumptive Water Use for U.S. Power Production (~1.8 L/kWh thermoelectric; 7.6 L/kWh aggregate).
- Home Water Works — average shower: 2.1 gal/min, ~8 minutes, ~60 liters. · EPA WaterSense — showerhead flow standards.
- Macworld — iPhone battery capacities in mAh and watt-hours.