# 500 mL for roughly 10–50 GPT-3 responses—not every prompt

Agent contract: 1.2.0

The 500 mL estimate modeled roughly 10–50 medium GPT-3 responses, including cooling and electricity water. It is not a universal per-prompt meter reading.

Canonical: https://brightaifuture.com/bright-check/water-per-prompt
Format: check
Source publication: Not established
Bright publication: 2026-10-24
Substantive update: None recorded
Evidence and review: Source interpretation checked 2026-10-24; not an independent replication.

## The claim

One ChatGPT query uses a 500-milliliter bottle of water.

The original paper estimated 500 milliliters for roughly 10–50 medium-length GPT-3-style responses, depending on location. It assumed 0.004 kilowatt-hours per request and combined onsite cooling with electricity-generation water. It did not meter today’s ChatGPT, a named facility, or every prompt.

## What the original study counted

The paper modeled about 500 mL of operational water consumption for roughly 10–50 medium GPT-3 responses, depending on place and time. Its boundary includes onsite cooling and electricity-generation water. This is not a meter reading of today’s ChatGPT or every prompt.

## A separate email example

The Washington Post and UC Riverside modeled about 519 mL for GPT-4 to write a 100-word email in 2024. That is a different model and task, not a correction of the GPT-3 paper or a universal current average.

## What could change

Model-, hardware-, token-, site-, hour- and cooling-specific metered energy and water data, with the electricity source and accounting boundary disclosed.

## Limitations

These estimates describe different models and tasks and cannot be combined into a current universal per-prompt average.

Local water source, season, consumption versus withdrawal, cooling design and electricity mix matter.

Bright checked source interpretation but did not independently replicate either model.

## Provenance and history

{}

## Original sources

- [Making AI Less “Thirsty”](https://arxiv.org/html/2304.03271v5)
- [Journal edition of the original study](https://doi.org/10.1145/3724499)
- [The Washington Post and UC Riverside: a bottle of water per email](https://www.washingtonpost.com/technology/2024/09/18/energy-ai-use-electricity-water-data-centers/)

## Continue exploring

- [Full Bright water guide](https://brightaifuture.com/guides/data-center-water-cooling)
- [Bright Check series](https://brightaifuture.com/series/bright-check)
- [Corrections](https://brightaifuture.com/corrections)