MTS

Wiki

The 500 mL Myth

The single most-cited datapoint in the AI-water discourse is some version of:

"ChatGPT drinks a bottle of water every time you ask it a question."

This is wrong. Tracing where it came from is the most useful single exercise in this wiki, because almost every downstream claim — "AI is draining the planet," "your emails are killing aquifers," etc. — depends on it.

The original paper

Pengfei Li, Jianyi Yang, Mohammad A. Islam, Shaolei Ren, "Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models" (UC Riverside / UT Arlington, April 2023). arXiv:2304.03271

The paper itself is honest scholarship. Its central empirical claim, in plain English:

A 500 mL bottle of water covers roughly 10 to 50 medium-length GPT-3 responses, depending on where and when the request is served, when running on Microsoft Azure data centers in 2022.

That works out to 10–50 mL per query, not 500 mL per query. The paper also separates direct (on-site cooling) water from indirect (electricity-generation) water, with the indirect share dominating.

How "10–50 mL per query" became "500 mL per query"

Three mistakes, often compounded:

  1. Reading "per 10–50 prompts" as "per prompt." Many summarising articles (and several headlines) inverted the ratio.
  2. Treating "per page of output" as "per query." GPT-3's reported energy figure was 0.004 kWh per page. Modern ChatGPT conversations are typically 1–2 pages, not 10–70. Sean Goedecke shows this single mistake accounts for ~10× inflation.
  3. Using GPT-3 / 2022 Azure as a present-day baseline. Hardware (H100 → GB200), model architecture (dense → sparse / MoE), serving stack (single-prompt → batched), and PUE (~1.5 → ~1.1) have all shifted by roughly an order of magnitude in efficiency since the measurement.

10× × 10× = 100×. That is roughly the gap between "5 mL per conversation" (Goedecke, Oct 2024) and "500 mL per query" (the viral version).

The Washington Post amplification

The October 2023 Washington Post piece "A bottle of water per email" was the dominant vector for the misreading. Its calculation required every one of the following to hold simultaneously:

  • 10 prompts per email (high)
  • Power served from a water-stressed grid (Washington state hydro)
  • Counting evaporation from hydroelectric reservoirs as data-center "consumption"
  • 2022 Azure efficiency, applied unchanged to 2023 services
  • Worst-case evaporative cooling tower assumption

Strip any one and the number collapses; strip all five and you land within an order of magnitude of Altman's 0.32 mL.

The Karen Hao / Empire of AI error

Karen Hao's 2025 book Empire of AI repeated the inflated figure and added several of its own. Andy Masley's response post documents at least one number off by a factor of 1,000. The book has since become the most-cited public source for AI-water alarm in mainstream press.

This matters because the misreporting has become self-sustaining: outlets cite Hao, Hao cites the WaPo piece, the WaPo piece misreads Ren, and Ren's actual paper reports a number ~50× lower than what gets attributed to it.

What Ren himself says

Ren has been clear in interviews that the headline framing of his work is wrong, and that water use varies by orders of magnitude across data center locations, time of year, and cooling design. The paper's policy recommendation — better disclosure and locational accounting — is not "AI is a water disaster." It is "we should be able to know."

Bottom line

The "500 mL per query" claim is dead on every honest reading:

Source Per-query estimate
Viral framing (WaPo, Hao, social media) ~500 mL
Ren et al. 2023 (correctly read, GPT-3) 10–50 mL
Ren et al. 2023 + scope-2 (one US-avg run) 16.9 mL
Goedecke Oct 2024 (modern model, modern stack) ~5 mL per conversation
OpenAI / Altman, Gentle Singularity (Jun 2025) 0.32 mL per query

Two to three orders of magnitude. It is not a close call.

Sources cited on this page