Uncovering AI & water usage

In this video online creator Hank Green, a prominent science communicator behind YouTube mainstays such as _SciShow_ and _Crash Course_ discusses AI and water usage and who is telling the story and how the way they count it matters.

- Sam Altman claims an average ChatGPT query uses about 0.000085 gallons* of water, roughly *1/15 of a teaspoon.

- Morgan Stanley projects AI data centres’ annual water use (cooling + electricity generation) could reach about 1,068 billion litres (~1 trillion litres)* by **2028**, an *11× increase from 2024 estimates.

- Both numbers can be “true” because they measure different parts of the lifecycle* and use *different boundaries for what counts as “water use”.

Why the numbers diverge so much

1. What’s included in the calculation?

- Per‑query figures* (like Altman’s) typically count only the *on‑site cooling water tied to that single inference.

- System‑level projections (like Morgan Stanley’s) include:

- Scope 1: On‑site data centre cooling.

- Scope 2:* Water used in *electricity generation (especially thermoelectric power plants).

- Scope 3:* Water for **semiconductor manufacturing**, which can require up to *5 million gallons of ultrapure water per day.

2. Training vs inference

- A large share of AI’s resource use comes from training* models, not just running queries. Training can account for around *50% of total resource use for a model.

- Per‑query stats often exclude training entirely, even though the model couldn’t exist without it.

- There’s also ongoing and future training for newer models, which is part of the build‑out driving the 2028 projections.

3. Power‑plant water accounting

- Thermoelectric power plants (coal, gas, nuclear) account for about 40% of all freshwater withdrawals in the US.

- Some AI water estimates attribute a share of this intake-and-return water to AI based on electricity demand, which massively inflates the “water use” number compared to counting only evaporated/consumed water.

- Only about 2–3% of that power‑plant water is lost to evaporation; the rest is returned to the source, albeit warmer.

Types of water matter

- Municipal (potable) water used for data centre cooling competes with other urban uses and relies on treatment infrastructure.

- Ultrapure water for chip manufacturing is far more energy‑intensive to produce than drinking water, even if the volume is smaller.

- Power plants and some data centres often use non‑potable water taken directly from rivers, lakes, or the sea, which is a different stressor than tapping municipal supplies.

Context: AI vs other water uses

- US corn production uses around 20 trillion gallons of water per year*, roughly *80× more than estimated global AI data centre use (~260 billion gallons).

- About 40% of US corn* goes to *ethanol for fuel, not direct human food, meaning tens of trillions of gallons are evaporated annually to make transport fuel.

- Residential lawn irrigation in the US also consumes trillions of gallons of (often municipal) water each year.

Key takeaways

- It’s easy to understate AI water use by:

- Counting only per‑query cooling,

- Excluding training, chip manufacturing, and power‑generation water.

- It’s also easy to overstate it by:

- Counting all water flowing through power plants (most of which is returned) without clarifying consumption vs withdrawal.

- The location matters as much as the volume: building water‑cooled data centres in arid regions is far more problematic than in water‑rich areas.

- The speaker is more concerned about AI’s power demand and carbon impact than its water use in aggregate, though local water stress can still be serious.

Photo by Patrick Pahlke on Unsplash