A FIELD GUIDE TO INFERENCE COOLING

Every prompt
drinks.

Large language models run on racks that get hot, and the cheapest way to cool most of them is to evaporate fresh water into the air. This is a plain accounting of how much, drawn from the few numbers companies and researchers have actually published.

ESTIMATED GLOBAL AI COOLING WATER
SINCE YOU OPENED THIS PAGE

0 liters

A rough illustration, not a live meter — modeled from published per-query estimates and public traffic figures. See methodology below.

01

The pour, per query

0.5 L
One bottle of water
the standard unit everyone already understands
0.519 L
A 100-word email drafted by GPT-4
roughly one bottle, per the UC Riverside / UT Arlington estimate
65 L
A 5-minute shower
a standard low-flow showerhead at 13 L/min
0.5 L
25–50 short chatbot exchanges
Ren et al.'s range for a single bottle's worth of cooling water
02

What one training run costs

GPT-3 (175B), trained in Microsoft US data centers
Iowa, U.S.
700,000 L

direct on-site cooling water only — excludes water used to generate the electricity

GPT-3, same run trained in Asia instead
hypothetical Asia-Pacific site
1,400,000 L

Ren et al.'s modeled figure if the same cluster ran in a hotter, more humid region

Llama 2 (70B), full training run
Meta data centers
435,000 L

estimate from Meta's own model card, on-site cooling only

03

What the disclosures say

Google

2023
6.4Bgallons withdrawn

up 17% year over year, tracked against 2022 in its own environmental report

Microsoft

2022
1.7Bgallons withdrawn

up 34% year over year, the steepest jump the company had disclosed, coinciding with its AI buildout

Neither company breaks out the AI-specific share of these totals — data centers also run search, cloud storage, and productivity software. The jump in growth rate is the clearest signal available of what generative AI added on top.

04

Why it's evaporated, not returned

Most large data centers cool their servers with evaporative towers: warm water is sprayed over a mesh, a fraction turns to vapor and carries the heat away, and the rest recirculates. It's cheap and energy-efficient — far less electricity than running chillers around the clock — but the water that evaporates doesn't come back down the drain. It leaves the local watershed as humidity.

That's the direct, "on-site" water. There's a second, larger pool of "off-site" water used upstream to generate the electricity itself, mostly at thermoelectric power plants that cool turbines the same way. Most public model-level estimates, including the ones on this page, cover on-site cooling only — so they understate the total.

05

Where the pressure shows up

The Dalles, Oregon
city officials fought a two-year legal battle to keep a local data center’s water use sealed before releasing it in 2022
Mesa, Arizona
new campuses draw on a basin already managed under a state-declared groundwater shortage
Zeewolde, Netherlands
a proposed hyperscale campus was blocked by the town council after public pressure over water and land use
Cerrillos, Chile
a planned data center was withdrawn after community opposition during the country’s worst drought in a decade
06

Podcasts on AI's water use