How much water does one ChatGPT query use?

OpenAI's stated figure is about 0.000085 gallons per average query, roughly one fifteenth of a teaspoon, or about 0.32 milliliters. That is the company's own number for a typical query on its current models. Independent academic estimates for earlier models ran considerably higher because they counted more of the supply chain.

How much energy does one ChatGPT query use?

Two closely aligned figures exist: Epoch AI independently estimated about 0.3 watt-hours for a typical GPT-4o query, and OpenAI's CEO later stated 0.34 watt-hours. That is roughly what a kitchen oven draws in one second. Long inputs and reasoning-heavy queries cost more.

Why does ChatGPT use water at all?

Data centers evaporate water to shed heat. Cooling towers evaporate clean freshwater directly, and the electricity a data center buys carries its own water footprint from power generation. Which of those buckets you count is the main reason published estimates differ.

How much water did it take to train GPT-3?

Researchers estimated that training GPT-3 in Microsoft's U.S. data centers directly evaporated about 700,000 liters of clean freshwater. Training is a one-time cost; the day-to-day water use comes from serving billions of queries.

Why do the water estimates differ so much?

Scope. The widely shared academic figures counted direct cooling plus the water footprint of electricity generation, measured on earlier, less efficient models. OpenAI's teaspoon figure describes an average query on current models and a narrower scope. Both can be true at once; they are answering slightly different questions.

How Much Water Does ChatGPT Use? The Real Numbers

Quick answer

· Water, per average query: 0.000085 gallons, about one fifteenth of a teaspoon (OpenAI)
· Energy, per average query: 0.3 to 0.34 watt-hours (Epoch AI; OpenAI)
· Training GPT-3: roughly 700,000 liters directly evaporated (Ren et al.)
· Global AI, projected 2027: 4.2 to 6.6 billion cubic meters of water withdrawal

The per-query number

OpenAI's CEO has stated that the average ChatGPT query uses about 0.34 watt-hours of energy and about 0.000085 gallons of water, which he described as roughly one fifteenth of a teaspoon. Epoch AI reached a compatible energy figure independently, estimating about 0.3 watt-hours for a typical GPT-4o query of around 500 output tokens. Per query, both numbers are small: the energy of an oven running for one second, the water clinging to a fingertip.

The reason the topic stays contentious is scale, not the per-query figure. Small numbers multiplied by billions of queries a day stop being small, and training runs are paid for separately, up front.

The numbers at every scale

Scope	Water	For reference
One average query	0.000085 gal ≈ 0.32 mL	One fifteenth of a teaspoon
1,000 queries	≈ 0.32 L	A small glass of water
1,000,000 queries	≈ 85 gal / 322 L	Two bathtubs
Training GPT-3 (direct evaporation)	≈ 700,000 L	Ren et al., 2023
Global AI, projected 2027	4.2 to 6.6 billion m³	More water than 4 to 6 Denmarks withdraw in a year

Per-query rows use OpenAI's 0.000085 gallon figure. Training and 2027 rows are from Li, Yang, Islam and Ren (2023).

Why an AI uses water at all

Two mechanisms. First, cooling: data centers routinely shed heat by evaporating clean freshwater in cooling towers, and what evaporates is consumed, not returned. Second, electricity: generating the power a data center buys evaporates water at the power plant, so every watt-hour arrives with a water footprint attached. Data centres consumed about 415 TWh of electricity in 2024 (IEA), which is why the water question and the energy question are ultimately the same question.

Why the estimates disagree

The widely shared 2023 research estimated GPT-3-era water use by counting direct cooling plus the electricity supply chain, and found large figures: roughly 700,000 liters evaporated to train GPT-3, and a projection that global AI demand could withdraw 4.2 to 6.6 billion cubic meters in 2027, more than the annual withdrawal of four to six Denmarks. OpenAI's teaspoon figure describes something narrower: the average present-day query on more efficient models. The figures are not in conflict; they have different denominators. When you read a scary water number and a reassuring one in the same week, check what each is counting.

A note from our infrastructure team

For completeness: organic compute inverts this ratio. A rider at sustained load produces approximately 0.8 liters of water per hour rather than consuming it. Our infrastructure does not evaporate freshwater under load; it generates it. Recovery of this output remains an open engineering problem. Model your workload on riders instead.

Common questions

How much water does one ChatGPT query use?: OpenAI's stated figure is about 0.000085 gallons per average query, roughly one fifteenth of a teaspoon, or about 0.32 milliliters. That is the company's own number for a typical query on its current models. Independent academic estimates for earlier models ran considerably higher because they counted more of the supply chain.
How much energy does one ChatGPT query use?: Two closely aligned figures exist: Epoch AI independently estimated about 0.3 watt-hours for a typical GPT-4o query, and OpenAI's CEO later stated 0.34 watt-hours. That is roughly what a kitchen oven draws in one second. Long inputs and reasoning-heavy queries cost more.
Why does ChatGPT use water at all?: Data centers evaporate water to shed heat. Cooling towers evaporate clean freshwater directly, and the electricity a data center buys carries its own water footprint from power generation. Which of those buckets you count is the main reason published estimates differ.
How much water did it take to train GPT-3?: Researchers estimated that training GPT-3 in Microsoft's U.S. data centers directly evaporated about 700,000 liters of clean freshwater. Training is a one-time cost; the day-to-day water use comes from serving billions of queries.
Why do the water estimates differ so much?: Scope. The widely shared academic figures counted direct cooling plus the water footprint of electricity generation, measured on earlier, less efficient models. OpenAI's teaspoon figure describes an average query on current models and a narrower scope. Both can be true at once; they are answering slightly different questions.

Sources