Why Does AI Use Water?

01 / The physical explanation

Why does artificial intelligence use water?

When you submit a question to ChatGPT or another AI service, specialized computer chips perform calculations in data centers. Their electrical energy ends up primarily as heat. Every cooling system must move that heat away. A facility may use air cooling, a closed liquid loop, evaporative cooling, or a combination; liquid cooling is not the same as consuming fresh water.

In an evaporative cooling tower, warm water circulates, some evaporates and carries heat into the air, and fresh make-up water replaces that loss. A portion is also drained to manage mineral build-up. Other facilities use dry coolers with little on-site water consumption, sometimes at the cost of greater electricity needs. The trade-off changes with climate and engineering design. The U.S. Department of Energy explains the mechanisms and WUE metric.

Four different water footprints people often mix together

The important distinction: water withdrawal is not water consumption. Water returned to the watershed after use is different from water evaporated or moved elsewhere. A high withdrawal number must not be relabeled “water used up.” See the 2026 water-accounting research.

02 / Numbers with context

How much water does one ChatGPT question actually use?

A June 2025 statement by Sam Altman put an average ChatGPT query at approximately 0.000085 US gallons, or 0.32 mL of water, and 0.34 Wh of electricity. It was a public estimate, not a full reproducible, model-by-model measurement with all lifecycle boundaries reported. Read the original statement.

Google’s more explicit May 2025 median Gemini Apps text prompt figure was 0.26 mL and 0.24 Wh. Google disclosed a methodology including serving hardware and overhead, but its cited water metric uses data-center cooling water rather than a full lifecycle total. Google explicitly warned that its prompt-level numbers are point-in-time and not independently verified. Read Google's method.

Meanwhile, Mistral reported 45 mL for a 400-token Le Chat response using a lifecycle-oriented assessment that includes broader upstream impacts and excludes the user device. That does not mean one model is automatically over 100 times thirstier than another. The measurements are not the same experiment. See Mistral’s disclosure.

Published figures: the comparison most headlines skip

The GPT-3 row is a range inferred from an estimate for a group of responses, not a measured ChatGPT query. The figures come from different dates, systems and scopes and should never be ranked as equivalent apples-to-apples measurements.

Why the numbers disagree

1. Different requests: a short text reply, a long reasoning trace and a video are not comparable workloads.
2. Different statistics: median is not the same as arithmetic average.
3. Different facilities: water cooling, dry cooling, temperature, grid and utilization vary.
4. Different boundaries: on-site cooling, electricity, manufacturing and training are different totals.
5. Different years: faster hardware and improved batching can lower per-query energy while total use rises.
6. Different confidence: a source-backed measured sample and a public estimate have different evidentiary weight.

04 / Original SXF explainer

SXF Claim Decoder: what does a water-per-prompt headline leave out?

Choose a published claim and examine its measurement boundary. This decoder does not rank models: it checks what the source actually supports and what remains unknown.

05 / SXF original interactive model

SXF Water Footprint Lab: build a transparent scenario

This is not a ChatGPT meter. Instead of assigning every prompt a fabricated universal footprint, change the assumptions and watch the calculation. The lab separates on-site cooling water from water consumed to generate electricity. It deliberately excludes manufacturing and training because estimating those needs more information.

The default inputs are round, hypothetical teaching values chosen to make the math inspectable. They are not Google’s, OpenAI’s or Mistral’s actual provider settings. Read about WUE and the original research.

06 / Energy explained

How much electricity does ChatGPT or AI use?

Electricity use is commonly expressed as watt-hours (Wh). One watt-hour is the electricity used by a one-watt load over an hour. A kilowatt-hour (kWh) is 1,000 Wh. A June 2025 public estimate put an average ChatGPT query at approximately 0.34 Wh. Google reported 0.24 Wh for its median Gemini Apps text prompt in May 2025. Both figures are snapshots, not rates that apply to every feature or model.

Microsoft’s 2026 production-aware research found a median estimate of 0.31 Wh per typical query across the systems and assumptions it modeled. It also found long reasoning and agentic tasks can use much more energy. Estimates for one short text answer should not be multiplied uncritically across image generation, video, web research and multi-step agents. Read Microsoft Research.

How much electricity do data centers use worldwide?

The International Energy Agency’s updated 2026 outlook estimated 485 TWh in 2025 and around 950 TWh in 2030 for all data centers worldwide in its central projection. These are sector-wide data-center figures, not an electricity meter for ChatGPT alone. IEA expects AI-focused capacity to grow faster than the overall total, but the infrastructure also supports storage, business software and other digital services. IEA’s updated assessment.

07 / The wider picture

Is AI bad for the environment?

AI has environmental costs; a single water-per-query figure cannot determine its net impact. Important outcomes include electricity demand, grid carbon intensity, land use, manufacturing, water withdrawal, local water stress and the possible environmental benefits of the task AI enables. Claims that AI is harmless or uniquely catastrophic from one number both miss this context.

Local water stress matters more than a global average

One liter consumed in a region experiencing drought can have different consequences from one liter consumed in a water-abundant location. Operators also differ in whether they withdraw potable freshwater or treat and reuse reclaimed water. Comparing only global “drops per prompt” can hide impacts on nearby communities and water systems.

Does dry cooling solve AI water consumption?

It can greatly reduce on-site water consumption, but may increase the electricity required for heat rejection under some conditions. The water and emissions effects depend on the local grid, outdoor temperature and engineering design. “Zero water for cooling” is not the same as “zero total water footprint.”

Training versus using a model

Training can consume energy and water over a concentrated period, while model inference consumes resources every time it is used. The balance depends on the model’s size, the duration of its deployment, the number of requests and whether a wider lifecycle is counted. Hardware manufacturing should also be examined, rather than automatically divided by a guessed number of queries.

What would responsible reporting look like?

• Name the specific model and workload, rather than saying “AI” without qualification.
• Give the date, region, sample size and median or average.
• Separate water consumption from withdrawal and discharge.
• Separate on-site cooling from water associated with electricity production.
• Show whether training, hardware and end-user devices were included.
• Publish energy, PUE, WUE, uncertainties and methods where possible.
• Discuss local water stress, not only global totals.

08 / Practical actions

What can AI providers and users actually do?

Providers can publish reproducible workload data; improve model and serving efficiency; use climate-appropriate cooling; consider reclaimed water where safe; select grid and cooling configurations jointly; and evaluate impacts on local watersheds. Reducing an isolated per-query number is not enough if total demand and local stress are ignored.

Users can choose the right tool for the task, avoid repeated unnecessary high-compute runs, and seek verifiable environmental reporting. Do not assume that a short question produces a particular amount of water or that every use of AI displaces zero-emission work. A better decision asks what actual benefit was obtained per unit of resource use.

For more technical decisions on hosted versus local systems, see SXF’s guide to running AI models locally, AI model selection guide and AI agent cost and total-ownership analysis. For an accessible explanation of why model answers can be unreliable, read why ChatGPT makes things up.

09 / Common questions

AI water and electricity: frequently asked questions

10 / Evidence record

Sources, evidence levels and how SXF keeps this guide current

This is an editorial synthesis, not a laboratory measurement of private provider systems. Company measurements and statements are attributed; modeled academic estimates are identified as estimates; the original interactive lab is explicitly hypothetical. Published figures are included with their dates so future changes do not silently rewrite history.