How does AI use water?
AI uses water mainly to remove heat from computing equipment and indirectly through electricity generation. Chip manufacturing adds another lifecycle path. Google’s point-in-time estimate puts a median Gemini text prompt at 0.26 mL under its comprehensive serving boundary, while LBNL’s U.S. data-center totals cover all workloads, not AI alone. No single water number applies to every model, location or prompt.
Why — the first-principles explanation
AI does not consume water because a model has a special thirst. It consumes resources because computation uses electricity and produces heat, and because the infrastructure supplying that electricity and hardware has its own water footprint.
On site, a data center may use water in cooling towers or other evaporative systems to move heat out of the building. Some water is recirculated; some is withdrawn, treated and eventually consumed as vapor or otherwise unavailable to the immediate supply. Other facilities use air cooling or closed-loop liquid cooling, which can reduce direct water use but changes the energy, equipment and location trade-offs.
Upstream, electricity generation can consume water, especially where thermal power plants use water for cooling. That water is not visible at the data-center fence, so a “zero water cooling” claim may describe only the site boundary, not the full electricity-related footprint. Across the lifecycle, semiconductor fabrication, construction and water treatment add additional inputs. These are shared by ordinary data centers too; AI changes the scale, density, workload mix and growth rate, not the underlying physics.
The number depends on the boundary. Google’s August 2025 methodology measured a median Gemini Apps text prompt at 0.24 Wh and 0.26 mL of water, using data from May 2025 and a comprehensive serving boundary that includes active and idle capacity, host systems, overhead and data-center water. It is a Google vendor measurement for one point in time, not a universal AI conversion rate. LBNL’s 2024 U.S. data-center report estimated about 66 billion liters of on-site water consumption and nearly 800 billion liters of indirect electricity-related water consumption in 2023; those totals cover U.S. data centers as a sector, not AI-only workloads.
Training and inference both create heat, but their water footprints depend on different workloads, hardware, utilization, electricity mix, cooling design, climate, water source and accounting method. Always ask whether a figure means withdrawal or consumption, direct or indirect water, one prompt or a fleet, and which year and location. The responsible conclusion is a range with a boundary—not a viral fixed number.
An example that makes it click
Imagine a bakery. Water can be used inside the bakery to wash equipment and cool a machine, upstream by the power station that supplies its electricity, and earlier in the factories that made the ovens. Counting only the water pipe behind the bakery understates the footprint; adding every bakery in a country does not tell you how much water one loaf used. AI accounting has the same boundary problem.
How to do it
- Name the activity: model training, one inference, batch processing, storage, chip manufacturing or the whole data-center fleet.
- State the boundary: on-site cooling, electricity-related upstream water, hardware lifecycle or a combined estimate.
- Check the metric: water withdrawal is water taken from a source; water consumption is the portion not returned to that source in the same condition or place.
- Record the model, hardware, utilization, location, climate, cooling design, electricity mix, time period and whether the number is measured or modeled.
- Use vendor measurements as scoped disclosures, not universal benchmarks. Compare two estimates only when their boundaries and units match.
- For lower impact, reduce unnecessary runs, use an appropriately sized model, cache or batch work where useful, prefer transparent providers and ask for local water-stress and replenishment information.
- When publishing a claim, label uncertainty and distinguish AI-only results from all-data-center or all-electricity totals.
Key facts
- Google’s August 2025 technical methodology estimated a median Gemini Apps text prompt at 0.24 Wh and 0.26 mL of water using data from May 2025; Google says the estimate includes the full serving stack and is not a universal rate.
- LBNL’s 2024 U.S. data-center report estimated roughly 66 billion liters of on-site water consumption and nearly 800 billion liters of indirect electricity-related water consumption in 2023; the figures cover all U.S. data centers, not AI-only activity.
- On-site water can support evaporative cooling, while upstream water can be associated with electricity generation; a site-only water figure does not describe the full lifecycle.
- Water withdrawal and water consumption are different metrics, and climate, location, cooling design, power mix and return flows can change both.
- Training, inference and hardware manufacturing have different workload and lifecycle boundaries; a study estimate for one model or facility cannot be applied to every prompt.
- Google’s measurement methodology includes active and idle machines, host CPU/RAM and data-center overhead, showing why active-accelerator-only estimates can understate serving impacts.
- AI shares data-center cooling and electricity infrastructure with non-AI workloads, so industry totals should not be presented as the water use of AI alone.
Choose an environmental claim with a boundary
Move from a water-footprint question to a verifiable AI workflow, tool comparison and responsible-use decision.
▶ The 60-second explainer (script)
How does AI use water? Mainly through the infrastructure that computes and cools it. Data centers may use water in evaporative cooling, while electricity generation can consume water upstream. Chip manufacturing and construction add lifecycle impacts. The number depends on the boundary. Google’s point-in-time measurement estimated a median Gemini text prompt at 0.26 milliliters under a comprehensive serving method using May 2025 data. LBNL estimated about 66 billion liters of on-site and nearly 800 billion liters of indirect electricity-related water for all U.S. data centers in 2023—not AI alone. Training, inference, climate, hardware, cooling, power mix and accounting method all change the result. Ask whether a figure is withdrawal or consumption, direct or indirect, measured or modeled. A scoped range is more honest than one viral number.
What authoritative sources say
People also ask
Does water touch the AI chips?
Usually the cooling loop is separated from the electronics by heat exchangers. Closed-loop liquid cooling can carry coolant near chips without continually evaporating it, while facility-level cooling and power generation still have their own boundaries.
Does every AI prompt use the same amount of water?
No. Prompt length, model, hardware, utilization, cooling, climate, electricity mix, location and accounting boundary all matter. Google’s 0.26 mL figure is a scoped median serving estimate from May 2025, not a universal conversion rate.
What is the difference between water withdrawal and consumption?
Withdrawal is water taken from a source. Consumption is the portion not returned to that source in the same condition or place, often because it evaporates or is incorporated into a product. Studies must state which metric they use.
Does air cooling eliminate AI’s water footprint?
It can reduce on-site water use, but it may require more electricity and shift impacts upstream. The total depends on the cooling design, power mix, climate and boundary; “waterless cooling” is not automatically “zero water.”
Does AI use more water than ordinary computing?
The cooling physics is shared. AI can increase demand through dense accelerators, high utilization and rapid serving growth, but a sector-wide data-center number cannot be attributed to AI alone without an allocation method.
How can a company reduce AI water use?
Measure a consistent boundary, right-size models, avoid unnecessary retries, batch or cache work where appropriate, choose efficient hardware and cooling, consider reclaimed water and local water stress, and publish uncertainty rather than one flattering number.
The same question, asked other ways
- Why does AI use water?
- Why does AI use so much water?
- How does AI waste water?
- Does AI waste water?
- How is AI using water?
- Why does AI need water?
- How does AI use water and energy?
- Does AI actually use water?
- Why do AI data centers need so much water for cooling?
- Why do data centers use water?
- Why do data centers use so much water?