AI uses water in two big ways: to run itself (cooling data centers and power plants) and to manage water in the real world (monitoring leaks, floods, and irrigation).

Quick Scoop: How AI Uses Water

1. Why AI needs water at all

Every time you send a prompt, your request runs on huge server farms that eat a lot of electricity and generate heat. That heat has to be removed, and the cheapest way in many places is: water.

  • Data centers spray or circulate water through cooling systems so servers do not overheat.
  • Power plants that generate the electricity for AI often use water for cooling, too.
  • Manufacturing the chips and servers behind AI (fabs, server plants) also consumes large volumes of water.

A UK government analysis breaks it into three buckets: direct cooling at data centers, indirect water for electricity generation, and water used to manufacture the hardware.

2. How much water are we talking about?

Numbers vary by location, design, and how “AI‑heavy” a data center is, but the trend is rough: more AI → more water.

Some snapshots:

  • A typical 100‑megawatt data center in the US can use about 2 million liters of water per day for cooling, roughly similar to the water use of around 6,500 households.
  • Researchers have estimated that training a large model like GPT‑3 at a modern US data center could consume on the order of 700,000 liters of clean freshwater.
  • One policy article notes that AI’s expansion could push global water usage from around 1.1 billion to 6.6 billion cubic meters by 2027, which is more than half of the UK’s total annual water use.
  • Some analyses estimate that a single mid‑length AI prompt (around 100 words) can indirectly equate to about a small bottle of water when you factor in cooling and power.

One article describes AI as having a “hidden thirst” behind every chat reply, image generation, video, playlist recommendation, or translation we casually trigger.

3. Where that water comes from (and why it’s controversial)

The controversy isn’t just the amount of water, but where it’s taken from.

  • Many new AI‑heavy data centers have been sited in already water‑stressed regions in the US, where there is high competition between cities, agriculture, and industry.
  • Bloomberg’s analysis found over 160 new AI‑oriented data centers in three years in US regions with scarce water, a jump of about 70% over the previous period.
  • Using potable drinking water for server cooling has triggered local pushback when communities are facing restrictions or drought.

On top of direct water use at the facility, a big part of AI’s “water footprint” is upstream in power generation: water used to cool thermal power plants that produce the electricity feeding the data centers.

4. How engineers actually cool AI systems

Cooling is where most of the direct water use happens.

Common approaches:

  • Evaporative cooling towers
    Warm water absorbs heat from servers, is sent to towers, and part of it is evaporated into the air, taking heat away. That evaporated water does not neatly condense back into the same reservoir; it becomes part of the broader atmosphere.
  • Direct evaporative cooling (air‑side)
    Outside air is pushed through wet pads; as the water evaporates, it cools the air that’s blown into server rooms. This can reduce electricity use, but it increases water use because the evaporated water is lost to the air.
  • Closed‑loop liquid cooling
    Water or coolant circulates in sealed pipes or direct‑to‑chip systems, transferring heat to a secondary system that can sometimes use less water overall.
  • Alternative sources: recycled or non‑potable water
    Some large providers now commit to using reclaimed or recycled water rather than drinking water. For instance, one major cloud company announced expanding recycled water use to around 100 US data centers, aiming to preserve hundreds of millions of gallons of drinking water in surrounding communities.

The trade‑off engineers juggle is: electricity vs water. Some designs save power but use more water; others save water but use more energy or cost more to build.

5. The “water footprint” of AI: how it’s calculated

Organizations like the OECD treat AI’s water impact as a water footprint you can estimate with a formula that ties together server energy use and cooling efficiency.

One formula described is:

WaterFootprint=ServerEnergy×WUEonsite+ServerEnergy×PUE×WUEoffsiteWaterFootprint=ServerEnergy\times WUE_{onsite}+ServerEnergy\times PUE\times WUE_{offsite}WaterFootprint=ServerEnergy×WUEonsite​+ServerEnergy×PUE×WUEoffsite​

  • ServerEnergyServerEnergyServerEnergy is the energy used by the IT equipment.
  • WUEonsiteWUE_{onsite}WUEonsite​ is water‑use efficiency at the data center itself (liters per kWh consumed for direct cooling).
  • PUEPUEPUE (power usage effectiveness) indicates how much extra electricity is spent on cooling and other overhead beyond pure computing.
  • WUEoffsiteWUE_{offsite}WUEoffsite​ measures how much water is used per kWh at power plants supplying the electricity.

This is how researchers get from “X kWh of AI compute” to “Y liters of water,” both inside the data center fence and upstream in the grid.

6. AI is also being used to save water

Ironically, AI is not just using water; it’s also being deployed to protect and optimize water systems.

Examples include:

  • Digital twins of water networks that forecast demand and detect leaks in pipes so utilities can fix them faster.
  • AI models that improve irrigation scheduling in agriculture by combining weather, soil moisture, and crop data to cut water use.
  • Flood prediction and early‑warning systems that analyze rainfall, river levels, and terrain to reduce damage from extreme weather.
  • Condition monitoring in treatment plants so pumps, filters, and reservoirs operate more efficiently and with less waste.

The big open question is whether AI’s benefits in water management can offset its own expanding water footprint, especially as demand for larger, more capable models grows.

7. What people are saying online (forum flavor)

Forum and social threads around “how does AI use water?” usually circle a few recurring themes:

  • Confusion about evaporation vs “reusing” water: some users assume that evaporated cooling water condenses right back into the same local system, but others point out that it disperses into the wider atmosphere and is effectively lost to that local basin in the short term.
  • Debates about whether we should just “build in rainy or cold places” to minimize water use, with pushback that land, grid access, and latency to users complicate that.
  • Skepticism over viral claims that “each prompt uses X bottles of water,” and requests for math checks on those numbers.

One government sustainability blog goes as far as calling AI’s water appetite a major driver of future global water demand if data‑center growth continues unchecked.

8. Efforts to shrink AI’s water footprint

There is active work on making AI less water‑hungry.

Some approaches:

  • Better siting
    Placing data centers in cooler climates or near abundant non‑potable water sources so they don’t compete as directly with households and farms.
  • Recycled water and alternative sources
    Using treated wastewater, industrial effluents, or seawater (with appropriate engineering) instead of freshwater can significantly reduce pressure on drinking water supplies.
  • More efficient cooling
    Direct‑to‑chip liquid cooling, closed‑loop systems, and hybrid designs can cut water use while controlling temperature for increasingly dense AI chips.
  • Model and hardware optimization
    Designing more efficient models (smaller, task‑specific, or better trained) and more efficient chips reduces total energy per query, which brings both energy and water use down.
  • Transparency and reporting
    Governments and NGOs are pushing for standardized disclosures of water use alongside carbon metrics so buyers and regulators can compare providers more fairly.

One analysis suggests that by 2030, AI infrastructure could consume on the order of 1.1 to 1.7 trillion gallons of freshwater annually if current trajectories continue, comparable to the yearly residential water use of a large US state.

9. Big picture: “How is AI using water?” in one line

AI uses water directly to cool the servers and chips that run models, indirectly through the power plants and factories behind them, and constructively when those same models are applied to conserve and manage water systems in the real world.

Information gathered from public forums or data available on the internet and portrayed here.