AI uses a surprising amount of water , but the exact number depends on what you count (cooling, electricity generation, chip manufacturing) and whose estimates you trust.

Quick Scoop

  • Global AI data centers are projected to use on the order of trillions of liters of freshwater per year by the late 2020s.
  • Rough estimates suggest each AI prompt uses a few to ~10 milliliters of water when you include cooling and electricity generation, though figures vary.
  • The same systems that run AI also power clouds, search, and video streaming, so AI is only one slice of overall data‑center water use.
  • There’s an active debate: some experts say “per‑prompt” water use is tiny, others say it’s large when you include training, chip production, and the growing scale of AI.

How much water are we talking about?

Different studies and organizations give different numbers, but they all point in the same direction: AI’s water footprint is big and growing.

  • The World Resources Institute estimate (reported by industry outlets) suggests AI infrastructure could consume around 1.1–1.7 trillion gallons of freshwater annually by 2030 , comparable to all household water use in California in a year.
  • A Morgan Stanley analysis cited in late 2025 put AI data centers’ global annual water consumption at about 1,068 billion liters by 2028 , and noted this was far above previous projections.
  • Other work cited by sustainability writers projects 4.2–6.6 trillion liters of water for global AI demand by 2027 , including direct and indirect uses.

To make that less abstract, think of hundreds of millions of people’s annual household water use being comparable to the amount consumed by AI‑intensive infrastructure.

Water per kWh and per prompt

Because AI runs on power‑hungry hardware in big data centers, many analyses look at water per unit of energy.

  • Typical figures: around 1–2 liters of water per kWh for data centers, with some locations and setups going up to roughly 9 liters per kWh , plus several more liters indirectly for power generation.
  • One estimate for a major cloud provider put direct data‑center water use around 1 liter per kWh , with indirect “Scope 2” water use ~3.5 liters per kWh , giving a conservative ~4.5 L/kWh total when you include the grid.

Translating that into user experience:

  • A popular claim from OpenAI’s CEO is that a ChatGPT query uses about 0.000085 gallons of water (roughly a fifteenth of a teaspoon) per query.
  • Other analyses that spread training and infrastructure costs over all queries, and include grid water use, arrive at about 10 milliliters of water per prompt as an upper estimate.
  • With around 1 billion prompts per day for a large chatbot ecosystem, this rough upper‑bound gives ~10 million liters of water used per day , or billions of liters annually.

So: each single question to an AI assistant does not “drink a bottle of water,” but at scale the water adds up quickly.

Training vs. everyday use

AI’s water use comes from several stages, not just the live chat you see on screen. 1. Training big models

  • Training large models (like GPT‑class systems) runs thousands of GPUs over weeks or months, using huge amounts of electricity and cooling water.
  • Some studies have estimated that training a single frontier‑scale model can consume millions of liters of water , depending on where the data center is and how power is generated.

2. Serving queries (inference)

  • After training, every prompt you send still requires power and cooling, but per‑query water use is far smaller than training.
  • However, if billions of prompts are sent daily across many services, the cumulative water footprint becomes comparable to that of smaller cities.

An analogy: training is like building a giant stadium (huge one‑time resource cost), and inference is like hosting events there every day (smaller but continuous cost).

Where the water actually goes

AI’s water footprint isn’t just “taps at the data center.” It appears at multiple layers.

  • Data center cooling: Many centers use evaporative cooling systems that consume water (not just circulate it) to keep servers from overheating.
  • Power generation: If the electricity is from coal, gas, or nuclear plants, those plants may use large amounts of water for steam and cooling, adding an indirect water cost to each kWh feeding AI chips.
  • Chip and hardware manufacturing: Producing advanced semiconductors is water‑intensive; this “embedded water” often gets overlooked in “per query” numbers, but it exists in the background.

Location matters a lot:

  • In some Microsoft data centers, AI workloads consume 1.8 to 12 liters of water per kWh , with big differences between regions like Ireland and Washington State.
  • Even an “average” 15 MW data center can use as much water as three hospitals or more than two 18‑hole golf courses each year, while hyperscale AI‑focused sites (150 MW and up) are in another league.

Why the estimates don’t agree

If you’ve seen wildly different answers to “how much water does AI use?”, you’re not imagining it. Experts disagree because they measure different things:

  • Some only count on‑site cooling water at the data center.
  • Others include indirect water use from power plants and sometimes chip manufacturing.
  • Some divide total water by just the queries , others spread it across training plus queries , and some even include models that are still in development.

That’s how you can have:

  • A claim of a tiny fraction of a teaspoon per query.
  • A different estimate that equates a chunk of ChatGPT usage to multiple liters per person once you add training, power, and hardware.

Science communicators have pointed out that both “small” and “big” numbers can be technically correct , if you choose different boundaries. The debate is now shifting toward full‑lifecycle accounting so people see the entire picture.

Is AI’s water use “a big deal” or overblown?

You’ll find multiple viewpoints in news, think‑tank reports, and forums. Arguments that it’s a serious concern

  • AI demand is growing fast, and projections into the late 2020s show multi‑trillion‑liter annual water use , often in regions already facing water stress.
  • Activists and local communities have protested new data centers, arguing that AI shouldn’t compete with drinking water and agriculture in drought‑prone areas.
  • If you consider the full lifecycle (training, energy, chips), AI becomes a non‑trivial piece of the tech sector’s water footprint.

Arguments that it’s being exaggerated

  • Some analysts note that the entire AI sector’s water use is still small compared to agriculture, heavy industry, or global household use.
  • Forum defenders of AI often argue that everyday human activities (diet, clothing, showers) use more water than your AI queries , and targeting AI alone misses larger systemic issues.
  • Industry voices highlight that per‑query use is very low , especially compared to things like a single load of laundry or a hamburger.

The nuanced view: AI is not the main driver of global water scarcity right now, but in some local hotspots and future scenarios it can meaningfully add pressure, especially if large AI data centers cluster in already dry regions.

What companies and researchers are doing

Because this is now a public and regulatory issue, big players are experimenting with ways to shrink AI’s water footprint.

  • Alternative cooling : Closed‑loop systems and direct‑to‑chip cooling can drastically reduce evaporative water loss and rely more on recirculated fluids.
  • Location choice : Placing data centers in cooler, wetter regions or near non‑potable sources (like seawater with special cooling designs) reduces competition with local drinking supplies.
  • Scheduling and demand response : Running the most intensive training jobs at times or locations where power is cleaner and water is less constrained.
  • Efficiency improvements : More efficient chips and algorithms mean fewer kWh per query , which directly cuts water use tied to both cooling and electricity.

Some companies have set goals to become “water positive” , meaning they plan to replenish more water than they consume , but how effectively those pledges translate into local outcomes is still being scrutinized.

Latest news and forum vibes

In the last year or two, AI and water has become a trending topic in tech and climate discussions.

  • News outlets and explainers (including broadcasters) now regularly cover how AI uses our drinking water and why data‑center build‑outs have sparked protests.
  • Commentary pieces and newsletters debate whether recent forecasts (like Morgan Stanley’s trillion‑liter projections) show that analysts previously underestimated AI’s water demand.
  • On forums, you see two recurring threads:
    • Posts asking whether viral claims like “a bottle of water every few prompts” are real or exaggerated , with users doing back‑of‑the‑envelope math based on kWh and liters per prompt.
* Counter‑posts insisting that **daily human habits still dwarf AI’s water use** , and urging people to focus on agriculture or energy reform first.

One illustrative breakdown in a community calculation: assuming ~10 ml per prompt , 1 billion prompts per day leads to around 10 million liters/day , which sounds huge, but is still small compared with water use in a large metro area —yet it’s concentrated in specific data‑center regions.

TL;DR (bottom line)

  • On a per‑question basis, AI uses a few milliliters of water at most , depending on how you count.
  • At global scale , AI data centers are on track to use hundreds of billions to several trillion liters of water per year over the next few years.
  • The real issue is where and when that water is drawn: in water‑stressed regions, large AI build‑outs can become a real environmental and political flashpoint.

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