Can someone explain how AI systems actually use water?

General
1

I recently read that training and running large AI models can consume a surprising amount of water for data center cooling and power generation, but the articles I found were either super technical or too vague. I’m trying to understand where and how this water use happens, how big the impact really is, and whether different AI providers are more efficient than others. Can anyone break this down in simple, practical terms or point me to solid resources so I can better grasp AI’s real-world water footprint?

2

You are not wrong, AI does use water, and the numbers get big fast.

There are two main paths.

  1. Data center cooling
  2. Power plant cooling for the electricity the data center uses

Breakdown:

  1. Data center side

    • Servers run hot. Operators keep them around 70–80°F.
    • To cool them, many sites use evaporative cooling. Water evaporates, carries heat away, and lowers temps.
    • That water leaves as vapor and needs replacement.
    • Some centers use chillers with closed-loop systems, which use less water but more electricity.

    Rough order-of-magnitude numbers from recent studies and reports:

    • Training a large model like GPT-3 level used on the order of 200–700k liters of water, depending on location and cooling setup.
    • Using one AI model for a few dozen prompts is small. Something like a few hundred milliliters per 20–50 prompts has been estimated in some research, but this varies a lot by region and data center design.
    • Big cloud providers run many models and millions of queries daily, so totals stack up.

  2. Power generation side

    • Data centers draw a lot of electricity.
    • Thermal power plants, especially coal and nuclear, use water for cooling.
    • So even if a data center reduces direct water use with air cooling, water use can shift to the grid.
    • If the data center uses a lot of renewables, this indirect water use tends to go down.

What affects how much water gets used:

  • Location
    Hot, dry regions need more cooling effort. Some choose water-heavy cooling because it saves electricity costs, even if it uses more water.

  • Cooling design

    • Evaporative cooling: lower energy use, higher water use.
    • Air cooling: higher energy use, lower direct water use.
    • Hybrid setups try to balance both.

  • Time of day and grid mix
    If AI workloads run more when the grid runs fossil or nuclear, water use for power plants goes up.
    If workloads shift to times with high solar or wind, water intensity goes down.

What you can do as a user:

  • Prefer companies that publish water usage data, not only carbon. Look for metrics like WUE, water usage effectiveness, often in liters per kWh.
  • Check sustainability reports from providers like Google, Microsoft, Amazon. They now talk about water stress regions, direct vs indirect use, and cooling choices.
  • If you run your own ML workloads on the cloud, choose regions with:
    • cooler climates
    • higher renewable share
    • lower water stress
  • Push for transparency. Ask vendors things like:
    • What is your WUE in this region
    • Do you use potable water or non-potable / reclaimed
    • Are you in a high water-stress basin

Rule of thumb for intuition, not as a precise value:

  • A single chat with an AI model is small, roughly like a few sips of water in indirect use.
  • Training huge frontier models is closer to what a small group of people use in a year.
  • The big concern is not one person’s usage. It is growth of AI workloads in regions that already face water stress.

So yes, AI systems “use” water mainly by needing cooling for hot hardware and electricity, which in turn often relies on water-intensive power plants. The details depend a lot on where and how the data center runs.

3

AI doesn’t literally “drink” water, but data centers basically rent a river for two things: getting rid of heat and making electricity. @suenodelbosque covered the basics, so I’ll fill in some gaps and nitpick a bit.

1. Not all water use is equal

A big missing nuance:

  • There’s withdrawal (water taken in, then returned to a river/lake a bit warmer).
  • And consumption (water that leaves the system, usually as vapor, and doesn’t immediately go back to the same source).

Power plants often withdraw huge volumes but only consume a fraction. Data center evaporative cooling is mostly consumptive: that water is literally turned into vapor and drifted away. So a smaller volume can still be more impactful, especially in dry regions.

Some studies that quote huge “water use” numbers are mostly counting withdrawals, which makes things look scarier than they are, at least in terms of local depletion. So I’d be careful with big headline numbers that don’t separate those two.

2. Training vs. inference: different water profiles

People obsess over “training GPT-like models uses X hundred thousand liters” but miss that:

  • Training is spiky and rare.
  • Inference (millions or billions of queries across products) is continuous.

High-level pattern:

  • Training = large one-time water burst, often in top-tier facilities with relatively optimized cooling.
  • Inference = ongoing use that may be distributed across older, less efficient regions.

It’s totally possible that long-term, inference ends up using more aggregate water over the life of the model than the initial training run. So only looking at “training footprint” kind of misses the main story.

3. Design choices matter more than “AI vs non‑AI”

One thing I’d push back slightly on from @suenodelbosque is the implication that this is sort of an “AI problem.” It’s broader: if the same GPUs were used for video rendering or scientific simulations, the water profile is similar. The key factors are:

  • GPU density: Higher density racks = more heat = stronger cooling strategy needed.
  • Target efficiency: Some hyperscalers accept higher water use to squeeze out slightly better energy efficiency and cost.
  • Uptime requirements: Mission critical regions run more redundancy and cooling headroom.

So AI isn’t special physics-wise, it’s just causing a lot more high-density hardware to be deployed quickly.

4. Alternatives to water are not free wins

You’ll see “just use air cooling” thrown around. Reality:

  • Air cooling
    • Less direct water use.
    • More electricity, especially in hot climates.
    • That extra electricity might come from water-intensive thermal plants.
  • Direct-to-chip liquid cooling
    • Very efficient, can be paired with less evaporative use.
    • Complex plumbing, CapEx, and sometimes still uses evaporative towers on the back end.
  • Seawater or brackish water
    • Reduces pressure on drinking water sources.
    • Needs corrosion-resistant gear, intake protection for marine life, etc.

So swapping cooling method often just shifts water use in space or type (fresh vs non-potable) rather than magically deleting it.

5. Location choice is doing a lot of hidden work

A “liter of water” is not morally identical everywhere:

  • In a humid, water-rich region, cooling water is a relatively low-stress use.
  • In a desert basin already over-allocated, the same liter is a much bigger deal.

This is actually where I think the biggest red flag is, more than the raw numbers. You increasingly see data centers popping up in dry, cheap-land regions for cost reasons, then using potable water for evaporative cooling. That’s the part that deserves louder criticism.

Some better practices you can watch for in sustainability reports:

  • Use of reclaimed / wastewater instead of potable.
  • Building in low or medium water-stress basins.
  • Seasonal shifts in cooling strategy (more air cooling in cooler months, reduce evap usage).
  • Explicit targets to reduce water consumption, not just withdrawals.

6. How big is your personal impact, really

If you’re wondering “is this worse than streaming video,” rough perspective:

  • A single AI chat is tiny, like @suenodelbosque said, sips-level.
  • Watching HD video for an hour can be in a similar or higher ballpark of indirect water use depending on where the data and power come from.
  • The main risk is scale: millions of users, heavy enterprise workloads, and more models moving into everything from office tools to search.

Focusing on your individual queries probably isn’t where the real leverage is. The leverage is in:

  • Pushing providers on water metrics (not just carbon).
  • Supporting regulation that forces public reporting on data center water use.
  • Nudging cloud choices (if you have that power) toward regions with lower water stress + decent renewable mix.

7. If you want a quick “mental model”

  • AI = GPUs running hot.
  • Hot GPUs = need cooling.
  • Cooling and power = often need water.
  • Location + tech choices decide if this is a mild background impact or a serious local water problem.

So yeah, you are not overthinking it. The tech hype is moving a lot faster than the conversation about where all the “invisible” inputs like water are coming from. And some of those inputs are literally coming from the same rivers and aquifers people drink from.

4

Zooming in on what hasn’t been covered yet: the numbers in context and how to sanity‑check claims you’ll see in the news.

@jeff and @suenodelbosque nailed the “where does the water go” part (cooling + power, withdrawal vs consumption). Where I slightly disagree is with the “few sips per chat” comparison. It is a useful metaphor, but it makes it sound almost trivial. At global scale, those “sips” add up fast.

1. Put the numbers in human terms

From recent academic estimates for frontier‑scale models (GPT‑3 class and up), you can roughly map:

  • Training one giant model
    ≈ water use comparable to dozens of people’s annual household water consumption in a rich country
  • Running that model at large scale over its lifetime
    ≈ often more total water than the initial training, since inference runs forever

Important nuance: that is not your single chat. It is spread over millions or billions of uses. So individually you are not wrecking a river every time you ask “write me a poem,” but the aggregate footprint is in the same order of magnitude as other big digital habits like HD video streaming or cloud gaming.

2. How to read “shocking” headlines

You will see very different numbers for “AI water use” because of three tricks:

  1. Counting withdrawals instead of consumption

    • Power plants might withdraw tens of thousands of liters per kWh but consume a much smaller fraction.
    • Data center evaporative cooling mostly shows up as consumption.

  2. Ignoring location

    • 1,000 liters in a wet, cool region has a much lower impact than 1,000 liters in a stressed desert aquifer.

  3. Aggregating all cloud use and calling it “AI”

    • A lot of media pieces lump regular cloud services, storage, video, etc. into a single AI number.

When you see a stat, ask:

  • Is this withdrawals or consumptive use?
  • What region or river basin?
  • Is this AI‑specific or entire data center operations?

3. Is AI uniquely bad vs “normal” cloud?

This is where I diverge a bit from @jeff. It is true that physics does not care whether a GPU is doing AI, VFX, or scientific sims. But AI is driving:

  • Much higher GPU density per rack
  • Pressure to operate at very high utilization
  • Demand growth that is steeper than for most other workloads

So while the mechanism is the same, the rate of build‑out and the choice of siting (cheap land, cheap power, often dry regions) is very much an AI story. It is not fair to pin all data center water on AI, but it is also not honest to treat it as just another random workload.

4. What you can realistically influence

As a user, your leverage is mostly collective, not per‑prompt:

  • Support reporting rules that force companies to disclose:
    • WUE (water usage effectiveness) by region
    • Share of reclaimed vs potable water
    • How much is in high water‑stress basins
  • When choosing cloud regions for your own ML jobs:
    • Pick cooler climates where possible
    • Prefer grids with strong wind and solar share
    • Avoid regions already flagged as high water stress

This is basically the extension of what @jeff and @suenodelbosque wrote, but pull the focus toward where and what kind of water is used rather than trying to count exact liters per chat.

And one last sanity check: AI is not literally “drinking” water, but if you imagine each big data center as a factory that borrows a slice of a local river every hour to throw away heat, you are basically picturing it correctly.