AI & Water
A Strange Connection
Here is something that sounds made up but is completely real: when you ask an AI model to write you an email, water evaporates in a data center somewhere on the planet.
Not metaphorically. Literally. Actual water, turning into actual vapor, leaving an actual cooling tower, drifting into the actual sky.
Most people who use ChatGPT, Gemini, Claude, or Midjourney have never thought about this. Why would they? You type a prompt, you get a response. There is no splash sound. No dripping faucet icon. Nothing in the interface hints that your request just consumed water.
But it did. And as AI scales up — more users, bigger models, longer conversations, more images and videos generated — the water bill is scaling up with it.
This is the story of AI and water. It is not a doomer piece. It is not AI propaganda either. It is just an honest look at a real resource cost that most people have never heard of, and what it means going forward.
Why Does AI Need Water in the First Place?
To understand this, you need to understand one thing about computers: they generate heat. A lot of heat.
Your laptop gets warm when you are running heavy applications. Now imagine a building filled with thousands of processors, each one far more powerful than your laptop, all running at full capacity, 24 hours a day, 365 days a year. That is a data center.
The processors used for AI — primarily GPUs (graphics processing units) and increasingly custom AI chips like Google's TPUs — run especially hot. Training a large language model means running these chips at near-maximum capacity for weeks or months straight. Even after training, every time someone uses the model (which is called "inference"), those chips fire up and generate heat.
Heat is the enemy of electronics. If chips get too hot, they slow down, produce errors, and eventually fail. So data centers need cooling systems. Lots of cooling.
There are several ways to cool a data center, and most of them involve water.
How Data Center Cooling Works
Let me walk through the main cooling methods, because this is where the water comes in.
Evaporative Cooling (Cooling Towers)
This is the most common method and the biggest water consumer.
The basic idea: hot water from the data center is pumped to a cooling tower. In the tower, the water is spread over a large surface area and exposed to air. Some of the water evaporates, and evaporation absorbs heat — the same principle that makes you feel cool when you step out of a swimming pool on a windy day.
The cooled water is pumped back into the data center to absorb more heat. The water that evaporated? It is gone. Into the atmosphere. That water needs to be replaced with fresh water.
This is where the water consumption happens. The water does not get "used" in the sense that it gets dirty and needs to be treated (though some does). It literally disappears into the air.
A single large data center can evaporate millions of liters of water per day through its cooling towers.
Direct Liquid Cooling
Some newer data centers pump liquid coolant directly to the chips, rather than cooling the air in the room. This is more efficient because liquid absorbs heat much better than air. But the heat still needs to go somewhere eventually, and that "somewhere" is often an evaporative cooling tower outside.
So direct liquid cooling reduces water use, but does not eliminate it.
Air Cooling
In cooler climates, data centers can sometimes use outside air directly — just pulling in cold air from outside and pushing hot air out. This uses very little water. But it only works in places that are consistently cool, and it cannot handle the heat densities of modern AI hardware.
Closed-Loop Systems
Some facilities use closed-loop systems where the coolant circulates without evaporating. These use significantly less water but are more expensive and less efficient in hot climates.
The bottom line: most data centers today use evaporative cooling as their primary or secondary cooling method, and evaporative cooling consumes water. A lot of water.
Okay, But How Much Water Are We Actually Talking About?
This is where the numbers get interesting — and a bit unsettling.
Per-Query Numbers
Researchers at UC Riverside published a widely cited study in 2023 estimating the water footprint of AI. Here are some of their findings:
- A conversation of 20-50 exchanges with GPT-4 consumes roughly 500 milliliters of water (about a standard water bottle) in direct cooling.
- A single image generated by an AI image model consumes roughly the equivalent of a few sips of water.
- These numbers vary significantly depending on the data center location, time of year, and cooling method.
And that is only direct water consumption (the water that evaporates in cooling towers). It does not include:
- Water used to generate the electricity that powers the data center
- Water used in manufacturing the chips
- Water consumed in the supply chain
Training vs. Inference
Training a large model is extremely water-intensive because it runs GPUs at maximum capacity for extended periods.
Estimates suggest that training GPT-3 consumed roughly 700,000 liters of water — enough to fill a small swimming pool. GPT-4, being much larger, likely consumed several times that amount, though OpenAI has not published exact figures.
But here is the counterintuitive part: inference (using the model) collectively consumes far more water than training. Training happens once. Inference happens billions of times. Even though each individual query uses a tiny amount of water, the sheer volume of queries adds up to a much larger total.
It is like the difference between building a factory (expensive one-time cost) and running the factory (cheaper per unit, but the units never stop). The running cost dominates over time.
The Electricity-Water Connection
There is a second layer of water consumption that most people miss entirely.
Data centers consume enormous amounts of electricity. That electricity has to be generated somewhere. And many forms of electricity generation — particularly coal, natural gas, and nuclear — consume water for cooling at the power plant.
So the water footprint of AI is not just the water used at the data center. It also includes the water used at the power plant that supplies the data center's electricity.
When you add this "indirect" water consumption, the total water footprint roughly doubles or triples compared to the direct cooling numbers alone.
According to estimates, the global AI industry's water consumption (direct and indirect) is in the range of billions of liters per year, and growing rapidly.
Where Is This Water Coming From?
This is the part that turns a technical discussion into a social and environmental one.
Data centers do not create water. They draw it from local water supplies — municipal water systems, wells, rivers, or reservoirs. The same water sources that supply drinking water to homes, irrigation water to farms, and flow to ecosystems.
And here is the problem: many data centers are located in places that are already experiencing water stress.
The American West
A significant number of the world's largest data centers are in the western United States — Arizona, Nevada, Oregon, and parts of California and Texas. These regions have been dealing with drought conditions for years. The Colorado River, which supplies water to 40 million people across seven states, has been at historically low levels.
Microsoft's data center in Goodyear, Arizona, drew national attention when residents discovered it was consuming millions of gallons of local water annually. In a desert. During a drought. For AI.
Google's data center in The Dalles, Oregon, consumed about 12.7 million liters of water in a single month (July 2022, one of the hottest months). The city had to weigh the economic benefits of the data center against the water needs of its 16,000 residents.
Global Examples
It is not just an American problem. Data centers are being built globally, and many of them are in water-stressed regions:
- India: New data centers in Chennai, Mumbai, and Hyderabad — cities that have experienced severe water shortages.
- Middle East: Major data center projects in Saudi Arabia and the UAE, some of the most water-scarce regions on Earth (though some plan to use desalinated seawater).
- Chile: Data centers in Santiago, where the metropolitan region has faced water rationing.
- South Africa: Data center growth in the Western Cape, which experienced severe drought in 2017-2018.
What Are the Big Companies Saying?
To their credit, the major tech companies have started acknowledging the issue. But their responses vary.
Microsoft
Microsoft's 2024 sustainability report showed that its water consumption increased by 34% from 2021 to 2022, largely driven by AI growth. The company has pledged to become "water positive" by 2030 — meaning it will replenish more water than it consumes globally. They are investing in water replenishment projects like wetland restoration and watershed management.
But "water positive" is a global metric. Restoring a wetland in one country does not help the specific community in Arizona whose water the data center is consuming.
Google reported that its data centers consumed approximately 21.2 billion liters of water in 2022. Like Microsoft, Google has pledged to replenish 120% of the water it consumes by 2030. Google has also been investing in air-cooled data centers and more efficient cooling technologies.
OpenAI
OpenAI has been less transparent about water consumption. Most of their infrastructure runs on Microsoft's Azure data centers, so the water consumption shows up in Microsoft's reports rather than OpenAI's own.
Meta
Meta's data centers consumed about 8.6 billion liters of water in 2022. They have also made water replenishment commitments.
The common theme: everyone acknowledges the problem, everyone is making pledges, but consumption keeps going up year over year because AI demand is growing faster than efficiency improvements can offset it.
Wait, How Does This Compare to Other Industries?
This is an important question, because context matters.
Agriculture consumes roughly 70% of all freshwater globally. Industry (manufacturing, mining, etc.) takes about 20%. Domestic use is about 10%. Data centers, including AI, are a relatively small slice of the industrial portion.
So in absolute terms, AI's water footprint is a fraction of agriculture's. A single almond farm in California might use more water than a data center.
But there are two reasons why AI's water use deserves attention despite being relatively small:
1. Growth rate. Agriculture's water use is relatively stable. AI's water use is doubling every couple of years. If the trajectory continues, AI could become a significant consumer within a decade.
2. Location. A farm uses water where food is grown, which is usually where water is naturally available (or historically was). Data centers are built where land is cheap, electricity is abundant, and tax incentives are attractive — which does not always align with water availability.
The concern is not that AI is the biggest water user today. The concern is what happens if we do not plan for its growth.
What Can Be Done About It?
This is not a hopeless situation. There are real solutions being developed and deployed. Some are already working.
Better Cooling Technology
Immersion cooling: Servers are submerged in a non-conductive liquid that absorbs heat directly. This can reduce water consumption by 90% or more compared to evaporative cooling. Microsoft, Google, and others are experimenting with this at scale.
AI-optimized cooling: Ironically, AI itself can help. Google's DeepMind developed an AI system that optimized data center cooling and reduced cooling energy (and therefore water) by 30%. The AI monitors hundreds of sensors and adjusts cooling parameters in real time.
Geothermal and deep lake cooling: Some facilities are exploring pumping heat deep underground or using cold water from deep lakes. These approaches use almost no water.
Location Matters
Building data centers in cool, humid climates dramatically reduces water needs. The Nordic countries (Sweden, Finland, Norway, Iceland) are becoming popular data center locations because outside air can handle much of the cooling, and water is abundant.
Iceland has the added advantage of geothermal energy — abundant, cheap, and water-friendly.
Model Efficiency
Smaller, more efficient models require less computation, which means less heat, which means less cooling, which means less water.
The trend toward model distillation (creating smaller models that approximate the performance of larger ones), quantization (reducing the precision of model weights), and more efficient architectures all help reduce the water footprint per query.
When you use a lightweight model for a simple task instead of throwing GPT-4 at it, you are indirectly saving water.
Water Recycling and Reuse
Some data centers are starting to use recycled wastewater (treated sewage water) instead of fresh water for cooling. This does not reduce total water consumption, but it reduces the burden on drinking water supplies.
Meta's data center in Prineville, Oregon, uses reclaimed water for cooling. Google has committed to using non-potable water sources where possible.
Transparency and Accountability
Perhaps the most important change is simply tracking and reporting water use. When companies publish their water consumption numbers, it creates pressure to reduce them. Several advocacy groups are pushing for mandatory water reporting for large data center operators.
The Ethical Dimension
There is a deeper question here that goes beyond engineering: who gets to decide how water is allocated?
When a tech company builds a data center in a water-stressed community, they are implicitly making a claim on that community's water. The company gets economic output (and profits). The community gets jobs and tax revenue. But the community also loses water that could have been used for other purposes.
Is that trade-off fair? It depends on who you ask.
A farmer in Arizona might see it differently than a tech worker in San Francisco. A resident of Chennai whose taps run dry for hours each day might see it very differently from a Silicon Valley investor.
The uncomfortable truth is that AI's water consumption disproportionately affects communities that are already water-stressed. The people benefiting most from AI (largely in wealthy, water-secure regions) are not the same people bearing the water costs.
This does not mean we should stop building AI. But it does mean we should be honest about the costs and intentional about how they are distributed.
What Can You Do as an Individual?
Honestly? The impact of individual choices here is small. This is primarily a systemic issue that requires corporate and policy-level solutions.
But there are a few things worth being aware of:
Use the right model for the task. If you need a quick factual answer, you do not need a 1-trillion-parameter model. Lighter models consume less energy and less water. If your AI tool offers model selection, choosing a smaller model for simple tasks is marginally better.
Be mindful of unnecessary usage. Generating 50 variations of an AI image "just for fun" has a real resource cost. That does not mean you should never do it — but awareness is the first step.
Support transparency. When companies report their environmental impact, pay attention. Reward transparency. When they do not report, ask why not.
Think about it. Seriously, that is meaningful. Most people have never thought about the physical infrastructure behind the chat window. Just understanding that AI has a material cost — in water, energy, minerals, and land — changes how you relate to the technology.
Putting It in Perspective
I want to end with some honest perspective, because this topic can easily tip into fear-mongering or dismissiveness, and neither is helpful.
AI's water footprint is real and growing. It deserves attention, especially in water-stressed regions. Companies should be transparent about it, invest in solutions, and be thoughtful about where they build.
But AI also has the potential to help solve water problems. AI systems are being used to detect leaks in water infrastructure (which globally loses about 30% of treated water to leaks), optimize irrigation in agriculture, predict droughts, model climate change impacts on water systems, and design more efficient water treatment processes.
The technology that consumes water could also help us conserve it. That is not a justification for ignoring the problem — it is a reason to be thoughtful rather than reactionary.
The worst outcome would be pretending the water cost does not exist. The second worst would be using it as an excuse to halt progress on technology that could genuinely help humanity.
The right path is somewhere in between: build AI, but build it responsibly. Measure the costs. Invest in solutions. Be honest about trade-offs. And never forget that behind every floating chat bubble, there is a physical machine, generating real heat, consuming real water, in a real community, on a real planet.
This one was harder to write than my usual posts because the data is scattered and the companies are not always forthcoming. If you spot something I got wrong or have better numbers, let me know. I will update it.