Every time you ask an AI a question, a building somewhere wakes up. It has racks of chips, miles of cable, and a cooling system working harder than most people realize. That building needs two things in huge amounts: power and water.
I work around data center infrastructure every week, and I can tell you this plainly. The people who understand the physical side of AI are going to have a lot of options in the next ten years. You do not need a computer science degree to see it. You need to look at where the pressure is building.
So let’s look at the real numbers, and then talk about what a regular person with a good idea and a strong work ethic can do with them.
Why AI needs so much water
Chips turn electricity into computing, and they also turn it into heat. A lot of heat. If you do not pull that heat out, the chips slow down or fail. Pulling heat out is called cooling, and for years the cheapest way to do it was to use water. Some of that water evaporates, the way sweat cools your skin.
The newest AI chips run hotter than the old ones, which means the cooling job keeps getting bigger. That is the whole story in one sentence. More AI means more heat, and more heat means more cooling.
The numbers, in plain terms
The Environmental Law Institute published a fact sheet in January 2026 with figures drawn from national research. Here is what stood out.
U.S. data center water use, direct
Water used on site, mostly for cooling.
Billions of liters per year. That is about a 211% increase in nine years, or roughly 181 million liters a day in 2023. Source: ELI Data Centers and Water Fact Sheet, January 2026.
Now here is the part most people miss. The water a data center uses on site is only a small piece. The bigger piece is hidden in the electricity. Power plants use water too, and data centers pulled about 176 terawatt hours from the grid in 2023.
The hidden water behind the power
Billions of liters in 2023. Indirect water use is roughly 12 times the direct amount. Source: ELI, January 2026.
Two more facts matter for anyone thinking about business. The ELI sheet notes that roughly two thirds of the data centers built since 2022 sit in water-stressed regions. And it cites projections that data centers could use between 6.7% and 12% of U.S. electricity by 2028.
The cooling menu
Not all cooling is equal. The choice a builder makes here decides how much water a site needs.
How each cooling method uses water
| Cooling method | Water profile | What it means |
|---|---|---|
| Air cooling | Low to none | Uses more electricity instead |
| Evaporative cooling | High | About 80% of the water evaporates, 20% is discharged |
| Direct-to-chip liquid | Low to none | Coolant goes right to the chip in a closed loop |
| Immersion cooling | Low to none | Servers sit in a special fluid |
Source: ELI Data Centers and Water Fact Sheet, January 2026.
There is a trade-off hiding in that table. Evaporative cooling is very energy efficient but thirsty. Air cooling saves water but pulls more electricity. Liquid cooling to the chip can cut water use a lot, and it is the direction the newest AI builds are moving. Every option solves one problem and creates another, and that is exactly where opportunity lives.
You may also hear the term WUE, which stands for water usage effectiveness. It is simply water used divided by the energy the computing equipment uses. A lower number is better. If you ever want to sound sharp in a room full of data center people, ask what a site’s WUE target is. It shows you did your homework.
What about one AI question?
People love to quote a number for a single chatbot reply. Estimates vary a lot depending on the model, the building, the weather and how you count. Some published estimates put a text response somewhere between 10 and 50 milliliters of water, which is a few sips at the high end. I would treat any single number with care. The honest takeaway is that one question is small, and billions of questions are not.
What builders can learn from this
I am not a trillionaire and I will not pretend I have a magic formula. But I have watched enough industries to see a pattern, and this one is clear.
1. Big waves create boring, valuable problems
When gold rushes happen, the people selling picks, boots and water often do very well. AI is a gold rush. Chips get the headlines. But somebody has to move heat, treat water, monitor flow, run pipes, secure land and keep it all running at 2 in the morning. Boring problems pay well when they are urgent.
2. Look for where demand outruns supply
Data centers want to be built fast. Water, power and land are slow. The gap between those two speeds is where services, tools and clever ideas get hired. Think water reuse and recycling, leak and flow monitoring, heat recovery, site planning, permitting help, training for technicians, and clear communication with the communities next door.
3. Trust is a product
Communities are asking fair questions about water. The companies that answer honestly, with real numbers, are going to win approvals faster. If you can explain a hard topic simply and kindly, you are valuable. That is a skill you can start building today, and this post is a small example of it.
4. You do not need to be the smartest person in the room
You need to be the one who shows up, learns the vocabulary, asks good questions and follows through. Most people stop after the first no. If you keep going, you will be surprised how few people are still in the room.
Your week-one plan
- Pick one corner of this world: water, power, land, cooling or communities.
- Read one trade article a day for seven days and write down every word you do not know.
- Message three people who work in that corner. Ask what keeps them up at night. Then listen.
- Write down one problem you heard twice. That is your first idea.
Let’s build this together. Start small, start this week, and keep a notebook. In a year you will know more than most people who claim to be experts.
Quick answers
How much water do U.S. data centers use?
Using ELI figures, about 66 billion liters directly in 2023, up from 21.2 billion in 2014. Another 800 billion liters were used indirectly to generate the electricity they consumed.
Does liquid cooling use less water?
Direct-to-chip and immersion cooling are listed as low to none for water, compared with high for evaporative cooling. Electricity needs and costs still matter, so every design is a trade-off.
Where are new data centers being built?
The ELI fact sheet says roughly two thirds of those built since 2022 are in water-stressed regions, which is why water planning is now a major part of siting.
Can a small business or solo founder get into this space?
Yes. Services, software, training, monitoring, consulting and local partnerships all sit around data centers. Start with a narrow problem and get good at explaining it.
Numbers come from the Environmental Law Institute Data Centers and Water Fact Sheet (January 2026). Per-query estimates vary by source. This post is general information, not financial or investment advice.
Sources: ELI, Data Centers and Water Fact Sheet, January 2026 | AKCP, The Data Center Water Footprint, August 2026
