If you sit through enough data center meetings, you’ll hear a line like this:

“Our facility uses a closed-loop cooling system. It uses about as much water as a restaurant.”

That can be true. It can also be one of the most misleading things said in these meetings. The number isn’t wrong. The problem is the numbers they leave out.

A data center can use almost no water at its own fence line. But it can still be the reason hundreds of millions of gallons of water get boiled away every year. It can create a steady stream of chemical wastewater that a rural county has no way to handle. And more and more often, it brings its own power plant to town. The water and the waste don’t disappear. They just move to places nobody puts on the slide.

This post walks through how the cooling really works. It shows why the “restaurant” line is technically honest. And it shows why an honest count has to follow the electricity back to where it’s made, and the wastewater forward to where it ends up.

Two loops, not one

The words “closed loop” do a lot of quiet work. So let’s be clear. A modern AI data center has at least two cooling stages. They work separately from each other.

The inside loop is the water that actually touches the hot chips. (It’s usually water mixed with antifreeze, called glycol.) In today’s high-powered AI builds, this is “direct-to-chip” cooling. The coolant is pumped through small cold plates that sit right on top of the chips — the GPUs and CPUs. It carries the heat away to a heat exchanger, then circles back. The best picture is the cooling system in your car. The coolant pulls heat off the engine, sheds it through the radiator, and goes around again. You don’t refill it every day. It’s part of the machine. When operators say “closed loop,” this is the loop they mean. And for this loop, the claim is true: the same water keeps circling, and none is used up.

The outside stage is how the heat finally leaves the building and goes into the open air. This is where water is either used up or not. And it has nothing to do with whether the inside loop is closed. There are two basic choices:

  1. Evaporative cooling (cooling towers). The building dumps heat by boiling off water into the air. The boiling is how it cools, so this water is gone for good. This is the old, water-hungry way.
  2. Dry cooling (air-cooled chillers and dry coolers). The building blows the heat into the outside air through big radiators and fans. Nothing boils off, so this uses almost no water.

A building can run a truly closed inside loop and still throw off its heat through cooling towers that boil away millions of gallons a year. “Closed-loop liquid cooling” describes the inside loop. By itself, it does not tell you whether the building uses up water.

The one question that decides almost everything

Before any of the numbers, there’s a single question that sets a facility’s whole water and waste picture:

At the final stage, is the heat thrown off by boiling water, or by blowing it into the air?

That one fork decides four separate things:

  • On-site water use. Evaporative cooling uses up millions of gallons per megawatt each year. Dry cooling uses almost none.
  • The wastewater stream. Evaporative cooling makes a steady stream of chemically treated water called “blowdown.” Dry cooling only makes a small amount of used coolant, drained every few years.
  • Upstream water. Dry cooling uses more electricity, because it’s less efficient. More electricity means more water boiled off at whatever power plant feeds the site (more on this below).
  • Where the power plant lives. If the site makes its own power, the cooling and power choices decide whether the water and the air pollution land in your county or someone else’s.

If a developer talks about “closed loop” but won’t plainly answer the boil-it-or-blow-it question, that’s the thing to press on. Everything else follows from it.

When “almost no water” is actually true

To be fair, the better operators are now building truly dry facilities that don’t boil off water. For those, the on-site claim holds up. The inside loop is filled once during construction and then recirculates for years. With dry cooling, the only water the building uses is for bathrooms, sinks, and an occasional top-off. That really can land in the range of a restaurant or a small office — about a million gallons a year or less. A normal evaporative facility would burn through tens or hundreds of millions.

So when a developer says their closed-loop, dry-cooled building sips water on site, they may well be telling the truth. The dishonesty isn’t in that sentence. It’s in stopping there.

The water that moved upstream

A data center’s job is to run electricity through chips around the clock. That electricity has to be made somewhere. And most of the American grid still makes power the same basic way: boil water into steam, use the steam to spin a turbine, then cool the steam back into water. Cooling that steam boils off water. This is true of coal, natural gas, and nuclear plants alike.

There are two different water numbers here. Mixing them up is how these debates go sideways.

  • Withdrawal is water pulled from a river or lake and then mostly put back. The numbers look huge — the U.S. power sector pulls about 11,600 gallons for every megawatt-hour — but most of that water goes right back. Quoting withdrawal as if it were “used up” is the kind of overreach that gets your argument thrown out.
  • Consumption is water that’s actually boiled off and not returned. This is the honest, apples-to-apples number. It’s the same thing the data center’s own cooling-tower number measures.

On consumption, the well-established figure from the National Renewable Energy Laboratory is about 0.47 gallons of fresh water boiled off per kilowatt-hour delivered to the end user. By technology, the range runs from about 200 gallons per megawatt-hour for efficient natural-gas plants up to roughly 600–820 for nuclear. Wind and solar use essentially zero. Texas, on the ERCOT grid, leans heavily on natural gas, with a fast-growing share of wind and solar. So the water used per unit of power here is probably toward the lower end of that range. But it is not zero, because the gas plants doing the heavy lifting still boil off water to run.

A worked example you can check

Numbers are more honest when you can check them yourself. So here’s the math out in the open. Take a single 100-megawatt facility. That’s on the smaller side for the AI projects being proposed in Texas today.

  • Running around the clock: 100 MW × 8,760 hours = 876,000 MWh of computing energy per year.
  • Cooling, fans, and losses add overhead. At a typical factor of 1.3 (dry cooling tends to run higher), the total pull from the grid is about 1.14 million MWh per year.
  • Now apply the upstream water range above:
    • At about 200 gal/MWh (gas-heavy, low end): about 230 million gallons boiled off per year.
    • At about 470 gal/MWh (national average): about 535 million gallons boiled off per year.

So a “zero-water” closed-loop data center — the kind that honestly uses restaurant-scale water at its own gate — can still be the reason 230 to 535 million gallons of water get boiled off every year at the power plants feeding it.

Here’s what that means locally. A typical American household uses about 110,000 gallons of water a year. That puts this one facility’s upstream water footprint at about 2,000 to 4,900 households — every year, with no end date. That’s the household water of a good-sized town, tied to a building that reports almost nothing on its own water bill.

What happens when the power plant moves to your county

Here’s the change that redraws the map. The Texas grid can’t connect all the proposed data centers fast enough. Getting hooked up now takes years. So developers are more and more choosing to bring their own power — building their own generation right at the site. This has gone from a rare workaround to a mainstream plan. Analysts now expect that a quarter to a third of all new data center demand through 2030 will be met this way. They call it “behind-the-meter” power. And most of these projects were announced in just the last year. Because Texas has so much natural gas, gas is the fuel of choice. At least one large West Texas campus has already signed up for on-site gas engines as its main power, with no grid hookup at all.

Here’s the honest way to think about it: bringing gas on-site doesn’t always change the total water much — it moves the water and the air pollution into the host county. The water I described as boiling off “at some power plant upstream” can now boil off in your community, pulled from your aquifer. But how much water depends on the type of generation. This is where it pays to be exact instead of overstating:

  • Reciprocating gas engines — the kind being installed in West Texas right now — are radiator-cooled and closed-loop. They use almost no water to run. If a project proposes these, water is not the issue. The real local impacts are air pollution (nitrogen oxides, carbon monoxide, formaldehyde, fine soot), lower fuel efficiency, and noise.
  • Simple-cycle gas turbines do use water. They need clean water to hold down NOx pollution, and to cool the incoming air for more output in hot weather. And that need gets worse in a Texas summer — exactly when water is shortest. (Turbines with special low-NOx burners can skip the NOx water.)
  • Combined-cycle plants with evaporative cooling are where the local water draw is real and large — about 200 gallons per megawatt-hour, now used in your county instead of someone else’s.

The point isn’t “on-site gas means a flood of local water use.” It’s that a project pitched as a data center quietly becomes a power plant too. The community now hosts the power plant’s pollution, fuel burn, and — depending on the cooling — its water. Tie the claim to the technology and it can’t be knocked down.

“But those are just backup generators”

Developers often call on-site generators “backup.” The thing that matters is how often they run — what engineers call the duty cycle. A true emergency generator runs maybe fifty hours a year for testing. It uses tiny amounts of water and fuel, no matter what it burns, simply because it barely runs. The same equipment run eight thousand hours a year is a power plant, and should be judged as one — for water, air, and everything else.

So the question to ask isn’t “do you have generators?” It’s how many hours a year will they run, and what does the air permit allow? Equipment permitted as once-in-a-while backup but run as full-time power deserves the scrutiny of a power plant, not a generator shed.

The wastewater nobody plans for

The water going in is only half the story. Something has to happen to the water coming out. And in a rural county, that may be the bigger problem.

It comes back to the boil-it-or-blow-it fork:

  • A truly dry, closed-loop facility makes no steady discharge. Its chemical waste is the coolant drained out every few years — the glycol, rust blockers, and germ killers — which a licensed waste hauler collects and carries off. Real, but only now and then.
  • An evaporative facility makes cooling-tower blowdown all the time. As water boils off, the minerals and treatment chemicals left behind build up in the loop. So part of it has to be drained and replaced constantly. That blowdown carries concentrated minerals plus the germ killers, scale blockers, and rust blockers used to keep the system running.

A quick word on the “chemical buildup that has to be scrubbed off,” since it’s worth getting right. Dry radiators get dirty on the air side, with dust and pollen, and get washed down. The heavy mineral scale that needs acid to remove is mostly found on evaporative equipment, where boiling concentrates minerals on the wet parts. So the steady, concentrated, chemical-filled wastewater is mainly an evaporative-cooling problem. Tie the wastewater worry to whether the facility is evaporative, and it holds up.

These are real, permitted industrial discharges in Texas. One existing data center near Houston, for example, holds a state permit to discharge cooling-tower blowdown and condensate up to 150,000 gallons a day. The question is where that stream goes. And every option is harder in a rural county:

  • A sewer line to a treatment plant is the default in a city. In rural Leon County, it simply doesn’t exist — there’s no city sewer system to send it to. And even where one exists, a big new blowdown stream can overwhelm a small plant.
  • Dumping it into a creek or river needs a state water-pollution permit, a stream with enough flow to dilute it, and limits on pollutants like leftover chlorine. A creek that runs only part of the year may not qualify.
  • Spraying it on land needs a permit focused on protecting groundwater, soil limits, and setbacks. Spread over the Carrizo-Wilcox aquifer — which many rural families pump their well water from — it’s a direct path to groundwater pollution.
  • Reusing it has an easy path only if the blowdown stays under 2,000 mg/L of dissolved solids. Concentrated blowdown often goes over that, which kicks it into stricter testing and approval rules.
  • Hauling it off-site works only if it’s done carefully and enforced. A rural county has little ability to check on it — on top of the truck traffic it adds.

And here’s the part that deserves the most attention: a septic system can’t take any of this. A septic system is permitted and sized for household waste only. The germ killers in the cooling water are made to kill the exact bacteria a septic system needs to work. The glycol would overload the drain field. The acids and concentrated salts don’t belong there at all. Routing industrial cooling waste into a septic drain field isn’t just against the rules. It’s a straight line into the shallow groundwater that feeds the neighbors’ wells.

That’s the imbalance that makes a rural site worse than a city one, not better. The host community runs on groundwater with no city backup. There’s no industrial-pretreatment program and no local oversight. And a single failure — a torn pond liner, a hauling shortcut, too much sprayed on a field — lands in the same aquifer the residents drink from, with no other supply to fall back on.

The tradeoff nobody mentions

Here’s the catch that ties it together: the very design that saves water on site is the one that uses more electricity. Dry cooling avoids boiling off water, but it’s less efficient than evaporative cooling — especially in a Texas summer. More energy means more power pulled. And if that power comes from plants that boil water, that means more water boiled off upstream, plus more pollution. Choosing dry cooling and then making your own power with gas doesn’t make the water or the pollution vanish. In many cases it just moves them into the host community, and may even raise the total.

This is also why you should read the industry’s favorite number carefully: “Water Usage Effectiveness,” or WUE. It counts only the water used at the building. By design, it leaves out every gallon boiled off to make the power. A company can post an impressive WUE and a “water positive by 2030” pledge while the real water use sits at a power plant that may now be right down the road.

What an honest accounting looks like

None of this means data centers are uniquely evil, or that closed-loop cooling is a scam. Dry cooling is a real improvement over evaporative towers, and operators who choose it deserve credit. The problem is one-sided accounting. An honest conversation counts water from source to sink — not just at the fence line. And it counts the waste on the way out, not just the supply on the way in.

If you’re sitting in a commissioners court or a public hearing, these are fair and specific questions:

  • How is heat thrown off at the final stage — by boiling water (evaporative) or by blowing air (dry)? This one answer drives on-site water, the wastewater stream, and the upstream energy penalty. “Closed loop” does not answer it.
  • Where does the power come from? If it’s the grid, which plants, and whose water? If it’s made on-site, what technology, what fuel, and what do the air and water permits say?
  • For any on-site generators, how many hours a year will they run? Backup run fifty hours a year and full-time power run around the clock are completely different animals.
  • What happens to the wastewater? Where does the cooling-tower blowdown or used coolant go, when there’s no city treatment plant here — and what protects the aquifer our wells draw from if the plan fails?
  • Does any water number you’ve given include the water used to make your electricity? If the answer is no, the headline number is incomplete.

A facility that uses restaurant-scale water on site, half a billion gallons upstream, and makes a chemical wastewater stream the county can’t process is not lying when it cites the first number. It’s just hoping you won’t ask about the rest. The most useful thing a community can do is ask.

Sources

Upstream / power-plant water consumption

Cooling technology and on-site water

On-site generation (behind-the-meter gas)

Wastewater and disposal

Household comparison

Figures are presented as order-of-magnitude ranges. Actual upstream and on-site water consumption depends on the specific generation and cooling technology; wind and solar consume essentially none, reciprocating engines use near-zero process water, while evaporatively-cooled thermal generation evaporates water to run.