Nothing in a data center hearing gets louder faster than water.
It makes sense. Water is local, finite, visible, and easy to picture. A million gallons a day sounds enormous, and in a rural district it is. Electricity is abstract; a dropping well is not.
It is also the topic where both sides most reliably quote figures that don’t mean what they’re being used to mean. So let’s take it apart, because the way out of this argument is not picking a side. It’s getting the units right.
The distinction that changes everything
Water withdrawn is water taken from a source. Water consumed is water evaporated and lost from the local system. Withdrawn water is largely returned. Consumed water is gone.
Those are wildly different quantities, and which one a speaker is using determines whether their number is alarming or trivial.
At a large evaporative-cooled facility, roughly 78 to 85 percent of withdrawal is consumed. At a closed-loop or air-cooled site, the consumed fraction is very small. Same headline withdrawal number, completely different local impact.
So when someone says “this facility will use two point eight million gallons a day,” the first question isn’t whether that’s a lot. It’s: use how?
Cooling design determines everything
Almost every water question in a data center hearing is actually a cooling design question wearing a water costume.
Air-cooled designs cut water use by more than ninety percent. They also carry a real energy penalty, which means more power draw, which means a larger grid impact and more indirect water consumed at the power plant. Nationally, indirect water use at generation runs roughly twelve times direct on-site use.
That trade is worth saying out loud, because a community that wins dry cooling has traded a local water impact for a larger regional energy impact. That might be the right trade. It’s still a trade, and a facilitator who lets a room believe otherwise isn’t doing the room any favors.
Two other things worth knowing:
“Closed-loop” is not zero-water. Blowdown and makeup are real. And direct-to-chip liquid cooling — increasingly common — often pairs with a heat-rejection stage that may itself be evaporative. Ask what happens at every stage, not just the one on the slide.
Reported water-use-effectiveness figures span more than tenfold between operators. So an industry average tells you nothing about the facility in front of you.
Both sides are overstating
Here’s where I annoy everyone, which is roughly the job description.
The advocacy side. You’ve seen the figure: a hundred-word AI response consumes a bottle of water. It got enormous circulation. It was subsequently revised down by its own authors by roughly two orders of magnitude. If you are still quoting the original, you are quoting a number its own researchers have retracted, and the moment someone in the room knows that, everything else you said becomes suspect.
The industry side. “A few drops per prompt” excludes the larger term — the water consumed at the power plant generating the electricity. It’s a true number about an incomplete boundary, which is a more sophisticated way of being misleading.
And the aggregate defense. “Data centers are under one percent of U.S. freshwater withdrawals” is an aggregation fallacy in a local debate. Nobody is worried about national withdrawals. They’re worried about their district’s summer peak. A national percentage is a change of subject dressed as a rebuttal.
The only formulation that works locally
Ask for water draw as a percentage of the utility’s permitted capacity and its summer peak-day production. Never as an absolute number.
That one reframing settles more arguments than any study.
Two point eight million gallons a day is a rounding error in a large county system. It is existential for a rural district that produces four million on its hottest day. The absolute number is identical. The answer is completely different. And the absolute number, quoted alone, is designed — by whichever side is quoting it — to make you feel a particular way rather than to inform a decision.
Once the question is a percentage of peak-day production, everyone in the room is arguing about the same thing, and the argument becomes finite.
What goes in the permit
Then convert the answer into terms, because a number in a presentation protects nobody:
- A hard cap on withdrawals, with the source identified and the cooling technology specified.
- A contractually committed water-use-effectiveness figure, verified after construction rather than projected before it.
- A prohibition on potable water for cooling — narrower and more defensible than “we’ll minimize water,” and the industry’s own leader has now publicly committed to it, which makes it awkward to refuse.
- Monthly public reporting of withdrawal and consumption, separately stated. Press hardest here. Reporting is cheap, verifiable, and now industry-endorsed, which means an applicant who refuses it is telling you something.
- Curtailment obligations if usage exceeds the cap, with a consequence attached.
Notice what that list does. It stops requiring anyone to trust a projection. “We will use water-efficient cooling” is a vibe. “We will withdraw no more than X gallons per day from source Y, monitored quarterly, with curtailment obligations if exceeded” is a clause.
Why a neutral is useful here specifically
I said earlier that this is the job description, and I meant it fairly literally.
If I only ever corrected the developer’s numbers, I’d be an advocate with better manners. If I only ever corrected the opposition’s, same thing in the other direction. The reason to have someone in the room whose fee doesn’t move with the outcome is that they can say both of those things in the same session, and be believed on the third question because of how they handled the first two.
Water is the best place to demonstrate that, because both sides genuinely are overstating, and everyone in the room half knows it.
Get the units right. Ask for the percentage, not the gallons. Put the number in the permit with a verification method attached.
Then go argue about noise, which is where the community case is actually strongest — and where almost nobody is asking the right question at all.