The fight over AI data centres has become one of the most heated local-politics stories in the country, and for good reason. A recent Gallup poll found around 70% of Americans oppose a data centre being built in their area, with water use near the top of the list of worries. In the first quarter of 2026 alone, at least 75 projects worth a combined $130 billion were disrupted by local opposition. But underneath the placards is a genuine engineering trap that almost nobody in the shouting match acknowledges: when it comes to cooling these things, you cannot win. You can only choose which resource to burn.
The paradox, in plain terms
Data centres run hot, and all that heat has to go somewhere. There are broadly two ways to shed it. The first is evaporative cooling, essentially giant swamp coolers, which is wonderfully energy-efficient and is why modern facilities boast such impressive power-usage numbers. The snag: it works by evaporating enormous quantities of water, which is how a single campus ends up drinking like a small town.
The second option is to cool with less water, using dry or closed-loop systems that recirculate rather than evaporate. Do that and your water footprint drops sharply, which is exactly what worried communities are demanding. The catch, and it is a big one: those systems use significantly more electricity to do the same job. So the water you save comes straight back as extra power draw, and that power, unless it is genuinely clean, brings its own environmental bill (including, awkwardly, the water used to generate it).
That is the trap. Optimise for power efficiency and you spend water. Optimise for water and you spend power. There is no free cooling, only a swap. And it gets worse when you count the indirect footprint: the water used to generate the electricity, and the water used to manufacture the chips, can exceed what the data centre sips on-site. The number on the fence-line is not the whole number.
Why this matters for the fight, and the investment case
For communities, the lesson is to interrogate which trade a developer is actually making, because “we use very little water” often really means “we use a lot more power,” and vice versa. A promise to be water-thrifty in a region whose grid runs on gas is not the clean win it sounds like. The honest question is not “water or power” but “which one can this specific place actually afford to give up.”
For investors, the paradox is a real driver, not a footnote. It is why cooling technology, water recycling, and above all cheap, abundant, clean power are becoming the genuine bottlenecks of the AI build-out. Companies that crack low-water, low-power cooling, or that lock up clean energy at scale, are selling the one thing every data centre operator now desperately needs: a way out of the trade-off. That is where a lot of the next decade’s infrastructure money is going, and it is a more durable bet than any single chip. (Not investment advice. Sources: Gallup polling and 2026 data-centre reporting.)
Related: the global race to build AI compute.