By Colin Harper

The greatest aphorism I’ve ever heard came from my Southern grandmother.

“There are really three sides to every story, Colin,” Sue Sue told me. “What one person says, what the other says, and what actually happened.”

You can apply this Sue Sue-ism to almost any disagreement, but it seems particularly salient for The AI Problem – particularly, the question of AI’s ecological impact.

The AI transition has whipped up a fury of anti data center sentiment, with offshoots of America’s environmental movement finding a new boogeyman for their pitchforks and torches. Except, it’s not really fair to pin this sentiment solely on kombucha-swilling greenniks; middle Americans – the salt of the earth, the MAGA-heads, the country bubbas – have also become data center NIMBYs.

Horseshoe theory in action, this trend provides a rare moment of right-left unity in our wildly polarized society. 

To be clear, left-leaning Americans are more likely to oppose data centers, but the divide isn’t as cleanly cut as other political schisms on national issues. A Spring 2026 Gallup poll found that 71% of respondents opposed the construction of a data center in or near their community, with 58% of Democrats, 48% of Independents, and 39% of Republicans strongly opposing the measure. A Marquette Law School poll reported that 76% of Democrats, 73% of Independents, and 62% of Republicans think the costs outweigh the benefits of parking a data center in their community’s backyard.

In the Gallup poll, 50% of respondents cited resource concerns, with water and energy equally topping concerns in this category.

Now, to tech and data center executives, this is little more than a witch hunt fed by misinformation, or outright lies (and no doubt, some of the concern is likely astroturfed by foreign adversaries like China, as Bitcoin Policy Institute research points out).

But these executives obviously have a financial interest to downplay these concerns, and it’s understandable that locals would be leery of the techno-megaliths that power this newfangled technology if they don’t know how these complex systems tick.

There are always three sides, but in this case, the executives are cocksure for a reason. The data is on their side. 

Data centers use much less water than you thinkData centers use much less water than you think

If we only look at numbers without any other context, data centers do appear to use a lot of water.

Citing EPA data, MOST Policy Institute reports that data centers in the US consumed 17.3 billion gallons of water in 2023, and MOST extrapolates this number to estimate that they could consume anywhere from 38-73 billion gallons by 2028.

That seems like a crazy amount, but the 2023 figure is equivalent to 160,000 American households. (This isn’t a totally fair comparison, though, because most of that water is recycled back into the system. But more on that later). 

It’s worth noting that MOST’s is measuring “consumption” – that is, water that does not return to the local source. Consumption is different from “withdrawal,” which measures the gross amount of water taken from a source; in agriculture, some of this withdrawn water returns to the local source via ground absorption and runoff, while the total measured water burden of a crop is called “applied water.”

These distinctions are important, because some research might cite total withdrawal and pass it off as consumption, assuming readers won’t appreciate the difference.

Understanding consumption versus total withdrawal is also imperative if we want an apples-to-apples comparison of data centers’ relative thirst to other industries.

Almond farming is caught in the crossfire of this debate, as it quickly became the favored punching bag of AI proponents when steelmanning data center water use. It’s true, almond farms – most of which reside in a single state, California – annually soak up roughly 60-75x more water than data centers in the US, anywhere from 1.3-1.6 TRILLION gallons per year, according to combined data from the California Water Impact Network and the Pacific Institute.

This figure, though, is for applied water. After running some back-of-the-napkin math through ChatGPT – probably the equivalent of a few glasses of water – we get an estimate that this equates to roughly 1-1.4 trillion gallons consumed annually – that is, water that does not return to the local source. 

“But people eat almonds. You can’t eat data!” 

So true, king. But that logic is turtles-all-the-way-down; I could just as easily say that you’d be better off eating something more protein dense that needed less quenching to cultivate, like beef (or if you want to stay vegan with it, soybeans/tofu). 

But we can also compare data centers to something totally trivial: golf. Golf courses in the US alone use anywhere from 476-760 billion gallons of applied water annually. Running another round of back-of-napkin math with ChatGPT, this roughly translates to 300-600 billion gallons of consumption each year.

Those Studio Ghibli memes don’t seem that thirsty in comparison, do they?

AI is changing data center water habitsAI is changing data center water habits

Perhaps ironically, AI data centers could consume even less water, even if they proliferate at a faster rate than traditional data centers.

A majority of large modern data centers still use some form of evaporative (open-loop) cooling because it’s energy-efficient, especially in hot climates. Open-loop sites can evaporate 1–9 liters per kWh and consume up to 80% of the water they withdraw. A closed-loop site, by contrast, may consume only 5–10% of withdrawn water.

Data centers are increasingly moving toward closed loop systems, which use less water at the expense of using more electricity. And in the case of AI data centers, hyperscalers are moving toward closed loop liquid cooling with dielectric fluid, which requires virtually no water at all. 

Closed-loop liquid cooling can use dramatically less water than the evaporative cooling it replaces. Microsoft’s next-generation design, for example, targets near-zero on-site water and expects to save 125M+ liters per site per year.

Microsoft says zero-water cooling will be standard on all new builds from late 2027. So as AI forces the shift to liquid, per-site water intensity is trending down, not up — the exact opposite of what the panic assumes.