Are Data Centers Bad for the Environment?

A data center is a machine for turning electricity into heat. Tens of thousands of servers convert nearly every watt they draw into warmth, and that warmth has to go somewhere. Everything else about the industry’s environmental footprint — the water, the emissions, the local politics — follows from that single fact.

The heat leaves by one of three routes. Air cooling moves it with fans and chillers: little or no water at the site, but more electricity. Evaporative cooling absorbs heat by evaporating water, usually through cooling towers: much less electricity, but the water is gone — consumed, not returned. Liquid and closed-loop cooling circulates coolant in a sealed loop past the chips themselves: the least water of all, at a modest energy penalty, and increasingly the default for the dense AI racks now being built.

That trade-off is the whole story in miniature. Operators can minimize electricity or minimize water, and in hot, dry places the two goals directly conflict. Which one a facility picks determines whether it shows up in a carbon ledger or a watershed dispute.

Start with the scale. The International Energy Agency estimates the world’s data centers consumed 415 terawatt-hours in 2024 — about 1.5% of global electricity — and projects that figure to more than double, to roughly 945 TWh by 2030, driven overwhelmingly by AI. Electricity demand from AI-optimized facilities alone is expected to quadruple over the same period.¹ For perspective: data center growth accounts for about a tenth of all expected growth in global electricity demand this decade — less than industrial motors, less than air conditioning, less than electric vehicles.¹

The United States is the epicenter. The US Department of Energy’s most comprehensive national assessment, from Lawrence Berkeley National Laboratory, found US data centers went from a flat 58 TWh in 2014 to 176 TWh in 2023 — 4.4% of all American electricity — and projects 325 to 580 TWh by 2028, as much as 12% of the national supply.² The IEA expects data centers to account for nearly half of all US electricity-demand growth to 2030.¹

On emissions, the picture is similarly two-sided. Globally, data centers are responsible for roughly 180 million tonnes of CO₂ a year — about 0.5% of energy-sector emissions, because nearly 60% of their electricity still comes from fossil fuels (renewables supply 27%, nuclear 15%).⁴¹ The IEA judges the emissions increase from data center growth to be “small in the context of the overall energy sector” — while cautioning that this assumes grids keep decarbonizing on schedule.¹

But averages conceal the geography. In Ireland, data centers consumed 21% of all metered electricity in 2023, and credible projections put the share near a third within a few years; growth around Dublin has been curtailed by grid constraints.¹⁶⁴ In Virginia, home to the largest cluster on Earth, data centers already draw about a fifth of the state’s electricity, and the state’s own legislative auditor projects that unconstrained data center demand would double Virginia’s total power consumption within ten years — with new infrastructure costs flowing into every customer’s bill, an estimated $444 per household per year by 2040 if tariff rules don’t change.¹⁷

Nationally, data centers are a rounding error. Locally, they can be the largest customer a power grid — or a watershed — has ever seen.

Efficiency, meanwhile, has stalled. Power usage effectiveness — the ratio of total facility power to the power that actually reaches servers — improved dramatically in the early cloud era, from 2.5 in 2007 to about 1.65 by 2013. Since then the industry average has barely moved: 1.56 in 2024, essentially flat for eleven years.³ The hyperscalers do far better — Google reports a fleet average of 1.10⁵ — but the long tail of older enterprise facilities keeps the mean stuck, and AI’s dense racks are now straining cooling systems designed for a gentler era.

Energy is the impact people argue about. Water is the impact people feel. Electricity can be generated anywhere and shipped over wires; water has to come from the watershed the building sits in, and in summer it has to come at exactly the moment the watershed can least spare it.

The best national accounting comes from the same Berkeley Lab study. US data centers consumed about 66 billion liters of water on-site in 2023 — 17.4 billion gallons — triple the 21.2 billion liters of 2014, with hyperscale and colocation facilities responsible for 84% of it.² The same study puts the off-site figure — water consumed by power plants generating the sector’s electricity — at nearly 800 billion liters, an average of 4.52 liters embedded in every kilowatt-hour.²

Globally, the IEA estimates data centers consumed about 560 billion liters of water in 2023, two-thirds of it indirectly through electricity generation, and projects the total to roughly double, to about 1.2 trillion liters, by 2030.¹

What the companies disclose

The hyperscalers’ own reports confirm the trajectory. Google’s company-wide water consumption reached 24.2 billion liters in 2023 — up 17% in a single year — with 95% of it going to data centers.⁵ Microsoft’s consumption jumped 34% in fiscal 2022, to 6.4 billion liters, a spike the company attributed to AI and cloud growth; it rose another 22% the following year, and roughly 40% of it occurred in water-stressed regions.⁶⁷ Amazon, notably, does not publish a comparable company-wide consumption figure.

The per-query fight

No environmental statistic of the AI era has been more contested than the cost of a single chatbot query. Researchers at UC Riverside estimated in 2023 that training GPT-3 evaporated about 700,000 liters of freshwater in Microsoft’s US data centers, and that a conversation of 10 to 50 responses carries a footprint of roughly 500 ml — a bottle of water — once power-plant water is counted.⁸ The Washington Post later put a 100-word GPT-4 email at about 519 ml on the same accounting.¹⁰

The industry’s rejoinders are narrower in scope, and worth reading carefully. Sam Altman says an average ChatGPT query uses 0.34 watt-hours and about 0.32 ml of water — a fifteenth of a teaspoon.¹¹ Google measured its median Gemini prompt at 0.24 Wh, 0.26 ml, and 0.03 g of CO₂.¹² Both figures count only on-site cooling water — they exclude the power plants, which is precisely the larger half of the footprint that Berkeley Lab measured. The two sides are not contradicting each other; they are drawing the boundary in different places.

WUE: the metric that decides where the water goes

The industry measures on-site water intensity as water usage effectiveness — liters evaporated per kilowatt-hour of computing. Berkeley Lab puts the 2023 US fleet average at roughly 0.36 L/kWh, but that mean blends thousands of small air-cooled rooms with vast evaporative plants; the hyperscalers’ own disclosures show both the spread and the direction of travel. Google reports a fleet average of 1.15 L/kWh. Microsoft drove its average from 0.49 in 2021 to 0.30 L/kWh in fiscal 2024 — and every Microsoft facility designed since August 2024 uses a sealed, zero-evaporation loop the company says saves more than 125 million liters per building per year.²¹²¹³

The environmental case against data centers rests on specific places, and the strongest version of it is very strong. Consider three facts. First, a peer-reviewed assessment found US data centers already rank among the ten largest water-consuming commercial and industrial activities in the country.¹⁸ Second, the consumption spikes in summer — peak cooling demand coincides with peak water stress, by physical necessity. Third, the water drawn is usually potable or potable-grade, because cooling towers need clean water to avoid fouling. Data centers are not competing with swimming pools; they are competing with drinking taps and irrigation ditches.

The industry’s water problem is not that it uses a lot of water. It is that it uses drinking-quality water, in stressed places, at the hottest times of year — and has fought to keep the amounts secret.

The Dalles, Oregon, a city of 16,000 on the dry side of the Cascades, is the canonical case. Google’s facilities there consumed 355 million gallons in 2021 — 29% of the city’s entire water supply, triple their 2017 draw. The company funded the city’s legal effort to keep those figures secret as a trade secret; a local paper sued, won, and published them.⁹ In West Des Moines, Iowa, Microsoft’s cluster drew 11.5 million gallons in a single month — July 2022, the month before GPT-4 finished training there — about 6% of the water district’s total use, and the local water utility has since said it will only approve future Microsoft projects that demonstrably cut peak consumption.⁷

There is a governance failure underneath the resource one. Google litigated to keep The Dalles’ numbers quiet. Amazon declines to publish fleet-wide water figures at all. Site-level permitting routinely treats consumption data as confidential. Communities are being asked to host the most resource-intensive commercial buildings ever constructed while being told, in effect, that the meter readings are none of their business. That posture — more than any single gallon figure — is what has turned water into the industry’s most volatile political liability, and it is self-inflicted.

Now the other side, argued honestly. Several of the loudest claims about data centers do not survive contact with the numbers.

The global water footprint is small. Agriculture accounts for about 70% of all freshwater withdrawn by humanity; industry as a whole is a fifth.¹⁵ In the United States, crop irrigation alone consumes about 73 billion gallons a day. Every data center in the country, meanwhile, consumes on the order of 17 billion gallons a year on-site.¹⁴² Run the arithmetic: American farms evaporate in under six hours what all US data centers evaporate in twelve months. Even counting the indirect power-plant footprint, data centers are a rounding error in the national water budget. The harm is real but concentrated — a siting crisis, not a supply crisis.

Efficiency gains were real, and may not be finished. Between 2010 and 2018, global computing demand exploded while data center electricity use stayed nearly flat — one of the great unheralded efficiency achievements in industrial history, visible in the 2014–2018 plateau in Figure 1. The easy gains are gone, but the hyperscalers’ 1.10 PUEs show what the frontier still looks like versus the 1.56 average.³⁵

The water problem is mostly an electricity problem. Twelve of every thirteen liters in the US data center water footprint are consumed at power plants, not on-site.² Wind and solar consume essentially no water to operate. Every kilowatt-hour of the data center buildout supplied by renewables instead of a thermal plant shrinks the sector’s true water footprint by an amount no cooling-tower redesign can match. The industry is also among the largest corporate buyers of renewable power on Earth, which does not erase its fossil draw but does mean its growth is directly financing new clean capacity.¹

The technology is moving the right way. Zero-water closed-loop designs are now standard in new Microsoft builds; liquid cooling is becoming mandatory for AI racks regardless, because air can no longer carry the heat.¹³ And siting is finally being treated as an engineering variable: the same training run scheduled in cool, wet Iowa instead of Phoenix costs a fraction of the water — a point Microsoft’s own researchers make explicitly.⁷

And the per-query panic is misplaced. Even on the fullest accounting, a chatbot conversation costs a bottle of water; on the operators’ measurements, a few drops. Either way, an hour of streaming video, a hamburger, or a single load of laundry dwarfs a day of AI chatting. Individual guilt over prompts is a distraction from the actual policy questions: where facilities go, what powers them, and who is told the numbers.

Every credible projection points the same direction. The IEA’s base case has global consumption more than doubling to 945 TWh by 2030, with sensitivity scenarios ranging from 700 TWh to 1,700 TWh by 2035 depending on AI adoption and efficiency.¹ Berkeley Lab sees US consumption reaching 325–580 TWh by 2028 and direct water use roughly doubling to quadrupling.² Uncertainty is enormous — the spread between scenarios is itself a finding — but no scenario shows the curve bending down.

Three forces will decide what that growth costs. The first is grid mix: the IEA projects renewables rising from 27% to about half of data center electricity by 2030, with gas filling much of the rest — a fork that determines both the emissions and most of the water footprint.¹ The second is cooling architecture: whether the industry’s sealed-loop, near-zero-WUE designs become the default fast enough to matter. The third is transparency, and here regulation is arriving: the European Union now legally requires data centers above 500 kW to report energy and water performance into a public database each year, the first mandatory disclosure regime of its kind.¹⁹ Similar reporting pressure is building in US states. What gets measured, consistently, tends to get managed.