What the concern is actually about
Every Google search, every Netflix stream, every ChatGPT answer ends in a building. These buildings — warehouses of stacked computers called data centers — run hot, and keeping them running takes electricity and water in industrial quantities. As artificial intelligence moved from research demo to consumer staple in 2023, the question stopped being academic: what is all this computing actually costing?
The honest answer comes in three parts. On energy, data centers are a large but not dominant electricity user, and their carbon emissions are softened by the greening grid and hyperscaler-scale efficiency. On water, the picture is worse than most people realize — and most of it is hidden, evaporated invisibly through cooling towers rather than returned to its source. And on accountability, the industry has a documented transparency problem that makes even basic questions hard to answer.
This report walks through what the data actually shows: the scale of consumption, where it is genuinely bad, where it is overstated, and what is changing as AI demand accelerates.
A big number that is not a big share
Data centers use roughly 1.5–4.5% of electricity in the regions that host them. The number is growing fast — but it is still small next to heating, transport, and industry.
Start with the most cited figure. The International Energy Agency (IEA) estimates that data centers consumed around 415 terawatt-hours (TWh) of electricity globally in 2024 — about 1.5% of total world electricity.[1] In the United States, the Lawrence Berkeley National Laboratory (LBNL) put 2023 data-center consumption at 176 TWh, roughly 4.4% of national electricity, and projected that share to climb to between 6.7% and 12% by 2028.[2]
So data centers are a meaningful slice, not a dominant one. For comparison, global electricity demand is itself rising about 3.6% a year through 2030, driven by heat pumps, electric vehicles, air conditioning, and industry — data centers are one pressure among several, not the whole story.[3]
"For all the attention they get, data centers are still roughly one-hundredth of global final energy consumption."
— paraphrasing the IEA's framing of the 1.5% figure[1]
Efficiency is real — and hitting a wall
The standard yardstick for energy efficiency is PUE, Power Usage Effectiveness: the ratio of total facility power to the power that actually reaches the IT equipment. A perfect PUE is 1.0; everything above 1.0 is overhead, mostly cooling.
The hyperscalers are extraordinary here. Google's fleet averages a PUE near 1.10, meaning only about 10% of power is wasted on cooling and overhead.[4] But the broader industry is much worse. The Uptime Institute's 2024 global survey found an industry-average PUE of about 1.58, and that figure has been flat for five consecutive years after two decades of steady improvement.[5] The easy efficiency gains are mostly captured. What's left is harder.
That flatness matters. Efficiency used to absorb demand growth — servers got more powerful per watt, cooling got smarter, utilization rose. That offset is now largely exhausted just as AI workloads arrive with much higher power density per rack. The result is the growth curve in the next chart.
Carbon: the strongest part of the case for data centers
If energy is the headline, carbon is where the data-center industry has its best story. The IEA estimates that electricity-related emissions from data centers sit around 180 million tonnes of CO₂ today, rising to roughly 300 Mt by 2035 in the base case — and the agency projects emissions will peak around 2030 and then decline, even as electricity use keeps climbing.[6] That decoupling is driven by grid decarbonization: as power grids add renewables, every kilowatt-hour a data center draws gets cleaner.
To put that in global context: data centers today account for well under 1% of global energy-related CO₂, and are projected to reach about 1% by 2030 in the central scenario.[7] For all the headlines, computing's carbon footprint remains a small fraction of the sectors that actually drive climate change.
Three forces keep the carbon number lower than the electricity number suggests:
Why emissions lag electricity
1. Cleaner grids. The biggest operators buy renewable energy at enormous scale, and the underlying grids they sit on are decarbonizing regardless. 2. Better hardware. Compute per watt keeps rising — not as fast as it used to, but enough to matter. 3. Colocation economics. Operators are fanatical about utilization because idle servers cost the same as busy ones, which pushes work toward fuller, more efficient facilities.
The caveat is location. A data center on a coal-heavy grid produces many times the carbon of the same facility on a hydro- or nuclear-heavy one. This is why siting decisions — Virginia vs. Quebec vs. Iceland — matter more than any incremental efficiency gain inside the building.
The hidden number, and the real concern
Water is where data centers do their quietest damage — most of it evaporates, and most of it was never disclosed.
Electricity gets the attention because it is metered, traded, and priced. Water is different. It is cheap, often unmetered at the facility level, and the bulk of what a data center "uses" does not come back — it evaporates through cooling towers and drifts away on the wind. That distinction matters and is widely misunderstood.
Withdrawal vs. consumption — the distinction that changes everything
Two terms get conflated, and the difference is the whole story:
- Withdrawal is water taken from a source. A nuclear plant withdraws enormous volumes of water for cooling, but returns nearly all of it (warmer) to the river. Withdrawal is not the same as use.
- Consumption is water that is withdrawn and not returned — evaporated, incorporated into a product, or otherwise removed from the local watershed. This is the figure that depletes aquifers and rivers.
Data centers, despite their small withdrawal numbers, are almost entirely consumptive users. Evaporative cooling — the dominant method because it is energy-efficient — turns liquid water into vapor and sends it somewhere else. That water is gone from the basin where the facility sits.[8]
A data center can be frugal with water on paper and a net drain on the town it sits in.
— the core tension of the water story
The standard metric: WUE
The industry's yardstick is WUE, Water Usage Effectiveness, defined by The Green Grid as annual water consumed for cooling and humidification (liters) ÷ annual IT energy (kWh).[9] Lower is better. A typical evaporative-cooled facility runs 0.5–1.0 L/kWh; a closed-loop or liquid-cooled one can approach zero. Microsoft reported an average WUE of 0.27 L/kWh in 2025, well below the industry pack.[10]
Translated: a typical hyperscale facility consumes roughly 1–2 gallons of water per kWh of IT load. Multiply that by the hundreds of megawatts a large campus draws, 24 hours a day, and the numbers become municipal-scale.
Per-query: a bottle of water for a conversation
The most vivid figure in this whole debate comes from a 2023 study by Shaolei Ren's group at UC Riverside: training and running GPT-3 consumed roughly a 500 ml bottle of water for every 10–50 responses, depending on the weather and where the servers sat.[15] That is not a measure of one prompt's worth of electricity — it includes the cooling water evaporated to shed the heat the chips produce, plus the water consumed upstream at the power plant generating the electricity.
The same study estimated that training GPT-3 alone consumed around 700,000 liters of fresh water.[15] That number has only grown with larger models. The point is not that a single ChatGPT query will dry up a river — it is that the cost is real, invisible, and additive at scale.
Figure 2 · Where US water goes — and where data centers sit
US freshwater withdrawals by major use, compared with estimated data-center consumption (billions of gallons per day).
Sources: USGS 2015 water-use estimates (most recent published compilation); data-center figure from EESI (~449M gal/day in 2021).[16][17] Bars are withdrawals, not consumption — data-center withdrawal is also nearly all consumptive.
Two things are true at once, and both matter. Nationally, data centers are a rounding error in US water use — about 0.14% of withdrawals, against thermoelectric power and agriculture which together account for over 80%. If the question is "are data centers draining the country's water," the answer is plainly no.
But water is not a national problem. It is a watershed problem. A facility that is invisible at the national scale can still be the single largest water consumer in its county, and that is exactly what happens in the places data centers like to build.
What changes when inference goes generative
The pre-AI data-center story was one of remarkable efficiency. From 2010 to 2018, global data-center compute grew roughly sixfold while electricity use barely moved — the gains were absorbed by virtualization, better cooling, and the shift from small server rooms to hyperscale facilities.[1] That era is over.
AI workloads break the old model in three ways:
1. Density
A traditional rack draws 5–10 kW. An AI rack full of GPUs can draw 50–100+ kW, and the next generation pushes higher.[18] That heat cannot be moved by air alone, which is forcing a shift to liquid cooling — and changing the water math.
2. Inference at scale
Training a big model is expensive but finite. Inference — answering every user query, forever — is open-ended. ChatGPT-style services now run millions of queries per minute, each consuming power and cooling water, with no saturation in sight.
3. Water intensity
Higher-density AI cooling can consume 10–50× more water per unit of compute than traditional air-cooled facilities under comparable conditions.[18] Combined with siting in hotter, drier climates (which demand more evaporative cooling), the AI buildout is concentrating water load in some of the worst possible places.
The efficiency offset that absorbed data-center growth for fifteen years is exhausted. From here, growth shows up directly in the grid and the watershed.
The IEA's base case now has global data-center electricity more than doubling from 415 TWh (2024) to about 945 TWh by 2030, with a high-growth scenario that pushes higher still.[1]
The problem isn't the total. It's the where.
Water and electricity share a feature that makes them politically explosive: they are local. Averaged across a country, data centers are a footnote. Inside a single watershed, they can be the dominant user. That gap between the macro and the micro is where the real conflict lives.
Bloomberg's 2025 investigation found that more than 160 new AI data centers were built in the US between 2022 and 2025 in regions with high competition for scarce water.[19] A separate analysis from Ceres projected that data-center growth could increase water stress in already-strained basins by up to 17% annually, with sharper spikes during peak (hottest) cooling months.[20]
Three places the story turned into a fight
The Dalles, Oregon
Google's campus in The Dalles (population ~16,000) consumed roughly 550 million gallons in 2024 — about 40% of the entire city's water.[14] The figure only came out because the city spent thirteen months in court fighting a public-records lawsuit from The Oregonian to keep it secret; Google had contractually forbidden city officials from disclosing its water use when seeking permits to expand.[21]
Council Bluffs, Iowa
Google's Iowa data-center complex drew roughly 1.3 billion gallons in 2024 (3.8 billion liters), the largest single-site figure publicly disclosed.[13] Iowa is not as water-stressed as the desert Southwest, but at this scale a single campus is the equivalent of a small city's entire demand.
The desert Southwest & arid West
Arizona, New Mexico, and Utah have become major data-center corridors despite sitting on some of the most water-stressed basins in North America. Fortune's 2026 reporting on "America's data centers are thirsty" documented tanked water pressure and contested groundwater in rural host communities in both Arizona and Georgia.[22] Siting in the desert is cheap for land and power — and expensive, invisibly, in cooling water.
A 0.14% national share means nothing to a town watching its reservoir drop while a server farm next door runs its cooling towers flat out.
The counterarguments, taken seriously
A fair accounting includes the case for data centers — and several common criticisms don't survive contact with the data.
Genuinely worth worrying about
- Watershed-level water load. Data centers can dominate local water budgets in arid regions where it matters most.[19]
- Opacity. Most operators still won't disclose site-level water figures; Google fought disclosure in court for over a year.[21]
- Consumptive loss. Evaporative cooling removes water permanently from the basin — it doesn't come back.[8]
- The AI growth curve. Doubling electricity by 2030 with no offsetting efficiency gains ahead.[1]
Often overstated
- "Data centers will boil the planet." They're <1% of global CO₂ and projected to peak this decade even as electricity rises.[6][7]
- "They drain the country's water." ≈0.14% of US withdrawals — agriculture and power plant cooling dwarf them.[16][17]
- "Efficiency is impossible." Hyperscalers run PUE near 1.10; the technology exists.[4]
- "Nothing is being done." Microsoft launched zero-water-cooling designs in Aug 2024; both Google and Microsoft pledge to be water-positive by 2030.[23][24]
The strongest counterarguments, in detail
1. Efficiency gains are real and large
The 2010–2018 plateau in data-center electricity — sixfold compute growth, flat energy — was one of the great untold efficiency stories of the century. Hyperscalers pushed PUE from ~1.5 toward ~1.1, and academic estimates suggest that better IT and cooling practices can still cut a typical facility's energy use by 20–40%.[25] The technology is not the bottleneck; deploying it across the long tail of older, smaller facilities is.
2. Cooling technology is changing fast
The water intensity of data centers is not fixed — it is a design choice trading water against electricity. The three main approaches tell the story:
| Cooling method | Typical WUE (L/kWh) | Energy cost | Trade-off |
|---|---|---|---|
| Evaporative (cooling towers) | 0.5–1.0+ | Lowest (best PUE) | Highest water consumption |
| Closed-loop (chillers / dry coolers) | ≈0 (top-up only) | Higher | Up to ~70% less water, more electricity |
| Immersion / direct-to-chip liquid | ≈0 | Lowest | Capital cost; best for high-density AI racks |
Microsoft's August 2024 launch of a zero-water-cooling datacenter design built for AI workloads is the most concrete sign that the water problem is an engineering choice, not a law of physics.[23] The trade-off — using more electricity to use less water — is acceptable on a clean grid and painful on a dirty one. There is no free lunch, but the menu is expanding.
3. Replenishment commitments (with caveats)
Both Google and Microsoft have pledged to be "water positive" by 2030 — replenishing more water than they consume. Google's commitment is to replenish 120% of consumption; Microsoft has reported reaching water-positive status in FY2025.[24] Google projects its restoration projects will return over 19 billion gallons annually by 2030.[24]
The caveat is where. Replenishing a watershed in Costa Rica does nothing for a drying reservoir in Arizona. The accounting is real; whether it counts as solving the problem depends entirely on whether the replenishment happens in the same basins bearing the consumption load. Mostly, it doesn't.
4. Siting on clean grids
The carbon case improves markedly when data centers locate on hydro, nuclear, or wind-heavy grids — Quebec, Iceland, Norway, the Pacific Northwest. The same megawatt-hour that produces ~800 g CO₂ on a coal-heavy grid produces near-zero on hydro. Siting is, quietly, the single biggest carbon lever the industry has.
The next five years will settle the argument
Three things are now in motion simultaneously, and how they interact will determine whether the data-center story of the late 2020s is a climate problem or an engineering success.
The demand curve is locked in. Whatever happens to efficiency, the AI buildout is happening — LBNL's projection of 6.7–12% of US electricity by 2028 is not speculative; the facilities are permitted and under construction.[2] The IEA's doubling-to-945-TWh base case is the conservative path.[1]
The grid is decarbonizing, unevenly. US power-sector emissions have been falling for fifteen years and the IEA expects data-center emissions to peak around 2030 even as electricity rises.[6] But that average hides wide variance: a data center on a coal-heavy grid is still a coal plant by proxy.
The water question is the open one. Closed-loop and immersion cooling can cut water consumption to near zero. Replenishment commitments exist. The technology and the pledges are in place. What is missing — and what would change the whole debate — is mandatory, site-level disclosure. Today the public cannot find out how much water most data centers consume, in most of the places they operate. Until that changes, every national average hides a local fight.
The data center of 2030 will use twice the electricity of 2024. Whether it uses twice the water is a decision, not a forecast.
So: are data centers bad for the environment? At the global scale, no — they are a small and shrinking share of the climate problem. At the watershed scale, in the arid West, behind a wall of corporate secrecy — yes, genuinely, and getting worse. The honest answer depends on the denominator you choose, and on who gets to see the numbers.