Climate series, part 8

The Cloud Next Door

Globally, data centres are a rounding error on the grid. In the towns that host them, they are the biggest thing on the meter. This essay follows the numbers from a town's water system out to the planet, and tags each one by how well anyone actually knows it.

Sources at the end. Every figure carries a tag: ✓ measured · ≈ estimate · ? disputed

scroll ↓

The town

A third of the water

In 2021 a reporter for The Oregonian asked the city of The Dalles a simple question: how much of the town's water goes to Google? The city hosts three of the company's data centre sites, on the Columbia River in Oregon.

The city sued him. Google's water use was a trade secret, it argued in December 2021. The county district attorney sided with the paper. The city settled in November 2022, released ten years of records, and paid the paper's legal fees ✓.

The records showed why the number was worth hiding. Google's use had nearly tripled in five years. It now took more than a quarter of all the water the city pumped. By 2024 it was a third, about a million gallons a day, up from 12% in 2012 ✓. The town has about 16,200 people ✓.

This essay is about what a data centre takes: the electricity, the water, the ground it sits on. It is also about who gets to see the meter. Unlike the other essays in this series, this one cannot simply trust its numbers: the data is sparse, contested, and changing month to month. So the grade of a figure matters here as much as the figure, and each one is tagged. ✓ is measured: a meter, a court record, an audited filing. ≈ is an estimate: a model, with a method behind it. ? is disputed: a number that changes with who does the counting. Watch the tags as the essay zooms out. They are the story.

What's metered

Ireland and Virginia, straight off the meter

Ireland keeps the cleanest count anywhere. Its statistics office reads data centre power straight off the network meters and publishes it. In 2015 data centres used 5% of the country's metered electricity. In 2024, 22%, or 6,969 GWh. In 2025, 23% ✓. That is more than every urban household in Ireland put together ✓. (Count all households, urban and rural, and they still edge ahead at 28%.) The gap is closing fast: data centre demand has grown about 10% a year lately, everyone else about 3%.

The grid operator has already drawn the line. Since 2022 EirGrid takes no new data centre connections in the Dublin region until about 2028 ✓. Around Dublin, data centres are reported to draw about half the electricity ≈.

Virginia counts too. JLARC, the state legislature's audit arm, spent 2024 on it. Northern Virginia is the largest data centre market in the world, and 80% of the state's data centres sit in three counties, with Loudoun County alone holding about half ✓. Virginia's power demand was flat from 2006 to 2020. JLARC's forecast now has it doubling within ten years, data centres the main reason, and a typical Dominion home paying $14 to $37 more a month by 2040 in today's money ≈. The data centres pay their own way today; the rise is what the whole system has to build to keep up. And Dominion's 2023 plan says as much in its own words: it expects to keep "existing carbon-emitting units beyond [the 2045] statutory retirement deadlines".

Up close, none of this is subtle. The shares are big, the trends steep, the numbers metered and public. Sometimes only after a lawsuit.

What's modelled

Zoom out, and every number becomes a model

Nobody meters the planet. Once you leave the metered countries, every number is a model.

The IEA's central figure puts data centres at about 415 TWh in 2024, roughly 1.5% of the world's electricity, heading for about 945 TWh by 2030, just under 3% ≈. Push on to 2035 and its scenarios spread from about 700 TWh to more than 1,700, a range wider than the thing it measures. In the United States, Lawrence Berkeley National Laboratory builds the number up from server shipments: 176 TWh and 4.4% of US electricity in 2023, 192 TWh and 4.7% in 2024 ≈, and somewhere between 325 and 580 TWh by 2028, which is 6.7% to 12% of US power ≈. The width is honest. It rides on GPU shipments, how hard the machines run, and how they are cooled, none of which anyone can see yet.

The recent past is why people are nervous. From 2010 to 2018 the world's data centre compute grew about 550% while the energy to run it grew about 6%, to around 205 TWh, near 1% of global electricity ≈. For a decade, efficiency ate the growth. The shipment data show that run ending around 2017, when the GPU-heavy servers arrived; US use more than doubled from 2017 to 2023 ≈. The machines still get more efficient every year. They just stopped getting efficient faster than they multiplied.

The whole argument fits one line of the IEA's base case. Data centres are less than 10% of the world's electricity demand growth to 2030, and more than 20% of it in the rich economies, with the US and China taking about 80% of the data centre share ≈. Both sides of the shouting match are holding real numbers. "It's only 1.5%" is true of the planet. "It will eat the grid" is nearly true in Dublin and Loudoun County. The gap between them is just the zoom level. And not one of the global figures is a meter reading.

Five distances

How big is the cloud? Depends where you stand

The same industry, seen from five distances. Each bar is the share data centres take of the local total. Each tag says how well the number is known.

The top row is water; the rest are electricity. The comparison is between shares, not substances. As you pull back, two things shrink at once: the share, and how well anyone knows it.

What's disclosed

Water, on the companies' word

Water has no CSO and no JLARC. Here the meter belongs to the company, and the tags need reading with that in mind.

Start with what "using water" means, because it stands for two numbers, not one. A data centre cools its machines by running water round a loop and over cooling towers. The loop throws off heat by letting some of that water evaporate, so it has to be topped up with fresh supply, constantly. Withdrawal is the total a site draws from the river or the mains. Consumption is the part that evaporates and never comes back. It is not a sealed loop and it is not water pulled out and gone; it is a loop that leaks to the sky. An air-cooled design consumes almost none, which is the whole point of the redesigns further down. And none of these figures count the water spent upstream at the power station to make the electricity, which is often larger than what the site uses itself.

Google says it used 8.1 billion gallons of water across its data centres and offices in 2024, up 28% in a year ✓. Microsoft reports 8,170 megalitres for its 2025 financial year, up 22% and more than double what it was five years before ≈. That one is an estimate on purpose: since 2024 Microsoft no longer meters the water, it works it out from an efficiency ratio, and says so. Amazon went one further. Until June 2026 it published only the ratio, 0.15 litres per kWh, and no volume at all; its first real figure, about 2.5 billion gallons in 2025 ✓, counts water withdrawn, not water consumed, and some of that goes back to the river. Three companies, three numbers, no way to line them up.

Where the water comes from is a local question, and the local story keeps repeating. Tucson's council voted 7 to 0 in August 2025 against hooking up Project Blue, a campus tied in reporting to Amazon ✓; the developer moved onto county land and sank wells cleared for about 31 million gallons a year ≈. In Aragón, three Amazon data centres are licensed for up to 755,720 cubic metres a year, enough to irrigate more than 200 hectares of maize ≈. In Chile, a court threw out part of Google's permit for its Cerrillos site in September 2024 over the water; Google had already redrawn the plans for air cooling ✓. In Uruguay, after the 2023 drought left the capital short of drinking water, Google redrew a site there for air cooling too; the first design was reported to want 7.6 million litres a day. Four places, and each time the pressure worked. Each time, the disclosure came after it.

What's disputed

The numbers nobody can give you

Now the numbers you have actually seen shared: the per-question ones. This is where the tags run out of road.

How much does one AI query cost? Google says a typical Gemini text prompt takes 0.24 Wh, emits 0.03 grams of CO₂e, and uses 0.26 millilitres of water ?. Epoch AI, working from the outside, put a typical GPT-4o query near 0.3 Wh ?. A 2023 estimate by de Vries, widely quoted, said about 3 Wh ?. That is a twelvefold spread, and almost none of it is measurement. It is where each study draws the line. Google's own paper makes the point: count only the chip and the number halves, to 0.10 Wh. Mistral, counting the whole life of the model, gets 1.14 grams of CO₂e and 45 millilitres of water for one 400-token reply ?, nearly forty times Google's carbon figure, because it folds in the training, the making of the hardware, and the grid as it really runs. None of these is wrong. They answer different questions under the same headline.

The same split runs through the carbon books, with more at stake. Meta's 2024 scope 2 emissions are either 1,358 tonnes or 5,967,348, a factor of 4,400, depending which column of the same audited filing you read ✓. The small number is market-based: renewable certificates set against the power drawn. The big one is location-based: the grid as it actually was. How much do the certificates flatter the books? One study of 115 companies with science-based targets found reported scope 2 cuts of 30.7% from 2015 to 2019 fall to 9.9% once you take the certificates out ≈. Two-thirds of the progress was paperwork. Google, to its credit, keeps the strict score on itself: matched hour by hour on the grids it actually runs on, it was 66% carbon-free in 2024 ✓, not the 100% its annual certificate maths has claimed since 2017.

That is why the global argument never ends. At planetary scale the numbers are built, and you can build them either way.

A question next to a flight

The comparison that won't hold still

You have seen the format: one AI question equals a phone charge, or a bottle of water, or a light left on. It is a tempting way to make a vague cost feel real. On this subject it is a trap, because the number underneath it is the disputed one. In September 2026 OpenAI's Sam Altman gave the trap its purest form: 38,000 queries, he said, cost the water of a single almond, and a large data centre draws no more than an office building running its taps ?. The almond is solid, about 1.1 gallons of it in a California drought; the query beside it is the disputed figure in a costume, a per-question number no one outside the company can check, and which the researchers who have tried say has no public data to check it against. Take three anchors instead. Two are solid. One is a range.

One AI text prompt 0.03 to 1.14 g CO₂e ?
One mile in an average car about 270 g CO₂e ≈
One short-haul return flight, London to Barcelona 289 kg CO₂e ≈

Work the equivalence and it falls apart in your hands. A mile of driving is worth somewhere between a couple of hundred and several thousand prompts, depending whose figure you trust. One short-haul flight is worth between a quarter of a million and nearly ten million. The spread is not wobble around a true value. It is the same line-drawing problem as before, blown up. The comparison tells you more about the accountant than the query.

Step back from the prompt to the person, though, and the fog clears. Ask a chatbot ten questions a day for a year, about 3,650 prompts, and the carbon comes to between roughly a hundred grams and four kilograms ?: half a mile of driving at one end, a short errand at the other, and either way nothing beside a single holiday flight. At the scale of one person typing, AI is not where the emissions are.

The weight is somewhere else. Training one frontier model is a different order of thing. Mistral put the training of one of its models at about 20,000 tonnes of CO₂e ?, tens of thousands of those flights, paid once and then spread thin over every answer the model ever gives. That shared training is most of why the whole-life figure sits so far above the running one.

So the flight comparison points the wrong way. Your questions are nothing, the training is huge but shared, and neither is a meter reading. The number you can read, and argue over, is the one from earlier: the substation in Loudoun County, the reservoir above The Dalles. That is where the cost stops being an estimate and turns into a bill.

Back to the meter

Where the numbers are real

So the real fight is where the numbers are: at the meter, in the court file, at the council vote. One tracker counts $64 billion of US data centre projects blocked or delayed by local opposition through March 2025 ?. The tag is fair, since the tracker will not say who funds it and its dollars are announced, not spent. But you can see the direction without the total: 7 to 0 in Tucson, a court in Santiago, a connection freeze in Dublin, an air-cooled redesign in Uruguay.

And The Dalles. The town that sued to keep one water figure secret is now planning a $70 million dam and a bigger reservoir, sized for an unnamed industrial user who needs at least a million gallons a day ✓. You can argue with the plan. You cannot argue about the number any more, because this time it is on the record.

Across the planet, data centres are a modelled 1.5% of the electricity. In Dublin and Loudoun County they are the biggest load on the grid. In The Dalles they are a third of the water. The closer you stand, the clearer the number gets; the further off, the more of the haze is put there on purpose. At the end of one pipe, in a town of sixteen thousand, the meter reads what it reads.

All climate posts

Sources

The Dalles
Reporters Committee for Freedom of the Press, settlement record (Nov 2022); OPB, city water and dam expansion reporting (to Jan 2026).
Ireland
CSO Ireland, Data Centres Metered Electricity Consumption, 2024 and 2025 releases (ESB Networks meter data); EirGrid Dublin connection statement; CRU direction on data centre connection policy.
Virginia
JLARC, Data Centers in Virginia, Report 598 (Dec 2024), which also quotes Dominion Energy's 2023 Integrated Resource Plan.
Global and US models
IEA, Energy and AI (2025) and its Key Questions update; LBNL, 2024 United States Data Center Energy Usage Report and 2025 Update (DOI 10.71468/P1RP4F); Masanet, Shehabi, Lei, Smith & Koomey, Science 367:984–986 (2020).
Corporate water
Google 2025 Environmental Report; Microsoft 2026 Environmental Data Fact Sheet (incl. its WUE-estimation note); Amazon sustainability disclosures via Latitude Media (Jun 2026).
Local water conflicts
AZ Luminaria and KJZZ (Tucson, Project Blue); SourceMaterial and The Guardian (Aragón, Apr 2025); Climate Litigation Database (Cerrillos tribunal ruling, Sept 2024); Developing Telecoms (Uruguay, Teros).
Per-query estimates
Google, arXiv:2508.15734 (Aug 2025); Epoch AI (Feb 2025); de Vries, Joule 7 (2023); Mistral AI lifecycle assessment (Jul 2025).
The almond claim
Altman on the Sources podcast with Alex Heath, via Tom's Hardware (Sept 2026); CalMatters fact-check quoting Shaolei Ren (UC Riverside) and Michael Kiparsky (UC Berkeley).
Carbon accounting
Meta / EY Independent Accountants' Review Report FY2024; Bjørn et al., Nature Climate Change 12:539–546 (2022), DOI 10.1038/s41558-022-01379-5; Google 2025 Environmental Report (24/7 CFE).
Community pushback
Data Center Watch (10a Labs), through Mar 2025; NBC News (Jun 2026).

Tag discipline: ✓ (measured) means metered, court-released or audited primary data; ≈ (estimate) means a model with a stated methodology; ? (disputed) means the value is contested or boundary-dependent. Corporate disclosures are graded on their own stated basis, which is why Microsoft's water is an estimate (calculated from a ratio since FY24) while Google's is measured (self-reported meter and invoice data). Ireland's percentages use the CSO's metered-consumption denominator, which differs slightly from EirGrid's share-of-demand figures. Water consumption and withdrawal are different quantities and are never compared directly. The per-query figures are shown only as a spread with their boundaries named, never as a point estimate. Full working: the claim register in the site repository (tools/datacenters/research/).

Get new essays by email

No spam, unsubscribe anytime · or use the Atom feed