A Read ·
How much water does AI really use?
Towns are fighting data centres over water. How much does AI use, and where does it matter?
In April 2026, with Georgia in drought, people in Fayette County were asked to water their lawns only two days a week. Around the same time, a county letter came to light. It said a data centre being built nearby had used more than 100 million litres of county water without being billed.
Santiago, Chile, had its own fight two years earlier. A court there put a planned Google data centre on hold until its effect on the city’s aquifer, the underground water it would draw on, was assessed again. That aquifer was already promised to more users than it could supply.
And in The Dalles, Oregon, the city went to court to keep Google’s water use secret. It sued the newspaper that asked, and Google paid the legal bills. When the records finally came out, they showed Google taking about a quarter of the city’s water in 2021.
Put stories like these side by side and it sounds as if AI is drinking the world dry. You may have heard that every email a chatbot writes uses up a bottle of water. But look closer and the three fights are very different: one is over very little water, one over an aquifer with none to spare, and one over a lot. Telling them apart starts with a simpler question. What does a data centre do with water in the first place?
AI runs hot
Every answer a chatbot gives is worked out on computer chips in a data centre, a warehouse of computers that runs AI along with the rest of the internet. Those chips turn almost all the electricity they use into heat, and the heat has to go somewhere. Many data centres get rid of it the way your skin does on a hot day: they let water evaporate, because evaporating carries a lot of heat away.
If the water evaporates, doesn’t it just rain back down? It does, on average about nine days later, but usually hundreds or thousands of kilometres away. Even a basin as big as the Mississippi’s gets back less than a third of the water that evaporates from it. So for the town that supplied it, the water is gone. It has left the river or the aquifer in the season it was taken, and an aquifer can take years, or far longer, to refill.
The electricity costs water too. Coal, gas and nuclear power stations turn water into steam to spin their turbines, and many of them cool that steam by evaporating more water; wind and solar use almost none. Add it all up and almost two-thirds of AI’s water is used up at power stations rather than at data centres, sometimes far from the computers themselves.
Most of the cooling water goes up as vapour
And the power station behind it uses up even more
Source: This piece’s water model (v1.3); methods below See the numbers: Most of the cooling water goes up as vapour
Hide the numbers
| What | Share | Note |
|---|---|---|
| Share of the water drawn that evaporates, at Google’s data centres | 78% | Google fleet 78%, Microsoft 62% and Meta 55% (whole companies, offices included); Google sites drawing over 190 million litres a year 44–91% (median 77%) |
| Share of AI’s water used up at the power stations that supply it (2025) | 64% | This piece’s water model (v1.3): power stations’ share of the water per message |
The rest, the data centre’s own share, mostly comes from the local supply: the same rivers, reservoirs and wells as the town around it. At Google’s data centres, about 81% of it is clean enough to drink.
Your share
How much of that is down to you?
How much water does one conversation with a chatbot use?
Drag the water in the bottle to your guess. The bottle holds 500 mL.
Your guess: about half a glass (125 mL)
Up and down arrow keys move the water a little, Page Up and Page Down ten steps, Home and End go to 1 mL and to full. Enter checks your guess.
3.2 mLabout two-thirds of a teaspoon
An average conversation, about four questions and answers. The water counted is what’s used up at the data centre and at the power stations that supply it.
Source: This piece’s water model (v1.3); methods below See the numbers: How much water does one conversation with a chatbot use?
Hide the numbers
| Figure | Water |
|---|---|
| Our estimate | 3.2 mL |
| Likely range | 1.2–8.6 mL |
| A teaspoon, for scale | 5 mL |
| A glass | 250 mL |
| A bottle | 500 mL |
The bottle-per-email figure came from a 2024 Washington Post graphic, which assumed that writing one email took 140 watt-hours of electricity. That looks far too high even for then: Google later said its own typical chatbot prompt in May 2024 used about 8 watt-hours. But the world changed too. A year later, by Google’s own figures, the energy for its typical prompt had fallen about 33-fold, mostly because the models had become more efficient. Counted in multiples rather than millilitres, roughly half of the gap between a bottle per email and the few drops a single question uses today is that real improvement, and most of the rest is the poor guess.
The answer above is for an average conversation, and most are shorter: a question or two uses under a millilitre. Longer chats cost more with every reply, because before each answer the chatbot rereads everything said so far.
Some jobs cost far more. A few seconds of AI video uses about half a glass, and AI agents left to work through a big job on their own can use much more, as the end of this piece shows. For everyday chatting, though, it’s a teaspoon here and there.
Even a year of it stays small. In mid-2025, ChatGPT’s weekly users sent about four messages a day on average, which over a year comes to about a litre of water each, roughly the irrigation water behind the beans for one cup of coffee.
Now multiply it by everyone
But you’re not the only one chatting. Hundreds of millions of people use AI every week, and their teaspoons add up.
Start with a single day. In mid-2025, ChatGPT’s users sent about 2.6 billion messages a day between them, which used up roughly an Olympic swimming pool of water (somewhere between a third of a pool and two pools).
Then widen it to all of AI, and ChatGPT’s chats turn out to be a small part. By my rough estimate they come to less than 1% of AI’s water. The rest goes to other chatbots, to AI built into search, feeds, apps and businesses, and to training new models. Altogether, AI used up about 127 billion litres in 2025 (likely between 84 and 188 billion).
Shared out across everyone on Earth, all of AI’s water would come to about 15 litres each a year, about a week of drinking water. But it isn’t shared out. It lands in a few places.
Is that a lot?
Next to the oceans, that cube is nothing, but the oceans are the wrong comparison. AI’s cube is almost all fresh water, taken from rivers, lakes and aquifers and used up: the same kind people drink and farms irrigate with. Against all the fresh water people use up in a year (a 1996–2005 average, the latest global count), it’s about one part in 8,000, and most of the rest goes to farming.
Is it worth the water? Some uses are easier to defend than others. Counting the same kind of water, the world’s beef cattle use about 280 times as much as AI, and they feed people. That doesn’t make AI’s water free, because a bigger user somewhere else doesn’t refill a town’s aquifer. Nor can money settle it. A litre earns far more in industry than on a farm, but that says nothing about which is needed more.
Could it ever get bad enough to threaten food? Not worldwide. Even in my thirstiest sketch for 2030, built on the International Energy Agency’s highest case, AI would use about one part in 1,100 of the fresh water people use up. From today, it would have to grow about 80 times just to reach 1%.
Still, the total is heading up. A short answer keeps getting cheaper in energy and water, but people and companies are asking for far more answers, and newer reasoning models and AI agents work at length before they reply (more on agents at the end). Google’s data centres used up twice as much water in 2025 as in 2022.
In all four scenarios I sketched for 2030, from efficient and dry to a building boom in dry places, AI’s water grows, by somewhere between two and seven times. How fast demand grows matters most for the total. Where the new data centres go decides where it hurts, and by one recent count a quarter of today’s already sit in water-stressed places. The laws now arriving mostly make companies report their water, not limit it.
Locally, the question is who gives way when a drought comes. In Taiwan in 2021, as reservoirs ran dry, the government stopped irrigating 74,000 hectares of rice to save water, while homes and industry, chip makers among them, competed for what was left. It had moved farm water to homes and industry the same way in six earlier droughts since 2002.
A cube taller than the Eiffel Tower, and a speck beside all the fresh water people use up: both are AI’s water, and most arguments pick one. What decides the harm is where it lands.
It depends where, and when
Which brings us back to the three places. A data centre doesn’t use water a teaspoon at a time. It takes it all together, in one place, and if it cools with water it takes the most in summer, just when everyone else needs it too. Whether that’s a problem depends on how much water the place has to spare. None of the three has yet run short because of a data centre, and each ended differently.
Fayette County’s was mostly a scare. The water was for construction, such as concrete and dust control, rather than for cooling computers, and it came to under 1% of what the county supplies. The lawn limits were a voluntary request under the state’s drought rules. The company paid the back bill at double the normal rate, and the county later said its own letter had been “unclear”.
Santiago’s limit was real. The court agreed the aquifer was over-promised, and Google has since dropped the original plan, saying it will start again with a design cooled by air. No new plan had been approved by October 2026.
The Dalles is where it bites. In 2024 Google took about 30% of the city’s water, for all its computing rather than only AI, and it took more again in 2025. The city’s own plan projects a small shortfall by 2034 as demand from businesses grows, and the squeeze comes in late summer, when demand peaks just as the city must leave water in its small main river.
In The Dalles, Google took almost a third of the city’s water in 2024
The city’s water in 2024, from the average daily demand in its water plan
Water taken in, including what’s later put back. Across the whole US, it would be too thin a slice to draw.
Source: This piece’s water model (v1.3); methods below See the numbers: In The Dalles, Google took almost a third of the city’s water in 2024
Hide the numbers
| Water | Share |
|---|---|
| Google’s data centres (all their computing, not only AI) | 30% |
| Everyone else | 70% |
| Google’s water here, as a share of all US fresh water drawn (2015 US total) | 0.00045% |
Tucson, Arizona, said no at the start. In 2025 its city council voted unanimously to end talks with a data centre that wanted city water. “We need to be the city that says not here, not now and not like this,” one council member said. The data centre is going ahead anyway, just outside the city, designed to be cooled by air and to draw on its own wells.
AI isn’t draining the world’s water. In towns like The Dalles, though, data centres take a share worth fighting over, and more are being built. For each new one, the question is whose water, and in which month.
What if this is wrong?
These numbers are estimates, and four things could move them.
- None of the big tech companies says how much water its AI uses in total. They report data centres as a whole, so AI’s share here is worked out from its share of their electricity.
- Much of the measuring is Google’s own. Its figures for the electricity a question takes haven’t been checked by anyone outside.
- Place matters a lot. Leaving out the lakes behind dams, the same conversation uses about 70 times as much water in Mumbai as in Stockholm, because of the heat and how the power is made. Count those lakes and hydro-heavy Stockholm comes out higher than Mumbai.
- What counts as “using” water is a choice. This piece counts water that’s used up, as the researchers behind the bottle did. Count water that’s borrowed and put back as well, mostly by power stations, and the figures grow about 16 times.
All the checking for this piece was done by AI agents, not a human expert.
How the numbers were worked out
The figures come from a model built for this piece (version 1.3). The water for a question is the electricity it takes, times the water used up per unit of electricity at the data centre and at the power stations that supply it. A conversation adds up its questions and answers, including the cost of reading the conversation again each time. Each input is a range taken from company reports, government statistics and published research, and the model tries 50,000 random combinations of them to give the ranges shown.
The lakes behind hydroelectric dams are left out, because the same lakes also supply farms and towns. Counted at the median published rate for evaporation from those lakes, they would roughly double these figures. At a much higher published rate, an average across 35 dams of 245 litres per kilowatt-hour, the figures would be about 17 times bigger.
The famous figures, and where they come from:
- Half a litre per 20 to 50 questions (2023). From a University of California, Riverside paper on GPT-3, with 4 watt-hours per answer.
- A bottle per 100-word email (2024). From a Washington Post graphic using the same method, with 140 watt-hours per email, scaled up from a smaller model rather than measured.
- 0.26 mL per Gemini prompt (2025). From Google’s own estimate, counting only its data centres’ cooling water.
- “One fifteenth of a teaspoon” per ChatGPT question (2025). From Sam Altman, OpenAI’s chief executive, with no method given.
- 45 mL per Le Chat answer (2025). From Mistral, for its large model, on a whole-life count, though making the servers is only about a twentieth of it.
This piece’s own footprint, about 15 kilowatt-hours and 22 litres (likely 3–190 litres), comes from the logs of the AI sessions that made it, using energy measured on open models.
Every claim was traced back as far towards its original source as it would go. Separate checks re-opened the sources for nearly 270 of them, and an independent rebuild of the model checked the arithmetic.
What this piece cost in water
I made this piece with a lot of help from AI: agents that read reports, checked claims, built the model and drafted the graphics, sometimes working on their own for an hour or more. Anthropic, which makes the AI I used, doesn’t publish how much energy its answers take, so this is a rough estimate: by the afternoon of 22 October 2026, making it had used up about 22 litres of water (somewhere between 3 and 190), about as much as 6,900 average conversations.
Most of that, by my estimate, went on rereading. Over about 4,100 model calls, the agents reread everything gathered so far (notes, sources, code and drafts) before each step, about 275,000 tokens each time, roughly the length of a long novel. A typical ChatGPT user, chatting the usual way, would take about 20 years to use as much water.
This piece used as much water as about 6,900 conversations
Each dot is one average conversation, about 3 mL
Source: This piece’s water model (v1.3); methods below See the numbers: This piece used as much water as about 6,900 conversations
Hide the numbers
| What | Water | Conversations |
|---|---|---|
| One average conversation (about four questions and answers) | 3.2 mL | 1 |
| A year of an average weekly ChatGPT user’s messages (mid-2025) | 1.0 L | 326 |
| This piece’s AI agents (4,116 model calls; likely 3–190 L) | 21.8 L | 6,861 |
Work like this is where AI use is heading. Google says the number of tokens it processes each month, the word fragments AI reads and writes, grew more than 300-fold in the two years to May 2026. On OpenRouter, a large marketplace for AI models, models that reason at length went from a sliver of all tokens to over half in the year to November 2025. For a sense of scale: if each of ChatGPT’s weekly users set AI one job this size a week, that alone would use up about six times as much water as all of AI did in 2025, near the top of my sketches for 2030. Few of them would notice, because nothing shows them the meter.
I can’t tell you where that water was taken from, or in which month. After all this, that’s the part that matters.
Sources
The 109 sources behind this piece, in the order they come up.
Show all 109 sources Hide the sources
The opening scenes
- Fayette County. Fayette County Water System declares Level 1 drought response. Georgia, 2026.April 2026 request to water lawns two days a week.
- E&E News. Georgia residents seethe over 30M gallons of missing water. 2026.The county letter on unbilled water used by the data centre.
- Segundo Tribunal Ambiental. Sentencia, causas R-270-2020 y R-271-2020. Chile, 2024.Court ruling suspending the planned Google data centre.
- The Oregonian, via The Daily Astorian. The Dalles settles public records lawsuit over Google’s data centers. The Daily Astorian, 2022.The city’s lawsuit, and Google paying the legal costs.
- The Oregonian. Google’s water use is soaring in The Dalles, records show, with two more data centers to come. 2022.Google’s share of the city’s water in 2021.
- City of The Dalles. Water system master plan update. 2024.The city’s water demand, the late-summer peak, the river it must leave water in, and the 2034 shortfall.
AI runs hot
- van der Ent & Tuinenburg. The residence time of water in the atmosphere revisited. Hydrology and Earth System Sciences, 2017.Evaporated water stays in the air about nine days.
- van der Ent & Savenije. Length and time scales of atmospheric moisture recycling. Atmospheric Chemistry and Physics, 2011.Evaporated water travels hundreds to thousands of kilometres.
- Tuinenburg, Theeuwen & Staal. High-resolution global atmospheric moisture connections from evaporation to precipitation. Earth System Science Data, 2020.How little evaporation a basin gets back as rain.
- U.S. Geological Survey (Dieter et al.). Estimated use of water in the United States in 2015. USGS Circular 1441, 2018.Definition of consumptive use; US freshwater withdrawals.
- Gleeson, Befus, Jasechko, Luijendijk & Cardenas. The global volume and distribution of modern groundwater. Nature Geoscience, 2015.How little groundwater is modern, so slow to refill.
- World Resources Institute (Reig et al.). Guidance for calculating water use embedded in purchased electricity. 2020.Water per kWh of electricity, by grid and fuel.
- Shehabi et al. (Lawrence Berkeley National Laboratory). 2024 United States data center energy usage report. eScholarship, University of California, 2024.US data-centre electricity, power overhead (PUE), water per kWh (WUE) and power-station water.
- Siddik, Shehabi, Rao & Marston. Water IMPACT v1.1: US grid water intensity by balancing authority. HydroShare, 2024.US power-station water per kWh, with and without hydro.
- Siddik, Shehabi & Marston. The environmental footprint of data centers in the United States. Environmental Research Letters, 2021.Direct and indirect water of US data centres.
- Department for Energy Security and Net Zero. Digest of UK energy statistics (DUKES): electricity. GOV.UK, 2026.UK electricity generation mix, 2025.
- Byers, Qadrdan, Leathard, Alderson et al. Cooling water for Britain’s future electricity supply. Proceedings of the ICE: Energy, 2015.British power-station cooling water factors.
- Defra and Environment Agency. ENV15: water abstraction tables for England. GOV.UK, 2018.Water abstracted for power generation in England.
- Drax Group. Full-year results 2025 (regulatory announcement). FCA National Storage Mechanism, 2026.Drax power station output, for the GB grid’s water.
- U.S. Energy Information Administration. Thermoelectric cooling water data. 2024.US power stations by cooling type and water source.
- National Energy Administration of China. 四十年煤电脱胎换骨与新使命 (Forty years of coal power’s transformation, and its new mission). 2018.Freshwater use per kWh at Chinese thermal power plants.
- Ember. Yearly electricity data. 2025.Country electricity mixes for grid water estimates.
- Central Electricity Regulatory Commission (India). Order in petition 597/MP/2020. 2021.India’s water limits for thermal power plants.
- Google. 2026 environmental report. 2026.Google’s data-centre water used up and returned, its sources, The Dalles’ use, growth since 2022, and prompt energy in 2024–25.
- Microsoft. 2026 environmental data fact sheet. 2026.Microsoft’s water withdrawn, discharged and where it goes.
- Meta. 2025 environmental data index. 2025.Meta’s water sources, discharge and WUE.
- Equinix. Sustainability data summary 2024. 2025.Equinix water discharged to sewers; non-potable share.
- Sickinger, Van Geet, Belmont, Carter & Martinez (NREL). Thermosyphon cooler hybrid system for water savings in an energy-efficient HPC data center: results from 24 months and the impact on water usage effectiveness (NREL/TP-2C00-72196). OSTI.gov, 2018.Where cooling blowdown goes, and a data centre’s summer water peak.
- Amazon. 2025 Amazon sustainability report. 2026.Amazon data centres’ water per kWh (WUE).
- Microsoft. 2025 environmental sustainability report. 2025.How Microsoft defines its water per kWh (WUE).
Your share
- Zheng et al. LMSYS-Chat-1M: a large-scale real-world LLM conversation dataset. arXiv, 2023.Turns and tokens per chatbot conversation.
- Zhao et al. WildChat: 1M ChatGPT interaction logs in the wild. arXiv, 2024.Turns per chatbot conversation.
- Karnam et al. Bowling with ChatGPT: on the evolving user interactions with conversational AI systems. arXiv, 2026.Turns per conversation in donated ChatGPT histories.
- Zaman & Hoque. Vulnerabilities in personalization: assessing health privacy risks in ChatGPT logs and memory. arXiv, 2026.Prompts per conversation in donated ChatGPT histories.
- Yan et al. ShareChat: a dataset of chatbot conversations in the wild. arXiv, 2025.Turns and reply length in shared chatbot conversations.
- Microsoft Azure. Azure public dataset. GitHub, 2024.Context length in real conversation traces.
- DeepSeek. Open infra index. GitHub, 2025.Share of input tokens served from cache.
- Jegham et al. How hungry is AI? Benchmarking energy, water, and carbon footprint of LLM inference. arXiv, 2025.Energy cost of input tokens; energy per prompt.
- Epoch AI. How much energy does ChatGPT use? 2025.Energy per prompt and per input token.
- OpenAI, Anthropic, Google and DeepSeek. API pricing for cached input tokens. Provider documentation, 2025.Rough guide to the cost of rereading a conversation.
- Pranshu Verma & Shelly Tan. A bottle of water per email: the hidden environmental costs of using AI chatbots. The Washington Post (archived), 2024.The bottle-per-email figure and its 140 Wh assumption.
- Andy Masley. I think I figured out exactly how the “AI used a bottle of water per prompt” miscalculation happened. Andy Masley’s blog (archived), 2026.How the bottle figure’s energy guess was inflated.
- Elsworth et al. (Google). Measuring the environmental impact of delivering AI at Google scale. arXiv, 2025.Google’s energy and water per Gemini prompt.
- ML.ENERGY. The ML.ENERGY leaderboard (v3). 2026.Measured energy for chat, images and video on open models.
- Delavande, Pierrard & Luccioni. Video killed the energy budget: characterizing the latency and power regimes of open text-to-video models. arXiv, 2025.Energy to generate a short video.
- Chatterji et al. How people use ChatGPT. National Bureau of Economic Research, 2025.ChatGPT messages per day and weekly users, 2025.
- Mekonnen & Hoekstra. The green, blue and grey water footprint of crops and derived crop products (Value of Water report 47). Water Footprint Network / UNESCO-IHE, 2010.Irrigation water behind coffee beans.
- Chapagain & Hoekstra. The water needed to have the Dutch drink coffee (Value of Water report 14). UNESCO-IHE, 2003.Coffee per cup, for the coffee comparison.
Now multiply it by everyone
- FINA. Facilities rules 2017–2021. 2017.Volume of an Olympic swimming pool.
- International Energy Agency. Energy and AI. 2025.AI’s share of data-centre electricity; 2030 cases; PUE.
- Patterson et al. The carbon footprint of machine learning training will plateau, then shrink. Computer (IEEE), 2022.Inference uses more energy than training at Google.
- Wu et al. Sustainable AI: environmental implications, challenges and opportunities. Proceedings of Machine Learning and Systems 4 (MLSys 2022), 2022.Inference’s share of machine-learning energy at Meta.
- Hoekstra & Mekonnen. The water footprint of humanity. PNAS, 2012.All fresh water people use up, and farming’s share.
- Our World in Data, from the IEA. Data centres’ share of electricity demand. Our World in Data, 2026.Data-centre electricity by country, 2024–25.
- World Health Organization. Domestic water quantity, service level and health, second edition. 2020.People need about two litres of drinking water a day.
- European Food Safety Authority. Scientific opinion on dietary reference values for water. EFSA Journal, 2010.Daily water needs from drinks and food.
Is that a lot?
- Mekonnen & Hoekstra. The green, blue and grey water footprint of farm animals and animal products (Value of Water report 48). Water Footprint Network / UNESCO-IHE, 2010.Fresh water used up by the world’s beef cattle.
- UN-Water and FAO. SDG indicator 6.4.1: change in water-use efficiency over time. United Nations, 2023.A litre earns more in industry than on a farm.
- Aubakirova et al. State of AI: an empirical 100 trillion token study with OpenRouter. arXiv, 2026.Reasoning models’ growing share of tokens.
- Sundar Pichai. I/O 2026: welcome to the agentic Gemini era. Google blog, 2026.Growth in tokens Google processes each month.
- Rosa, Sangiorgio, Scanlon, Mazzotti et al. One quarter of global data centers are located in water-stressed areas. Environmental Science & Technology, 2026.A quarter of data centres sit in water-stressed places.
- CalMatters Digital Democracy. AB 2469 (2025–26 session). CalMatters, 2026.California law on data-centre water assessments.
- Holland & Knight. California’s 2026 data center laws: what you need to know. 2026.California data-centre water reporting rules.
- Utah Division of Water Rights. Data centers. 2026.Utah’s data-centre water reporting requirement.
- Beveridge & Diamond. Environmental highlights from Virginia’s 2026 legislative session. 2026.Virginia’s data-centre water reporting law.
- Minnesota Legislature. Laws of Minnesota 2025, 1st special session, chapter 12. Office of the Revisor of Statutes, 2025.Minnesota’s water review for large data centres.
- European Commission. COM(2026) 500 final: report on data centres’ energy performance. 2026.EU data-centre water reporting and average WUE.
- European Commission. Draft delegated regulation on a data-centre rating scheme, annexes, Ares(2026)3247482. Green Web Foundation (copy), 2026.Proposed EU water-efficiency labels for data centres.
- USDA Foreign Agricultural Service. Taiwan drought results in rice area reduction and crop loss. USDA GAIN report TW2021-0037, 2021.Taiwan halting irrigation of 74,000 hectares in 2021.
- Council of Agriculture, Taiwan. 104年第一期稻作停灌補償作業說明 (2015 first rice crop: compensation for halting irrigation). Ministry of Agriculture, Taiwan, 2015.Earlier droughts when Taiwan halted irrigation.
- Taipei Times staff writer, with CNA. TSMC touts water saving efforts and emissions goals. Taipei Times, 2021.TSMC preparing water trucks during the 2021 drought.
It depends where, and when
- Fayette County. Facts about the QTS Fayetteville Data Center Campus Project. Georgia, 2026.County statement: construction water, back bill, “unclear” letter.
- Atlanta News First. Records raise questions about Georgia’s largest data center and its water use. 2026.The data centre’s share of county water; the outcome.
- The Citizen. Behind Fayette’s QTS water controversy: a missed meter, 8,000 workers and a massive construction project. 2026.County officials on the lawn-watering request.
- Andreas Knobloch. Potential environmental impact: Google plans new data center in Chile. heise online, 2024.Google dropping the original Santiago plan.
- OPB. As Google’s water demands grow, The Dalles aims to pull more from Mount Hood forest. 2026.The city’s lawsuit over Google’s water records.
- City of Tucson. Project Blue development agreement, final draft for Mayor and Council review. 2025.Water the proposed Tucson data centre would have used.
- City of Tucson. Citizens’ Water Advisory Committee minutes, 3 June 2026. 2026.Tucson Water’s total supply.
- KJZZ. Project Blue would have used millions of gallons of Tucson water. Residents, city leaders said no. 2025.The council vote and later meter revocation.
- City of Tucson. Motions adopted, Mayor and Council study session, 6 August 2025. 2025.Council’s unanimous vote to end talks.
- AZ Luminaria. Tucson City Council rejects Project Blue data center amid intense community pressure. 2025.The council member’s “not here, not now” quote.
- Katya Mendoza. Arizona water officials approve wells tied to Project Blue data center. AZPM, 2026.The data centre built outside the city, air-cooled, on wells.
What if this is wrong?
- Microsoft. Datacenter efficiency. Microsoft Datacenters, 2025.Microsoft’s water per kWh (WUE) by region.
- Amazon Web Services. AWS Cloud. Amazon Sustainability, 2025.AWS water and power efficiency by region.
- Amazon. FAQ: why does AI use water? Amazon Sustainability, 2025.AWS sites that use little or no cooling water.
- Macknick, Newmark, Heath & Hallett. Operational water consumption and withdrawal factors for electricity generating technologies: a review of existing literature. Environmental Research Letters, 2012.Water per kWh by power-station type; hydro reservoir rate.
What this piece cost in water
- Chung et al. The ML.ENERGY benchmark: toward automated inference energy measurement and optimization. arXiv, 2025.Measured energy per token, for this piece’s own footprint.
The model’s inputs and the famous figures
- Mekonnen & Hoekstra. The blue water footprint of electricity from hydropower. Hydrology and Earth System Sciences, 2012.The 245 litres per kWh hydro rate.
- Li, Yang, Islam & Ren. Making AI less “thirsty”: uncovering and addressing the secret water footprint of AI models. arXiv, 2023.The half-litre per 20 to 50 questions figure.
- Sam Altman. The gentle singularity. Sam Altman’s blog, 2025.OpenAI’s “one fifteenth of a teaspoon” and energy per query.
- Mistral AI. Our contribution to a global environmental standard for AI. 2025.Mistral’s 45 mL per Le Chat answer.
- Google. Power usage effectiveness. Google Data Centers, 2025.Google’s PUE; industry-average PUE.
- Oviedo et al. (Microsoft). Energy use of AI inference, efficiency pathways, and test-time scaling. arXiv, 2025.Energy per prompt and per reasoning answer.
- Microsoft. Sustainable by design: next-generation datacenters consume zero water for cooling. Microsoft Cloud blog, 2024.Microsoft’s zero-water data-centre designs.
- Scott Guthrie (Microsoft). Inside the world’s most powerful AI datacenter. Official Microsoft Blog, 2025.Closed-loop cooling in new AI data centres, for 2030 WUE.
- Infocomm Media Development Authority. Green data centre roadmap. Singapore, 2024.Water per kWh at Singapore data centres.
- Digital Realty. Impact report. 2025.Digital Realty’s WUE, for Asia-Pacific.
- NAVER. 2025 integrated report. 2025.NAVER’s data-centre WUE, for Asia-Pacific.
- Alibaba Group. Environmental, social and governance report (SEC filing exhibit). U.S. Securities and Exchange Commission, 2026.Alibaba Cloud’s WUE, for China.
- Tencent. 2025 environmental, social and governance report. 2026.Tencent’s data-centre WUE, for China.
- Baidu. 2025 environmental, social and governance report. 2026.Baidu’s data-centre WUE, for China.
- European Commission, DG ENER. Assessment of the energy performance and sustainability of data centres in EU: first technical report. Publications Office of the EU, 2025.Reported WUE of EU data centres.
- techUK. Balancing bytes and drops: understanding data centre water use in England. 2025.Share of surveyed English colocation sites cooled without water.
- MOSL. Data centre study calls for new reporting framework to manage water use. 2026.UK data-centre water use, for the UK WUE.
- Houses of the Oireachtas. Dáil Éireann written answers, 28 July 2026 (PQ 58369/26). 2026.Irish data centres’ share of public water supply.
- Khazna Data Centers. ESG report 2025. 2025.UAE data-centre WUE.
- U.S. Department of Energy, FEMP. Best management practice #10: cooling tower management. 2024.How much cooling-tower water evaporates versus drains.
- U.S. Geological Survey. Water use across the conterminous United States, water years 2010–20 (Professional paper 1894-D). 2025.US water used up, and power stations’ freshwater share.
- Ember. Grids for data centres in Europe. 2025.European data-centre electricity, for regional totals.