Every time you ask an AI a question, somewhere on Earth, water evaporates. Not metaphorically — literally. This is the story of how the intelligence we are building is quietly draining the planet’s most precious resource, and the science of what we can do about it.
A Question Nobody Was Asking!
People often debate the environmental cost of AI, and carbon emissions usually dominate the conversation. But there is a quieter crisis running alongside it — one that threatens something more immediately essential to human life: fresh water. Only 3% of Earth’s water is freshwater. Of that, a mere 0.5% is accessible and safe for human use. That thin sliver is what billions of people, farms, and ecosystems depend on entirely, and increasingly, it is what keeps AI running.
The Controversy: Silicon Is Thirsty
The numbers, when you sit with them, are staggering. A single large data center can consume up to 5 million gallons of water per day — roughly the daily water use of a town with 10,000 to 50,000 residents. Google reported using over 5 billion gallons of water across all its data centers in 2023, with nearly a third drawn from watersheds already experiencing water scarcity. A University of California, Riverside study estimated that around 20 AI queries consume water equivalent to a small bottle of drinking water.
What makes this a controversy, rather than simply an engineering challenge, is a combination of choices that many find hard to justify.
The Location Problem: The majority of new AI data centers built in the United States since 2022 have been located in regions already grappling with high water stress — Arizona, Texas, where aquifers are shrinking, and drought conditions are worsening. OpenAI’s $100 billion Stargate AI infrastructure project is being anchored in Abilene, Texas, a city where local hydrologists describe a “water-energy nexus crisis.” Water has historically been treated as an afterthought in data center siting decisions because, compared to land and electricity, it is cheap.
The Transparency Problem: Many companies are less than forthcoming about how much water their facilities use. A 2021 survey found that only 51% of data center operators tracked their water usage at all — and just 10% tracked it across all their facilities.
The Evaporation Problem: Approximately 80% of the fresh water withdrawn by data centers evaporates during the cooling process. It does not become wastewater that can be treated and returned. It simply vanishes into the atmosphere — gone from local water systems permanently.
The Science: Why Do Computers Need So Much Water?
When electricity flows through a conductor, some of that energy is converted into heat — a phenomenon described by Joule’s Law. Modern AI processors, particularly the GPUs used to train and run large language models, push enormous amounts of current through extremely dense arrangements of transistors. The result is intense, concentrated heat that must be continuously removed or the chips will throttle, degrade, and ultimately fail.
The traditional solution — still dominant in roughly 95% of the world’s data centers — is evaporative cooling. Industrial cooling towers spray water over large surface areas, where it evaporates and carries heat away. It is effective, relatively inexpensive, and consumes massive volumes of fresh water in the process. Approximately 60% of total data center water consumption is actually indirect, originating from the fossil fuel power plants supplying their electricity, not from on-site cooling alone.
The Next Frontier: When You Dip Electronics in “Water”
You may have seen videos of circuit boards submerged in liquid, running perfectly. This is not water — contact with live electronics would cause immediate short circuits and component failure. What those boards are submerged in is a dielectric fluid: a liquid engineered to be electrically non-conductive while being an excellent conductor of heat.
This technology, known as immersion cooling, submerges electronics in a bath of dielectric fluid. The fluid absorbs heat directly from the surface of chips and carries it to a heat exchanger, where it is cooled and recirculated. The same volume of dielectric fluid can absorb approximately 1,500 times more heat than the same volume of air.
Single-phase immersion cooling uses hydrocarbon-based mineral oils or synthetic esters. The fluid absorbs heat and remains liquid throughout, circulating through a heat exchanger in a continuous loop. Two-phase immersion cooling uses engineered fluorocarbon-based fluids with deliberately low boiling points. As chips generate heat, the fluid literally boils, transitions to vapor, rises, condenses on a cooled surface, and drips back down as liquid. This passive thermodynamic cycle requires no pumps and can handle heat fluxes exceeding 40 watts per square centimetre.
Unlike fresh water, dielectric fluid is not drawn from rivers or aquifers. Mineral oil comes from crude oil refining. Fluorocarbon-based fluids are synthesised by fluorinating hydrocarbon compounds. Synthetic esters, the most environmentally benign option, are derived from vegetable oils through esterification. None involves fresh water as a raw material, which is precisely the point.
So, Why Isn’t Everyone Doing This?
Immersion cooling setups cost three to four times more than traditional air or evaporative cooling infrastructure upfront. Existing servers are not designed to be submerged — fans must be removed, components need compatibility verification, and power supplies require modification. The entire data center layout must be rethought around liquid-filled tanks. There are also no universally adopted industry standards yet.
There is also a complication with the most advanced fluids. Fluorocarbon-based dielectrics contain PFAS — per- and polyfluoroalkyl substances, sometimes called “forever chemicals” — because they do not break down in the natural environment. The European Union has been moving to restrict several PFAS compounds, and 3M has announced it would phase out Novec production. The industry is working on PFAS-free alternatives using biodegradable synthetic esters, but mainstream adoption will take time.
The Incomplete Solution: Why the Water Crisis Remains
Even if every data center in the world switched to immersion cooling tomorrow, the water crisis would not be fully resolved. Data centers draw from regional electrical grids substantially powered by thermal generation, gas turbines, coal plants, and nuclear reactors — all of which require large volumes of cooling water. Immersion cooling can eliminate on-site evaporative losses while the power generation side of the problem persists as long as the grid remains fossil-fuel dependent.
The real solution is a combination: immersion or closed-loop cooling to eliminate on-site waste, paired with a transition to renewable energy sources, solar and wind, which require virtually no operational water, to address the indirect footprint. Countries like Iceland and Norway, which power data centers almost entirely with geothermal and hydroelectric energy in naturally cold climates, demonstrate what this combination can look like at scale.
The Bottom Line!
There is something quietly paradoxical about the AI revolution. We are building systems of extraordinary intelligence that, at their physical foundation, require some of the most basic things on Earth: electricity, land, and water.
The path forward is not to stop building AI, nor to pretend the environmental costs do not exist. Dielectric immersion cooling is not a magic solution, but it is a real one, grounded in solid thermodynamics. Its wider adoption, combined with a cleaner grid and genuine corporate transparency, represents the clearest technical pathway toward AI infrastructure that does not compete with human beings for the water they need to survive. The next time you send a query to an AI, somewhere, cooling equipment hums. Whether what it costs the planet is necessary and proportionate is, increasingly, a question that cannot be deferred.
References & Further Reading:
- Lincoln Institute of Land Policy — “Data Drain: The Land and Water Impacts of the AI Boom” (February 2026) — lincolninst.edu
- Environmental and Energy Study Institute (EESI) — “Data Centers and Water Consumption” — eesi.org
- Brookings Institution — “AI, Data Centers, and Water” (November 2025) — brookings.edu
- Bloomberg — “The AI Boom Is Draining Water From the Areas That Need It Most” (May 2025) — bloomberg.com
- EthicalGEO — “The Cloud is Drying our Rivers: Water Usage of AI Data Centers” (July 2025) — ethicalgeo.org
- ScienceDirect / Cell Press — “The carbon and water footprints of data centers” (December 2025) — sciencedirect.com
- Undark — “How Much Water Do AI Data Centers Really Use?” (December 2025) — undark.org
- IEEE Xplore — “Compact Two-Phase Immersion Cooling With Dielectric Fluid for PCB-Based Power Electronics” — ieeexplore.ieee.org
- Supermicro — “What Is Immersion Cooling?” — supermicro.com
- LiquidStack — “Understanding Approaches to Immersion Cooling” — liquidstack.com
- The Green Grid — Water Usage Effectiveness (WUE) metric documentation — thegreengrid.org
Similar articles:
When Water Turns to Conflict: A Global Story of Scarcity, Security, and Violence
The AI Nuclear Renaissance: How Diablo Canyon Became the Backbone of California’s Power Boom
Farrukh Ahmed is pursuing a BS in Space Sciences at the Institute of Space and Technology (ISST), University of Karachi. An aspiring space scientist and member of the Karachi Astronomers Society, he is passionate about science communication and astronomy outreach in Karachi. He also works as a digital news curator and writer.
