Every time you send a message to an advanced AI model, a server farm somewhere is consuming fresh water to keep its chips from overheating. The relationship between AI computation and water consumption is not abstract or theoretical. It is a direct, measurable physical consequence of the laws of thermodynamics, and it is scaling at the same rate as AI adoption itself.
The data center industry is now under serious regulatory and public pressure over this consumption. Google, Microsoft, Amazon, and Meta have each made public commitments to reduce or neutralize their water footprints. But commitments on a sustainability page and engineering changes in a server room are different things. What Google is actually building, as demonstrated most recently by the June 2026 groundbreaking at Horndal, Sweden, represents a genuine architectural shift away from water-based cooling.
This piece breaks down the problem from the physics up, explains Google's three-part engineering response, and examines the trade-off they are making to get there.
The Core Problem | AI's Hidden Thirst
To understand why data centers consume so much water, you need to understand what happens inside a server rack when an AI model runs. A large language model inference request requires hundreds or thousands of processor operations per second across clusters of custom AI chips. Those operations are electrical, and electricity moving through resistive materials generates heat. At AI cluster scale, that heat is not trivial. It is a constant, enormous thermal output that must be managed continuously or the hardware degrades and fails.
The dominant method the industry developed to manage this heat is evaporative cooling. The principle is the same as sweating: water absorbs heat as it evaporates. In a data center cooling tower, water is sprayed into an airstream moving past the server intake. The water evaporates, carrying the heat energy with it into the atmosphere. The cooled air then flows through the server racks, absorbing more heat, and the cycle repeats.
The operational scale of this process is staggering. A standard hyperscale data center consumes between 300,000 and 500,000 gallons of fresh water per day. That is comparable to the daily water consumption of a town of several thousand residents, drawn continuously from local municipal supplies or groundwater aquifers. A campus of three or four such buildings, which is typical for hyperscale deployments, can draw over 1.5 million gallons per day from a single water district.
The AI era has accelerated this problem significantly. Research from the University of California Riverside published in 2023 estimated that a single prompt-and-response exchange with a state-of-the-art large language model can require the equivalent of roughly 500ml of water to cool the servers processing the query. Multiply that by billions of queries per day across the industry and the aggregate water demand becomes a politically and ecologically significant number.
Regulators have begun to respond. Ireland and the Netherlands placed effective moratoriums or severe restrictions on new data center permits in regions where the facilities were straining grid and water capacity. Arizona faced public backlash over data center expansions during drought conditions. Google, Microsoft, and Amazon have each faced local opposition to proposed builds based on water consumption projections.
Google's Three-Part Engineering Response
Google's answer to this constraint is not a single technology change. It is a coordinated architectural shift across three distinct systems, each addressing a different aspect of the thermal management problem.
Part 1 | Pure Air-Cooled Architecture
In Google's newest facility designs, including the Horndal, Sweden data center, the primary cooling infrastructure is air-based rather than water-based. Instead of evaporating water into an airstream to lower temperatures, large industrial fan arrays move high volumes of ambient outdoor air directly through or across the server infrastructure.
This approach drops water consumption close to zero for general compute workloads. The key enabling factor is geography. In Dalarna County, Sweden, ambient outdoor air temperatures are cold enough for a substantial portion of the year to provide meaningful cooling without any additional chilling. Google's engineering teams specifically selected the Horndal site in part because its climate makes air-only cooling viable at a scale that would not be achievable in warmer geographies.
The engineering challenge is airflow volume. To move enough air to cool a hyperscale building without the density bonus that evaporative cooling provides, the fan infrastructure must be significantly larger and more powerful. That has a direct impact on electricity consumption, which is the key trade-off discussed later in this piece.
Part 2 | Closed-Loop Liquid Cooling
Air cooling works well for general compute, but Google's AI clusters run custom TPU chips that generate heat at densities that air simply cannot manage at acceptable efficiency. For these workloads, Google deploys closed-loop liquid cooling systems.
In a closed-loop design, chilled fluid is circulated through pipes or cold plates that make direct contact with the chip surface. The fluid absorbs heat from the chip, is then routed through insulated lines to an outdoor radiator or heat exchanger, releases the thermal energy into the outside air, and returns to the chip cooled and ready to absorb more heat. The loop is fully sealed.
The critical distinction from evaporative cooling is that no fluid escapes the system. There is no evaporation, no water consumption at the cooling stage, and no dependency on a local water supply to maintain cooling capacity. The fluid circulates indefinitely within the closed system, with only minimal top-up required to compensate for minor losses over time.
Closed-loop liquid cooling is not new. It has been used in supercomputing and specialized industrial settings for decades. What is new is the deployment of this architecture at hyperscale across AI training and inference clusters as a default rather than an exception.
Part 3 | Municipal Heat Recovery (Industrial Symbiosis)
Both air-cooled and closed-loop systems remove heat from the servers, but they still produce that heat as a byproduct. In conventional designs, this thermal energy is vented to the atmosphere as waste. Google's newest facilities, including Horndal, take a different approach.
By installing heat exchangers at the point where the cooling infrastructure releases its thermal load, Google captures that energy and routes it directly into the surrounding municipality's district heating network. District heating is a centralized infrastructure system common in Scandinavian cities where heat is generated at one source and distributed to residential and commercial buildings through insulated underground pipes.
The Horndal facility will provide this captured heat to Avesta Municipality at no charge to residents. Homes and businesses in the surrounding area will be warmed by thermal energy that is an unavoidable byproduct of running AI workloads. The data center's waste heat becomes a community asset.
This type of arrangement, sometimes called industrial symbiosis, has been pioneered by data center operators in Finland and Denmark over the past decade. It serves three functions simultaneously: it eliminates the waste heat problem from an environmental accounting standpoint, it creates a direct and tangible economic benefit for the local community, and it gives the municipality a financial interest in the data center's continued operation rather than opposition to its existence.
The Engineering Trade-Off | Water for Electricity
Google's three-part cooling architecture solves the water problem. But it does not solve the thermodynamic problem itself. The heat still exists and still must be managed. The trade-off is that eliminating water from the cooling loop requires significantly more electrical power to move the volumes of air and fluid necessary to achieve equivalent thermal performance.
The comparison table below captures this trade-off directly:
| Cooling Method | Water Consumption | Energy Efficiency | Grid Impact |
|---|---|---|---|
| Traditional Evaporative | High — 300,000 to 500,000 gal/day per facility | Very high PUE | Lower electrical draw, extreme water strain |
| Google Air-Cooled (Modern) | Near zero | Moderate PUE | Significantly higher electrical draw from fan infrastructure |
| Closed-Loop Liquid | Near zero — sealed system | High PUE for dense AI chips | Higher electrical draw than evaporative for equivalent density |
By eliminating water, Google's data centers require significantly more raw electrical power to operate the fan grids, fluid pumps, and radiator systems that replace the evaporation process. This is precisely why Google has signed over 700 megawatts of Nordic wind power purchase agreements to support the Horndal facility alone.
The strategy is explicit: Google is trading a water problem for an energy problem, and betting that it can solve the energy problem through clean grid procurement. In Sweden, where the grid is already predominantly hydropower and wind, that bet is reasonable. In regions where the grid is still majority fossil fuel, the same trade-off produces a worse environmental outcome overall, which is why site selection for air-cooled facilities is inseparable from the clean energy procurement strategy.
Why This Engineering Shift Matters Beyond Google
Google is not the only company making this architectural transition. Data Center Dynamics has documented similar pivots underway at Microsoft and at several large colocation providers responding to customer pressure. But the speed at which operators can make this transition is constrained by capital expenditure cycles and the availability of sites in climates that make air cooling viable at hyperscale.
The deeper implication is that the AI infrastructure build-out is driving a permanent restructuring of where data centers are built. Water-constrained geographies, regardless of their other advantages, become structurally less attractive as the default cooling architecture shifts away from evaporation. Cool, renewable-energy-rich regions like Scandinavia, Scotland, Iceland, and parts of Canada become structurally more attractive.
Sightline Climate's 2026 infrastructure report identifies this geographic reorientation as one of the three primary structural forces reshaping digital infrastructure investment over the coming decade. The other two are grid decarbonization timelines and the consolidation of AI compute around a small number of chip architectures that have specific power density requirements those traditional air-cooled designs cannot meet, hence the necessity of closed-loop liquid cooling as a parallel track.
Google's Horndal facility is the most complete public demonstration of all three parts of the new architecture running simultaneously in a single facility. Its operational data, once the site comes online in 2028 or 2029, will become a reference case for every other hyperscale operator planning facilities through the end of the decade.
Frequently Asked Questions
How much water does a typical AI data center use per day?
A standard hyperscale facility using evaporative cooling consumes between 300,000 and 500,000 gallons of fresh water per day. A campus of three to four buildings can exceed 1.5 million gallons daily, drawn from local municipal or groundwater supplies.
How much water does a single AI prompt use?
Research from the University of California Riverside estimates that a single prompt-and-response exchange with a state-of-the-art large language model requires the equivalent of roughly 500ml of water, equivalent to a standard drinking bottle, to cool the servers processing the query. The figure varies by model size and hardware generation.
What is closed-loop liquid cooling?
Closed-loop liquid cooling circulates chilled fluid through sealed pipes or cold plates in direct contact with AI chips. The fluid absorbs heat, routes it to an outdoor radiator, and returns cooled. Because the loop is fully sealed, no fluid evaporates or is consumed. It is the standard cooling method for ultra-dense AI chip clusters like Google's TPU farms.
Why does Google need 700 megawatts of wind power for a water-free data center?
Eliminating water from the cooling process requires significantly more electricity to power the fan arrays and fluid pumps that replace evaporative cooling. Air-cooled facilities draw more raw electrical power per unit of compute than evaporative facilities. Google secured 700MW of Nordic wind to offset this higher electrical demand with clean energy.
What is district heating and why is Google providing it for free?
District heating is a centralized system that generates heat at one source and distributes it to buildings through insulated pipes. Google captures the waste heat from its Horndal servers and routes it into Avesta Municipality's district heating network at no cost. It eliminates waste heat as an environmental liability, creates a community benefit, and builds local political support for the facility's operation.