Data center climate siting: why location changes cooling and water

Climate changes how much cooling a data center needs. In cool, dry places, outside air or seawater can remove heat for most of the year, which cuts both electricity and water use. In hot or humid places, operators lean on chillers or evaporation. But climate is only one of several factors in choosing a site, and power often wins.

This article explains the mechanisms, uses the published numbers that exist, and is clear about what they do and do not show. Figures are labeled by source and date. Nothing here ranks places or operators.

Key numbers at a glance

• 1.09: Google's reported fleet-wide annual power usage effectiveness (PUE) for 2025, with quarterly values of 1.08 to 1.10 in 2026 so far (Google efficiency page, company statement, accessed 19 October 2026).
• 1.54: the Uptime Institute 2025 survey average PUE that Google uses as its industry comparison (same page).
• 1.4: LBNL's estimated average PUE of US data centers in 2023, down from 1.6 in 2014 (Lawrence Berkeley National Laboratory, LBNL, report dated December 2024, modeled estimate).
• 0.05 and 1.10: AWS's reported 2025 water usage effectiveness (WUE) in liters per kWh for Europe and for Asia Pacific excluding China (AWS sustainability page, company statement, accessed 19 October 2026).
• 0.06 and 2.85: AWS's reported 2025 WUE for its US East (Northern Virginia) region and for Jakarta (same page).
• 1.57: AWS's reported 2025 WUE for Singapore (same page).
• 0.12: AWS's reported global WUE for 2025, from 0.15 in 2024 (same page).
• 0.30: Microsoft's reported FY2024 average WUE in liters per kWh (Microsoft blog, December 2024, company statement).
• 98 percent: carbon-free energy share Google reports for Finland in 2023 (Google Hamina page, company statement, accessed 19 October 2026).
• 2011: the year Google says it opened its first Hamina, Finland data centre, which uses seawater from the Bay of Finland for cooling (same page).
• 2 to 5 percent: cooling energy savings for every 1 degree Celsius rise in data center operating temperature, according to Singapore's Infocomm Media Development Authority (IMDA Green Data Centre Roadmap, 2024, government publication).
• 2.0 cubic metres per MWh: the WUE that Singapore aims for new and existing data centers to reach or beat within 10 years, against a 2021 median of 2.2 for large water users (IMDA, 2024).

The basic idea: heat needs somewhere cooler to go

Cooling a data center is not about making heat vanish. It is about moving heat from the servers to somewhere cooler, then letting it go. Anywhere that is cooler than the servers can serve as that sink: outside air, a lake, a river, the sea, the ground, or the water in a cooling tower after it evaporates.

The cooler and drier the surroundings, the easier it is. When the outside air is colder than the air the servers need, you can simply bring it in or use it to cool a water loop, with fans and pumps doing the work. Engineers call this free cooling, because the compressors that make cold artificially, the biggest electricity users in a cooling plant, can stay off.

When the surroundings are hot, you must either run compressors to make cold, or spray water to cool by evaporation. Evaporation is energy-efficient but consumes water. Compressors consume electricity. This is the same energy and water trade-off covered in the guide to data center water use. Climate sets how hard the trade-off bites.

What the modeling shows

LBNL's 2024 US report gives the clearest public view of how climate enters the numbers. It simulated power usage effectiveness (PUE, total facility electricity divided by electricity used by the computing equipment) and water usage effectiveness (WUE, liters of water consumed per kWh of computing energy) using hourly weather data from stations across the country. It describes its ranges as running from "data centers with best practice efficiency values in favorable climates" at the low end to "poor efficiency values in hot, humid climates" at the high end.

Three findings from the report are worth keeping in mind.

First, climate matters most for specific cooling systems. Airside economizers, which use outside air directly, and waterside economizers and dry coolers, which use outside air to cool a water loop, give most of their benefit in climates where outdoor conditions are favorable for much of the year. LBNL notes that in favorable climates the backup chiller in some designs "may not actually be deployed" at all.

Second, humidity is a separate issue from temperature. Bringing in humid air, or air that is too dry, requires conditioning. The LBNL report mentions occasional humidification as a source of the small water use in some waterless-leaning designs. In humid tropical places the question is how much moisture the equipment can tolerate, which is why Singapore wrote its own standard for it (see below).

Third, the national average hides the spread. LBNL's average PUE across all US data centers was an estimated 1.4 in 2023, and it expects 1.15 to 1.35 by 2028. Those are averages across many climates and facility types, mainly pulled down by the shift to large hyperscale and colocation facilities rather than by any one climate.

What operators report by region

Company disclosures give a second view, though with caveats about definitions and boundaries.

AWS publishes WUE by region for 2025. Its global figure was 0.12 liters per kWh. By geography it reports 0.05 for Europe, 0.08 for North America, 0.09 for Central and South America, and 1.10 for Asia Pacific excluding China. At city level, its US East region in Northern Virginia was 0.06, its US West region in Northern California was 0.39, Singapore was 1.57, Tokyo was 1.22, and Jakarta was 2.85. The spread from 0.05 to 2.85 is a factor of more than fifty.

It would be tempting to read that as a pure climate effect. The page does not say so, and two cautions apply. The values reflect the cooling designs chosen for each region and the way regional operations were run that year, not just the weather. And the same page shows movement from one year to the next that is not about climate: Frankfurt's value moved from 0.01 in 2024 to 0.17 in 2025, and Tokyo's from 0.91 to 1.22. Regions with less than a full year of operation show as unavailable. So treat the pattern as indicative only. Even so, the broad shape fits the physics: the warmest and most humid regions listed, Jakarta, Singapore and Tokyo, report the highest values.

Google's efficiency page reports PUE for its whole fleet rather than by site. Its 2025 annual figure was 1.09, and the quarterly figure has stayed between 1.08 and 1.10 in the periods shown for 2026. Google compares that with a 1.54 industry average from an Uptime Institute survey and says that corresponds to 83 percent less overhead energy. These are company statements, and the page notes that its overhead boundary includes items such as substations and transformers. A low fleet PUE is the result of many things at once, including site selection, design, scale and operating practice.

Microsoft's reported FY2024 average WUE was 0.30 liters per kWh, down from 0.49 for 2021. It also reports that cooling updates at its Phoenix datacenters produced a 23 percent year-over-year WUE reduction in FY2025. Phoenix is a notably hot, dry place, which is a good illustration of how design and controls, not only the weather, move the numbers.

Cool places: free cooling, seawater and heat reuse

Cool climates have long attracted data centers, because outside air or cold water can do much of the cooling work for free.

A well documented example is Google's site at Hamina, Finland. Google says it bought a former paper mill there in 2009 and opened its first data centre at the site in 2011. It states the site uses "a cooling system with seawater from the Bay of Finland." The page gives no figures for seawater temperature, flow or energy savings, so the benefit cannot be quantified from it. Google also says that heat from the site is recovered by an off-site heat recovery facility, with an estimated 80 percent figure stated alongside a collaboration with the local energy company Haminan Energia. The page does not define precisely what that 80 percent measures, so it should not be read as the share of the site's heat that is captured.

Heat reuse deserves a note. A data center produces a steady stream of warm water or air. In places with district heating networks, which pipe hot water to homes and buildings, that heat can have a customer. The Open Compute Project lists heat reuse as one of the five areas of its Cooling Environments project, with work on reference designs, economics, policy and a global projects database. The project's starting point is that almost 100 percent of the energy used in a processor turns into heat. Warmer water from liquid cooling is generally easier to use in a heating network than lukewarm air. No authoritative figure for how much data center heat is reused worldwide has been published.

Cool places are not free of constraints. Cold-climate sites may be far from the largest population centers and the busiest network routes. Grid connection can be the limiting factor even where the weather is ideal. And a cool climate does not by itself guarantee an abundant supply of clean power.

Hot places: chillers, evaporation and closed loops

Hot regions rely more on mechanical cooling or evaporation. That does not stop operators building there, for reasons covered in the next section, and it has pushed design in a specific direction: cut water use even at an energy cost.

Microsoft's zero-water cooling design is the clearest published example. The company says that since August 2024 all its new datacenter designs use chip-level cooling in a closed loop that "consumes zero water for cooling," and that it plans pilot sites in 2026 in Phoenix, Arizona and Mt. Pleasant, Wisconsin, with new sites expected to begin coming online in late 2027. It states the trade-off openly: replacing evaporative cooling with mechanical cooling raises PUE, offset in part by warmer-temperature cooling and economizing chillers, with a nominal increase in annual energy use compared with evaporative designs. The two pilot locations, Arizona and Wisconsin, are in quite different climates.

The LBNL report points the same way from the modeling side. Liquid-cooled IT equipment can run at warmer coolant temperatures, so even in warm places dry coolers can carry more of the load across the year, with chillers needed only for the hottest hours. That is one of the reasons the liquid cooling shift matters for climate flexibility. The equipment itself is covered in air versus liquid cooling for AI servers.

Singapore shows a third route: change the operating point rather than the machine. Its Infocomm Media Development Authority (IMDA) says in its 2024 Green Data Centre Roadmap that its Tropical Data Centre standard lets facilities run safely at higher temperatures and humidity levels, and that 2 to 5 percent of cooling energy can be saved for every 1 degree Celsius increase in operating temperature. The roadmap describes an operator raising a data hall's set points by 1 degree at a time over four months, aiming for 26 degrees Celsius and above. It also sets an aim for all data centers in Singapore to reach a PUE of 1.3 or lower at full IT load within 10 years. In a hot, humid city where free cooling is scarce, running warmer is one of the few climate levers left.

Why climate is not the whole story

If cool climates are so efficient, why are so many data centers in warm places? Several reasons stand out, in no particular order.

• Power availability. A site needs a grid connection with enough capacity, and for AI campuses that is often the scarcest item. See what a data center needs to connect to the grid and why AI labs and cloud vendors pick one country for data centres.
• Latency and customers. Services that need to respond quickly sit near large user populations. Northern Virginia and Singapore have large data center clusters that are driven largely by connectivity and customers, not weather.
• Power price and mix. Electricity price, the share of carbon-free generation and the water intensity of the grid all differ by location. LBNL's indirect water estimate varies by region because grid mixes do.
• Land, permits and water rights. Local rules and community acceptance can decide a project. See how long it takes to build an AI data center.
• Existing clusters. Fiber, skilled labor and equipment suppliers concentrate where earlier facilities were built.
• Regulation. Some jurisdictions set efficiency or water targets. Singapore's IMDA says it will work with the national water agency, PUB, to help new and existing data centers reach a WUE of 2.0 cubic metres per MWh or lower over 10 years, and notes that cooling towers account for up to 97 percent of a data center's water use. See Singapore's data centre capacity rules.

A practical consequence is that PUE and WUE comparisons between sites in different climates can mislead. PUE does not account for outside climate, so colder locations need less cooling energy and PUE alone cannot fairly compare sites in different climates. Singapore's roadmap makes a related point, saying PUE alone is insufficient because it does not consider how efficiently the IT equipment itself uses power. A PUE of 1.1 in a cool place and 1.3 in a hot one may reflect equally good engineering.

Climate, water stress and local impact

One more layer matters: the local water situation. A warm, dry place may offer abundant solar energy and cheap land and also face limited water supply. In that case, operators and regulators may favor dry cooling or reclaimed water even at an energy penalty. A cool, wet place may have the opposite situation, with water available but grid capacity scarce.

This is why national averages are poor guides to local impact. LBNL's direct water estimate of 66 billion liters for the US in 2023 is a national total. A single large campus in a small basin is a different matter, and operators rarely publish site-level data that would show it. The best a reader can do is ask for the specifics: what cooling design, what water source, what electricity mix, and what the site reports.

Climate change itself

An open question is how a warming climate affects cooling needs. No published source cited here quantifies it. Hotter summers could mean more hours when free cooling is unavailable, and designs based on historical weather data may need updating. LBNL's simulations rely on typical meteorological year weather data from stations across the US, which by construction reflect typical past conditions. This is a reasonable question, not a finding.

What is firm and what is uncertain

Firm enough to rely on:

• The mechanism: cooler, drier surroundings allow more free cooling and less evaporation, and the energy and water trade-off is real. LBNL and Microsoft both state it.
• The modeling range: LBNL's low and high bounds correspond to favorable and hot, humid climates.
• Company-reported regional WUE values from AWS, as company statements for 2025.
• Google's description of seawater cooling and heat recovery at Hamina, and Microsoft's description of its closed-loop design and pilot locations.

Uncertain:

• How much of the difference between regions is climate and how much is design choice. The reported figures do not separate them.
• Whether regional or fleet figures are comparable across companies, given different boundaries.
• The size of heat reuse worldwide.
• The effect of a warming climate on cooling demand.
• Whether targets such as Singapore's WUE and PUE aims are met within their 10 year windows.

What to watch next

• Whether Microsoft's 2026 pilots in Phoenix and Wisconsin publish measured water and energy results after they come online.
• Whether more operators publish site-level PUE and WUE with the cooling design, water source and electricity mix beside them.
• Whether regulators tie permits to cooling design or WUE targets.
• Growth in heat reuse projects, including those tracked by the Open Compute Project's database.
• Whether liquid cooling's warmer coolant temperatures let more hot-climate sites use dry heat rejection.

For the software that runs in these facilities, Safeguard offers SBOM and dependency scanning, with a free scan.

Quick check

Sources

• Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report (eScholarship), December 2024, read 19 October 2026
• Amazon sustainability, AWS cloud page, accessed 19 October 2026
• Google data center efficiency page, accessed 19 October 2026
• Google data centers, Hamina Finland, accessed 19 October 2026
• Microsoft, next-generation datacenters that consume zero water for cooling, accessed 19 October 2026
• Microsoft sustainability report page, accessed 19 October 2026
• Open Compute Project, Cooling Environments project, accessed 19 October 2026
• Infocomm Media Development Authority, Green Data Centre Roadmap, 2024, accessed 19 October 2026

Last reviewed: 19 October 2026.