Image from Unsplash
The technological leaps in data centers have been massive and a great distance from the US military-funded Electronic Numerical Integrator and Computer (ENIAC) completed at the University of Pennsylvania in late 1945. From basic to complex computing workloads, AI data centers crunch more complicated tasks including machine learning. Currently, most facilities around the world are still the digital warehouses that we know, but the new ones being built specialize in AI. The latter infrastructure runs on additional hardware, uses more power and therefore generates extreme heat, necessitating advanced cooling solutions to function effectively.
Cooling technologies
AI data centers use a combination of cooling technologies such as air- and liquid-based systems. The air method is still the most common, but liquid cooling is a fast-growing approach in dealing with AI hardware’s massive heat output.
Air cooling systems
Under this category are computer room air conditioners (CRACs) and computer room air handlers (CRAHs), with one using a refrigerant and the other chilled water. Either of these options can use between 33 and 40 percent of a facility's electricity.
Containment is another method, with server rows arranged in order to separate cold intake air and hot exhaust air and make cooling much more efficient. In addition, cooling units are placed between racks or closer to the heat.
Liquid cooling systems
There is a limit to what air cooling can handle as AI racks are now aiming for 40 to 250kW compared to 5 to 20kW per rack in most older facilities. To fill in this gap, liquid cooling is the solution that is gaining traction.
In this approach are rear-door heat exchangers, which equip with back of racks with coolant-filled doors that remove heat from server exhaust before it enters the room. They are a common bridge for upgrading existing air-cooled rooms.
Cold plates are another option. These are metal pieces that sit directly on CPUs and GPUs, with coolants flowing through them carrying heat to a coolant distribution unit (CDU). This has become the leading method for AI and hyperscale data centers as it maintains rack layouts even at high densities. Nvidia has specified this kind of cooling for its GB200 systems.
There is immersion cooling as well. This technique entails servers submerged in tanks of nonconductive (dielectric) fluid. It handles over 100kW per rack and needs no fans. However, it requires redesigned hardware and new maintenance procedures, so it's used less widely than direct-to-chip cold plates.
Another approach
Between air and liquid cooling systems are heat rejection and free cooling techniques. They involve chillers and cooling towers. Chillers make cold water while cooling towers release the heat, using a lot of water through evaporation. Economizers or free cooling, meanwhile, use cool outside air or water instead of mechanical chillers whenever the weather allows such as in cold climates like the Nordics. And there is evaporative or adiabatic cooling, with air cooled by evaporating water into it. This saves energy but uses a lot of water, which is a growing concern in dry regions. And then, there are dry coolers or fans that release heat to ambient air with no water use. They are becoming more practical as chips can run on warmer coolant.
Going hybrid
Cooling methods are almost never used in isolation. Instead, they are layered together as a chain of heat capture, transport and rejection. Even liquid cooling only captures a certain percentage of heat, so it still relies on complementary air cooling for the remainder.
Here’s a breakdown of what each system does at what stage.
Heat capture at the chip or rack level is done using server fans, room air conditioners and handlers, rear-door heat exchangers, cold plates and immersion tanks.
Heat transport through the building involves coolant distribution units (CDUs), chilled-water or warm-water loops, pumps and heat exchangers.
Heat rejection to the outdoors is achieved through chillers, cooling towers, evaporative or adiabatic coolers, dry coolers, economizers, sometimes seawater or lake water, or heat reuse.
As such, even immersion cooling can't stand alone. The tank captures the heat, but it still needs a CDU and an outdoor system to get rid of it.
And this is how air and liquid cooling work together:
Cold plates usually cover the CPUs and GPUs, which is the source of most of the heat, but memory, power supplies, storage and network gear still need air. That's why liquid-cooled AI halls still have air handlers, containment or rear-door heat exchangers.
Facilities have mixed rooms, usually with air-cooled racks for ordinary computing and storage and liquid-cooled AI racks.
Using both methods allows retrofitting of older sites. The addition of rear-door heat exchangers or liquid-to-air CDUs to an air-cooled room enables running of liquid-cooled racks without rebuilding the plant.
The weather can help switch outdoor systems to advantage. In cool weather, for instance, a site can run fully on free cooling, while partial free cooling with chiller help is enough in mild weather. On hot days, there are chillers to help cool everything down. Adiabatic or hybrid coolers, meanwhile, run dry most of the year, spraying water only on the hottest days. This means trading a little water for a lot of energy savings. Warm-water liquid cooling is another option that can let a site use dry coolers year-round in many climates, reserving chillers only for the air-cooled part of the load or dropped.
Usually, data centers build in redundancy, typically one spare unit or a fully duplicated system or may also keep thermal storage like chilled-water tanks or ice in case of a power or equipment failure. That means more than one cooling path often exists for the same load down the line.
Stacking up against the environment
There are solutions for every major impact, that is, on energy, water, refrigerants and waste heat, but adoption is uneven. Measures that save operators money or are required by law are spreading. Voluntary ones are lagging, and progress overall has stalled in some areas.
Energy use
Adoption of solutions is partial at best as progress has slowed.
Power usage efficiency (PUE) is measured with total facility energy divided by the energy used by the computers themselves. According to Uptime Institute Global Data Center Survey 2025 of 800+ operators, the average PUE is 1.54, or almost unchanged for six years in a row. This is attributed to old infrastructure and hot climates. Newer sites do better, however, with those built since about 2020 averaging 1.48, with the largest hyperscale operators reporting much lower figures. So, despite the well-established efficiency methods, many older facilities haven't been upgraded.
Water use
Adoption of available techniques on water conservation is growing among large operators, including Microsoft and Amazon, but it is weak elsewhere. Water use is still rising in absolute terms because so much capacity is being built. Google's 2026 environmental report says its water consumption rose about 34 percent in 2025, to 10.9 billion gallons.
Industry-wide tracking is improving, though. In Uptime's survey, the share of operators measuring water use rose from 43 percent in 2024 to 47 percent in 2025.
But there's a hidden loop that cannot be ignored. Cutting water use often raises electricity use, and power plants consume water, too. According to the International Energy Agency’s estimates, about two-thirds of data center-related water consumption comes from electricity generation, versus about a quarter from on-site cooling. So, solving the data center climate conundrum requires striking a balance between these resources simultaneously instead of addressing water or electricity use in isolation.
Refrigerants and cooling fluids
Currently, the use of these chemicals is driven by regulation, including the EU’s revised F-gas Regulation (2024/573) tightening limits on high-warming refrigerants and aiming to phase out HFCs by 2050. In the US, there’s the American Innovation and Manufacturing (AIM) Act stipulating that new data center cooling equipment must use refrigerants with a warming potential below 700 from January 1, 2027.
However, adoption is limited to new equipment as older equipment keeps its old refrigerants until it is replaced. This is an expensive option because switching means new compressors and valves, not just a new fluid.
Waste heat
The solution is to feed server heat into district heating networks or use it to heat buildings, greenhouses or pools. The other option is to adopt warm-water liquid rather than air cooling to produce more usable heat. This practice, however, is common in Northern Europe only but is not practical elsewhere as many data centers are not near anyone who needs the heat.
Reporting
The EU now requires annual reporting of energy, water and heat-reuse figures for data centers of 500kW or more, with Germany keeping a public registry. Outside Europe, reporting is largely voluntary and patchy. According to the Uptime Institute Global Data Center Survey 2025, sustainability reporting stopped improving in 2025, and this is attributed to pressures to build AI capacity and looser rules in some regions.
References
https://www.enconnex.com/data-center-history-and-evolution/
https://rcrwireless.com/20250327/fundamentals/ai-data-center-difference
https://www.datamintelligence.com/research-report/data-center-liquid-cooling-market
https://spectrum.ieee.org/data-center-liquid-cooling
https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2025.Annual.Survey.Report.pdf?version=0
https://sustainability.google/reports/google-2026-environmental-report/
https://intelligence.uptimeinstitute.com/resource/uptime-institute-global-data-center-survey-2026
