Data Center Cooling Systems Explained: 2026 Complete Guide

Data centers produce enormous amounts of heat, and cooling is the single largest energy expense after the IT load itself.

The Uptime Institute’s 2024 Global Data Center Survey found that cooling accounts for roughly 30% to 40% of a typical facility’s total energy consumption.

That translates to millions of dollars per year for a single campus.

If you work in data centers, or plan to, understanding how cooling systems operate is a core skill that shows up in interviews, on-the-job decisions, and career advancement conversations.

This guide covers every major data center cooling technology in use today: air cooling, computer room air conditioners (CRAC), computer room air handlers (CRAH), airflow and containment strategies, chilled water systems, liquid cooling, and immersion cooling.

You will also learn how to measure cooling efficiency, how these systems connect to building HVAC, and where the industry is heading in 2026 and beyond.

Overview of data center cooling concepts

Data center cooling is the process of removing heat generated by servers, storage devices, networking equipment, and power distribution systems so that hardware operates within safe temperature ranges.

ASHRAE’s Thermal Guidelines for Data Processing Environments recommend inlet air temperatures between 18°C and 27°C (64°F to 80.6°F) for most enterprise equipment, with an allowable range up to 35°C for some hardware classes.

The basic principle behind every cooling method is the same: move heat from where it is generated (the IT equipment) to somewhere it can be rejected (the outside air or a cooling plant).

The difference between cooling technologies comes down to what medium carries the heat and how efficiently it does so.

Air cooling systems use circulated air as the heat transfer medium.

Cold supply air is pushed through server racks, absorbs heat from the equipment, and the resulting hot exhaust air is returned to a cooling unit for processing.

Liquid cooling systems use water or specialized coolants that can absorb far more heat per unit of volume than air.

A liter of water can carry roughly 3,500 times more thermal energy than a liter of air at the same temperature differential, according to Schneider Electric’s Data Center Reference Design guides.

Cooling Medium	Heat Capacity	Best For	Typical Density Supported
Air (traditional)	Low	Standard density racks, 5-15 kW per rack	Up to 15 kW/rack
Air (high-efficiency containment)	Low-Medium	Medium density, 15-25 kW per rack	Up to 25 kW/rack
Rear-door heat exchangers	Medium	Retrofitting air-cooled rooms for higher density	Up to 35 kW/rack
Direct-to-chip liquid	High	High-performance computing, AI/GPU clusters	Up to 80 kW/rack
Full immersion cooling	Very High	Ultra-high density, AI training clusters	100+ kW/rack

Modern data centers are increasingly deploying 40 kW, 60 kW, and even 100+ kW racks to support AI workloads.

Dell’Oro Group reported in 2025 that data center infrastructure spending reached $350 billion globally, driven largely by AI-related capacity.

This density shift is the primary reason liquid cooling has moved from niche technology to mainstream consideration.

Air cooling overview

Air cooling is the most common method for removing heat from data center equipment and has been the default approach since the industry’s earliest days.

The concept is simple: cold air is supplied to the front of server racks, passes through the equipment to absorb heat, and the hot air expelled from the rear of the racks is captured and returned to a cooling unit.

Traditional air cooling works well for rack densities below 15 kW per rack.

Most enterprise data centers built before 2020 were designed around 6-10 kW per rack averages, which air cooling handles comfortably.

CBRE’s 2024 North America Data Center Trends report noted that the average rack density across the existing US data center stock is still around 8-10 kW per rack, meaning air cooling remains the dominant installed technology.

Air cooling does have limitations.

As rack densities climb above 15 kW, the volume of air required to remove heat becomes impractical.

Fans need to move faster, which consumes more energy.

Hot spots form when airflow cannot reach every server uniformly.

Energy efficiency drops because moving large volumes of air is inherently less efficient than moving smaller volumes of liquid.

For these reasons, operators running AI infrastructure with GPU-dense racks are increasingly supplementing or replacing air cooling with liquid alternatives.

Free cooling is an air cooling strategy that uses outside air (or outside air temperatures) to cool the data center when ambient conditions are cool enough, reducing or eliminating the need for compressor-based refrigeration.

Facilities in climates like Scandinavia, the Pacific Northwest, or parts of Canada can use free cooling for 8,000+ hours per year, cutting cooling energy costs by 50% or more according to Uptime Institute research.

Evaporative cooling is another variation that uses the natural cooling effect of water evaporation to lower air temperatures before the air enters the data center.

This approach is popular in hot, dry climates like Phoenix, Arizona, where Google and Microsoft both operate large campuses that rely on evaporative cooling systems.

Computer room air conditioners (CRAC)

A computer room air conditioner, or CRAC unit, is a self-contained cooling system that uses a direct expansion (DX) refrigerant cycle to cool data center air.

CRAC units are the traditional workhorse of data center cooling and are found in thousands of facilities worldwide.

CRAC units work by drawing warm return air from the data center, passing it over a refrigerant-filled evaporator coil that absorbs the heat, and then blowing the cooled air back into the room, typically through a raised floor plenum.

The refrigerant carries the absorbed heat to a condenser (usually located outdoors) where it is rejected to the outside air.

This is the same basic refrigeration cycle used in a home air conditioner, just at a much larger scale.

A single CRAC unit can deliver between 30 kW and 200 kW of cooling capacity, depending on the model and manufacturer.

Schneider Electric, Vertiv, and Stulz are the three largest CRAC unit manufacturers serving the data center market.

Most data center floors deploy multiple CRAC units in an N+1 or 2N redundancy configuration to protect against unit failure.

CRAC units pair well with raised floors. Cold air is pushed into the plenum below the raised floor and delivered to the server rows through perforated floor tiles.

This under-floor delivery method has been the standard approach for decades.

The raised floor height (typically 24 to 36 inches) and tile placement directly affect cooling performance.

Poorly placed tiles or cable obstructions under the floor are one of the most common causes of hot spots in CRAC-cooled data centers.

Computer room air handlers (CRAH)

A computer room air handler, or CRAH unit, cools data center air by passing it over a chilled water coil rather than a direct expansion refrigerant coil.

The CRAH unit itself does not contain a compressor or refrigeration system. Instead, it relies on a central chilled water plant to supply cold water, typically at 7°C to 12°C (45°F to 54°F).

CRAH units are generally more energy efficient than CRAC units for large deployments.

Because the chilled water plant can use high-efficiency chillers, economizer modes, and variable-speed pumping, the overall system efficiency is better than running individual compressors in each CRAC unit.

The Uptime Institute has noted that facilities using chilled water CRAH systems typically achieve PUE values 0.1 to 0.3 points lower than comparable DX CRAC installations.

Feature	CRAC (DX Refrigerant)	CRAH (Chilled Water)
Cooling medium	Refrigerant (R-410A, R-134a)	Chilled water from central plant
Compressor location	Inside each unit	Central chiller plant
Typical capacity per unit	30-200 kW	50-500+ kW
Energy efficiency	Moderate	Higher (with efficient chiller plant)
Capital cost per unit	Lower	Higher (requires chiller infrastructure)
Best suited for	Small to mid-size rooms	Large enterprise and hyperscale facilities
Scalability	Add individual units	Scale chiller plant capacity
Maintenance complexity	Per-unit refrigerant management	Centralized water system maintenance

Large colocation providers like Equinix and Digital Realty use CRAH-based chilled water systems in most of their enterprise facilities.

Hyperscale operators such as Google and Microsoft design their own CRAH configurations optimized for their specific server hardware and facility layouts.

If you work as a data center technician or facilities engineer, understanding the difference between CRAC and CRAH systems will come up in interviews and daily operations at almost every employer.

Airflow and containment strategies

Airflow management is the practice of controlling how air moves through a data center to maximize cooling efficiency and prevent hot exhaust air from mixing with the cold supply air before it reaches the servers.

Poor airflow management is one of the fastest ways to waste cooling energy and create dangerous hot spots.

The most important concept in data center airflow is the hot aisle / cold aisle layout.

Servers are arranged in alternating rows so that the fronts of all racks in one row face the fronts of racks in the adjacent row (creating a cold aisle), and the backs of racks face each other (creating a hot aisle).

Cold supply air enters the cold aisles, passes through the servers, and hot exhaust air exits into the hot aisles.

This arrangement prevents hot air from recirculating to the front of the racks, which is the single biggest cooling problem in data centers without containment.

The 7×24 Exchange has published operational guidelines showing that proper hot aisle/cold aisle orientation alone can improve cooling efficiency by 10% to 15% before any physical containment is added.

Cold aisle containment

Cold aisle containment is a physical enclosure system that seals the cold aisles using doors at each end and a ceiling panel (or curtain) above the aisle.

This keeps the cold supply air contained in the aisle until it passes through the servers.

Cold air cannot escape to the ceiling or mix with hot exhaust air before doing useful work.

The benefits of cold aisle containment are significant.

Schneider Electric’s white paper library documents that cold aisle containment can improve cooling efficiency by 20% to 30%, reduce fan energy consumption, and allow operators to raise the cold aisle set point temperature by 3°C to 5°C without affecting server inlet temperatures.

Raising the set point by just 1°C can reduce cooling energy costs by roughly 2% to 4%, according to ASHRAE research.

Common retrofit challenges include fitting containment around existing cable management, fire suppression systems, and lighting.

Facilities with raised floor cooling need to match perforated tile placement to the contained aisle dimensions.

Data center operators report that containment retrofit projects typically cost $1,000 to $3,000 per rack depending on the configuration, with a payback period of 12 to 24 months through energy savings.

Hot aisle containment

Hot aisle containment works in the opposite direction: it encloses the hot aisles and routes all hot exhaust air directly back to the cooling units.

The rest of the data center floor becomes a large cold air plenum.

This approach is common in newer facilities where the cooling units are positioned above the rows or use overhead ducting.

The choice between cold aisle and hot aisle containment depends on the facility layout, the cooling unit type, and the fire suppression system design.

Both approaches produce similar energy savings when properly implemented.

The decision often comes down to practical factors: hot aisle containment works better with overhead cooling distribution, while cold aisle containment pairs naturally with raised floor delivery.

Chilled water systems and cold water loops

A chilled water system is the backbone of cooling infrastructure in most mid-size and large data centers.

It uses a central plant with mechanical chillers to produce cold water, which is then pumped through a distribution loop to CRAH units, fan coil units, or other heat exchange devices throughout the facility.

The major components of a chilled water loop include: chillers (which use a refrigeration cycle to cool the water), cooling towers (which reject heat to the outside air), pumps (which circulate the water), and piping that connects everything.

The chilled water supply temperature is typically maintained between 7°C and 12°C (45°F to 54°F), and the return water comes back from the CRAH units at 12°C to 18°C (54°F to 64°F) after absorbing heat from the data hall.

Chilled water plants are sized in tons of refrigeration.

One ton equals 3.5 kW of cooling capacity.

A 10 MW data center might require a chiller plant rated at 3,000 to 4,000 tons depending on climate, PUE targets, and redundancy requirements.

Trane, Carrier, York (Johnson Controls), and Daikin are the major chiller manufacturers serving the data center market.

Chilled water system design considerations

Redundancy is a critical design factor.

Most enterprise data centers require N+1 chiller redundancy at minimum, meaning one more chiller than needed to handle the full cooling load.

Uptime Institute Tier III and Tier IV facilities require 2N redundancy for all cooling infrastructure, which doubles the chiller plant capacity.

Pump and piping layout directly affects system efficiency.

Variable-speed pumps adjust water flow to match the actual cooling load, which can reduce pumping energy by 30% to 50% compared to constant-speed designs.

Primary-secondary and variable primary flow configurations are the two most common piping layouts, and the choice between them affects both capital cost and operating efficiency.

Liquid cooling and rack-level solutions

Liquid cooling is any data center cooling method that brings a liquid medium (water, glycol, or engineered coolant) into direct or close contact with heat-generating components.

As rack densities climb past 30 kW and into the 60-100+ kW range driven by NVIDIA GPU clusters and AI training infrastructure, liquid cooling has shifted from an exotic technology to a practical requirement.

Rear-door heat exchangers (RDHx)

A rear-door heat exchanger is a cooling device mounted on the back door of a server rack. It uses chilled water flowing through a coil to absorb heat from the hot exhaust air as it exits the rack, before that air enters the data center room.

The result: the air leaving the rear door is close to room temperature, which reduces the load on the room’s CRAH or CRAC units.

RDHx units are a popular retrofit option because they connect to the existing chilled water infrastructure and do not require changes to the servers themselves.

A well-designed RDHx can remove 60% to 80% of the rack’s heat load at the point of generation.

Vertiv, CoolIT, and Motivair are leading manufacturers of rear-door heat exchangers for data centers.

Direct-to-chip liquid cooling

Direct-to-chip cooling, also called cold plate cooling, brings liquid directly to the hottest components inside the server.

A cold plate attaches to the CPU or GPU, and chilled water (or coolant) circulates through the plate to absorb heat.

The heated liquid returns to a coolant distribution unit (CDU) where the heat is transferred to the building’s chilled water loop.

This method can handle rack densities of 50 to 80+ kW per rack and is the approach used by major cloud providers deploying NVIDIA H100 and H200 GPU clusters.

Microsoft has publicly stated that its Azure AI data centers use direct-to-chip liquid cooling for GPU-intensive workloads.

Google and Meta have disclosed similar strategies in their sustainability and infrastructure reports.

Immersion cooling

Immersion cooling is the most aggressive form of liquid cooling.

Servers are fully submerged in a non-conductive dielectric fluid that absorbs heat directly from all components.

There are two types: single-phase immersion (the fluid stays liquid) and two-phase immersion (the fluid boils at a low temperature and the vapor is condensed back into liquid).

The data center cooling market for liquid solutions, including immersion, is growing rapidly. MarketsandMarkets projects the data center liquid cooling market will grow from $4.3 billion in 2024 to $15.5 billion by 2030, representing a CAGR of over 20%.

Immersion cooling can handle rack densities exceeding 100 kW per rack and eliminates the need for server fans entirely, which reduces noise and saves the energy those fans would consume.

Early adopters include cryptocurrency mining operations, high-performance computing (HPC) labs, and, increasingly, AI training clusters. GRC (Green Revolution Cooling), LiquidCool Solutions, and Submer are three of the largest immersion cooling vendors.

Adoption in mainstream enterprise data centers is still limited in 2026, but pilot programs are expanding quickly.

Liquid Cooling Method	How It Works	Density Supported	Retrofit Difficulty	Adoption in 2026
Rear-door heat exchanger (RDHx)	Chilled water coil on rack door absorbs exhaust heat	Up to 35 kW/rack	Low (uses existing infrastructure)	Common in colocation
Direct-to-chip (cold plate)	Cold plate on CPU/GPU, coolant loop to CDU	50-80+ kW/rack	Medium (requires server modifications)	Growing fast for AI workloads
Single-phase immersion	Servers submerged in dielectric fluid	100+ kW/rack	High (specialized tanks, fluid handling)	Pilot/early adoption
Two-phase immersion	Servers submerged; fluid boils and condenses	100+ kW/rack	High (sealed enclosures, condenser systems)	Experimental/pilot

Cooling efficiency and monitoring

Cooling efficiency measures how much energy you spend on cooling compared to the useful work your IT equipment performs.

The most widely used metric is Power Usage Effectiveness (PUE), defined as total facility energy divided by IT equipment energy.

A PUE of 1.0 would mean zero overhead, which is impossible. A PUE of 2.0 means you spend as much energy on cooling, lighting, and other overhead as you do on the IT equipment itself.

The Uptime Institute’s 2024 survey reported a global average PUE of 1.58, down from 2.5 in 2007.

The most efficient hyperscale facilities operated by Google, Meta, and Microsoft report annual PUE values below 1.10. Google reported a fleet-wide trailing twelve-month PUE of 1.10 in its 2024 Environmental Report.

Other cooling-specific efficiency metrics include:

Cooling system efficiency measures the kW of cooling delivered per kW of energy consumed by the cooling plant. Higher numbers are better.

Modern chiller plants with variable-speed drives and economizer modes can achieve cooling efficiencies of 0.3 to 0.6 kW per kW of cooling, meaning the cooling system uses 30% to 60% of the energy value of the heat it removes.

Return temperature index (RTI) measures how well cold supply air and hot return air are separated. An RTI of 100% means perfect separation.

Most data centers with containment achieve 80% to 95% RTI.

Environmental sensors are the foundation of cooling monitoring.

Temperature and humidity sensors placed at the front and rear of every rack, at CRAH supply and return points, and at critical infrastructure locations give operators real-time visibility into thermal conditions.

DCIM (Data Center Infrastructure Management) software platforms from vendors like Schneider Electric (EcoStruxure IT), Vertiv (Trellis), and Nlyte collect sensor data and provide dashboards, alerts, and capacity planning tools.

Telemetry from servers themselves is becoming increasingly important.

Modern servers report CPU and GPU junction temperatures, fan speeds, and power consumption through IPMI, Redfish, or proprietary interfaces.

Operators can use this data to detect hot spots before they become outages and to optimize cooling set points dynamically.

Integration with building HVAC and air conditioning

The interface between data center cooling and building HVAC is a critical design consideration that affects energy efficiency, redundancy, and maintenance.

In a standalone data center, the cooling system is purpose-built and self-contained.

In a colocation or enterprise data center located within a larger building, the data center’s cooling plant may share infrastructure with the building’s comfort cooling system, or it may operate independently.

Data hall cooling and building HVAC serve fundamentally different purposes.

The data hall cooling system must operate 24 hours a day, 365 days a year regardless of outdoor temperature or building occupancy.

Building HVAC systems for office spaces and common areas operate on schedules, respond to occupancy, and typically shut down at night and on weekends.

Mixing these two systems without proper controls creates reliability risks.

Coordination with facility HVAC controls matters most during economizer operation.

When outside air is cool enough for free cooling, the data center cooling system may open dampers or switch to dry coolers that reject heat directly to the outside air.

The building HVAC system may simultaneously be running in heating mode (in winter, for example).

Proper control sequences prevent these two systems from fighting each other, which wastes energy and confuses building management systems.

For operators working in colocation facilities or multi-tenant buildings, understanding this interface is part of the job. Equinix, Digital Realty, and CoreSite all have dedicated facilities engineering teams that manage the boundary between data hall cooling and building mechanical systems.

Design best practices and operational guidance

Designing and operating cooling systems well comes down to a handful of principles that the best operators apply consistently.

Containment is non-negotiable for new builds. Every major data center design standard now recommends hot aisle or cold aisle containment.

The energy savings (20% to 30% reduction in cooling costs) pay for the containment infrastructure within 12 to 24 months.

Designing without containment in 2026 is leaving money on the table.

Target the highest supply temperature your equipment will tolerate. ASHRAE A1 class servers can handle inlet temperatures up to 32°C (89.6°F) for short periods.

Running your cold aisle at 24°C to 27°C instead of the traditional 18°C to 20°C dramatically reduces chiller energy consumption.

For every 1°C increase in supply air temperature, cooling energy drops by 2% to 4%.

Balance chilled water temperature set points carefully. Lowering chilled water supply temperature improves cooling capacity but increases chiller energy use.

Raising it reduces energy use but may limit capacity during peak loads or hot weather.

The optimal set point depends on your climate, chiller efficiency curve, and current IT load.

Many operators find the sweet spot between 8°C and 12°C for chilled water supply.

Schedule routine maintenance for cooling equipment.

Dirty coils, clogged filters, and degraded fan bearings reduce cooling capacity and increase energy consumption.

AFCOM’s State of the Data Center report consistently identifies deferred cooling maintenance as one of the top five causes of unplanned temperature excursions.

A quarterly cleaning and inspection cycle for CRAH/CRAC coils, filters, and fan assemblies is standard practice at well-run facilities.

Cost, sustainability, and future trends

Cooling infrastructure represents a significant portion of data center capital expenditure and ongoing operating expense.

A new 10 MW air-cooled data center’s cooling plant (chillers, cooling towers, CRAH units, piping, controls) typically costs $15 million to $25 million to build, according to Turner & Townsend’s 2025 International Construction Cost Survey.

Liquid cooling infrastructure can add 20% to 40% to that cost, depending on density and technology choice.

On the CAPEX versus OPEX question: liquid cooling systems cost more to install but can reduce operating costs by 20% to 40% through lower PUE, reduced fan energy, and smaller footprint requirements.

The payback period for liquid cooling over air cooling in a high-density AI data center is typically 3 to 5 years, assuming rack densities above 40 kW.

Sustainability is pushing the cooling conversation forward.

Data centers consumed an estimated 460 TWh of electricity globally in 2024, according to the International Energy Agency, and that number is projected to more than double by 2030.

Cooling accounts for 30% to 40% of that consumption.

Reducing cooling energy through higher efficiency systems, free cooling, and liquid cooling directly reduces carbon emissions and water consumption.

Water use is an increasingly scrutinized metric. Cooling towers consume significant water through evaporation.

Meta reported that its data centers used 2.6 billion liters of water in 2023 for cooling.

Dry cooling and closed-loop liquid cooling systems use zero water for heat rejection, which is why operators in water-stressed regions (Phoenix, parts of Texas, the Middle East) are moving toward these technologies.

Heat reuse is an emerging trend. Data center waste heat can be captured and used to warm nearby buildings, greenhouses, or district heating systems.

Equinix operates heat reuse programs in several European locations, and Microsoft’s planned Stockholm facility will supply heat to the local district heating network.

The amount of waste heat available from a 100 MW data center is equivalent to heating roughly 20,000 homes, creating both sustainability benefits and potential revenue.

Summary and next steps

Data center cooling systems range from traditional air-cooled CRAC and CRAH setups to advanced liquid immersion solutions.

The right choice depends on your facility’s rack density, budget, climate, and growth plans.

Three things to remember: containment is mandatory for energy efficiency, chilled water systems outperform DX systems at scale, and liquid cooling is no longer optional for AI-density workloads above 30 kW per rack.

Your next step depends on your role.

If you are evaluating cooling for a new facility, run a lifecycle cost analysis comparing air cooling with containment, rear-door heat exchangers, and direct-to-chip solutions for your projected rack densities.

If you manage an existing facility, audit your current cooling performance by measuring PUE, checking containment integrity, and benchmarking against ASHRAE recommended temperature ranges.

If you are building a career in data centers, cooling knowledge is one of the most valuable technical skills you can develop.

Employers from Equinix to AWS are hiring people who understand these systems inside and out.

Check data center technician job description for roles that list cooling expertise as a requirement, or browse the data center engineer salary guide to see what cooling specialists earn.

Frequently asked questions

What is the most efficient data center cooling system?

Chilled water systems with free cooling economizers are the most energy-efficient air-based cooling approach, achieving PUE values below 1.2 in favorable climates. For high-density AI workloads, direct-to-chip liquid cooling is the most efficient per-kW option because it removes heat at the source with minimal fan energy. The best choice for your facility depends on rack density, local climate, and capital budget.

What is the difference between CRAC and CRAH units?

CRAC units use a built-in direct expansion refrigerant cycle to cool air, while CRAH units use chilled water supplied from a central plant. CRAH units are more energy efficient at scale because the central chiller plant can use economizer modes and high-efficiency compressors. CRAC units cost less per unit and work well for smaller deployments where a full chilled water plant is not justified.

How does liquid immersion cooling work in a data center?

Liquid immersion cooling submerges servers completely in a non-conductive dielectric fluid that absorbs heat directly from all components. The fluid transfers heat to a heat exchanger connected to the building’s cooling loop. Single-phase systems keep the fluid in liquid state, while two-phase systems allow the fluid to boil and condense. Immersion cooling supports rack densities above 100 kW and eliminates the need for server fans.

What temperature should a data center be kept at?

ASHRAE recommends data center inlet air temperatures between 18°C and 27°C (64°F to 80.6°F) for most enterprise equipment under the A1 classification. Many modern operators target 24°C to 27°C for cold aisle temperatures to reduce cooling energy costs. Running at the upper end of the recommended range can cut cooling energy by 10% to 20% compared to the traditional 18°C to 20°C set points.

How much does data center cooling cost?

Cooling infrastructure for a new 10 MW data center typically costs $15 million to $25 million in capital expenditure, according to Turner & Townsend. Operating costs depend on PUE, local electricity rates, and climate. A facility with a PUE of 1.4 spending $0.07 per kWh on electricity will pay roughly $2.5 million to $3.5 million per year on cooling energy for a 10 MW IT load. Improving PUE from 1.4 to 1.2 can save $500,000 to $1 million annually in cooling costs alone.