Data Center Cooling Systems Explained

Quick answer

Data center cooling systems do two jobs. They capture heat at the IT equipment, using room air (CRAC or CRAH units), aisle containment, rear-door coils, direct-to-chip cold plates or immersion. Then they reject it outdoors through chillers, cooling towers, dry coolers or free cooling. Conventional racks run on air; AI racks near 100 kW need liquid.

Data center cooling systems are usually described as a list of products. It is easier to see them as a chain: heat moves from the chips into air or liquid, then to a heat exchanger, then out of the site through a chiller, cooling tower or dry cooler.

This guide walks that chain and sorts the main data center cooling technologies by the rack density they suit. Figures come from the Department of Energy (DOE) and its labs, ASHRAE's guidelines as summarized by DOE, and Uptime Institute's survey. We left out the kW "limits" quoted online that we could not trace to a source.

How does data center cooling work?

Data center cooling moves heat from the IT equipment to the outdoors in two stages. PNNL notes that most of the electricity IT equipment uses is turned into heat, so the cooling system has to carry away close to what the racks draw.

Capture. Heat leaves the chips and enters a medium: room air, air passing through a coil at the rack, liquid in a cold plate, or a fluid the server sits in.
Transport. Air handlers or coolant distribution units (CDUs) pass that heat into a facility water loop.
Rejection. A chiller, cooling tower, dry cooler or outside air takes the heat out of the loop and releases it to the atmosphere.

The target at the server inlet comes from ASHRAE. Its recommended range for air-cooled IT equipment is 18 to 27 °C (64.4 to 80.6 °F), and DOE's design guide says to design and run to it.

Cooling is the largest overhead. Lawrence Berkeley National Laboratory (LBNL) says cooling dominates facility infrastructure energy, and infrastructure fell from 36% of US data center electricity in 2018 to 31% in 2024, helped by more AI servers in liquid-cooled facilities.

Air cooling: CRAC, CRAH and aisle containment

Room air cooling is still the default for conventional racks. Cold air reaches the front of the racks, server fans pull it through, and warm air leaves the back.

Two kinds of unit do the work:

CRAC (computer room air conditioner). A packaged unit with its own compressor and refrigerant. DOE calls packaged direct-expansion units the most common cooling equipment in smaller data centers. Heat is rejected through an air-cooled condenser, a dry cooler or a cooling tower.
CRAH (computer room air handler). Fans and a coil fed with chilled water from a central plant. There is no compressor in the unit itself. Large sites usually pair CRAHs, often in rooms next to the hall, with a central chilled water plant.

Central systems tend to waste less. DOE's guide compares two large facilities with similar IT loads. The one with a central air handler system put almost two-thirds of its input power into IT equipment. The one with many distributed CRAC units put in just over one-third.

Containment helps either kind. Racks face each other in alternating cold and hot aisles, and gaps are sealed so the air streams don't mix. With separation, return air often comes back at 85 °F or more, which DOE notes extends the hours a site can use free cooling.

Close-coupled cooling: in-row coolers and rear-door heat exchangers

Close-coupled cooling moves the air-to-liquid heat exchange from the edge of the room to the rack itself. DOE groups these as localized air-to-liquid heat exchangers.

Rear-door heat exchangers (cooling doors) replace a rack's back door with a water coil. Passive doors rely on server fans; active doors add their own.
In-row coolers sit between racks and pull air from the hot aisle.

The servers stay air-cooled, so this suits hardware with no liquid connections. It is often how a hall handles racks denser than its room units were designed for.

Liquid cooling: direct-to-chip and immersion

Direct liquid cooling takes heat into a fluid at the chip, before it ever reaches room air. DOE says liquid can serve higher heat densities and be much more efficient than air, because water carries heat far better.

Direct-to-chip (cold plates). The heat sinks on processors are replaced by plates with liquid channels, fed through a CDU. Most of the heat goes to liquid. The rest, from parts without plates, still goes to air, so these halls are hybrids.
Single-phase immersion. Servers sit in a tank of nonconductive fluid that is pumped through a heat exchanger.
Two-phase immersion. An engineered fluid boils at the hot components, and the vapor condenses on a coil and drips back.

High-performance computing adopted liquid early: DOE's guide notes racks of 60 kW in 2013 and recent ones above 125 kW. Warmer coolant also allows more free cooling and, in some cases, dry coolers in place of towers.

For the full comparison of cold plates against immersion, read our guide to liquid cooling for AI data centers.

Data center cooling methods by rack density, water and tenant input
Cooling method	Rack density it suits	Effect on site water use	What the tenant must provide or specify
Room air: CRAC or CRAH units	Low; conventional racks of a few kW	None in the room; set by the heat rejection plant	Air-cooled hardware with front-to-back airflow; inlet range
Hot or cold aisle containment	Low to medium; helps air reach the 10-30 kW range	Lowers water and energy by allowing more free cooling hours	Rack layout, blanking panels, cable openings sealed
Rear-door or in-row heat exchangers	Medium to high air-cooled racks	Chilled or warm water to the row; plant decides water use	Rack type and door fit; space for piping; power per rack
Direct-to-chip cold plates	High; AI racks of 100 kW and up	Warm coolant can allow dry coolers and more free cooling	Heat-to-liquid share, coolant class, flow and pressure, CDU and manifold ownership
Single- or two-phase immersion	High; dense racks with immersion-ready hardware	Warm coolant can allow dry coolers and more free cooling	Immersion-rated servers, approved fluid, tank layout and service access

Heat rejection: chillers, cooling towers, dry coolers and free cooling

Heat rejection is the outdoor half of the system, and it decides most of a site's water use.

Rejection option	How it rejects heat	Water used on site	Where it fits
Chiller (water- or air-cooled)	A compressor cycle cools the facility water loop, then passes heat to a tower or to outside air	Depends on what the chiller rejects to	Large facilities and warm or humid climates; DOE calls an efficient water-cooled chiller plant typically the most efficient option for large sites
Cooling tower	Evaporates part of the water to cool the rest	High: evaporation, blowdown and drift	Large chilled water plants; DOE notes towers can beat air-cooled systems on energy in dry climates
Dry cooler	Fans blow outside air over a closed coil, like a car radiator	Little to none	Warm coolant loops, such as liquid-cooled halls; mild or cool climates
Water-side economizer	Tower water cools the loop through a heat exchanger without running chillers	Uses tower water	Climates with low wet-bulb temperatures for many hours a year
Air-side economizer	Filtered outside air goes straight into the hall	Little, unless air is humidified or evaporatively cooled	Cool climates with clean air; hard to add to an existing building

Data center cooling towers reject heat by evaporation. DOE gives the water balance as make-up water equals evaporation plus blowdown plus drift. Blowdown is water drained to stop minerals building up, and it is the part operators control most.

Free cooling uses cold outside air or water instead of compressors. DOE says a water-side economizer suits climates with wet-bulb temperatures below 55 °F for 3,000 or more hours a year. Its list of steps for sustainable sites puts free cooling first and rejecting remaining heat to dry coolers third.

For the numbers on water, see our guide to data center water usage.

Which cooling system fits which rack density?

Rack density is the main thing that decides the capture method. The sourced anchors are:

A few kW. LBNL says a rack of conventional servers used about 3 to 5 kW. Room air with CRAC or CRAH units handles this.
About 10 to 30 kW. Uptime Institute's 2025 survey says average densities are rising slowly, driven by more racks in the 10 to 30 kW range, and that few facilities exceed 30 kW. Containment, and then rear doors or in-row coolers, carry air cooling into this range.
100 kW and up. LBNL says AI racks use as much as 100 kW. A December 2025 Pacific Northwest National Laboratory (PNNL) report for DOE puts AI training racks above 100 kW and present-day GPU racks at 120 kW. These are liquid-cooled, usually with cold plates.

There is no single number where air stops working. DOE's guide points to an ASHRAE white paper that ties the move to liquid to the power of each chip, not rack totals. In practice the hardware spec settles it: a server built with cold plates needs liquid at the rack.

Most dense halls end up mixed: cold plates for accelerators, air for the rest. To size the heat behind each density, see how much power a data center uses.

What should a tenant specify for cooling?

If you bring your own hardware, the cooling questions start with your servers. Settle these in writing before a hall is built:

Rack power, now and later. Day-one kW per rack and the density expected over the lease.
Air, liquid or both. For cold-plate servers, the share of heat that goes to liquid.
Coolant temperature class. ASHRAE names liquid classes by upper supply temperature in °C: W17, W27, W32, W40, W45 and W+. Warmer classes allow more free cooling.
Flow, pressure and fluid your hardware maker allows.
CDU and manifold ownership. Who supplies and maintains them.
Resilience. Spare cooling units, and whether pumps and fans ride on UPS. PNNL notes this is increasingly common in AI sites, because dense racks heat up fast.
Water limits. Whether the site can run towers or must reject heat dry.

That split is the core of GPU colocation: you own the servers, and the provider delivers power, cooling and space to your specification.

Where Accretion Compute fits

Accretion Compute builds AI data center capacity at the source of energy, with on-site, behind-the-meter power, and leases it to AI companies on multi-year terms. We are early: no site is built or operating, and we have no capacity available today. We are evaluating sites in Canada and the United States. Cooling is designed for the tenant's density, air or liquid, and halls are built to the tenant's specification, with room to expand on the same site. See how we approach capacity and site selection.

Good fit if

You are an AI company with power-dense hardware, air or liquid-cooled, and you want cooling designed around your racks on a multi-year lease.
You bring and own your hardware and want the energy, building, power, cooling and connectivity handled for you.
You need firm on-site power and don't want to wait for a grid connection.

Probably not a fit if

You want GPU-hours, cloud instances or rented GPUs. We are a capacity provider, not a cloud, and we do not own or rent out GPUs.
You need capacity live today. None is built yet.
You are shopping for cooling equipment. We design and run cooling as part of a leased site; we don't sell it.

Talk to us

Frequently asked questions

How are data centers cooled?

In two stages. Inside, heat is captured by room air from CRAC or CRAH units, by coils on the racks, or by liquid in cold plates or immersion tanks. Outside, it is rejected by chillers, cooling towers, dry coolers or outside air. Conventional racks use air; dense AI racks use liquid at the chip.

What is the difference between CRAC and CRAH units?

A CRAC (computer room air conditioner) has its own compressor and refrigerant, and is common in smaller data centers. A CRAH (computer room air handler) has no compressor: it blows air over a coil fed with chilled water from a central plant. Large sites usually use CRAHs with a central plant.

What temperature should a data center be kept at?

ASHRAE recommends a server inlet air temperature of 18 to 27 °C (64.4 to 80.6 °F) for air-cooled equipment. That is the air entering the servers, not the room average. The wider allowable ranges show what hardware tolerates, not what keeps it reliable.

Do data centers use cooling towers?

Many large data centers do, usually to reject heat from a chilled water plant or to run a water-side economizer. A tower cools water by evaporating part of it, so it uses water continuously. DOE's guidance puts make-up water as evaporation plus blowdown plus drift, and says raising cycles of concentration from three to six cuts make-up water by 20% and blowdown by 50%. Sites short of water can use dry coolers instead.

What is free cooling in a data center?

Free cooling uses cold outside air or water instead of compressors. An air-side economizer brings filtered outside air into the hall. A water-side economizer uses a cooling tower to chill the facility loop. DOE says many cool-climate data centers run with no chillers at all.

At what rack density do you need liquid cooling?

There is no single cutoff; it depends on how much power each chip draws and on the hall design. As anchors, conventional racks use about 3 to 5 kW and run on air, while AI racks reach 100 kW or more and are liquid-cooled. Containment and rear doors extend air in between.

Data Center Cooling Systems: Every Method, by Rack Density