Data center cooling systems pull heat away from servers and release it outdoors. The three main approaches are air cooling, liquid cooling, and chilled water. Air suits lower rack densities. Liquid handles high-density AI racks. Chilled water carries heat from the building to the plant. Most facilities combine them.

Say a chiller fails in the middle of the night. Warm air starts reaching the servers, their fans speed up, and the processors slow down as temperatures rise. There may not be enough time for someone to spot the issue on a dashboard and respond. A data center cooling system needs another cooling path ready to handle situations like this.

Keeping data centers running is getting more expensive each year. According to Lawrence Berkeley National Laboratory, data centers accounted for around 4.4% of all electricity used in the US in 2023. That figure could reach anywhere from 6.7% to 12% by 2028 (US Department of Energy, December 2024). Nearly all the electricity these facilities consume turns into heat, which their cooling systems then have to remove.

Wootz.work builds skids, heat exchangers, and enclosures for these systems through a network of 400+ qualified manufacturing partners. The later sections on equipment and sourcing come from the shop floor.

What do data center cooling systems do?

Servers turn almost every watt they draw into heat. A data center cooling system removes that heat around the clock. It keeps the air or liquid reaching the equipment inside a safe window.

ASHRAE Technical Committee 9.9 sets out the temperature guidelines for server equipment. It recommends keeping inlet temperatures between 18 and 27°C (64.4 to 80.6°F) for classes A1 through A4. Class A1 hardware can handle temperatures from 15 to 32°C, as noted in the ASHRAE TC 9.9 reference card. Cooling the servers more than needed adds to electricity costs, but letting them get too hot increases the chances of failure.

The data center cooling requirements for a new hall begin with a short list:

  • Heat load in kilowatts, per rack and for the whole hall
  • Target inlet temperature and humidity range
  • Redundancy level, such as N+1 (one spare unit beyond what the load needs)
  • Availability of water and electricity, along with the local climate
  • Floor space and ceiling height needed for the data center cooling setup

These details help determine the cooling system you'll need. The hyperscale infrastructure guide explains how to plan for them when building in phases. Low-density racks in cooler regions may only need air cooling, while GPU-heavy racks generally require liquid cooling directly at the chip. The chilled water plant also needs enough capacity to handle the cooling load.

How does data center cooling work?

Heat takes a six-step trip from the chip to the outdoors. Each step hands it to a cooler medium.

  1. Processors, memory, and power supplies turn electricity into heat.
  2. A heat sink and fan pass that heat to air, or a cold plate passes it to liquid coolant.
  3. The air or coolant carries it to a heat exchanger: a computer room air handler (CRAH) coil, a rear door, or a coolant distribution unit (CDU).
  4. A chilled water loop carries the heat from that exchanger to the chillers.
  5. The chillers push it into a condenser loop, or straight into outdoor air.
  6. A cooling tower or dry cooler releases it outside.

Energy leaks at the hand-offs. In an air system, hot exhaust that mixes with cold supply air forces the plant to supply colder air than the servers need. Designers fight that with hot aisle and cold aisle layouts. The cold aisle faces the server intakes. The hot aisle faces the exhausts. Physical containment keeps the two streams apart. Treat data center cooling air as a resource you can ruin by mixing it.

Liquid systems avoid much of that problem because the coolant touches the heat source directly. The air and liquid paths split at the server and rejoin at the chilled water loop. Everything after that point is shared plant.

Types of data center cooling systems

The three types in the title overlap. Air and liquid describe how heat leaves the server. Chilled water describes how it travels from the hall to the plant, and it often sits behind both. Free and evaporative cooling are ways to reject heat for less energy. Engineers group the types of data center cooling systems by where the heat gets removed. The table below sorts the main data center cooling methods and data center cooling technologies the same way.

Type How it moves heat Best fit Main limit
Air cooling (CRAC, CRAH, in-row) Fans push cooled air through the racks Low to moderate rack density Air carries little heat per volume, so fan energy climbs with density
Direct-to-chip liquid Cold plates on CPUs and GPUs carry heat into coolant High-density AI and HPC racks Needs CDUs, piping to every rack and leak management
Rear door heat exchanger A liquid coil on the rack door cools exhaust air Retrofits of existing air-cooled rows Still depends on a chilled water supply
Immersion Servers sit in dielectric fluid Ultra-dense, purpose-built halls Limits hardware choice and slows servicing
Chilled water Pumps move cold water between chillers and room or rack units Medium and large facilities Plant capacity and redundancy drive cost
Free and evaporative cooling Outdoor air or evaporating water removes heat Cool or dry climates Climate and water use

‍

Start by choosing the cooling equipment based on your rack density, whether that's an air handler, rear door, or cold plate. Once that's decided, size the chilled water plant to meet the cooling needs. Most data centers use a mix of air and liquid cooling. Air handles storage and network equipment, while liquid cooling takes care of racks that generate more heat. Both systems can run from the same plant.

Air cooling: CRAC, CRAH and containment

Air cooling is still common in many data center halls. The two main types of cooling units are CRAC and CRAH, and although they serve the same purpose, they work differently.

CRAC units cool air using a compressor and refrigerant built into each unit. CRAHs don't have that setup. They get chilled water from a central plant and pass it through a coil, with fans moving air over the surface. A CRAC is generally easier to install. In a large data hall, though, CRAHs can be more practical because multiple units run off the same chilled water system.

Once cooled, the air moves through ducts or beneath a raised floor before entering the cold aisle through floor tiles or overhead diffusers. In-row cooling units are placed between server racks, so the air doesn't have to travel as far to reach the equipment.

Keeping the hot and cold air apart also makes a difference. Doors and roof panels around the aisle help stop the two air streams from mixing. With less mixing, operators can raise the supply air temperature closer to the upper end of ASHRAE's recommended range, allowing the chillers to use less energy.

Air has a ceiling. Wootz.work’s hyperscale infrastructure guide puts the practical limit near 20 kW per rack, depending on containment and ceiling height. Past that point, the fans move so much air that noise and fan power become the problem. Designers then turn to in-row cooling, rear doors, or liquid.

A CRAH or in-row unit is mostly sheet metal, coils, fans, and piping. That makes every data center cooling system that uses air a fabrication problem as much as a thermal one.

Liquid cooling: direct-to-chip, rear door and immersion

A given volume of water absorbs far more heat than the same volume of air. That is why liquid reaches rack densities that air cannot. Liquid cooling in a data center takes three forms.

Direct-to-chip cooling mounts a metal cold plate on each CPU and GPU. Coolant flows through the plate, picks up heat and returns warm. The plates connect to rack manifolds through quick-disconnect couplings. The manifolds connect to a coolant distribution unit. The CDU holds pumps and a heat exchanger. It keeps the server-side coolant apart from the facility water on the other side.

A rear door heat exchanger has a liquid coil mounted at the back of a rack. It cools the hot air coming out of the rack before that air enters the room. Since it doesn't require major changes to existing rows, it works well for data centers that aren't planning to use cold plates just yet.

Immersion submerges servers in a tank of dielectric fluid. It handles extreme density. It also limits hardware choice, and a full tank is slow to service.

Inside the hall, these loops are closed. The same coolant circulates between the cold plates and the CDU, and heat leaves through the exchanger. Closed-loop data center cooling consumes no water at the rack. Water use depends on how the plant downstream rejects the heat.

Liquid cooling comes with higher upfront costs and more connections where leaks can occur. Operations teams also need to understand fluid chemistry and know how to detect leaks. The hyperscale guide points to another concern: if a coolant distribution unit (CDU) fails, high-density chips have very little thermal mass to help them manage the heat. This makes backup pumps and secondary cooling loops necessary.

Every part of a liquid loop is built to a pressure and cleanliness requirement. That means stainless tube, welded manifolds, and clean fittings. The quality of that hardware decides whether a data center cooling system built on liquid stays leak-free in service.

Chilled water systems: chillers, pumps and cooling towers

Chilled water is the backbone behind most large halls. A data center chiller cools water to a set supply temperature. Pumps send that water through the building to CRAH coils, rear doors and CDUs. It returns warmer, and the chiller pulls the heat back out. Chilled water data center cooling works at the scale of the whole facility. That is why plant capacity sets the ceiling for everything downstream.

The chiller still has to put that heat somewhere. Two loops share the job.

  • The chilled water loop is closed. The same water circulates between the chillers and the terminal units.
  • The condenser water loop carries heat from the chiller to the outdoors. Many plants send it through data center cooling towers, where a share of the water evaporates and takes heat with it. Others use dry coolers. These reject heat to outdoor air through finned coils and need no evaporative water.

Towers reject heat efficiently but consume water. Dry coolers avoid that water use but work harder on hot days. When outdoor air is cold enough, a waterside economizer lets the plant bypass the compressors for part of the year.

Chiller plants add spare units instead of duplicating everything. If the load needs three chillers, an N+1 plant installs four. The piping needs isolation valves so any one unit, pump or valve can come out for repair while the hall stays online.

The plant is where much of the fabrication volume sits. Wootz.work’s data center offering covers chilled water and condenser skids, cooling towers and pump sets, precision cooling controls, and CRAC and in-row systems. All of it is delivered through its partner network.

How to choose between air, liquid and chilled water

Choose data center cooling solutions by rack density first. Then check climate, water, and project type. The table compares the three on the factors that change the decision most.

Factor Air cooling Liquid cooling Chilled water plant
Rack density Low to moderate, practical limit near 20 kW per rack High, and the route for GPU racks Serves both and sets total capacity
Climate and water Cool climates allow more free cooling Less tied to room air temperature Towers need water, dry coolers need cooler air
Retrofit effort Easiest to extend Rear doors are the lighter lift; cold plates need piping to every rack Plant additions are major projects
Upfront cost Lower Higher Scales with total heat load
Operations skills Familiar to most teams Fluid chemistry and leak detection Established mechanical trades
Failure mode Hot spots build as airflow drops Fast thermal shutdown if coolant flow stops Needs spare chillers, pumps and valves

‍

For cooler locations, air cooling with proper containment and a chilled water plant is usually enough for racks below 20 kW. Once rack loads go beyond that, it's worth adding liquid cooling for the racks that generate the most heat, with room to expand later. If you expect rack density to reach the limits of air cooling within the next two years, install the liquid cooling pipes while the facility is being built. Adding them after the data hall is operational will cost much more.

What is inside a data center cooling system, and who builds it?

Behind every cooling plant sits hardware that someone has to cut, weld, test, and ship. Most data center cooling equipment falls into five groups.

  • Skids: chilled water, condenser and CDU skids that bundle pumps, valves, piping, instruments and a frame
  • Heat exchangers: coils, plate units and U-tube exchangers
  • Piping and valves: stainless prefabricated assemblies, fittings and isolation valves
  • Tanks and frames: reservoir tanks, base frames, jacket piping and flanges
  • Enclosures: racks, cabinets and air handler casings in sheet metal

Wootz.work supplies this equipment through its partner network. The capability list includes U-tube heat exchangers up to 40 ft with aluminum or copper fins, designed to ASME, PED, and TEMA requirements. It also includes sanitary valves, fittings, and prefabricated pipe assemblies built to ASME BPE and 3-A requirements for chilled water loops, direct-to-chip loops, and condenser water. Stainless grades run from SS304 to SS316L, with electropolishing available. Those standards apply to the parts and to the partner facilities that make them. Wootz.work itself holds ISO 9001:2015, ISO 14001:2015 and ISO 45001:2018 certificates.

Every quote starts with a design for manufacture (DFM) review. Engineers check weld access, tolerances, and material callouts before anyone prices the job. The first quote includes that review at no charge. The same team handles fabrication, from frames and tanks to finished assemblies.

Skids concentrate the integration risk, so they ship pre-assembled, with pumps, valves and controls mounted and tested. For background on assembling systems before they leave the factory, see Wootz.work’s guide to system integration in manufacturing. In one recent program for an industrial gases market leader, a chemical dosing skid shipped plug-and-play. It had a 48-hour factory acceptance test, no hot work on site, and a tenfold cut in instrumentation and control manpower. It is not a cooling skid, but the delivery method carries over to chilled water and condenser skids.

Paperwork matters as much as welds. Each shipment carries mill test reports with heat numbers, dimensional reports from CMM inspection, and weld documentation. A buyer’s quality team can verify the work without traveling. The quality assurance and transparency page shows how that record is kept.

How to source data center cooling equipment

Buyers comparing data center cooling manufacturers tend to start with unit price. The cost of a miss arrives later, as a late skid, a failed weld test, or a missing certificate. This checklist catches most of those problems before the purchase order.

  1. Send a complete drawing package with materials, standards, and test requirements.
  2. Ask for design feedback before the price.
  3. Confirm which standards the partner facility holds, not only the trading company.
  4. Request sample documents: mill test reports with heat numbers, dimensional reports, and weld records.
  5. Agree the factory acceptance test scope and who witnesses it.
  6. Check packing for ocean freight and confirm delivery terms, DAP or DDP.
  7. Ask who supports commissioning and spares after delivery.
  8. Check current duties and add a lead time buffer before you award.

Wootz.work handles design review, documentation, factory acceptance testing, and DAP or DDP delivery under one contract. The first quote arrives in under 24 hours. The Skid Solutions catalog on the resources page shows typical skid builds for data center cooling equipment.

Offshore sourcing is not always the right call. A single unit needed in a few weeks is better bought locally. So is a project with domestic-content rules, or work that needs the original manufacturer’s team on site for commissioning.

Efficiency and maintenance of data center cooling systems

Power usage effectiveness (PUE) measures the overhead. It is total facility energy divided by IT energy, so 1.0 is the floor and lower is better. Cooling is a major part of that overhead. A well-run data center cooling system shrinks it.

Four steps can help reduce it:

  • Keep the aisles contained and close any gaps, including empty rack spaces and cable openings.
  • Once containment is working properly, gradually increase the supply air temperature toward the upper end of ASHRAE’s recommended range.
  • Run waterside or airside economizers whenever outdoor conditions allow.
  • Monitor inlet temperatures, coolant flow, and leak sensors. Alert on trends instead of waiting for alarms.

Maintenance follows the hardware. Cooling towers need water treatment and inspection. Pumps, valves, and strainers need scheduled service. Liquid loops need coolant sampling. Keep a spare on the shelf for every critical part with a long lead time. A waiting pump turns a repair into an outage.

Choosing and building data center cooling systems

Good data center cooling systems match the cooling method to the rack density. Then they size the chilled water plant behind it. Air with containment covers many halls. Liquid takes over where GPU racks push past the air limit. A chilled water plant feeds both. Its pumps, valves, and skids decide whether the system holds up in service. Plan the piping for liquid early. Specify the documents you expect with each shipment. Test skids before they leave the factory.

Ready to price the hardware? Send us your drawing. Your first quote comes with a free DFM review. Contact Wootz.work

FAQs

What is the most common type of data center cooling system?

Air cooling with CRAH units on a chilled water loop remains the most common data center cooling system in medium and large halls. Liquid cooling is spreading in new AI builds, where rack density passes what air can handle. Many facilities now run both, with one chilled water plant behind them.

How much cooling does a data center need?

A data center needs roughly 1 kW of cooling capacity for every 1 kW of IT load, with some extra capacity for backup. For example, a 1 MW data hall needs around 284 refrigeration tons, as one ton equals 3.517 kW. You'll also need spare chillers and pumps, such as a fourth unit in an N+1 setup.

Is liquid cooling better than air cooling?

That depends on how much power each rack uses. Liquid cooling is generally better for racks above 20 kW and GPU clusters. Air cooling costs less and is easier to manage in facilities with lower rack densities. Many data centers use both, relying on air cooling for regular racks and adding cold plates or rear-door cooling where more heat needs to be removed.

Are data center cooling systems closed loop?

Inside the hall, yes. Chilled water loops and liquid loops circulate the same fluid. The condenser loop differs. A cooling tower exposes that water to air, and some of it evaporates. Dry coolers keep the loop closed all the way to the outdoor coil and avoid evaporation.

What temperature should a data center run at?

ASHRAE recommends keeping server inlet temperatures between 18°C and 27°C (64.4°F to 80.6°F) for equipment classes A1 to A4. Temperature readings should be taken at the server inlet rather than the return air grille. Keeping temperatures closer to 27°C can help reduce chiller energy use, provided hot and cold air are properly separated through containment.

Can you source data center cooling equipment from India?

Yes, for fabricated parts such as skids, frames, tanks, piping, and heat exchangers. The partner facility must hold the right standards and ship documentation. Wootz.work routes these jobs across a network of 400+ qualified partners and includes a free design review with the first quote.

‍