Data center cooling systems collect the heat that servers, storage and network equipment turn their electricity into and carry it outdoors, through air handlers, chilled water plants, heat rejection equipment and, increasingly, liquid loops that run straight to the processors. Paul Industries installs the mechanical side nationwide: chilled and condenser water piping, chillers and pumps, coolant distribution units piped into the hall, flushing, pressure testing, commissioning support and service. This page explains how each part of the data center cooling infrastructure works and what installation quality decides.
How are data centers cooled?
Every watt delivered to IT equipment becomes heat, so a data center is cooled by moving that heat through a chain of transfers until it reaches outdoor air or water. The chain has three links. Capture happens at the equipment, in room air or in a liquid touching the chips. Transport carries the heat away, usually in chilled water, a glycol loop or a refrigerant circuit. Rejection dumps it outdoors through chillers with cooling towers, dry coolers or air-cooled condensers. A data center cooling diagram drawn at any scale shows the same three links, and data center thermal management is the discipline of keeping each one sized, controlled and redundant so that inlet temperatures stay steady while the IT load moves.
The scale is simple arithmetic. One kilowatt of IT load produces about 3,412 Btu per hour, and one ton of refrigeration is 12,000 Btu per hour, so each megawatt needs roughly 284 tons of cooling before fans, power losses or redundancy. Moving that heat in air takes a lot of it: at a 20 degree Fahrenheit rise across the servers, each kilowatt needs about 158 cubic feet per minute of airflow. Water carries several thousand times more heat per unit volume than air, which is why liquid takes over as racks get denser.
Types of data center cooling systems
The main approaches differ in where the heat is captured and how far air has to carry it. Most halls combine two or more, and the data center cooling technologies chosen for a site follow from rack density, climate, water availability and how much of the building already exists.
| System | How heat is captured | Where it fits | Mechanical scope we install |
|---|---|---|---|
| CRAC units | Room air across a refrigerant coil inside the unit | Smaller rooms and legacy halls; no central plant needed | Refrigerant or condenser connections, condensate drains, humidifier water |
| CRAH units and fan walls | Room air across a chilled water coil | Most enterprise and colocation halls | Chilled water branches, isolation and control valves, drains, leak detection points |
| In-row and rear-door coils | Air cooled within a few feet of the rack | Dense rows inside a room-cooled hall | Chilled water or secondary loop distribution to each row, hoses and valves |
| Direct to chip cold plates | Liquid through cold plates on processors and accelerators | High-density AI and high-performance computing | CDU setting, facility and technical loop piping, manifolds, flushing and fill |
| Immersion tanks | Servers submerged in dielectric fluid | Specialist and very high-density deployments | Facility water to each tank heat exchanger, drains, fluid handling provisions |
Data center cooling strategies are really choices about where to stop using air. A small server room cooling system can be a pair of dedicated split or packaged units; a large hall is almost always served by a central chilled water plant. A comfort-grade air conditioner for data center use falls short because it removes moisture as much as heat, is not built for continuous duty, and lacks the controls and redundancy IT equipment needs. At the dense end, data center water cooling systems deliver water or a water and glycol mix to the rack itself, and air handles only what remains.
Data center cooling system components
Whatever the architecture, the same families of equipment appear on the drawings, and each one has installation details that decide whether it performs.
| Component | Job in the system | Installation details that matter |
|---|---|---|
| Chillers | Remove heat from the chilled water loop by refrigeration | Rigging and setting, vibration isolation, evaporator and condenser piping with service clearances, flow proving |
| Cooling towers, dry coolers, condensers | Reject heat to outdoor air | Structural support, condenser water or glycol piping, freeze protection, drains |
| Pumps | Move chilled, condenser and technical loop water | Alignment, suction piping with straight runs, strainers, check valves, variable speed drives coordinated with controls |
| Plate heat exchangers | Free cooling and loop separation | Isolation valves for cleaning, strainers upstream, access to open the plate pack |
| CRAH units, fan walls, in-row coils | Transfer heat from air to water | Control valve sizing, condensate drains, leak detection, balancing |
| Coolant distribution units | Separate and condition the liquid cooling loop | Clean secondary piping, compatible wetted materials, filtration, fill and venting |
| Expansion tanks, air separators, makeup | Hold pressure and remove air from closed loops | Correct connection point, sizing for the full loop volume, controlled makeup |
Data center cooling equipment is bought from manufacturers who rate it under standard conditions. Chillers, for example, are rated under AHRI Standard 550/590 at 44 degrees Fahrenheit leaving chilled water and 85 degrees entering condenser water, and a data center plant often runs far from that point. The data center cooling units in the hall, the data center cooling pumps in the plant and every data center heat exchanger between them must be selected for the temperatures the site will actually run.
CRAH vs CRAC, fan walls and containment
Air cooling for data centers still carries most of the installed load: cold air is delivered to the server inlets, warmed by the equipment and returned to a coil. A computer room air handler, or CRAH, blows room air across a chilled water coil and has no compressor of its own. A computer room air conditioner, or CRAC, carries its own refrigeration circuit and rejects heat through a condenser or a water or glycol loop. The classic CRAC unit data center layout puts units around the perimeter blowing into a raised floor plenum, with perforated tiles in the cold aisles.
Larger halls now use data center air handlers mounted outside the white space, often as a data center fan wall: an array of plenum fans and a chilled water coil forming one wall of the room, delivering air at low velocity without a raised floor. The water stays out of the IT space, which simplifies leak control. Close-coupled options, in-row coolers between racks and rear-door coils on the cabinets, shorten the air path further, and in rack cooling with an enclosed cabinet and its own coil is used for isolated dense racks.
Data center airflow management decides how much of that equipment is needed. Hot or cold aisle containment, blanking panels and sealed cable openings keep supply air from mixing with exhaust, which lets supply temperature rise and fans slow down. That is where data center HVAC differs most from comfort air conditioning: the target is every server inlet, not a room full of people.
Data center chillers and the chilled water plant
The chilled water loop in a data center is closed and carries heat from CRAH coils, fan walls, in-row coils and CDUs back to the chillers. The machines are usually water-cooled centrifugal or screw machines on large campuses, rejecting heat through towers, and air-cooled screw or scroll units where water is limited or the site is small. A data center chiller plant groups the chillers with primary or variable primary pumps, a heat rejection system, free cooling heat exchangers, an expansion and air removal system, and controls that stage everything to the load. The general principles of loops, pumping and piping are covered on our chilled water systems page; the data center version differs in temperature, redundancy and continuity.
The data center chilled water temperature is the biggest single design lever. Comfort plants traditionally supply about 44 degrees Fahrenheit. Because IT equipment can accept inlet air anywhere in ASHRAE’s recommended range, data center plants commonly supply chilled water well above that, with supply temperatures of 18 to 20 degrees Celsius or more used to gain efficiency and free cooling hours. Higher water temperatures need CRAH coils selected for them, so the decision comes before equipment is ordered.
Heat leaves the plant through condenser water and data center cooling towers, or through air-cooled condensers and dry coolers. Towers reject heat close to the outdoor wet-bulb temperature and are the most energy-efficient option in most climates; their water treatment is covered on our cooling tower water treatment page and their mechanical upkeep on cooling tower cleaning and maintenance.
What is data center free cooling?
Data center free cooling means using cold outdoor conditions to cool the IT load with the chillers’ compressors off or throttled back. The common form is a waterside economizer: a plate heat exchanger between the condenser water or dry cooler loop and the chilled water loop, so that when outdoor conditions allow, the tower or cooler makes chilled water directly. With 60 degree chilled water and a 7 degree approach, for example, a waterside economizer can carry the load whenever ambient conditions reach about 53 degrees. Partial free cooling, pre-cooling return water ahead of the chillers, extends the hours further. See our plate heat exchanger page for the exchanger itself.
This is why data centers in cold climates run their compressors for only part of the year, and why warmer water temperatures matter so much. A data center dry cooler, a finned coil with fans and no water spray, tracks the outdoor dry-bulb temperature and uses no water but cannot cool as far as a tower. Adiabatic cooling for data centers wets pads or sprays mist ahead of the dry cooler coil during the hottest hours, gaining most of an evaporative cooler’s capacity for a fraction of the year. A direct or indirect evaporative cooling data center design does the same with the air itself, cooling supply air by evaporation instead of a chilled water coil.
Closed loop cooling systems for data centers combine these: a dry cooler or closed-circuit cooler on a glycol loop, a heat exchanger to the building water, and chillers only for trim. Water use is covered on our data center water and piping page; mechanically, a closed glycol loop needs freeze protection measured, not assumed, and outdoor sections need drains, vents and isolation for winter service.
How does liquid cooling work in data centers?
Liquid cooling works by putting a liquid in direct thermal contact with the hottest components, so heat moves into the liquid at the chip and never has to be carried by room air. In the dominant form of liquid cooling for data center halls, a cold plate, a metal block with fine internal channels, sits on each processor or accelerator. Coolant flows through the cold plates, through hoses to a rack manifold, and back to a coolant distribution unit. A direct to chip cooling data center therefore has two water loops: the facility water system from the plant, and the technical cooling system that runs through the servers.
The data center cooling distribution unit, more often called a coolant distribution unit or CDU, sits between them. It contains pumps, a plate heat exchanger that keeps the two loops separate, filtration and controls, and it holds the secondary loop at a supply temperature above the room dew point so nothing condenses on the servers. Liquid-to-liquid CDUs hand the heat to facility water; liquid-to-air CDUs reject it into the room and suit halls with no facility water near the racks. The data center cooling liquid in the technical loop is usually treated water or an inhibited propylene glycol mixture. The Open Compute Project’s PG25 guidelines describe an inhibited fluid of about 25 percent propylene glycol by volume, a concentration at which the fluid is considered bio-static, which matters because cold plate channels are fine enough to foul.
The facility side runs warmer than a traditional plant. ASHRAE’s liquid cooling classes, renamed in 2021 as W17, W27, W32, W40, W45 and W+, are named for the upper limit of facility water supply temperature in degrees Celsius, and the warmer classes are what make dry coolers and heat reuse practical. Every data center cooling hose and quick disconnect is a potential leak point, so racks use dripless couplings and leak detection. Water cooled data center racks also include rear-door heat exchangers, which keep the servers air cooled but capture their exhaust heat in a coil on the cabinet door before it reaches the room.
Liquid immersion cooling for data centers submerges servers in a dielectric fluid, single phase or two phase. Either way the building side is a facility water loop to each tank’s heat exchanger, built to the same cleanliness standard.
How are AI data centers cooled?
AI data centers are cooled mainly by direct liquid cooling, because racks built around current accelerators produce more heat than air can remove. Sources put the practical ceiling for air cooling somewhere between about 30 and 50 kilowatts per rack, and current AI training racks are designed around 100 kilowatts and more with direct liquid cooling as a baseline requirement. Servers still reject some heat to air from memory, storage and power supplies, so an AI hall runs a technical loop, a facility water loop and an air system at once.
Data center power and cooling are planned together for that reason: every megawatt delivered to the racks is a megawatt the plant must reject. Training clusters also run near full load for long periods and can swing quickly, which favors fast, stable pump and valve control, buffer volume, and redundant CDUs so one failure never cuts flow to a rack. Liquid-cooled AI halls are where data center cooling systems change fastest, and where installation cleanliness matters most.
Data center liquid cooling vs air cooling
The comparison is about density, efficiency and what the building has to provide.
| Factor | Air cooling | Direct liquid cooling |
|---|---|---|
| Practical rack density | Up to roughly 30 to 50 kW with good containment | 100 kW and above per rack |
| Supply temperature | Cold air to the inlets, so cooler chilled water | Warm facility water, which extends free cooling and allows dry coolers |
| Fan energy | Large room and server fans | Reduced, though some air cooling remains |
| Piping in the white space | None, or chilled water to perimeter or in-row units | Secondary loop to every rack, manifolds and hoses |
| Water quality demands | Normal closed-loop chilled water treatment | Strict cleanliness and fluid control in the technical loop |
| Retrofit effort | Low in existing halls | Needs CDU space, piping routes, floor loading and leak management |
Most existing facilities end up hybrid, with liquid loops added to specific rows. The installation questions are where the CDUs go, how secondary piping reaches the racks, and how new loops tie into a plant already serving live load. Hybrid data center cooling systems need both sides commissioned together.
Data center cooling requirements and design
Requirements start with the IT load and the inlet conditions the equipment needs. The ASHRAE TC 9.9 thermal guidelines give a recommended range of 18 to 27 degrees Celsius, about 64 to 81 degrees Fahrenheit, at the server inlet for the common equipment classes, with a dew point between minus 9 and 15 degrees Celsius. The temperature of data center supply air can sit anywhere in that band, and running near the top of it is one of the cheapest efficiency gains available, provided containment keeps hot air from recirculating.
Data center cooling design then adds growth phases, partial loads in early years, the climate’s design wet-bulb and dry-bulb temperatures, and energy targets. Data center cooling efficiency is reported through power usage effectiveness, defined in ISO/IEC 30134-2 as total facility energy divided by IT equipment energy, and new facilities are subject to ASHRAE Standard 90.4, the energy standard written specifically for data centers. Those are the main data center cooling standards a mechanical contractor works within, alongside the piping code, usually ASME B31.9 for building services or ASME B31.3 where a process code is specified.
Data center cooling redundancy is described as N, the capacity the load needs, plus spares: N+1 adds one extra chiller, pump or CDU, and 2N duplicates the whole system. The Uptime Institute’s Tier III asks for concurrent maintainability, meaning any component can be taken out for maintenance without shutting down the load. In piping terms, that means looped headers or dual feeds, isolation valves placed so every chiller, pump, coil and CDU can be valved out, and bypasses designed and tested before go-live.
Piping installation, testing and tie-ins
Data center chilled water piping is usually carbon steel, welded or roll-grooved, in the plant and mains, with insulation and vapor retarder wherever the water runs below the dew point. Technical cooling loops are commonly stainless steel or another material on the CDU and cold plate manufacturer’s wetted-materials list, because one incompatible fitting can introduce corrosion into a loop that must stay clean. Our stainless steel piping and process piping pages cover materials and joining in depth.
Installation quality shows up in four places. First, cleanliness: pipe is capped in storage, spools are shop fabricated, and every loop is flushed with equipment bypassed before CDUs or coils see flow. Second, testing: hydrostatic testing to the code, which for ASME B31.9 means at least 1.5 times the design pressure, followed by leak checks on every mechanical joint. Third, valve placement for concurrent maintenance and isolation of each row. Fourth, tie-ins to a live plant are planned as a shutdown or done by hot tap, with a written method and rollback. Chillers, CDUs and pumps are set by riggers and aligned by millwrights; see millwright and rigging services.
Commissioning, maintenance and troubleshooting
Commissioning proves the system works as one: flows balanced, controls sequenced and redundancy tested by failing equipment on purpose under load banks. We support that through commissioning and validation. After handover, data center cooling maintenance keeps those conditions true: strainer and filter changes, glycol and inhibitor testing, pump checks, exchanger cleaning and valve exercising, under a preventive maintenance plan.
Most data center cooling problems trace back to a small set of causes, and the symptom usually points to the cause.
| Symptom | Likely cause | Mechanical fix |
|---|---|---|
| Hot spots at some racks while others are cold | Mixing of supply and exhaust air, missing blanking panels, poor tile placement | Containment, airflow management, balancing |
| Chillers short-cycling at low load | Too little loop volume, oversized plant in early phases | Buffer tank, staging changes, smaller trim chiller |
| Low return water temperature at the plant | Coil bypasses, open three-way valves, overpumping | Valve correction, pump control, coil checks |
| Rising pressure drop across a CDU or exchanger | Fouling or debris from an unflushed loop | Strainer service, exchanger cleaning, loop flushing |
| Glycol or coolant loss | Leaking couplings, relief valves lifting, unmetered makeup | Leak search, makeup metering, relief checks |
| Free cooling hours below design | Chilled water setpoint too low, heat exchanger fouled, controls overriding | Setpoint review, exchanger cleaning, sequence correction |
Data center cooling best practices follow from that table: warm supply temperatures with good containment, generous delta T at the coils, clean loops, metered makeup, and maintenance done one redundant unit at a time. When something does fail, our emergency repair service covers piping, pumps and exchangers.
Data center cooling cost
The cost has two parts: the installed cost of the plant and distribution, and the energy and maintenance to run it for decades. Installed cost is driven by capacity and redundancy, the heat rejection method, the piping and connections in the hall, liquid cooling scope, technical loop materials, and whether work happens in a new shell or a live facility. Operating cost is driven mostly by compressor hours, which is why free cooling, warmer water and containment usually repay their first cost. Prefabricated spools and skid-mounted pump packages shorten field time.
How to choose a data center cooling contractor
Data center cooling companies fall into three groups: manufacturers of chillers, CRAH units, CDUs and heat rejection equipment; engineers who design the system; and contractors who install, start up and maintain it. The best cooling solutions for data centers come from all three working to one set of drawings, temperatures and acceptance tests. When choosing the installer, ask for welding procedures, flushing procedures for liquid loops, pressure test records, live-facility tie-in experience and a documentation package that matches the owner’s requirements. Data center cooling systems are only as reliable as the weakest of those.
What Paul Industries does
Paul Industries installs and services the mechanical systems behind data center cooling: chilled water and condenser water piping, chiller and pump installation, plate heat exchangers and waterside economizers, CDU piping and technical loops, glycol systems, flushing, cleaning and passivation of stainless loops, pressure testing, commissioning support, preventive maintenance and emergency repair. We also provide makeup water treatment equipment, described on our water treatment systems page, and process chiller installation for plants beyond the data hall. We do not manufacture cooling equipment and we are not an electrical contractor. Request a quote for data center cooling installation or service.
Standards referenced: ASHRAE TC 9.9 Thermal Guidelines for Data Processing Environments · ASHRAE Standard 90.4 · ISO/IEC 30134-2 · OCP PG25 Guidelines · AHRI 550/590 · ASME B31.9 · ASME B31.3 · Uptime Institute Tier Standard
Frequently asked questions
What is the best cooling system for a data center?
There is no single best system. Air cooling with containment and chilled water CRAH units or fan walls suits most enterprise and colocation halls; direct to chip liquid cooling is required for AI and other racks above roughly 30 to 50 kW; and the heat rejection choice, towers, dry coolers or a combination, follows from climate and water availability. The best design is the one sized to the real density and run at the warmest water temperatures the equipment allows.
Why do data centers need cooling?
Nearly all of the electricity a server uses becomes heat. Without continuous removal, inlet temperatures climb past what processors, memory and drives tolerate, equipment throttles or shuts down, and component life shortens. Cooling keeps inlet conditions inside the ASHRAE ranges the hardware is built for.
What happens during a data center cooling failure?
Temperatures rise fast, especially in dense halls, because the heat source keeps running. Redundant chillers, pumps and CDUs take over automatically in a well-designed plant, and some sites keep thermal storage or pumps on backup power to ride through a chiller restart. Failure drills during commissioning prove that the switchover actually works.
What is the difference between a CRAC and a CRAH?
A CRAC has its own compressor and refrigerant circuit and rejects heat through a condenser. A CRAH has no compressor; it blows air across a coil fed by chilled water from a central plant. CRAHs dominate larger facilities because the central plant is more efficient and easier to make redundant.
Why must liquid cooling supply temperature stay above the dew point?
Any surface colder than the room dew point condenses moisture, and water forming on cold plates, manifolds or hoses inside a rack is a failure risk. The coolant distribution unit holds the technical loop supply above the room dew point, which is one reason warm-water liquid cooling designs are favored.
What coolant is used in direct liquid cooling?
Most cold plate systems use treated water or an inhibited propylene glycol mixture of about 25 percent, as described in Open Compute Project guidance. The exact fluid, filtration level and wetted materials are set by the cold plate and CDU manufacturers, and the loop must be built and flushed to suit them.
Can liquid cooling be added to an existing air-cooled data center?
Yes, usually row by row. The work is finding space and power for CDUs or using liquid-to-air units, routing secondary piping to the racks, confirming floor loading, and tying into the existing chilled water plant without interrupting live load. Rear-door heat exchangers are a lower-disruption step for moderate densities.
What chilled water temperature do data centers use?
Higher than comfort cooling. Many data center plants supply chilled water well above the traditional 44 degrees Fahrenheit to gain chiller efficiency and free cooling hours, and liquid cooled halls can use facility water in ASHRAE's warmer W classes. The coil and CDU selections must match the chosen temperature.
How much cooling does a data center need?
Start from the IT load: each kilowatt is about 3,412 Btu per hour, so a megawatt of IT load is about 284 tons of cooling. Add the heat from fans, power distribution losses and lighting, then add redundancy, typically N+1 or 2N on major equipment.
How many hours of free cooling can a data center get?
It depends on climate and on how warm the facility water can run. Warmer supply temperatures, which liquid cooling allows, raise the number of hours an economizer can carry the load with the chillers off. Cold, dry sites get the most free cooling hours; hot, humid sites the fewest.
Are rear-door heat exchangers a form of liquid cooling?
They bring water to the rack, but the servers themselves stay air cooled. A coil on the rear door captures the hot exhaust before it enters the room, which raises the density a rack can support without changing the IT hardware.
What piping material is used for data center cooling?
Carbon steel, welded or grooved, is standard for chilled and condenser water mains. Technical cooling loops use stainless steel or other materials approved on the CDU and cold plate wetted-materials lists. Copper and plastics appear in some smaller or specific applications when the design and manufacturers allow it.
How is a new liquid cooling loop cleaned before servers are connected?
The loop is flushed at velocity with the CDU and IT connections bypassed, strainers are checked until they stay clean, stainless sections are cleaned and passivated where specified, and only then is the loop filled with the specified fluid and connected. Water samples and filter checks document the result.
Does Paul Industries supply the cooling equipment?
We procure and install the equipment a project specifies, but we do not manufacture chillers, CRAH units or CDUs. Our role is the mechanical installation: rigging and setting equipment, piping, flushing, testing, commissioning support and maintenance.
How often should the cooling equipment in a data center be maintained?
On a schedule set by the manufacturers and the facility's risk tolerance, typically monthly checks of filters, strainers and fluid condition, quarterly inspections of pumps and valves, and annual heat exchanger and chiller service. Redundancy lets each unit be maintained without taking cooling capacity offline.
Do data centers need a dedicated chiller plant?
Large facilities almost always have one, because the plant must run continuously, at data center temperatures and with redundancy that a shared building plant rarely provides. Small server rooms in mixed-use buildings often tie into the building chilled water with a dedicated backup unit.
Related: Data Centers · Plate Heat Exchanger · Cooling Tower Water Treatment · Process Chiller Installation & Service · Sanitary Heat Exchanger Installation & Service · Chilled Water Systems · Request a quote