How Are Data Centres Cooled? CRAC, Containment and Liquid Cooling Explained
How Are Data Centres Cooled? CRAC, Containment and Liquid Cooling Explained
CRAC & Cooling · Data Centre Guides
Data centres are cooled by continuously removing the heat servers produce and holding the room at roughly 18 to 27 degrees Celsius, the ASHRAE TC 9.9 recommended range. Most Australian facilities do this with CRAC or CRAH units feeding air through a hot aisle and cold aisle containment layout, while high-density AI racks are increasingly moving to liquid cooling. Indigi Power and Cooling designs, supplies and maintains these systems across Brisbane, Sydney, Melbourne and regional Australia.
Every server, switch and storage array in a data centre turns electricity into heat. A single modern rack can dissipate anywhere from 5 kilowatts to well over 50 kilowatts, and unlike an office, a data hall runs that load 24 hours a day, every day of the year. If the heat is not removed as fast as it is generated, temperatures climb within minutes, hardware throttles to protect itself, and eventually components fail. Cooling is not a comfort system in a data centre. It is critical infrastructure, on the same tier as the power feed and the UPS.
This guide explains how data centres actually stay cool, from the room-level air handling units down to the rack, why containment matters, what temperature the room should sit at, and how the rise of AI compute is pushing the industry toward liquid cooling. It is written for facility managers, IT and infrastructure buyers, and anyone specifying or maintaining a server room in Australian conditions.
The short answer: how data centre cooling works
At the simplest level, data centre cooling is a heat-transport chain. Heat leaves the chips, moves into the air (or a liquid) around them, is carried to a cooling unit that rejects it, and is finally dumped outside the building. The job of a cooling design is to make every link in that chain reliable and efficient. Here is the short version:
- Capture the heat at the rack. Cold air is delivered to the front of the servers and hot exhaust air is collected at the back, kept strictly separate so the two never mix.
- Move it with precision cooling units. CRAC or CRAH units condition the air, controlling temperature and humidity to a tight set point rather than just chilling the room.
- Reject the heat outside. A refrigerant circuit, chilled water loop or condenser water loop transports the heat to external plant such as a condenser, dry cooler or cooling tower.
- Hold a stable set point. The whole system runs continuously to a controlled band, typically 18 to 27 degrees Celsius, with redundancy so a single unit failure does not take the room down.
The rest of this guide walks through each stage in detail.
CRAC and CRAH units: the workhorses of precision cooling
The core piece of equipment in most data centres is the precision cooling unit. You will see two closely related terms:
CRAC (Computer Room Air Conditioner) units contain their own refrigerant compressor and cooling coil, similar in principle to a very precise, industrial split system. They are common in smaller data centres, server rooms and edge sites where a dedicated chilled water plant would be overkill.
CRAH (Computer Room Air Handler) units have no compressor. Instead they pass room air over a coil fed with chilled water from a central plant. In larger data centres a CRAH-based chilled water system is usually more efficient at scale, because one central chiller plant serves many air handlers.
Both types do the same fundamental job: they draw in hot air from the room, pass it across a cold coil, and blow conditioned air back out. What makes them "precision" cooling rather than ordinary air conditioning is the level of control. A CRAC or CRAH unit holds temperature and, critically, humidity within a narrow band. Too much humidity risks condensation and corrosion; too little invites static discharge that can damage electronics. Comfort air conditioning does not manage this, which is one reason ordinary split systems are a poor substitute in a real server room.
CRAC versus CRAH precision cooling units at a glance
| Feature | CRAC unit | CRAH unit |
|---|---|---|
| Cooling source | Onboard refrigerant compressor (DX) | Chilled water from central plant |
| Best suited to | Server rooms, edge and small sites | Large data centres at scale |
| Efficiency at scale | Good for smaller loads | Higher, shared central plant |
| Controls | Temperature and humidity | Temperature and humidity |
There is also a distinction in how the heat is finally rejected outdoors, which gives rise to the three common topologies: DX (direct expansion), CHW (chilled water) and CDW (condenser water). Choosing the right one depends on the size of the load, the site, the climate and the redundancy you need. We cover that decision in detail on the CRAC topology selector.
Hot aisle and cold aisle containment: capturing heat at the rack
Having a powerful cooling unit is only half the story. If cold supply air and hot exhaust air are allowed to mix in the room, you waste a large share of your cooling capacity and create hot spots that no amount of extra plant will fully fix. This is the problem containment solves, and it is why the LinkedIn post that brought you here highlighted "heat is captured right at the rack".
The standard layout arranges server racks in alternating rows. All the racks in one aisle face each other so their cold air intakes draw from a shared cold aisle. The racks in the next aisle sit back to back, exhausting hot air into a shared hot aisle. Cold air is delivered to the cold aisles, usually up through a raised floor or via overhead ducting, and hot air is collected from the hot aisles and returned to the CRAC or CRAH units.
Containment takes this further by physically sealing off the aisles with panels, doors and roof caps so the cold and hot air genuinely cannot mix. There are two approaches:
- Cold aisle containment encloses the cold aisle, keeping the chilled supply air trapped in front of the servers.
- Hot aisle containment encloses the hot aisle, capturing the exhaust and channelling it straight back to the cooling units.
Either way, the benefit is the same: the cooling system deals with air that is genuinely hot or genuinely cold, not a lukewarm mix. That lets units run at a higher, more efficient return temperature, lets you raise the supply set point, and often allows fewer units to do the same work. Good containment is one of the highest-return upgrades available in an older data hall, and it is frequently retrofitted to rooms that were built before it became standard.
What temperature should a data centre be? The ASHRAE range
A common misconception is that data centres should be kept cold, like a refrigerator. In reality, running too cold simply wastes energy. The widely used industry guidance comes from ASHRAE Technical Committee 9.9, which sets a recommended inlet air temperature range of 18 to 27 degrees Celsius measured at the front of the equipment, with an allowable range that extends higher for many modern devices.
Holding the room toward the warmer end of that band, when the equipment allows it, reduces the work the cooling plant has to do and improves overall efficiency, often measured as PUE (Power Usage Effectiveness). The catch is that a warmer set point leaves less thermal buffer, so it depends on reliable, well-maintained cooling and good containment. Humidity matters just as much as temperature: too damp and you risk condensation and corrosion, too dry and static becomes a hazard. Precision cooling units manage both at once, which is exactly why they, rather than comfort air conditioning, belong in a server room.
In Australian conditions the ambient climate does real work here. Through a Queensland or Northern Territory summer, external heat-rejection plant works harder and the margin for a cooling fault shrinks, which makes redundancy and preventative maintenance non-negotiable. You can size the load for your own room with our server room heat load calculator.
Why AI density is pushing sites toward liquid cooling
For decades, air was enough. Racks drew a few kilowatts, and CRAC units with good containment kept everything within range. AI has changed the maths. Dense GPU servers used for training and inference can push a single rack past 50, 80 or even 100 kilowatts, concentrations of heat that air struggles to remove no matter how well the aisles are contained. Air simply cannot carry heat away from that surface fast enough.
The response is liquid cooling, which uses fluid rather than air to carry heat directly away from the hottest components. Liquid holds far more heat per unit volume than air, so it can cope with densities that would overwhelm an air-cooled design. The main approaches are:
- Direct-to-chip (cold plate) cooling brings coolant through cold plates mounted directly onto the CPUs and GPUs, removing heat at its source. It is currently the most common route for high-density AI racks because it can be added to otherwise conventional servers.
- Rear-door heat exchangers replace the back door of a rack with a water-cooled coil that captures hot exhaust as it leaves, often as a bridge between pure air cooling and full liquid cooling.
- Immersion cooling submerges entire servers in a non-conductive dielectric fluid, the most extreme option, capable of the highest densities but requiring purpose-built infrastructure.
In practice most facilities will run hybrid designs for years to come: liquid cooling for the dense AI rows, and conventional CRAC or CRAH air cooling with containment for everything else. The precision cooling fundamentals in this guide do not disappear with AI; they sit alongside liquid loops and, in the case of chilled water plant, often share the same infrastructure.
Redundancy: why cooling is designed to never fully stop
Because a cooling failure can take a data hall offline as surely as a power failure, cooling systems are built with redundancy. You will see configurations described as N+1, N+2 or 2N, meaning there is at least one spare unit (or a fully duplicated system) beyond the minimum needed to carry the load. If a CRAC unit trips or is taken out for service, the remaining units pick up the work without the room temperature drifting out of range.
Redundancy only delivers if the equipment is maintained. Filters clog, compressors and fans wear, refrigerant charge drifts, and chilled water valves stick. A unit that has quietly lost part of its capacity is a redundancy you no longer have. Scheduled preventative maintenance, aligned with standards such as AS/NZS 3666 for air-handling hygiene, is what keeps the designed redundancy real rather than notional.
Data centre cooling across Australia
Indigi Power and Cooling designs, installs and maintains precision cooling and containment for data centres, server rooms and switchrooms nationwide.
- Brisbane and QLD — Brisbane CBD, Woolloongabba, Eight Mile Plains, Port of Brisbane, Gold Coast, Sunshine Coast, Ipswich, Townsville, Cairns. Cooling sized for humid, high-heat Queensland summers.
- Sydney and NSW — Sydney CBD, Parramatta, North Ryde, Macquarie Park, Western Sydney, Newcastle, Wollongong, Canberra (ACT). Servicing the country's densest data centre corridor.
- Melbourne and VIC — Melbourne CBD, Port Melbourne, Docklands, Dandenong, Tullamarine, Geelong, Ballarat, regional Victoria. Precision cooling and containment retrofits.
- Perth and WA — Perth CBD, Fremantle, Kalgoorlie, Pilbara, regional WA. Served from our national hubs.
- Adelaide and SA — Adelaide CBD, Port Adelaide, Salisbury, Mount Gambier, Whyalla, regional SA. Served from our national hubs.
- Darwin and NT — Darwin CBD, Palmerston, Katherine, Alice Springs, remote NT sites. Cooling built for extreme tropical heat loads.
Western Australia, Tasmania and Pacific Islands: we support cooling design, installation and scheduled maintenance for remote and island sites where local specialists are scarce. Contact us to discuss scheduling.
Frequently asked questions
How are data centres cooled?
Data centres are cooled by continuously removing server heat and holding the room at roughly 18 to 27 degrees Celsius. Cold air is delivered to the front of the racks and hot exhaust is collected at the back, kept separate by hot aisle and cold aisle containment. CRAC or CRAH precision cooling units condition that air, and the captured heat is rejected outside via a refrigerant, chilled water or condenser water circuit. High-density AI racks increasingly add liquid cooling.
What temperature should a data centre be?
The widely used ASHRAE TC 9.9 recommended range is 18 to 27 degrees Celsius, measured at the air inlet to the equipment, with allowable ranges that extend higher for much modern hardware. Running toward the warmer end of that band saves energy, provided the cooling and containment are reliable enough to keep the margin safe.
Do data centres use water for cooling?
Many do. Chilled water and condenser water systems circulate water to carry heat from CRAH units or cooling towers, and liquid cooling brings water or coolant close to the chips themselves. Smaller sites often use direct expansion (DX) CRAC units that use refrigerant rather than water, so it depends on the size and design of the facility.
What is the difference between a CRAC and a CRAH unit?
A CRAC unit has its own refrigerant compressor and cools air directly, suiting server rooms and smaller sites. A CRAH unit has no compressor and instead uses chilled water from a central plant, which is usually more efficient in large data centres. Both control temperature and humidity precisely.
Why are AI data centres switching to liquid cooling?
AI training and inference use dense GPU servers that can push a single rack past 50 to 100 kilowatts. Air cannot carry that much concentrated heat away fast enough, so operators turn to liquid cooling, direct-to-chip cold plates, rear-door heat exchangers or immersion, which move far more heat per unit volume than air. Most sites run hybrid designs, with liquid for dense AI rows and air cooling elsewhere.
What is hot aisle and cold aisle containment?
It is a rack layout that keeps cold supply air and hot exhaust air physically separated. Racks are arranged so intakes face a shared cold aisle and exhausts face a shared hot aisle, then panels and doors seal the aisles so the two air streams cannot mix. This lets cooling units run more efficiently and prevents hot spots.
Related services and guides
- Server room cooling design and installation
- CRAC unit maintenance services
- CRAC installation and commissioning
- DX, CHW or CDW: CRAC topology selector
- Server room heat load calculator
- Server room UPS and cooling
Planning, upgrading or maintaining a data centre cooling system?
Tell us your rack density, room size and location, and we will scope the right CRAC, CRAH, containment or liquid cooling solution. Indigi Power and Cooling is an Indigenous and Veteran-owned specialist, Supply Nation and ICN Gateway registered, serving Brisbane, Sydney, Melbourne and regional Australia. Contact Indigi Power and Cooling.