Modern data center with air and liquid cooling systems
A data center cooling system should not be selected from a product catalog alone. The first question is where the heat must be captured: in the room, at the row, at the rear of the rack, directly on the processor, or around the entire server. That decision affects rack density, piping, controls, pump sizing, maintenance procedures, and the amount of air cooling that remains necessary.
Traditional air cooling still works well for many enterprise and colocation environments. However, AI training clusters and other high-performance computing workloads can concentrate heat in a much smaller footprint. In these facilities, the practical answer is often a hybrid architecture rather than a single technology.
This guide explains the main types of data center cooling systems and the conditions that should drive the selection.
Puntos clave
- Room-level air cooling remains practical for conventional rack densities and predictable IT loads.
- Rear-door heat exchangers and close-coupled systems increase capacity without immediately redesigning the entire white space.
- Direct-to-chip liquid cooling captures heat at CPUs, GPUs, and other high-heat components, but residual air cooling is usually still required.
- Immersion cooling removes heat from complete servers placed in dielectric fluid and requires a different service model.
- Chilled-water pumps, secondary-loop pumps, controls, and redundancy should be designed as part of the cooling architecture, not added after the IT equipment is selected.
Start With the Heat-Removal Path
Cooling systems are easier to compare when they are grouped by the point of heat capture.
Cooling approach | Primary heat-capture location | Typical infrastructure impact | Best fit |
|---|---|---|---|
Room-level air cooling | Data hall air | Low to moderate | Conventional enterprise and colocation racks |
Close-coupled cooling | Row or rack | Moderado | Localized hot spots and retrofit projects |
Rear-door heat exchanger | Rack exhaust | Moderado | Higher-density racks where rear access and water distribution are feasible |
Direct-to-chip liquid cooling | CPU, GPU, and accelerator cold plates | Moderate to high | AI, HPC, and dense compute clusters |
Immersion cooling | Entire server chassis | High | Purpose-designed high-density environments |
Hybrid cooling | Multiple locations | Project-specific | Mixed-density facilities and phased upgrades |
The heat-capture point is only part of the design. The rejected heat must still move through heat exchangers, coolant distribution units, pumps, dry coolers, chillers, cooling towers, or other heat-rejection equipment. The best system is the one that connects these layers without creating avoidable operating complexity.
1. Room-Level Air Cooling
Room-level air cooling is the established baseline for data centers. Computer room air handlers (CRAHs) typically use chilled water, while computer room air conditioners (CRACs) commonly use a refrigerant-based cooling cycle. Both circulate conditioned air through the data hall.
The airflow strategy matters as much as the cooling unit. Hot-aisle or cold-aisle containment reduces mixing between supply and exhaust air. Blank panels, sealed cable openings, and controlled fan speeds also improve performance.
Where climate and water conditions allow, economizers or evaporative systems can reduce compressor runtime.
Air cooling is familiar, serviceable, and compatible with standard IT equipment. Its limitation appears when heat density rises faster than airflow can be delivered efficiently. Increasing fan speed indefinitely raises energy use and can still leave hot spots around dense GPU racks.
2. Close-Coupled Cooling
Close-coupled cooling moves the cooling capacity nearer to the heat source. Common formats include in-row units, overhead units, and rack-adjacent modules. Shorter airflow paths reduce recirculation risk and make it easier to target a dense zone without overcooling the rest of the room.
This approach is useful when only part of the data hall has elevated heat load. It can extend the life of an existing facility while avoiding an immediate conversion to liquid-cooled IT equipment.
Close-coupled systems still depend on careful hydraulic and airflow design. A new in-row unit may require chilled-water branch piping, isolation valves, balancing, condensate management, and controls integration. The design team should confirm whether the existing chilled-water pumps can handle the revised flow and head requirements.
3. Rear-Door Heat Exchangers
A rear-door heat exchanger (RDHx) replaces or supplements the rear door of an IT rack. Server exhaust air passes through a liquid-cooled coil before returning to the data hall. Passive designs depend on server fans.
Active designs add fans to increase airflow control.
RDHx systems are often a practical retrofit bridge because they capture heat at the rack without requiring cold plates inside every server. They can reduce the burden on room-level cooling and support localized density increases.
The tradeoff is operational. Water or another coolant must be distributed near the racks, leak management must be planned, and rear access clearances must remain workable. Coil pressure drop and branch balancing also affect pump selection.
4. Direct-to-Chip Liquid Cooling
Direct-to-chip (D2C) liquid cooling places cold plates on high-heat components such as CPUs, GPUs, and accelerators. Coolant absorbs heat near the source and transfers it through a secondary loop. A coolant distribution unit (CDU) typically separates the technology cooling system from the facility water system and provides pumping, heat exchange, monitoring, and control.
Vertiv’s CoolChip CDU product family supports direct-to-chip and rear-door cooling applications for high-density computing. The ASHRAE liquid cooling guidance is also useful when defining temperature, water-quality, and facility-interface requirements.
D2C cooling can remove a large share of server heat more efficiently than air alone. It does not automatically eliminate fans or room cooling. Memory, power supplies, storage devices, and other components may still reject heat into the air unless the server platform is designed otherwise.
When designing the liquid loop, specify:
- Required coolant flow for each rack and branch
- Supply and return temperature targets
- CDU heat-exchanger approach temperature
- Pump head across cold plates, hoses, manifolds, filters, and valves
- Redundancy strategy for pumps and CDUs
- Water-quality, filtration, and material-compatibility requirements
- Leak detection and isolation procedures
5. Immersion Cooling
Immersion cooling places servers in a tank filled with dielectric fluid. The fluid absorbs heat from the server components directly. Single-phase systems circulate liquid fluid through a heat exchanger.
Two-phase systems use a fluid that boils at a controlled temperature and condenses as heat is rejected.
Immersion can support high heat loads with less dependence on server fans. It also changes the physical and operating model of the data center. Server service procedures, rack layout, lifting equipment, fluid handling, hardware compatibility, and technician training all need review.
Submer y LiquidStack provide examples of commercial immersion-cooling platforms. These systems are best evaluated as complete operating environments rather than as replacements for individual air-conditioning units.
Where Free Cooling, Evaporative Cooling, and Chillers Fit
Not every cooling term describes the same layer of the system. Direct-to-chip cooling and immersion cooling describe how heat is collected from IT equipment. Chillers, dry coolers, cooling towers, and economizers describe how heat is rejected from the facility.
A chilled-water plant may serve CRAHs, in-row units, CDUs, or rear-door heat exchangers. In cooler climates, a water-side economizer or dry cooler may reject heat without running compressors for part of the year. Evaporative or adiabatic equipment can reduce energy use but introduces water-consumption and water-treatment considerations.
This distinction matters during procurement. A project can use liquid cooling at the rack while still depending on a conventional chilled-water plant outside the data hall.
How to Select the Right Architecture
Start with measured and forecast rack heat loads, not an assumed average for the entire building. Map the racks that will host AI accelerators, the expected growth timeline, and the redundancy requirement for each cooling zone.
Then review the facility constraints:
- Confirm available electrical capacity, chilled-water capacity, and heat-rejection capacity.
- Identify whether liquid piping can be safely routed into the white space.
- Check water temperatures and flow rates required by the IT equipment or CDU vendor.
- Calculate pressure drop through the complete hydraulic circuit.
- Define maintenance access, isolation, filtration, and leak-response procedures.
- Decide which residual air loads remain after liquid cooling is installed.
For a mixed environment, a hybrid design is often the most practical solution. Conventional racks can remain on room-level air cooling while high-density AI racks use D2C cooling and CDUs. RDHx units can be used in selected retrofit zones.
This staged approach avoids forcing every rack into the same operating model.
Pump Considerations for Liquid-Cooled Data Centers
Pumps are the connection between cooling strategy and delivered performance. Facility-water pumps may feed chilled-water branches or CDUs. Secondary pumps may circulate coolant through cold plates, manifolds, and heat exchangers.
Their duty points should be calculated from actual flow demand and total dynamic head.
For stable operation, review pump redundancy, variable-speed control, minimum-flow limits, NPSH margin, seal and material compatibility, filtration, and service access. Variable-frequency drives can help match flow to changing IT loads, but the control sequence should prevent low-flow conditions in critical branches.
Preguntas frecuentes
What is the most common type of data center cooling system?
Room-level air cooling using CRAC or CRAH units remains common because it supports standard server hardware and established maintenance practices. Higher-density zones increasingly use close-coupled or liquid-cooling systems alongside it.
Is liquid cooling always better than air cooling?
No. Liquid cooling is valuable when heat density, energy efficiency, or footprint constraints justify the additional piping and controls. Air cooling may still be the simpler and more economical choice for conventional racks.
Does direct-to-chip cooling eliminate air conditioning?
Usually not. Cold plates target the hottest components, while other server components can still reject heat into the air. The remaining air-cooling requirement should be calculated for the selected hardware.
What is the difference between a CDU and a chiller?
A CDU transfers heat between a technology cooling loop and a facility water loop while controlling coolant flow and conditions. A chiller removes heat from water through a refrigeration cycle. Some designs use CDUs without relying on compressor-based chilling year-round.
Can an existing data center be upgraded for liquid cooling?
Yes, but the retrofit must be evaluated carefully. Review piping routes, floor loading, water temperatures, heat-rejection capacity, pump duty, controls, redundancy, and maintenance procedures before selecting equipment.
Conclusión
The main types of data center cooling systems are not mutually exclusive. Air cooling, close-coupled cooling, rear-door heat exchangers, direct-to-chip loops, and immersion platforms solve different heat-density and operational problems.
The best design starts with the heat-removal path, rack-level load forecast, and facility constraints. Once the architecture is defined, size pumps, CDUs, piping, and controls as one hydraulic system. That sequence produces a cooling strategy that can scale with AI workloads without adding unnecessary complexity to every rack.