The main types of data center cooling systems fall into four broad categories: air-based systems (including CRAC and CRAH units, hot/cold aisle containment, and rear-door heat exchangers), free and economized cooling (airside and waterside economizers, evaporative systems), liquid cooling (direct-to-chip cold plates and single-phase or two-phase immersion), and heat rejection equipment (chillers, cooling towers, and dry coolers). Each suits a different power density range, and the rise of AI and GPU-dense workloads is pushing more facilities toward the liquid end of that spectrum.
Air-Based Cooling Systems
Air cooling is the default in most existing facilities and remains practical for general-purpose compute. Its ceiling is real, though: once rack density climbs past roughly 15 to 20 kW, the physics of air as a heat-transfer medium start working against you. Within that envelope, three types of air-based systems are in common use.
CRAC Units (Computer Room Air Conditioning)
A Computer Room Air Conditioning (CRAC) unit runs a self-contained direct-expansion refrigeration cycle, cooling and circulating air without any connection to a central chilled water plant. That independence makes CRAC units practical for smaller facilities, edge deployments, and sites where adding central chiller infrastructure is not viable. The tradeoff is efficiency: direct-expansion compressors cycle on and off based on demand, which wastes energy at partial load. For a detailed look at how these systems are deployed and sized, the guide to air cooling methods for data centers covers CRAC configurations alongside the full air cooling toolkit.
CRAH Units (Computer Room Air Handlers)
A Computer Room Air Handler (CRAH) is a precision air handler connected to a centralized chilled water plant rather than running its own refrigeration cycle. Because large central chillers are far more efficient than many small compressors, CRAH units are the standard choice for hyperscale and enterprise facilities. Variable-speed fans allow precise airflow control. The chilled water loop can also integrate economizer modes, which lets the facility reduce or eliminate mechanical chilling hours when outdoor conditions allow. Most purpose-built AI data centers use CRAH units for this reason.
Hot/Cold Aisle Containment
Containment is not a cooling unit in itself; it is an architectural discipline that determines how efficiently air cooling actually works. In a hot aisle/cold aisle arrangement, server racks face alternating directions so that cool air intake faces a supplied cold aisle and hot exhaust exits into a contained hot aisle. Physical barriers, curtains, or hard enclosures then prevent the two air streams from mixing. Cold aisle containment boxes in the supply air; hot aisle containment captures exhaust and returns it directly to the CRAH unit. Well-implemented containment typically improves cooling efficiency by 20 to 30 percent compared to an open floor. ASHRAE sets the thermal envelope standards that containment design must satisfy.
Rear-Door Heat Exchangers
A rear-door heat exchanger (RDHx) is a chilled water coil installed in the rear door of a server rack. As hot air exhausts from the servers, it passes through the coil and is cooled before entering the hot aisle. This captures a substantial fraction of rack heat at the source without modifying the servers themselves, which makes RDHx units one of the most practical retrofit paths for facilities facing density increases beyond what their room-level air systems can handle. They are most effective between roughly 10 and 30 kW per rack. The surge in AI data center power demand has made rear-door cooling a common transitional step for operators not yet ready for full liquid cooling infrastructure.
Free and Economized Cooling
Free cooling, more formally called economization, is any mode that uses outdoor temperature conditions to reduce or eliminate mechanical refrigeration. When conditions permit, these methods are the cheapest cooling available. They do not replace active cooling entirely; they reduce the hours that active systems must run.
Airside Economizers
An airside economizer routes outdoor air into the cooling stream when temperature and humidity fall within acceptable bounds. A direct airside system pulls outdoor air straight into the data floor; an indirect version uses a heat exchanger to transfer the cooling effect without mixing the two air streams, protecting hardware from outdoor dust, humidity spikes, and pollutants. Hyperscale facilities in Scandinavia, Ireland, and the northern United States operate on airside free cooling for hundreds of hours annually, which directly suppresses cooling energy costs. The tradeoff is geographical: sites in warm or humid climates simply cannot access enough free cooling hours to make airside economization the primary strategy.
Waterside Economizers
A waterside economizer pre-cools the chilled water loop using outdoor air, through either a cooling tower or a dry heat exchanger, before water reaches the mechanical chiller. When outdoor wet-bulb temperature drops far enough, the economizer can carry the full load and the chiller shuts off entirely. Microsoft, Meta, and Google have deployed waterside economization across large portions of their campus footprints. The annual energy savings scale directly with local climate and the design setpoint temperature chosen for the chilled water loop.
Evaporative and Adiabatic Cooling
Evaporative cooling lowers air or water temperature through the latent heat of evaporation. Adiabatic cooling applies water spray or pads to supply air before it enters the facility, reducing its temperature without mechanical refrigeration. Both approaches perform best in dry climates where there is thermodynamic headroom for evaporation. The central tradeoff is water consumption: facilities achieving outstanding energy efficiency through evaporative cooling may consume significant volumes of water per hour in warm months. Operators in water-stressed regions face a genuine tension between low power consumption and low water use that neither approach resolves cleanly.
Liquid Cooling Systems
Liquid cooling brings a fluid much closer to the heat source, exploiting the fact that water carries heat roughly 3,500 times more effectively than air per unit volume. This gap is why liquid systems handle rack densities that air simply cannot. Current GPU and AI accelerator deployments are the primary driver pushing facilities toward these architectures.
Direct-to-Chip Cold Plate Cooling
In a direct-to-chip (DTC) system, a metal cold plate carrying chilled water or a water-glycol mixture is attached directly to the processor package. Heat conducts from the chip surface into the fluid, which carries it to a coolant distribution unit and then to a facility-level heat exchanger. Server fans and standard air cooling still handle the residual heat from memory, storage, and power components, so DTC is a hybrid rather than a full liquid system. NVIDIA’s H100 and H200 platforms support cold plate configurations, and most purpose-built AI server chassis now ship with DTC-ready designs. Practical density range is roughly 20 to 80 kW per rack, depending on how much of the heat budget the cold plates carry versus residual air.
Single-Phase Immersion Cooling
In single-phase immersion, entire servers are submerged in a dielectric fluid that remains liquid throughout the thermal cycle. The fluid circulates to an external heat exchanger, where it gives up heat before returning to the tank. Because every component is in direct contact with the coolant, heat removal is highly uniform and there is no thermal stratification problem to manage. Capital cost is higher than DTC because it requires purpose-built immersion tanks and a supply chain of compatible hardware, but operational simplicity improves: fans are removed entirely, reducing noise and mechanical failure points. Density range runs from roughly 50 to 100 kW per tank bay.
Two-Phase Immersion Cooling
Two-phase immersion uses a dielectric fluid engineered to boil at low temperatures, typically in the range of 49 to 60 degrees Celsius. As hot components generate heat, the fluid vaporizes on their surfaces, absorbing latent heat far more efficiently than sensible heat transfer alone. The vapor rises to a condenser coil, re-liquefies, and falls back into the bath. This phase-change cycle is thermodynamically efficient and the resulting heat removal rates at component level are exceptional. Two-phase systems can handle the densest AI training racks currently in production, with some deployments exceeding 100 kW per tank bay. Hardware compatibility and fluid cost remain the main adoption barriers.
Heat Rejection Equipment
Every cooling system moves heat from servers to somewhere else. Heat rejection equipment is the final step: it puts that heat into the outdoor environment. The choice of rejection technology affects both operating cost and water consumption, and it couples directly with the rest of the cooling chain.
Chillers
A chiller runs a vapor-compression refrigeration cycle to produce chilled water, typically in the range of 6 to 12 degrees Celsius, which is then distributed to CRAH units or direct liquid cooling systems. Large centrifugal chillers from manufacturers including Carrier, Trane, and Johnson Controls achieve high efficiency coefficients of performance at full load. Partial-load efficiency requires sequencing multiple units carefully, which modern building management systems handle automatically. Chillers represent the largest single equipment capital cost in most mechanical cooling plants.
Cooling Towers
A cooling tower rejects heat from the chiller condenser water loop by evaporating a fraction of the water into outdoor air. Because evaporation exploits latent heat, cooling towers can reject heat even when outdoor air temperature approaches or exceeds the condenser water setpoint, which makes them highly effective in warm climates. The operating costs are water, chemicals for treatment against scaling and biological growth, and power for fans and pumps. Legionella risk management is an ongoing regulatory and operational requirement for any facility running open-circuit evaporative towers.
Dry Coolers
A dry cooler (or fluid cooler in closed-loop configuration) rejects heat through sensible heat exchange with outdoor air, consuming no water at all. This makes them increasingly preferred in water-scarce regions and for higher-temperature liquid cooling loops, where coolant return temperature may be 40 to 50 degrees Celsius rather than the lower temperatures a chiller produces. At ambient temperatures well below the loop setpoint, dry coolers can carry the full rejection load; as outdoor temperature rises, they lose capacity and a mechanical chiller must supplement. The U.S. Department of Energy documents efficiency benchmarks for heat rejection equipment across climate zones through its data center efficiency programmes.
Cooling Type Comparison
| Cooling Type | How It Works | Density Suitability | Main Tradeoff |
|---|---|---|---|
| CRAC Unit | Self-contained DX refrigeration cycle cools and recirculates room air | Up to 10-15 kW per rack | Less energy-efficient at partial load; no central plant integration |
| CRAH Unit | Precision air handler on a central chilled water loop with variable-speed fans | Up to 15-20 kW per rack with containment | Requires central chiller plant; higher upfront infrastructure |
| Hot/Cold Aisle Containment | Physical barriers separate supply and exhaust air streams to prevent mixing | Extends air cooling to upper end of density range | Not a standalone system; depends on CRAC/CRAH quality upstream |
| Rear-Door Heat Exchanger | Chilled water coil in rack rear door absorbs exhaust heat before it enters the hot aisle | 10-30 kW per rack; effective retrofit option | Still relies on facility chilled water plant; servers remain air-cooled internally |
| Airside Economizer | Outdoor air cools data floor directly or via heat exchanger when conditions allow | Low-to-medium density; best in cold climates | Fully climate-dependent; ineffective in warm or humid regions |
| Waterside Economizer | Pre-cools chilled water loop via outdoor air, reducing or eliminating chiller run hours | Any density; supplements mechanical cooling | Requires additional heat exchange equipment; benefit scales with climate |
| Evaporative / Adiabatic | Water evaporation lowers supply air or water temperature without mechanical refrigeration | Low-to-medium density; best in dry climates | High water consumption; unsuitable for water-stressed regions |
| Direct-to-Chip (Cold Plate) | Chilled water cold plate attaches directly to processor; residual heat handled by air | 20-80 kW per rack; dominant AI accelerator choice | Hybrid system; memory and storage still air-cooled; requires CDU and facility piping |
| Single-Phase Immersion | Servers submerged in circulating dielectric fluid; fluid to external heat exchanger | 50-100 kW per tank bay | Purpose-built tanks required; hardware compatibility narrows procurement options |
| Two-Phase Immersion | Low-boiling-point dielectric vaporizes on components, condenses on cooled coils, recirculates | 100+ kW per tank bay; highest density available | Higher fluid cost; specialist maintenance; limited hardware vendor ecosystem |
| Chiller | Vapor-compression cycle produces chilled water for distribution | Any density; core of most mechanical plants | Largest capital cost in the mechanical plant; partial-load efficiency requires sequencing |
| Cooling Tower | Evaporative heat rejection from condenser water loop to outdoor air | Any density; pairs with chillers | Water consumption and treatment cost; legionella management overhead |
| Dry Cooler | Sensible heat exchange with outdoor air; no water evaporation | Any density; pairs with higher-temperature liquid cooling loops | Less effective in warm climates; needs chiller supplementation when ambient is high |
Why AI and GPU Density Is Driving the Shift to Liquid
Air-based systems built to serve 5 to 10 kW racks are not architected for what modern GPU clusters demand. A rack of NVIDIA H100 servers can draw over 80 kW. At those levels, air simply does not carry enough heat per unit volume to keep chip surfaces within operating temperature, regardless of containment quality or fan power invested. The density curve has moved faster than most facilities anticipated, and the gap between what air can handle and what AI accelerators require has grown wide enough that it cannot be bridged incrementally.
The consequence is a technology mix that was unusual five years ago and is now standard practice in new builds. Facilities designed around AI training workloads specify direct-to-chip cooling as the baseline, with immersion reserved for the most extreme rack configurations. The heat rejection side of the equation changes too: higher coolant return temperatures from liquid systems pair better with dry coolers than with traditional chilled-water plants optimized for low supply temperatures. The Uptime Institute has tracked rising liquid cooling adoption in its annual global data center surveys, with the trend accelerating each year since 2022 in parallel with GPU infrastructure investment.
Air cooling is not disappearing. General-purpose compute, storage, and networking infrastructure still runs at densities where air is cost-effective and operationally familiar territory. The shift is happening at the high end, where the physics make liquid mandatory rather than optional.
Frequently Asked Questions
What are the main types of data center cooling?
The main types of data center cooling systems are air-based cooling (CRAC units, CRAH units, hot/cold aisle containment, rear-door heat exchangers), free and economized cooling (airside economizers, waterside economizers, evaporative systems), liquid cooling (direct-to-chip cold plates, single-phase immersion, two-phase immersion), and heat rejection equipment (chillers, cooling towers, dry coolers).
What is the difference between CRAC and CRAH?
A CRAC (Computer Room Air Conditioning) unit runs a self-contained refrigeration cycle with no central chiller plant required, which makes it simpler to deploy but less efficient at partial load. A CRAH (Computer Room Air Handler) connects to a centralized chilled water plant, allowing larger chillers to operate more efficiently and enabling economizer integration. Hyperscale facilities standardize on CRAH; smaller and edge sites use CRAC.
What is free cooling in a data center?
Free cooling, formally called economization, uses outdoor temperature conditions to reduce or eliminate mechanical refrigeration. Airside economizers route outside air into the cooling stream; waterside economizers pre-cool the chilled water loop, cutting chiller run hours or shutting the chiller off entirely. The energy savings are real, but the benefit scales directly with how many hours per year local climate falls below the required setpoint.
Which cooling type is best for AI servers?
Direct-to-chip cold plate cooling is the standard for AI servers drawing 20 to 80 kW per rack, covering most NVIDIA H100 and H200 deployments. For racks above 80 to 100 kW, two-phase immersion offers the highest heat removal capacity available today. Air cooling alone cannot sustain dense GPU workloads without throttling, no matter how good the containment.
What is a rear-door heat exchanger?
A rear-door heat exchanger (RDHx) is a chilled water coil mounted in the rear door of a server rack. Server exhaust passes through the coil and is cooled before it enters the hot aisle, capturing a large fraction of rack heat at the source. Effective from roughly 10 to 30 kW per rack, RDHx units require no server modifications and are one of the lower-disruption retrofits for air-cooled facilities.
