Data Center Cooling Systems: Air, Liquid and Immersion Compared

For twenty years the hard question in a data center was electrical: how many megawatts could reach the site, and how they would be backed up. Cooling was answered with more air handlers and a bigger chiller. Accelerated computing has ended that. Rack density has passed the point where moving air can carry the heat away, and the thermal design now sets the shape of the building, the floor structure, the water strategy and the electrical layout rather than following them.

This article compares air cooling, the close coupled technologies that extend it, direct-to-chip liquid, immersion and hybrid halls, with the metrics that judge them and the California code layer that catches owners out. Budlong’s mission critical practice works on both sides of this problem, and our MEP engineering services cover the coordination a density change sets off, as our expertise overview shows.

1. What Data Center Cooling Is, and Why It Became the Constraint

Data center cooling moves waste heat from a semiconductor die to the outside air, continuously and within a temperature envelope the equipment manufacturer will warrant. Every watt delivered to a server leaves as heat, so a 10 MW IT load is a 10 MW heat load rejected without interruption.

What changed is density. The U.S. Department of Energy reported that data centers consumed about 176 TWh in 2023, roughly 4.4 percent of U.S. electricity, up from 58 TWh in 2014, with a projected 325 to 580 TWh by 2028. That growth arrived not as more racks but as far more power in the same cabinet footprint.

The Uptime Institute global survey for 2025 puts average modal rack density near 9 kW, up from 8.3 kW a year earlier. The spread matters more: around 82 percent of facilities report a maximum below 30 kW per rack, 9 percent sit between 30 and 49 kW and 9 percent exceed 50 kW. Mainstream racks have barely moved; the top of the distribution has moved a very long way.

2. How Data Center Cooling Works, From Die to Atmosphere

Every approach performs the same four steps, and the differences are about where the medium changes from air to liquid. Heat capture happens through server fans drawing room air across heat sinks, through a cold plate, or through submersion in a dielectric fluid. Density decides that step, because air carries roughly one four thousandth of the heat per unit volume that water does. Transport moves heat out of the space, through the room or through pipe. Heat exchange passes it to the plant loop at the coolant distribution unit. Rejection pushes it into outside air.

Economisation and free cooling

Economisation rejects heat without running compressors whenever ambient conditions allow. An airside economiser brings filtered outside air directly into the hall when it is cooler than the return air. It is the most efficient option, but it exposes the white space to outdoor particulate and humidity swings, so filtration to at least MERV 13 and careful changeover control are essential.

A waterside economiser makes chilled water using a cooling tower or dry cooler without running the chiller, keeping the air loop closed. Lawrence Berkeley National Laboratory guidance puts its home in climates with wet bulb temperatures below 55 degrees Fahrenheit for 3,000 or more hours a year: met along the California coast and in the Bay Area, less so in the Central Valley, barely at all in the desert.

Adiabatic and evaporative assistance sits between the two, wetting media upstream of a dry cooler to depress entering air temperature toward the wet bulb. It extends free cooling substantially in hot dry climates, which is exactly where California water is most contested.

3. The Four Cooling Families

Air cooling

CRAC versus CRAH. A computer room air conditioner is a packaged direct expansion unit with its own compressor, rejecting heat to an outdoor condenser. A computer room air handler has no compressor: it is a fan and a chilled water coil fed from a central plant. CRAC units deploy quickly and need no chilled water, but each is an independent refrigeration machine with its own maintenance burden. CRAH units need a central plant, which then carries economisation, variable speed pumping and thermal storage across the whole load. Above a few hundred kilowatts, CRAH almost always wins.

Raised floor versus overhead. A raised access floor used as a supply plenum stays flexible for cabling, but capacity falls as underfloor obstructions accumulate and unsealed cutouts bleed supply air where it does no work. Overhead distribution avoids the plenum and delivers directly above the cold aisle, so it suits new build halls with a known rack layout. Retrofits with limited height often have no choice.

Containment. This is the highest value air side measure available: racks face front to front and back to back, blanking panels close every vacant rack unit, cutouts are sealed, and a barrier stops hot exhaust mixing with cold supply. LBNL notes that a sealed arrangement can return air at 85 degrees Fahrenheit or higher, which is what makes economisation worth having. Without it, supply temperature must be depressed to protect the worst rack in the room.

In-row units and rear door heat exchangers

These transition technologies shorten the air path rather than eliminating it. An in-row unit sits in the rack line rather than at the room perimeter, drawing hot aisle air across a coil and discharging into the cold aisle a few feet away, so fan energy falls and capacity lands where the load is. A rear door heat exchanger replaces the rack’s back door with a chilled or condenser water coil, so exhaust leaves at close to room temperature and the rack is thermally neutral. Rear doors are the least invasive route into liquid: only water at the rack.

Direct-to-chip liquid cooling

Direct-to-chip puts a cold plate in contact with the processor or accelerator and circulates coolant through it. Two loops are involved, and the boundary between them is the commercial fault line of the design. The technology cooling system loop is the clean, chemically controlled secondary loop running to the servers. It belongs to the IT vendor’s reference design and carries strict limits on temperature, pressure, flow, water quality and materials. The facility water system loop is the building’s chilled or condenser water. A coolant distribution unit sits between them: liquid-to-liquid CDUs exchange to facility water, while liquid-to-air CDUs push the load back onto the air system.

Single phase keeps the coolant liquid, typically an inhibited water and glycol mixture, and dominates in practice. Two phase lets it boil at the cold plate and condense in the CDU, carrying far more heat per unit of flow at near constant temperature, at the cost of pressure control and a fluid inventory with real handling implications.

Immersion cooling

Immersion submerges entire servers in a dielectric fluid, so almost all heat leaves as liquid and the hall needs very little air conditioning. Single phase immersion circulates a dielectric oil or synthetic fluid through a tank to a liquid-to-liquid heat exchanger, usually in horizontal tanks with servers mounted vertically. Two phase immersion uses a fluid with a low boiling point: server heat boils it, vapour rises to a condenser in the tank lid, and condensate falls back. Two phase achieves the highest heat flux available, but its fluids sit under regulatory pressure as part of the PFAS phase out.

The implications are structural and operational. A filled tank concentrates a large mass over a small footprint, so slab capacity is checked first. Serviceability changes completely: a technician hoists a wet server out, lets it drain and works on a drip tray. Fluid inventory, filtration, spill containment and disposal become facility responsibilities that did not previously exist.

Hybrid halls

Most new high density projects are hybrid. A direct-to-chip rack still rejects roughly 20 to 30 percent of its heat to air through memory, drives, network cards and power supplies, so the hall needs both a liquid loop and a working air system. The failure to avoid is undersizing that residual air system so badly that the liquid rows cannot be commissioned.

4. Which Approach Applies at Which Rack Density

Density bands are the cleanest way to frame the decision.

The triggers that force the conversation

Four events should prompt a cooling review regardless of headline density: a change of IT vendor or platform, because reference designs carry non negotiable coolant requirements; a rise in electrical service, because load calculations that rise land as heat; a tenant fit out in a shell built to a lower density; and any structural alteration, because filled liquid equipment changes floor loading.

5. What Cooling Asks of Each Building System

A density decision is never confined to the mechanical scope. It propagates.

Mechanical

Chiller selection and staging, supply water temperature strategy, economiser type, pumping topology, CDU placement and redundancy, secondary loop pipe sizing, filtration, and the residual air system a hybrid hall needs. Accurate load calculation methods matter more here than in almost any other building type, because the load runs continuously and oversizing carries a permanent penalty. Trade offs belong in HVAC system selection, not after procurement.

Electrical

Pump and CDU loads must sit on the right level of backup. An air cooled hall has minutes of thermal mass after a utility failure; a liquid cooled rack has almost none, so coolant pumps belong on uninterruptible power. That is an electrical engineering decision driven entirely by the cooling choice.

Plumbing, structural and fire

Make up water, treatment, filtration, tank and CDU drainage, spill containment for dielectric fluids and, where evaporative equipment is used, blowdown handling form a genuine plumbing engineering scope, not a mechanical subset. Filled immersion tanks impose concentrated loads well beyond a general purpose slab, and on retrofits that constraint most often kills an immersion proposal. Pipe and non conductive fluids also change the fire risk model, so leak detection and fluid compatibility with the suppression strategy need resolving with the authority having jurisdiction.

Controls and coordination

Supply water temperature reset, CDU sequencing, differential pressure control across the secondary loop, leak alarm integration and the interface between vendor monitoring and the building management system all have to be specified. MEP BIM coordination and clash detection is not optional at these pipe densities.

6. Standards and Metrics: ASHRAE, PUE and WUE

ASHRAE Technical Committee 9.9

ASHRAE Technical Committee 9.9 publishes the Thermal Guidelines for Data Processing Environments, the reference the industry designs to. It defines air classes A1 through A4 for inlet conditions and liquid classes W17 through W+ for facility supply water temperature. It is not a code, but IT warranties are written against it, so commercially it behaves like one.

Power Usage Effectiveness and its limits

PUE is total facility energy divided by IT equipment energy. It measures infrastructure overhead, not efficiency: replacing servers with better ones cuts IT energy and makes PUE look worse while the building uses less power. Industry PUE has stagnated, with Uptime Institute reporting a weighted average of 1.54 for 2025, a sixth consecutive year around 1.54 to 1.58, though sites above 20 MW average near 1.44. A liquid cooled hall also breaks the boundary: pumps and fans that were facility equipment move inside the IT enclosure and are metered as IT load, so PUE improves without a watt being saved.

Water Usage Effectiveness

WUE is annual site water use divided by IT equipment energy, in litres per kilowatt hour. It exists because PUE can be driven down by trading water for energy, so a hall with aggressive evaporative rejection can post an excellent PUE and a poor WUE at once. Any serious California comparison carries both, plus whether the water is potable, recycled or reclaimed.

7. California, Healthcare, Laboratories and Retrofit Halls

Title 24 Part 6 and computer rooms

Part 6 of Title 24, the California Energy Code, treats computer rooms as a covered process, and the definition catches more than owners expect: any space with an IT equipment power density above 20 watts per square foot qualifies, which routinely captures equipment rooms in buildings nobody calls a data center.

Section 120.6 carries the mandatory requirements: controls must prevent simultaneous heating and cooling and prohibit reheat, humidification must be adiabatic, and fan motor demand must not exceed 50 percent of design wattage at 66 percent of design fan speed for units above 60,000 Btu per hour. Section 140.9 carries the prescriptive requirements: economisers, a total fan power limit of 27 watts per kBtu per hour of net sensible cooling capacity, air containment and minimum UPS efficiencies.

The gap between the code and high density reality

Title 24 was written around air cooled computer rooms and has no separate provisions for liquid cooled IT equipment. Its economiser thresholds, fan power limits and containment language all assume heat leaves the rack as air. The California Energy Commission is examining this for the 2028 cycle, acknowledging the absence of liquid cooling requirements. In the meantime these projects end up on the performance path, so open the conversation with the authority having jurisdiction during schematic design rather than at plan check. A well built energy model is the instrument for it.

Water availability and drought policy

California adds a constraint most markets do not. Evaporative rejection is attractive in hot dry inland climates, which is precisely where supply is most contested and agencies are least willing to commit to a new industrial demand. Projects increasingly find that the will serve letter, not the energy model, decides the heat rejection strategy. Recycled water can resolve it where a purple pipe network exists, but that belongs to site selection. Our work on water conservation in MEP design covers the reuse strategies involved.

Healthcare, laboratories and retrofit halls

Hospitals and research buildings carry a surprising amount of computer room load, and both now sit inside the Title 24 definition. In healthcare facilities, imaging, server and communications rooms compete for capacity against infection control ventilation that cannot be compromised. In laboratories, high performance computing sits alongside process exhaust, and the 2025 code removed the laboratory exemption. Converting air cooled space to liquid is rarely wholesale: the realistic scope is a few rows, and the binding constraints are electrical capacity, floor loading and routing space, which is where mechanical engineering in complex existing buildings earns its keep.

8. How to Choose a Cooling Approach

The comparison below sets the four families against the factors that decide procurement. Capital cost is relative to contained air cooling for the same IT load.

The selection factors in order of weight

Rack density comes first: today’s number and the number the client will want in five years. IT vendor requirements come second, since reference designs specify coolant chemistry, temperature class, flow and materials as warranty conditions. Retrofit versus new build decides how much of the field is open, and retrofit narrows it to rear door and targeted direct-to-chip. Water availability is a permitting item for the supplying agency, not an engineering assumption. Maintenance skills are a real line item, because a team that has never handled dielectric fluid needs training. Redundancy expectations come last but reshape everything.

9. Commissioning, Redundancy and Concurrent Maintainability

Redundancy in the thermal path

Redundancy language written for chillers does not extend automatically to a liquid cooled hall. A facility can have N+1 chillers and still be single threaded at the coolant distribution unit, the secondary loop pumps, the isolation valves serving a row, or the quick disconnects on a server. Each needs a stated position in the redundancy scheme.

Thermal ride through

An air cooled hall has minutes of thermal mass in the room air and chilled water volume. A direct-to-chip rack has almost none, so if coolant flow stops, junction temperatures rise within seconds and equipment throttles or shuts down. Secondary loop pumps and CDU controls belong on uninterruptible power, and the sequences must restore flow first.

Commissioning and resilience

Liquid systems add activities air systems never required: flushing and pressure testing the secondary loop, coolant fill and chemistry verification, leak alarm testing at every zone, CDU failover under load, and integrated testing that proves the residual air system carries its share while the liquid loop is deliberately failed. The building commissioning process must be scoped for this from the start, and our commissioning services treat integrated testing as the point of the exercise. California sites also face heat waves, smoke events and grid stress at once, which is the subject of resilient MEP systems for climate adaptation.

Key Takeaways

• Establish the target density and the five year density before any other decision.
• Air cooling with proper containment remains correct below roughly 20 kW per rack.
• In-row units and rear door heat exchangers cover the 20 to 40 kW band and are the least disruptive retrofit route into liquid.
• Direct-to-chip removes heat only from cold plated components. Plan for the 20 to 30 percent that still leaves as air.
• Immersion is governed by structural loading, fluid handling and serviceability, not thermal performance.
• PUE flatters liquid cooled halls. Report WUE alongside it.
• Title 24 Part 6 catches any room above 20 watts per square foot of IT load, and the containment threshold is now 10 kW.
• Thermal ride through in a liquid cooled rack is seconds. Coolant pumps belong on uninterruptible power.

10. Where Budlong Works

Cooling strategy is site specific in a way few other MEP decisions are. Climate zone sets the free cooling hours, the water agency sets whether evaporative rejection is available at all, and local plan check culture sets how a covered process on the performance path is received. Budlong maintains teams across California so local knowledge sits behind each judgement, as our company story explains.

Sectors where cooling decisions carry the most weight

Mission critical facilities face the density question directly. Healthcare buildings carry computer rooms that now trigger containment alongside infection control ventilation. Laboratories combine high performance computing with process exhaust. Commercial buildings absorb tenant equipment rooms that quietly cross the covered process threshold.

11. Related Reading

12. Frequently Asked Questions