Cooling and Airflow

Every watt becomes heat

A server does not store energy: nearly all of the electrical power it draws is turned into heat by processors, memory, drives, fans and the power supply itself. A rack that draws 10 kW is a 10 kW heater that runs day and night. If that heat is not removed, inlet temperatures rise until the servers throttle their processors and finally shut themselves down to protect the hardware. In a densely loaded room without cooling, that can happen within minutes.

So the first rule of cooling design is simple bookkeeping: the cooling capacity must at least match the IT load, plus the heat from lighting, people and the power equipment (UPS losses, transformers) in the same space, plus redundancy. Facilities staff often think in British thermal units per hour (BTU/h) or in tons of refrigeration (1 ton = 12,000 BTU/h, about 3.5 kW), while IT staff think in kilowatts. The conversion is fixed: 1 W of heat equals 3.412 BTU/h.

# Measured IT load per rack in a small server room, in kilowatts
racks = {"A01": 4.2, "A02": 6.8, "A03": 11.5, "B01": 3.1, "B02": 7.4}

BTU_PER_WATT = 3.412    # 1 W of heat = 3.412 BTU per hour
BTU_PER_TON = 12_000    # 1 ton of refrigeration = 12,000 BTU/h

for name, kw in racks.items():
    btu_h = kw * 1000 * BTU_PER_WATT
    print(f"{name}: {kw:5.1f} kW -> {btu_h:7,.0f} BTU/h")

total_kw = sum(racks.values())
total_btu_h = total_kw * 1000 * BTU_PER_WATT
print(f"Room: {total_kw:.1f} kW -> {total_btu_h:,.0f} BTU/h"
      f" = {total_btu_h / BTU_PER_TON:.1f} tons of cooling")

A01:   4.2 kW ->  14,330 BTU/h
A02:   6.8 kW ->  23,202 BTU/h
A03:  11.5 kW ->  39,238 BTU/h
B01:   3.1 kW ->  10,577 BTU/h
B02:   7.4 kW ->  25,249 BTU/h
Room: 33.0 kW -> 112,596 BTU/h = 9.4 tons of cooling

The same numbers drive the power side: a cooling unit that has to remove 33 kW itself draws power, and that power has to come from the same utility connection and, for critical sites, the same generators. See Power: Utility, UPS and Generators for the electrical half of the story, and CPUs for thermal design power at component level.

How a server breathes

Almost all data center equipment is built for front-to-back airflow: fans pull cool air in through the front panel, push it over the heat sinks and blow it out of the back, typically 10–20 °C warmer. Everything in rack design follows from that one direction:

• Blanking panels close every empty rack unit. Without them, hot exhaust air is pulled straight back to the front through the gap and the servers above it breathe their own exhaust (recirculation).
• Cable management keeps cable bundles away from fan outlets at the back; a wall of patch cords behind a 1U server can raise its internal temperature noticeably.
• Gaps around and under racks are sealed with brush strips or grommets, so cold air does not escape where no equipment uses it (bypass).
• Network switches often come with a choice of fan direction (port-side intake or port-side exhaust). A top-of-rack switch mounted with its ports to the back must have the airflow that matches, or it will suck in hot-aisle air.

Recirculation and bypass are the two classic airflow faults. Recirculation makes individual servers run hot even when the room seems cold; bypass wastes cooling capacity because cold air goes back to the cooling unit unused. Both are cheap to fix and expensive to ignore.

Hot aisles and cold aisles

If racks all faced the same way, each row would blow its exhaust into the intake of the row behind it. The standard layout therefore alternates the direction of the rows. The front sides of two rows face each other across a cold aisle, the aisle where cooled supply air is delivered and drawn into the equipment intakes. The back sides face each other across a hot aisle, the aisle that collects the warm exhaust and leads it back to the cooling units.

In rooms with a raised floor, cold air is blown into the space under the floor (the plenum) and comes up through perforated tiles placed only in the cold aisles. Newer halls often skip the raised floor and supply air from overhead ducts, from the walls, or from cooling units placed in the rows themselves (in-row cooling).

Aisle containment

Even with the right layout, air mixes over the tops and around the ends of the rows. Aisle containment physically closes off one of the two aisles, with doors at the row ends and a roof or ceiling panels above, so that supply air and exhaust air are separated. There are two variants:

Both variants have the same effect on efficiency: because supply and exhaust air no longer mix, the cooling units can deliver air closer to the temperature the servers actually need and receive warmer return air, which they remove heat from more efficiently.

Cooling units and where the heat goes

The units that cool the air in the room come in two families. A CRAC (computer room air conditioner) contains its own refrigeration circuit with a compressor, much like a large air conditioner, and rejects heat through a condenser outside. A computer room air handler (CRAH) contains only fans and a coil filled with chilled water; the water is cooled elsewhere by central chillers, cooling towers or dry coolers. Large data centers mostly use CRAHs or in-row and rear-door units on a chilled-water loop, because one central plant is more efficient than dozens of small compressors.

Wherever the cooling unit sits, the heat ends up outside the building. In a cool climate such as that of the Netherlands, outside air is cold enough for much of the year to cool the water or the air without running compressors at all. This free cooling (using an economizer) is one of the main reasons modern sites need far less cooling energy than older ones. Some sites also deliver their waste heat to a district heating network; see Energy and Sustainability.

Cooling units are redundant just like power equipment: a room that needs four CRAH units typically has five (N+1) or more, so one can fail or be serviced. Fault-tolerant designs also keep fans and pumps on UPS power, so that cooling continues during the seconds before the generators take over. See Availability Tiers and Redundancy for what N+1 and 2N mean.

Recommended temperatures

The reference for air-cooled IT equipment is the thermal guidelines of ASHRAE (the American Society of Heating, Refrigerating and Air-Conditioning Engineers), Technical Committee 9.9. Their key points, as of 2026 (5th edition):

• The recommended inlet temperature range for the common equipment classes A1–A4 is 18–27 °C. Within this range, equipment runs reliably and efficiently.
• The allowable ranges are wider, from 15–32 °C for class A1 up to 5–45 °C for class A4. Equipment may run there for limited periods, at the cost of more fan power and potentially shorter life.
• The 5th edition added class H1 for high-density air-cooled equipment, with a narrower recommended range of about 18–22 °C.
• Temperature is measured at the air inlet, at the front of the rack, not at the cooling unit and not in the hot aisle. A rack typically has sensors at the bottom, middle and top, because the top often runs warmest.

Humidity matters too, but less than many people think. Very dry air raises the risk of electrostatic discharge (see Safety and Handling), and air close to its dew point can condense on cold surfaces. ASHRAE expresses its humidity limits as a combination of dew point and relative humidity; check the guideline itself or your site's specification for the exact values.

The old habit of keeping a server room at 20 °C or lower "to be safe" wastes energy: every degree of extra cooling costs compressor and fan power, and modern servers are specified for the higher range. Raising the supply temperature toward the upper end of the recommended range, with good containment, is one of the cheapest efficiency gains a site can make.

Rack density and the limits of air

Rack density is the electrical power drawn by the equipment in one rack, usually in kilowatts per rack. Because power becomes heat, it is also the amount of heat that has to be removed from that one footprint. Density has grown steadily:

• In Uptime Institute's Global Data Center Survey 2026, the average typical (modal) rack density passed 11 kW (9 kW the year before), or 7.8 kW when high-density outliers are left out.
• 24% of respondents have at least some racks of 30 kW or more, and about one in seven have racks above 50 kW. The majority still have no racks at or above 30 kW.
• AI training racks of more than 100 kW are already shipping, and vendor roadmaps reach several hundred kilowatts per rack.

Air has a low heat capacity, so removing more heat means moving more air or allowing a bigger temperature rise. The example below compares how much air and how much water it takes to carry away the heat of one rack.

# Approximate volumetric heat capacity near room temperature, kJ per m3 per kelvin
AIR = 1.2       # air
WATER = 4180    # water (1 m3 = 1,000 liters)
DELTA_T = 12    # temperature rise from server inlet to exhaust, in kelvin

print("Rack     Air needed      Water needed (10 K rise)")
for rack_kw in [5, 11, 30, 120]:
    air_m3_h = rack_kw / (AIR * DELTA_T) * 3600
    water_l_min = rack_kw / (WATER * 10) * 1000 * 60
    print(f"{rack_kw:4} kW  {air_m3_h:7,.0f} m3/h    {water_l_min:6.1f} L/min")

Rack     Air needed      Water needed (10 K rise)
   5 kW    1,250 m3/h       7.2 L/min
  11 kW    2,750 m3/h      15.8 L/min
  30 kW    7,500 m3/h      43.1 L/min
 120 kW   30,000 m3/h     172.2 L/min

Thirty thousand cubic meters of air per hour through one rack footprint is not practical: the fans would be loud, power-hungry and still struggle. The same heat is carried by under 200 liters of water per minute, through pipes a few centimeters wide. That is why air cooling remains the standard up to roughly 20–30 kW per rack, and liquid cooling takes over above that.

Liquid cooling

Liquid cooling brings the coolant to the heat source instead of cooling the whole room. Two approaches dominate.

Direct-to-chip

Direct-to-chip liquid cooling mounts cold plates directly on the hottest components, mainly processors and GPUs, and pumps coolant through them. A coolant distribution unit (CDU) separates the clean, tightly controlled loop that runs through the servers from the facility water loop and controls flow, pressure and temperature. The cold plates remove most of the heat; memory, drives, power supplies and network cards are still cooled by air, so a liquid-cooled rack also needs some air cooling. Because the coolant can be warm (well above room temperature) and still do its job, direct-to-chip systems combine well with free cooling and heat reuse.

Immersion

Immersion cooling submerges whole servers in a tank of dielectric fluid, a liquid that does not conduct electricity. In single-phase immersion the fluid stays liquid and is pumped through a heat exchanger. In two-phase immersion the fluid boils on the hot components and condenses on a coil above the tank. Immersion removes nearly all heat without fans, but it changes everything about operations: servers are lifted out of a tank for maintenance, not slid out of a rack, and not every component or warranty is suitable for it. Some two-phase fluids are fluorinated chemicals, which may be affected by EU restrictions on PFAS (per- and polyfluoroalkyl substances).

Cooling, energy and PUE

Power usage effectiveness (PUE) divides the total energy of a data center by the energy used by the IT equipment. A PUE of 1.5 means that for every kilowatt-hour the servers use, another half kilowatt-hour goes to overhead, and cooling is usually the largest part of that overhead. Uptime Institute's 2026 survey puts the weighted industry average at 1.52, down only slightly from 1.56 in 2024. The definition, its pitfalls and the EU reporting rules are on Energy and Sustainability.

The levers that lower the cooling share are the ones on this page: containment so that no cold air is wasted, higher supply temperatures within the ASHRAE range, free cooling for as many hours as the climate allows, and liquid cooling where density demands it.

When cooling fails

A typical scenario: one CRAH unit trips on a fault at night. With N+1 cooling, the remaining units ramp up and the monitoring system raises an alarm; nothing happens to the IT load. The danger starts when a second unit is already out for maintenance, or when the chilled-water supply itself stops, for example during a power transfer to generators. Then inlet temperatures can rise several degrees in minutes, starting with the densest racks.

What operations staff typically do, in this order:

1. Follow the site's incident procedure and call facilities; the cooling units themselves are their domain.
2. Watch inlet temperatures per rack, not the room average, and look for the racks that rise fastest.
3. Keep containment doors closed: opening them seems logical but mixes hot air into the cold supply.
4. If temperatures keep rising, shut down non-critical systems according to the agreed priority list to reduce the heat load, before the servers do it for you.