Data Center Cooling Systems
Cooling removes the heat that IT equipment generates, keeping hardware inside safe operating temperatures. It is usually the largest source of energy overhead in a data center.
How Data Centers Are Cooled
Air cooling (CRAC/CRAH)
Computer room air conditioners/handlers push cool air through a raised floor; hot and cold aisle containment keeps supply and exhaust air separated.
Liquid cooling
Rear-door heat exchangers, direct-to-chip, and immersion move heat with liquid — far more effective for dense GPU racks than air.
Free & evaporative cooling
Outside air or evaporation reduces mechanical cooling load in suitable climates, lowering PUE.
In-row & close-coupled
Cooling placed near the heat source shortens air paths and handles higher density.
From Air to Liquid: One Continuum, Not a Cliff
In a conventional hall, the cooling loop is simple to describe: CRAC or CRAH units supply cold air, servers pull it through their inlets and exhaust it hot, and containment — hot aisle or cold aisle — keeps the two air streams from mixing. Everything that goes wrong with air cooling is a version of that mixing: hot exhaust recirculating into inlets, cold supply bypassing the racks entirely, or a containment gap that forces the whole room to run colder than the hardware actually needs. Free cooling sits on top of this loop as an operating mode, letting outside conditions carry part or all of the load when the climate allows.
Rising rack density is what breaks the model. Air is a poor heat-transfer medium, so every step up in power per rack demands more airflow, colder supply, and tighter containment — until the economics, and eventually the physics, favor bringing liquid closer to the silicon. The progression is gradual: rear-door heat exchangers first, then direct-to-chip cold plates, then immersion at the extreme. Most real facilities run a mix, with dense liquid-cooled rows inside an otherwise air-cooled building.
That mix is why cooling is an operations problem, not just a mechanical one. Sensaka's DCOS layer collects inlet and cabinet temperatures alongside out-of-band hardware telemetry from BMCs (Redfish, IPMI, iDRAC, iLO), maps cooling units to the racks and U-positions they serve, and tracks energy and PUE — so a thermal event reads as "these servers are at risk," not just "a unit alarmed." For coolant loops specifically, see liquid cooling monitoring.
Air, Rear-Door, Direct-to-Chip, Immersion
| Dimension | Air (CRAC/CRAH) | Rear-Door HX | Direct-to-Chip | Immersion |
|---|---|---|---|---|
| How heat is removed | Room air over CRAC/CRAH coils | Coil in the rack's rear door cools exhaust air | Cold plates on CPUs/GPUs, coolant loop to a CDU | Servers submerged in dielectric fluid |
| Density fit | Standard enterprise racks | Dense racks in an otherwise air-cooled room | High-density GPU and AI training racks | Very high density, specialized deployments |
| Facility change required | Containment, floor tiles, airflow planning | Water to the rack rear door | Coolant distribution units, manifolds, plumbing per rack | Tanks replace racks; different handling and service model |
| Residual air cooling | It is the system | Reduced room load, air path unchanged | Still needed for memory, drives, PSUs | Little to none |
| Monitoring focus | Inlet temps, supply/return delta, containment pressure | Door coil temps, water flow, valve state | Coolant flow, pressure, leak detection, per-chip temps | Fluid temperature, level, condition |
Cooling Is a Hardware-Risk Problem
A single hot spot can force the whole room to be over-cooled, wasting energy — or, if missed, push a rack past safe temperatures and shorten component life. AI racks concentrate heat, so the margin for error shrinks. Cooling decisions need device- and cabinet-level temperature data, not just a room average.
Where Cooling Monitoring Earns Its Keep
AI retrofit into an air-cooled hall
A GPU cluster lands in a room designed for lower densities. Rear-door heat exchangers or direct-to-chip loops handle the new racks while the rest of the hall stays on air — and monitoring has to cover both regimes in one view.
Hot-spot hunting
Inlet temperatures at top-of-rack run hotter than the room average suggests. Sensor data by rack and U-position locates recirculation and bypass airflow so containment can be fixed instead of the setpoint lowered.
Free-cooling hours
Economizer mode saves energy only when it is actually engaged. Tracking mode transitions alongside outside conditions and PUE shows whether the facility is capturing the free-cooling hours the climate allows.
Failure-domain awareness
When a CRAH trips, the question is which racks lose cold air and how fast they heat up. Mapping cooling units to the racks they serve turns a mechanical alarm into a ranked list of at-risk IT assets.
Coolant loop assurance
Direct-to-chip loops add flow, pressure, and leak sensors to the monitoring estate. A slow leak or a degraded pump is an IT incident in the making, not just a facilities ticket.
Capacity planning for the next density step
Thermal headroom per rack and per row — measured, not assumed — determines where the next high-density deployment can land without new cooling investment.
Common Questions About Data Center Cooling
What is the difference between a CRAC and a CRAH unit?
A CRAC (computer room air conditioner) contains its own refrigeration circuit — a compressor and refrigerant coil — and rejects heat directly. A CRAH (computer room air handler) has no compressor; it passes room air over a chilled-water coil fed by a central plant. CRAHs dominate in larger facilities because a central chiller plant is more efficient at scale and can take advantage of free cooling.
At what point does a rack need liquid cooling?
There is no single threshold — it depends on containment quality, supply temperature, and airflow design — but the pattern is consistent: as rack power climbs from typical enterprise levels into dense GPU territory, air cooling first becomes expensive (more fan energy, colder supply air, tighter containment) and then becomes physically impractical. Modern AI training racks are generally specified with direct-to-chip liquid cooling from the start rather than retrofitted later.
Does liquid cooling eliminate the need for air cooling?
Usually not. Direct-to-chip systems capture heat from CPUs and GPUs at the cold plate, but memory, drives, power supplies, and NICs still shed heat into the air, so a reduced air-cooling system remains in place. Full immersion is the main exception, since the entire server sits in dielectric fluid.
What is free cooling and when does it work?
Free cooling (economization) uses outside conditions to cool the facility instead of running compressors — either by bringing filtered outside air in directly or by cooling the water loop through outdoor heat exchangers. It works best in cooler, drier climates and during nights and winters elsewhere. Most facilities run it as a mode, switching between free, partial, and mechanical cooling as conditions change, which is exactly why economizer state belongs in monitoring.
What should a cooling monitoring system actually watch?
At minimum: server inlet temperatures at multiple rack heights, CRAC/CRAH supply and return temperatures and unit status, differential pressure across containment, and chilled-water or coolant loop temperatures and flow where liquid is in use. Inlet temperature at the server is the measurement that matters most, because it is what the hardware actually experiences — a room average can look healthy while the top of one rack is overheating.
Connect thermal data to the hardware it protects
Sensaka collects inlet temperature, cabinet thermal data, and device health together, so cooling and IT risk live in one view.