Where the heat goes: how data center cooling works
Start with one kilowatt of electricity entering a server.
Some of it moves data, trains a model, renders an image, or answers a request. Energy does not vanish after the computing work is complete. Almost all of it ends up as heat.
One kilowatt in becomes roughly one kilowatt of heat that has to leave.
Scale that across a rack, a room, and an entire data center, and cooling becomes easier to understand. Its job is not to create cold. It is to move heat through a series of handoffs until the heat reaches somewhere the computing equipment no longer has to deal with it.
The first handoff might be from a processor into a metal heat sink or cold plate. The next may use air, water, refrigerant, or dielectric fluid. Fans or pumps move the heat beyond the server. Another system carries it out of the data hall. At the end of the chain, the facility releases it into outdoor air, loses some water through evaporation, or transfers part of it to a nearby heating system.
Lawrence Berkeley National Laboratory describes data centers as high-power-density facilities in which almost all electricity becomes heat. Cooling systems then consume additional energy, and sometimes water, to move that heat into the surrounding environment.
The familiar argument between air cooling and liquid cooling can therefore be misleading. Air and liquid are not always competing descriptions of a whole data center. They may describe different stages of the same thermal journey.
To understand how a cooling system works, follow the heat.
From the chip to the room
In a conventional air-cooled server, heat leaves the processor through a heat spreader and heat sink. Fans pull air across the hot components and push the warmed air out of the server.
The facility supplies cooler air to the front of the rack and collects hotter exhaust from the back. Rows are commonly arranged into cold aisles and hot aisles, with containment used to stop the two air streams from mixing before the heat can be removed.
Air then carries the heat toward computer-room air-conditioning or air-handling equipment. Depending on the facility, that equipment may transfer the heat into refrigerant, chilled water, or another part of the cooling system.
This architecture remains useful because it is mature and widely supported. Most server components are designed to operate in moving air, and many conventional workloads do not produce enough heat in one rack to justify a more complicated approach.
The limitation is density. Moving more heat requires moving more air. Fans become larger or faster. Ducts, aisles, and containment have to handle greater volumes. More fan power is consumed and hot spots become harder to control. Eventually, the heat concentrated inside a rack exceeds what an air-only system can remove practically.
AI has intensified that pressure. Modern accelerated computing places large numbers of power-hungry GPUs close together. ASHRAE’s current guidance describes AI environments reaching far beyond the rack densities for which many older facilities were designed. It recommends liquid or liquid-assisted cooling at higher densities while retaining air where it still makes sense.
The change is less like replacing air with liquid throughout the building and more like shortening the distance between the hottest components and a fluid capable of carrying their heat away.
In direct-to-chip cooling, cold plates sit on processors such as CPUs and GPUs. Coolant flows through channels inside the plates, absorbs heat close to where it is produced, and moves through pipes to a coolant distribution unit. The unit controls temperature, pressure, and flow while transferring heat into the facility-side system.
Rear-door heat exchangers intervene slightly later. Server fans still push hot air through the rack, but a liquid-cooled door captures much of that heat before it enters the room.
Immersion cooling removes the usual air path more completely. Server components sit in a dielectric fluid that absorbs heat without conducting electricity. Heat then moves through an exchanger into another circuit.
These systems have different consequences for hardware, maintenance, piping, leak management, structural loading, controls, and the temperature at which heat leaves the equipment. They should not be collapsed into one category simply because each uses liquid.
Direct-to-chip cooling has become a leading option for high-density AI and high-performance computing because it works with familiar server formats while removing heat directly from the components producing most of it. But reaching the coolant does not end the journey.
The heat still has to leave the building.
From the room to the outside world
This is the part of data-center cooling that everyday labels often hide.
A server can be described accurately as liquid-cooled while the facility ultimately rejects its heat through an evaporative cooling tower. Another liquid-cooled system may use outdoor air and dry coolers. A nominally air-cooled server room may transfer its heat into a large chilled-water system elsewhere in the building.
The label attached to the server does not tell us what happens at the end of the chain.
The U.S. Department of Energy illustrates this with two common heat paths. In a conventional chilled-water design, IT equipment heats room air. Air-handling equipment moves that heat into chilled water. A chiller transfers it into a condenser-water loop, which carries it to a cooling tower. Evaporation at the tower finally releases the heat into the atmosphere.
In a direct-liquid design, a recirculating loop captures heat closer to the IT equipment and carries it to a coolant distribution unit. The facility can still pass that heat into condenser water and onward to a cooling tower. Liquid cooling at the rack has changed part of the route, while final heat rejection remains evaporative.
A cooling tower takes advantage of water’s ability to carry away a large amount of heat when it evaporates. That can reduce the electricity needed for heat rejection, but water continually leaves the system and must be replaced.
A dry cooler works more like a large radiator. Fans move outdoor air across coils carrying a warmer fluid, transferring heat into the air without relying on evaporation as the normal mechanism. This can reduce direct water demand, though performance depends strongly on outdoor temperature.
Chillers use a refrigeration cycle to move heat from a cooler circuit to a warmer one. Compressors make it possible to reject heat when natural temperature differences are insufficient, but they consume electricity.
Many facilities combine these methods. A system may use outdoor conditions directly when the weather permits, mechanical refrigeration during hotter periods, and evaporation when additional heat-rejection capacity is needed.
Climate therefore belongs inside the cooling design. A system in a cool northern climate may spend long periods using outside air or dry heat rejection with limited compressor operation. The same equipment in a hot climate has a smaller temperature difference available for moving heat outdoors and may need more fan power, refrigeration, evaporation, or larger equipment.
ASHRAE’s AI infrastructure guidance favors warm-water liquid loops partly because warmer coolant can expand the number of hours when dry coolers work without chillers. It also notes that extreme heat can still require another form of assistance.
No cooling method escapes the environment around it. It establishes a different relationship with that environment.
The labels that cause the most confusion
Once the complete heat path is visible, several apparent contradictions disappear.
A liquid-cooled data center may consume little water at the site, or it may ultimately use an evaporative cooling tower. The phrase identifies a heat-transfer method somewhere in the system, not the amount of water crossing the facility boundary.
An air-cooled data center can still contain refrigerant and liquid circuits. IT equipment may release heat into air while other systems use liquid to carry it out of the room.
A closed loop recirculates its working fluid rather than continually replacing it. It does not prove that the full cooling system consumes no water. The distinction is explored in our closed-loop cooling article.
“Free cooling” is another term that benefits from translation. It does not mean cooling without equipment or energy. It usually means favorable outdoor conditions allow the facility to reduce or bypass compressor-based refrigeration. Fans, pumps, controls, and heat exchangers still operate.
Even “water use” can describe several conditions. Water may be present inside a sealed circuit, withdrawn from a local supply, consumed through evaporation, discharged and replaced, or used upstream in electricity generation. The wider water question depends on which of those boundaries is being measured.
The same caution applies to efficiency. Cooling consumes electricity through fans, pumps, compressors, chillers, and controls. Reducing that supporting load can improve Power Usage Effectiveness. But a lower PUE does not reveal whether the cooling system consumed more water to achieve it. PUE answers one part of the efficiency question.
AI has changed where the bottleneck appears
A conventional data center may have enough floor area and electrical capacity for another row of servers, yet still be unable to host a dense GPU system.
The missing capacity is thermal.
If server fans cannot remove enough heat from the components, or the room system cannot carry it away quickly enough, the hardware cannot operate reliably at its intended load. More power at the site does not solve the problem. It creates more heat for the same constrained system to remove.
Cooling capacity has therefore become part of compute capacity. Higher-density racks affect electrical distribution, pipework, structural loading, leak detection, control systems, and final heat-rejection equipment. Decisions at the chip level propagate through the entire site.
The water implications remain less direct than popular per-query estimates suggest. AI hardware increases power density and total heat. Cooling architecture determines how that heat is removed. Climate and electricity supply then shape the resulting energy and water footprint. Those relationships are examined in our AI water-use analysis.
Higher-density cooling can also make heat more useful. Traditional air cooling often produces a large volume of relatively low-temperature exhaust. Direct-to-chip and warm-water liquid systems can carry heat away at higher temperatures, making it better suited to space heating, hot-water preheating, district systems, or some industrial processes.
Berkeley Lab notes that advanced cooling can raise the temperature of recoverable data-center heat. That does not guarantee a viable heat-recovery project. A compatible user must exist nearby, need the right temperature, and require heat at roughly the times the data center produces it.
Cooling begins with the site
A cooling architecture can look excellent on a diagram and fail to fit the place where it must operate.
The workload determines how much heat is produced and how densely it is concentrated. Climate determines how easily the heat can be rejected outdoors. Available electricity affects how much power fans, pumps, and compressors can consume. Water availability matters if the design relies on evaporation. A nearby heat user can change the value of the thermal system. Maintenance capability affects which technology can be operated reliably.
These questions belong in data-center site selection, not in a mechanical discussion held after the location has already been chosen.
Policloud develops modular data-center infrastructure for defined sites. Its cooling system has to be considered with the workload, available power, networking, climate, site conditions, ownership, regulation, and operating model.
A modular unit may arrive as an integrated object. The thermal system still has to connect that object to the physical conditions around it.
The relevant question is not whether air, water, refrigerant, or another liquid appears somewhere inside the unit. It is whether the entire route from the processor to the final heat sink suits the work and the place.
Follow that route and the commercial consequences become visible: how much computing equipment the site can support, how much electricity cooling adds, whether the system needs continuing freshwater input, how it performs during the hottest part of the year, whether the heat has another use, and what equipment must be maintained.
Every kilowatt entering the IT equipment eventually becomes heat.
The best way to understand a data-center cooling system is not to ask whether it is air-cooled or liquid-cooled.
Ask to see the whole heat path.
