When people think about a data center, the visible story is usually about computing: servers, GPUs, storage, networking and the software running on top of them. The less visible story is what happens to the energy those machines consume.
Almost all of the electrical power used by computing equipment eventually becomes heat. That heat must move from the processor into a cooling medium, travel through the facility, and ultimately be rejected into the surrounding environment. As computing becomes more concentrated, that physical process is becoming one of the main constraints on how data centers are built.
This phenomenon is especially apparent with AI and high-performance computing. ASHRAE’s 2026 AI Data Center Energy Performance Framework says rack densities have moved from roughly 120 kW toward several hundred kilowatts, with megawatt-class racks anticipated in the near term. The organization describes this as a fundamental change to assumptions that previously guided power delivery, cooling and physical infrastructure. (ASHRAE)
That does not mean every data center now needs liquid cooling. Traditional CPU-based workloads remain important, and many facilities can continue using air cooling. The change is that cooling architecture increasingly has to follow the density of the computing hardware.
The Heat Starts Inside the Semiconductor
A processor does not generate heat somewhere separate from the computation. The electrical activity that allows transistors to perform calculations also produces thermal energy.
Consider the path in a conventional server. Heat generated by the CPU or GPU passes through its package and thermal interface into a heatsink. Fans move air across the heatsink, and the warmed air leaves the server. The facility then has to keep that exhaust air from returning to the server’s intake and eventually remove the heat from the room.
The complete path is therefore as follows:
Processor → thermal interface → heatsink → air → server exhaust → room airflow → cooling equipment → outdoor environment
Every stage matters. If airflow is poorly controlled, hot exhaust can mix with cool supply air. If the heatsink can’t transfer enough heat into the air, the processor may hit its thermal limits. When the room is unable to efficiently remove the accumulated heat, the problem shifts from an individual server to the facility.
This is why conventional data centers use hot-aisle and cold-aisle arrangements, containment and carefully controlled airflow. The goal is not simply to make the room cold. It is to establish a predictable route for heat.
ASHRAE’s current guidance recommends advanced air management, containment and precise airflow control even in AI-oriented facilities because not every component or workload will necessarily be liquid cooled. (ASHRAE)
Why Rack Density Changes Everything
The important variable is not only how much heat a data center produces in total. It is where that heat is concentrated.
Imagine two facilities that each consume 1 MW of IT power. In one, the load is spread across a large number of relatively low-power racks. In the other, much of the same load is concentrated into a smaller number of high-density AI racks.
The total electrical consumption can be similar, but the cooling problem is very different.
A high-density rack puts much more heat into a relatively small physical space. Moving enough air through that rack becomes increasingly difficult. Larger heatsinks, faster fans and higher airflow can extend the useful range of air cooling, but they do not change the underlying heat-transfer properties of air.
Uptime Institute reported in July 2026 that typical rack densities were shifting toward about 10 kW, with more than a quarter of surveyed operators reporting densities above that level. Its research also points to increasing interest in direct liquid cooling as rack power continues to rise. (Uptime Institute Journal)
The issue becomes particularly significant in AI infrastructure because accelerators are designed to perform enormous amounts of parallel computation while communicating rapidly with other accelerators. ASHRAE notes that AI’s demand for dense hardware and high-bandwidth interconnection is a major reason these systems are being packed into increasingly dense configurations. (ASHRAE)
The physical consequence is straightforward:
More power in the same space → more heat in the same space → greater heat-transfer requirements.
Air Cooling Has Not Suddenly Become Obsolete
It would be misleading to describe the current transition as “air cooling is being replaced by liquid cooling.”
Air cooling remains practical for many workloads. ASHRAE’s 2026 framework explicitly notes that AI-optimized facilities still need to accommodate lower-density CPU workloads and may therefore contain heterogeneous rack densities and cooling profiles. (ASHRAE)
The industry is instead moving toward a mixture of approaches.
| Cooling approach | Where heat is captured | Typical role |
|---|---|---|
| Conventional air cooling | Heatsinks and server airflow | General-purpose and lower-density computing |
| Rear-door heat exchanger | Server exhaust air | Reduces heat entering the room |
| Direct-to-chip liquid cooling | CPU/GPU cold plates | High-density processors and accelerators |
| Single-phase immersion | Hardware immersed in liquid | Specialized high-density systems |
| Two-phase immersion | Liquid changes phase around heat sources | Specialized applications requiring different thermal architecture |
A rear-door heat exchanger, for example, should not be confused with direct-to-chip cooling. It uses a liquid-cooled heat exchanger at the rear of a rack to remove heat from the server’s air exhaust. The processor itself is still being cooled through the server’s air path.
Direct-to-chip cooling moves the liquid much closer to the source.
Direct-to-Chip Cooling Changes the Heat Path
In a direct-to-chip system, a cold plate is mounted against a high-power component such as a CPU or GPU. Coolant flows through channels inside the cold plate and carries heat away from the component.
The heat path becomes:
Processor → thermal interface → cold plate → coolant → distribution system → heat exchanger → heat-rejection equipment
The major change is where the heat enters the cooling medium.
Instead of first transferring most of the heat into room air, the system captures a substantial portion directly with liquid. ASHRAE’s current framework identifies direct-to-chip liquid cooling as the leading approach for high-density AI and HPC systems. (ASHRAE)
This is also why a liquid-cooled server does not necessarily look dramatically different from a conventional server from the outside. Fans may still exist because some components can remain air cooled. Memory, storage, power electronics and other equipment may have different thermal requirements from the main processors.
The result can be a hybrid thermal architecture in which liquid handles the most concentrated heat while air handles the remainder.
That is often more practical than attempting to redesign every component around a single cooling technology.
The Coolant Becomes Part of the Computing Infrastructure
Once liquid is introduced into a high-density rack, cooling becomes a distribution problem.
The coolant needs to reach the appropriate servers, collect heat and return to a facility system. This can involve manifolds, pumps, valves, sensors, heat exchangers and coolant distribution units, commonly called CDUs.
A simplified architecture looks like this:
| Component | Technical role |
|---|---|
| Cold plate | Transfers heat from the processor into coolant |
| Manifold | Distributes and collects coolant within the rack |
| CDU | Manages circulation and heat exchange between loops |
| Facility loop | Transports heat through the building |
| Heat exchanger | Transfers thermal energy between cooling circuits |
| Heat rejection | Releases heat to the environment or another useful heat sink |
| Sensors and controls | Monitor temperature, flow and operating conditions |
The exact implementation varies between systems. Vendors can use different fluid arrangements, connection systems, temperatures and control architectures, so it would be misleading to assume that every liquid-cooled rack uses the same internal plumbing.
The underlying mechanism, however, is consistent: the coolant becomes another infrastructure pathway through which computational energy leaves the IT equipment.
The Cooling System Has to Remove Heat From the Building
Capturing heat at the processor is only half of the job.
Once coolant leaves a cold plate, it carries thermal energy elsewhere. The facility still needs to reject that energy.
Depending on the architecture and climate, the final stages can involve dry coolers, cooling towers, chillers, heat exchangers or other heat-rejection technologies. The objective is to move heat from the IT equipment to an environment or process capable of accepting it.
This is where higher-temperature liquid cooling can become particularly advantageous.
If the coolant can operate at temperatures sufficiently above outdoor conditions, a facility may be able to reject heat through a dry cooler without relying continuously on mechanical refrigeration. ASHRAE’s 2026 framework describes warm-water cooling and chiller-less approaches as important opportunities for reducing mechanical cooling requirements and, in suitable environments, water consumption. (ASHRAE)
The important point is that liquid cooling is not automatically efficient simply because it uses liquid. The complete chain matters:
IT hardware → pumps → coolant loop → heat exchanger → heat rejection → environmental conditions
A poorly designed liquid system can still consume substantial energy. The engineering advantage comes from using the physical properties of liquid and the temperature relationship between the cooling loop and its heat sink effectively.
AI Is Also Changing the Timing of the Heat Load
There is another complication that is easy to miss.
AI systems can produce highly concentrated computational workloads across large numbers of accelerators. The electrical and thermal load therefore becomes closely connected to workload behavior.
The cooling system must consider more than just the average annual IT load if the hardware can experience substantial changes in power consumption during operation.
This is one reason modern data-center designs increasingly rely on sensors, controls and continuous monitoring. ASHRAE recommends intelligent controls, real-time monitoring and continuous commissioning for high-density AI environments so that cooling systems can respond to changing operating conditions. (ASHRAE)
The resulting system creates a feedback loop:
Workload → processor utilization → electrical power → heat generation → temperature/flow measurements → cooling controls → pump/fan/valve adjustments
Cooling therefore becomes partly a control-system problem.
The facility is not simply providing a fixed amount of cold air or liquid. It is continuously maintaining operating conditions while the computing workload changes.
Why the Data Center Building Has to Change
Once rack densities become high enough, the implications extend far beyond the server.
A facility designed for conventional air-cooled racks may not have the piping, pumps, heat exchangers, electrical distribution or structural capacity required for much denser liquid-cooled systems.
ASHRAE’s current AI framework specifically identifies power and cooling as interdependent design problems and recommends that architectural, electrical and mechanical systems be considered together. It also emphasizes adaptability because processor roadmaps and cooling technologies are changing faster than traditional building-design cycles. (ASHRAE)
This creates a new planning problem.
A data center may operate for decades, while a processor generation can be replaced within a much shorter period. A cooling system designed around today’s rack density therefore has to be evaluated against what the facility may need several hardware generations later.
That is why modern data-center designs increasingly include liquid-ready infrastructure even where the initial deployment does not use the maximum possible cooling density.
Water Is Becoming a Design Question Too
Liquid cooling does not automatically mean high water consumption.
A closed-loop system can circulate coolant repeatedly rather than continually consuming and discharging water. Water use depends heavily on how the facility ultimately rejects heat.
Evaporative systems can consume water because evaporation is part of their heat-rejection mechanism. Dry cooling can reject heat without the same evaporative process, although it has its own performance and environmental constraints.
The U.S. Department of Energy’s August 2026 COOLERCHIPS 1.5 program illustrates where research is heading: its selected projects are developing and validating advanced, water-free cooling systems for high-power AI data centers, with testing aimed at cooling loads of up to 1 MW per rack. These are research targets, not a statement that ordinary data centers already operate at those densities. (The Department of Energy’s Energy.gov)
The broader lesson is that water strategy and cooling architecture are increasingly connected.
What About Storage and Data Movement?
Computing heat is not determined by storage capacity alone. A server with several terabytes of storage does not necessarily consume more cooling power simply because the storage capacity is larger.
What matters thermally is the electrical power consumed by the actual hardware and how that power is distributed.
This distinction is useful when reading data-center specifications because capacity units and power units describe different things. If a specification uses TB, GB, PB or other storage quantities and you need to compare the units, the TechTalksX Data Storage Unit Converter can help with those conversions:
It should not, however, be treated as a cooling calculator. Storage capacity and thermal load are separate engineering variables.
The New Constraint: How Much Heat Can the Facility Move?
The most important shift is that computing capacity can no longer be considered independently from thermal capacity.
A traditional planning question might have been:
How much electrical power can this facility deliver to the servers?
For a high-density AI deployment, the question becomes:
How much electrical power can the facility deliver and continuously convert into useful computing while removing the resulting heat?
That difference is fundamental.
Even if the electrical system can provide 100 MW, the facility has less than 100 MW of usable computing capacity if the cooling infrastructure cannot reject the corresponding thermal load.
The same principle applies at rack level. A rack capable of accepting a large electrical load is not necessarily capable of operating that hardware continuously unless the cooling system can capture and transport the heat.
Why Future Data Centers May Be Designed Around Cooling First
The industry is therefore moving toward a design process that plans computing, power, and thermal systems together.
ASHRAE’s 2026 framework describes AI data centers as an integrated system in which power and cooling decisions directly affect one another. It also warns that rapidly changing hardware makes long-term adaptability increasingly important. (ASHRAE)
That explains why the physical design of an AI data center can look very different from an older server facility.
The building needs enough electrical capacity to supply the computing equipment. The rack needs enough thermal infrastructure to capture the resulting heat. The facility needs enough distribution and heat-rejection capacity to move that heat away. The control system needs enough instrumentation to keep the whole chain operating within its limits.
The visible computer is only one part of that system.
The actual process is:
Electricity → computation → heat → heat-transfer medium → facility cooling loop → heat rejection → environment
As processors become more powerful and more tightly packed, the middle of that chain becomes increasingly important.
Understanding it also changes how data-center specifications should be read. A higher-powered accelerator is not simply a faster component. It can require a different rack, different cooling distribution, different facility infrastructure and different planning assumptions.
That is the real heat problem changing data centers: the limit on computing is increasingly determined not only by what the silicon can calculate, but by how efficiently the building can carry away the energy that calculation produces.
Frequently Asked Questions
Why can’t data centers simply use bigger fans?
They can increase airflow to a point, and improved air management can extend the practical range of air cooling. But as heat concentrates, the required airflow, heatsink capacity, and fan power also increase. Liquid can capture substantial heat closer to high-power components without requiring the same amount of heat transport through room air.
Does liquid cooling mean the entire server is filled with water?
No. Direct-to-chip systems typically use cold plates attached to selected processors. Other components can remain air cooled. Immersion cooling is a different architecture in which compatible hardware is placed directly into a specialized cooling fluid.
Does every AI data center need liquid cooling?
No. Cooling requirements depend on hardware, rack density, workload and facility design. AI-oriented facilities may contain both high-density liquid-cooled racks and lower-density air-cooled equipment. ASHRAE specifically recognizes the need to support heterogeneous rack densities. (ASHRAE)
Is liquid cooling automatically more energy efficient?
Not automatically. Pumps, controls, heat exchangers and heat-rejection equipment all consume energy. The potential advantage comes from the ability to capture and transport concentrated heat efficiently and, with suitable temperatures and environmental conditions, reduce dependence on mechanical refrigeration.
Why is rack density more important than total data-center power?
Total power determines the overall heat load, but rack density determines how concentrated that heat is. Two facilities can have similar total IT power while requiring completely unique cooling architectures if one concentrates much more power into individual racks. That is one reason the rapid increase in AI rack density is influencing data-center design so strongly. (Uptime Institute Journal)