Artificial intelligence is changing more than the way data centers process information. It is changing how those facilities must remove heat.
Traditional enterprise data centers were generally designed around CPU-based servers with comparatively moderate rack densities. AI training and inference clusters use large concentrations of GPUs and other accelerators, placing substantially more electrical power—and therefore more heat—into each rack.
According to the ASHRAE AI Data Center Energy Performance Framework, purpose-built AI facilities may need to support rack densities of 50 to 120 kilowatts or more. By comparison, many legacy CPU-oriented racks operate closer to 5 to 15 kilowatts.
That difference is why AI data centers are generating demand for direct-to-chip liquid cooling, coolant distribution units, rear-door heat exchangers, upgraded chilled-water plants, advanced controls, economizers, dry coolers, and other high-capacity thermal-management technologies.
Why AI Computing Produces Such Intense Cooling Loads
Almost all electricity consumed by computing equipment eventually becomes heat that must leave the data center. When a facility adds more computing power to the same physical area, the cooling system must remove more heat from that area.
AI accelerators create several related challenges.
Higher Power Density per Rack
GPU servers are optimized to perform many calculations simultaneously. A cluster may place multiple high-power accelerators, processors, networking components, and power supplies into a relatively small rack footprint.
A data hall that once supported a uniform collection of low- or medium-density racks may now contain several thermal zones:
- Conventional storage and networking equipment
- General-purpose CPU servers
- High-density AI inference systems
- Extremely dense AI training clusters
One cooling design may not be appropriate for every zone.
Sustained, Concentrated Workloads
AI training workloads may keep accelerators operating at high utilization for extended periods. The cooling system therefore must manage sustained heat production rather than only brief peaks.
Cooling capacity, redundancy, pumping, heat exchangers, controls, and electrical infrastructure must all be designed around the expected workload profile.
Thermal Throttling Can Reduce Computing Performance
Computer components protect themselves when operating temperatures become excessive. A GPU may reduce its operating speed to control temperature, a process commonly called thermal throttling.
A cooling problem can consequently become a computing-performance problem. Data-center operators are not simply trying to maintain a comfortable room temperature; they are protecting equipment availability and the productive output of expensive AI hardware.
How AI Changes the Complete Heat-Removal Path
Advanced data-center cooling is best understood as a chain:
Chip → coolant or air → rack or server loop → facility loop → heat-rejection equipment → outdoor environment
Improving only one part of this chain may not solve a system-level bottleneck.
For example, cold plates can collect heat efficiently from GPUs, but that heat must still travel through pumps, manifolds, a coolant distribution unit, facility piping, heat exchangers, and finally an outdoor heat-rejection system.
A complete design must coordinate the IT equipment with the building’s mechanical infrastructure.
| Cooling approach | How it removes heat | Typical application | Key consideration |
|---|---|---|---|
| Optimized air cooling | Moves conditioned air through servers using fans and controlled airflow paths | Lower- and medium-density computing zones | Requires containment and control of bypass or recirculated air |
| Rear-door heat exchanger | Captures heat as server exhaust passes through a liquid-cooled rack door | Retrofits and hybrid air-liquid installations | Reduces room heat load without eliminating server airflow |
| Direct-to-chip liquid cooling | Cold plates collect heat directly from processors and accelerators | High-density GPU and HPC racks | Needs compatible cold plates, piping, CDUs, controls, and leak detection |
| Immersion cooling | Submerges compatible electronic equipment in dielectric fluid | Specialized high-density deployments | Hardware, fluid, maintenance, and material compatibility must be evaluated |
| Chilled-water or DX cooling | Uses mechanical refrigeration to move heat toward outdoor equipment | Facility cooling plants, CRAH units, CRAC units, and support spaces | Refrigerant, compressor, efficiency, redundancy, and regulatory requirements apply |
Technical basis: ASHRAE AI Data Center Energy Performance Framework and the U.S. Department of Energy’s data-center cooling guidance.
Why Liquid Cooling Is Receiving So Much Attention
Liquid can collect and transport concentrated heat more effectively than air. Direct-to-chip cooling places cold plates on high-heat components and circulates coolant through those plates.
The coolant does not necessarily contact the electronics. Instead, it flows through a sealed circuit connected to manifolds and a coolant distribution unit, commonly called a CDU.
The CDU transfers heat between the technology cooling loop and the facility loop. It may also control fluid temperature, pressure, flow, filtration, and monitoring.
Liquid Cooling Can Support Warmer Facility Water
Capturing heat close to the processor may allow a system to operate with warmer fluid temperatures than a conventional chilled-air installation.
When outdoor conditions permit, a facility may be able to reject this heat through dry coolers or waterside economization without running chiller compressors continuously. Reducing compressor operating hours can lower cooling energy use.
Warmer liquid also produces more useful heat for potential recovery applications, including building heating, water preheating, or a campus thermal network. Heat reuse remains site-specific because the facility needs a practical nearby use for the recovered energy.
Most Facilities Will Still Need Air Cooling
Liquid cooling does not automatically eliminate air cooling.
Power supplies, memory, storage devices, network switches, and other components may continue releasing heat into the server or room air. Lower-density racks elsewhere in the building may also remain air cooled.
For this reason, many AI data centers use a hybrid design:
- Liquid cooling for the highest-density components
- Air cooling for residual server heat
- Containment for conventional racks
- Chilled water, dry coolers, cooling towers, or DX equipment for facility-level heat rejection
- Integrated controls to coordinate all cooling layers
The goal is not to replace every fan or refrigeration circuit. It is to match each thermal load with an appropriate heat-removal method.
Advanced Cooling Is Also About Energy and Water
Cooling-system design affects both electricity and water use.
The 2025 update to Lawrence Berkeley National Laboratory’s U.S. Data Center Energy Usage Report, published in June 2026, estimates that data centers could account for approximately 11.8% of total U.S. electricity consumption in 2030. Its modeled scenario range is 9.5% to 15.3%.
Those projections include IT equipment and cooling-system performance. As AI capacity expands, cooling efficiency can materially affect a facility’s power requirements, operating cost, and ability to secure enough electrical capacity.
Air Systems Use Fan Energy
Poorly controlled air systems may waste energy by:
- Mixing hot exhaust with cold supply air
- Moving more air than the current IT load requires
- Operating at unnecessarily low supply temperatures
- Allowing conditioned air to bypass server inlets
- Running fans at excessive speed
Hot-aisle or cold-aisle containment, variable-speed fans, rack-level sensors, and improved control sequences can make air cooling more effective.
Evaporative Systems Use Water
Cooling towers and evaporative systems can reduce electrical consumption under suitable conditions, but they consume water and require water-treatment programs.
Dry coolers can reduce or eliminate evaporative water consumption, although their capacity and efficiency depend on outdoor temperature and the required fluid temperature.
The correct solution varies with climate, water availability, electricity costs, redundancy requirements, and the facility’s permitted operating temperatures.
Where Refrigerants Fit into AI Data Center Cooling
One important distinction is that liquid cooling does not always mean refrigerant cooling.
A direct-to-chip loop commonly circulates a formulated coolant or treated water solution. That fluid collects heat from electronic components and transfers it through a CDU.
A separate facility system may then use:
- A refrigerant-based chiller
- A direct-expansion computer-room air conditioner
- A dry cooler
- A cooling tower
- A waterside economizer
- A heat pump
- A combination of these technologies
Refrigerant may therefore remain essential even when liquid is carrying heat away from the chips.
Refrigerant Selection Must Follow the Equipment
A data center should never select a refrigerant simply because another product appears to offer lower cost, higher availability, or a lower global warming potential.
The refrigerant must be approved for the specific chiller or DX system. Compressors, lubricants, heat exchangers, expansion devices, pressure controls, relief systems, seals, sensors, and safety requirements are engineered around particular operating characteristics.
Retrofitting or substituting refrigerant without manufacturer approval can create performance, reliability, safety, and warranty problems.
U.S. Regulations Are Influencing New Equipment
EPA’s current Technology Transitions restrictions by sector establish a 700-GWP limit for new data-center, computer-room air-conditioning, and information-technology cooling systems, with an installation compliance date of January 1, 2027.
This does not mean every existing data-center cooling system must immediately change refrigerants. Applicability depends on factors including the equipment category, manufacturing or installation date, and the specific regulatory provisions.
New projects should evaluate:
- EPA requirements in effect for the installation date
- State and local mechanical and fire codes
- Refrigerant safety classification
- Equipment-room ventilation and detection requirements
- Charge size and equipment location
- Technician training and service-tool compatibility
- Long-term refrigerant and component availability
- Manufacturer support throughout the equipment lifecycle
Organizations maintaining legacy refrigerant-based equipment should also monitor procurement conditions. The Get Freon article on how refrigerant supply chains affect HVAC contractors explains why availability, planning, and service timing can become operational issues.
Why Existing Data Centers Cannot Always Add AI Racks Easily
Adding AI servers to a conventional data center is not as simple as placing new hardware in an empty rack.
The existing facility may have enough floor area but lack:
- Electrical distribution capacity
- Rack-level cooling capacity
- Adequate chilled-water flow
- Space for new piping and CDUs
- Sufficient outdoor heat-rejection capacity
- Redundancy for the additional thermal load
- Controls capable of managing mixed-density zones
A retrofit assessment should follow the entire heat path. Increasing cooling at the rack will not help if the central plant, distribution piping, or outdoor equipment cannot reject the additional heat.
Rear-door heat exchangers can sometimes provide a practical intermediate step. They capture heat close to the rack while allowing much of the existing air-cooled infrastructure to remain in service.
For still higher densities, direct-to-chip cooling may be necessary. However, the conversion requires coordinated mechanical, electrical, controls, IT, water-quality, and commissioning work.
What Advanced Cooling Demand Means for HVAC Contractors
AI data centers create opportunities for contractors, engineers, commissioning providers, control specialists, and equipment suppliers. They also require knowledge beyond conventional comfort cooling.
Important capabilities include:
- Hydronic-system design and balancing
- Coolant quality and corrosion control
- CDU installation and commissioning
- Refrigerant-system service
- Variable-speed pumping
- Leak detection and fluid containment
- Heat-exchanger maintenance
- Mission-critical redundancy
- Building-management and data-center infrastructure controls
- Continuous performance monitoring
Documentation is especially important. A data center may contain multiple coolant formulations and multiple refrigerant systems, each with its own compatibility and service requirements.
Cylinders, recovery equipment, hoses, pumps, lubricants, and service tools must be identified and managed correctly. Mixing fluids or using an unapproved replacement can damage equipment and introduce contamination.
Frequently Asked Questions
Do all AI data centers require liquid cooling?
No. Cooling requirements depend on rack density, server design, climate, workload, and facility architecture. Optimized air cooling may remain appropriate for lower-density areas, while the highest-density AI racks may require liquid-assisted or direct-to-chip systems.
Does direct-to-chip cooling use refrigerant?
Not necessarily. Many systems circulate treated water or a formulated coolant through cold plates. Refrigerant may still be used elsewhere in the facility, such as in a chiller or DX cooling unit.
Can liquid cooling eliminate chillers?
In some climates and operating conditions, warm-water liquid cooling can increase the number of hours when dry coolers or economizers handle the load without compressor operation. Other facilities will continue needing chillers during hot weather or for loads requiring lower temperatures.
Is immersion cooling the same as direct-to-chip cooling?
No. Direct-to-chip systems circulate coolant through sealed cold plates attached to selected components. Immersion systems place compatible electronic hardware directly into a dielectric fluid.
Why is cooling efficiency so important for AI facilities?
Cooling consumes electrical power that does not directly perform computing work. Efficient heat removal can reduce facility energy use, support higher computing density, control operating costs, and help prevent performance loss from excessive component temperatures.
The Future Is a Coordinated Thermal System
AI data centers are creating new cooling demand because they concentrate unprecedented computing power into individual racks and data halls. Conventional air conditioning will remain part of many facilities, but it can no longer be treated as the only heat-removal strategy.
The emerging model is a coordinated thermal ecosystem: cold plates or immersion fluids at the server, CDUs and pumps at the row or rack level, facility water loops, refrigerant-based equipment where needed, economizers when outdoor conditions permit, and intelligent controls across the entire system.
Successful projects will not simply install “more cooling.” They will select the right cooling method for each heat load while balancing reliability, electricity consumption, water use, refrigerant regulations, maintainability, and future expansion.
0 comments