Data center cooling is the engineered removal of heat generated by servers, networking gear, storage arrays, and power infrastructure inside a data center environment. It is a critical discipline that directly determines uptime, hardware lifespan, and operating cost. Cooling can account for up to 40% of total data center energy use (Source: U.S. Department of Energy, 2026), making it the single largest variable most operators can optimize.
The global data center cooling market was valued at USD 26.31 billion in 2025 and is projected to reach USD 128.31 billion by 2033, growing at a CAGR of 22.3% (Source: Grand View Research, 2025). That growth is fueled by AI workloads, rising rack densities, and regulatory pressure to phase down high-GWP refrigerants under the AIM Act.
This guide breaks down every major cooling approach, from traditional air-based systems to single-phase immersion tanks, and helps you decide what fits your facility. Whether you are building a modular edge data center in a parking lot or retrofitting a colocation hall, the physics and economics covered here apply.
How Does Data Center Cooling Work?
Data center cooling works by transferring heat from IT equipment to a rejection medium, typically outdoor air or a water loop, through a chain of heat exchangers. Air-cooled systems blow conditioned air across server inlets, absorb heat at the exhaust, and reject it outside. Liquid-cooled systems use water or dielectric fluid to capture heat directly at the chip or board level, then move it to an outdoor heat rejector.
Every cooling architecture has the same three stages:
- Heat capture at the source (server inlet, cold plate, or immersion tank)
- Heat transport through air ducts, chilled water pipes, or refrigerant lines
- Heat rejection to the outdoor environment via condensers, cooling towers, or dry coolers
The efficiency of each stage is what separates a facility running a PUE of 1.1 from one limping along at 1.8. Most waste comes from moving air long distances, running compressors harder than necessary, or rejecting heat against high ambient temperatures without economizer assistance.
ASHRAE TC 9.9 publishes thermal guidelines that define recommended inlet temperature ranges for enterprise-class equipment. Operators typically stay within these recommended envelopes to maintain a buffer for temperature fluctuations rather than pushing upper limits. The common misconception that a data center must be kept extremely cold is a relic of older hardware generations. Modern servers tolerate warmer inlet temperatures, and running cooler than necessary simply wastes energy.
What Are the Main Types of Data Center Cooling Systems?
The main types of data center cooling systems are air-based cooling, liquid cooling (direct-to-chip and immersion), evaporative cooling, and hybrid approaches that combine two or more methods. Each type addresses different rack densities, climate zones, and capital constraints. No single method is universally “best” because the right choice depends on workload, location, and budget.
Air-Based Cooling
Air-based cooling remains the most widely deployed method. It uses computer room air conditioners (CRACs) or computer room air handlers (CRAHs) to supply cold air, usually through a raised floor plenum or overhead ducts, into a cold aisle. Hot exhaust air returns to the cooling unit through a hot aisle.
Key variations include:
- Raised-floor supply with perforated tiles for legacy facilities
- In-row or row-based cooling (e.g., Vertiv Liebert CRV units rated 10-50 kW) that shortens the air path and reduces mixing
- Rear-door heat exchangers that capture heat at the rack exhaust before it enters the room
- Containment (hot-aisle or cold-aisle) to prevent recirculation
Air cooling is practical for rack densities up to roughly 15 kW per rack. Beyond that, the volume of air required becomes impractical, and fan energy climbs steeply.
Liquid Cooling
Liquid cooling leverages the fact that water and dielectric fluids transfer heat far more effectively than air. Direct liquid cooling (DLC) is projected to be 25 times more efficient at heat transfer than traditional air cooling, and immersion fluids can be up to 3,000 times more effective than air (Source: HyperFRAME Research, 2026).
Two primary liquid cooling approaches dominate:
- Direct-to-chip (cold plate) cooling: Liquid circulates through cold plates mounted on CPUs and GPUs. The rest of the server is still air-cooled. This is the most common retrofit path for existing facilities.
- Immersion cooling: Servers are submerged in a tank of non-conductive dielectric fluid. Single-phase immersion keeps the fluid below its boiling point; two-phase immersion boils the fluid at the chip surface for even higher heat flux.
The liquid cooling market is expected to roughly double in 2025, reaching close to USD 3 billion in revenue, and scale toward approximately USD 7 billion by 2029 (Source: Dell’Oro Group, 2026). Despite that momentum, only 19% of data centers had implemented liquid cooling by 2025, though 36% planned to adopt it within one to two years (Source: AFCOM State of the Data Center Report, 2025).
Immersion cooling uses dielectric fluids that are non-conductive, non-toxic, and biodegradable. The persistent myth that liquid immersion is unsafe or causes corrosion is unfounded. These engineered fluids actually protect components from oxidation.
Evaporative Cooling
Evaporative cooling for data centers exploits the latent heat of water evaporation to reject heat without running a compressor. A data center cooling tower or direct evaporative cooler can dramatically reduce energy consumption in dry climates. The trade-off is water consumption and limited effectiveness in humid regions.
Modern AI data centers utilizing closed-loop liquid cooling can operate at less than 0.2 L/kWh for water usage, far more efficient than older evaporative designs (Source: HyperFRAME Research, 2026).
Hybrid Cooling
Hybrid approaches combine air and liquid cooling within the same facility. A common pattern is air cooling for low-density racks (storage, networking) and direct-to-chip liquid cooling for high-density GPU racks. This lets operators match cooling investment to actual heat load rather than over-provisioning the entire floor.
Air Cooling vs. Liquid Cooling: How Do They Compare?
Air cooling is simpler to deploy and maintain but hits a practical ceiling around 15 kW per rack, while liquid cooling handles densities of 40 kW and beyond with significantly less energy. Choosing between them depends on rack density, capital budget, operational skill, and long-term workload trajectory.
| Factor | Air Cooling | Direct-to-Chip Liquid | Immersion Cooling |
|---|---|---|---|
| Typical rack density | 5-15 kW | 15-60 kW | 40-140+ kW |
| Heat transfer efficiency | Baseline | ~25x air | Up to 3,000x air |
| PUE potential | 1.3-1.6 | 1.1-1.3 | Below 1.1 |
| Water usage | High (if evaporative) | Moderate (closed loop) | Low to none |
| Capital cost | Lower upfront | Moderate | Higher upfront |
| Retrofit complexity | N/A (existing) | Moderate (piping + cold plates) | High (tanks, fluid, new racks) |
| Maintenance | Filter changes, fan replacement | Leak detection, pump service | Fluid quality monitoring |
| Noise | Higher (fans) | Lower | Lowest |
Rack power densities have jumped from 5-10 kW per rack a decade ago to 15 kW, 20 kW, 30 kW, and even 40 kW in standard deployments. Some NVIDIA H100-based racks reach 140 kW (Source: Tom’s Hardware, 2025). At those densities, air cooling is not physically viable. Leading-edge GPUs are projected to exceed 4,000 W per chip by 2029, which will push liquid cooling from optional to mandatory for AI workloads.
Immersion cooling solutions can enable building data centers for as much as 60% less than traditional air-cooled operations, with continuous savings on energy and maintenance. For operators planning any AI or HPC buildout, liquid cooling should be evaluated from day one rather than treated as a future retrofit.
What Causes Hot Spots, and How Do You Eliminate Them?
Hot spots occur when the air temperature at server inlets approaches or exceeds the ASHRAE recommended inlet temperature of 27 degrees C (80.6 degrees F), typically caused by poor airflow management, inadequate containment, or mismatched cooling capacity. The fix is almost always better airflow engineering, not more cooling tonnage.
Common causes and solutions:
- Bypass airflow: Cold air escaping through cable cutouts or unsealed floor tiles. Seal every penetration with brush grommets or blanking panels.
- Recirculation: Hot exhaust mixing back into the cold aisle. Install hot-aisle or cold-aisle containment.
- Uneven rack loading: A 30 kW rack next to a 5 kW rack creates pressure imbalances. Redistribute loads or add targeted in-row cooling.
- Over-provisioned fans: Running too many cooling units actually causes turbulence and reduces effective airflow to IT equipment. Right-size the cooling plant to the actual load.
- Blanking panel gaps: Empty rack U-spaces allow hot air to short-circuit to the cold aisle. Fill every open U-space.
A counterintuitive truth: providing excessive cooling and airflow can deliver less airflow to servers, not more, while driving up energy costs. Computational fluid dynamics (CFD) modeling or even basic temperature mapping with wireless sensors will identify hot spots faster than adding tonnage.
Recommended Equipment for This Application
– Mitsubishi 24000 BTU Mini Split AC Wall Mount Indoor Unit | WX-Series R454B (MSZ-WX24NL): Ideal for supplemental cooling in edge closets or small server rooms up to 2 kW of IT load, using the next-generation R-454B refrigerant.
– ACiQ 24000 BTU Mini Split Heat Pump AC System | Heats Down to -22 F | Single Zone | R454B: Full heat pump system rated for extreme cold, well-suited for edge deployments that need both cooling and freeze protection year-round.
– Mitsubishi 36000 BTU Mini Split Heat Pump AC Condenser | Inverter GX-Series R454B (MUZ-GX36NL): Higher-capacity inverter condenser for server rooms handling 3-4 kW rack loads where a single outdoor unit simplifies installation.
– ACiQ 48000 BTU Mini Split Heat Pump AC System | Heats Down to -22 F | Single Zone | R454B: 4-ton capacity for larger edge rooms or multi-rack enclosures, with extreme-temperature heat pump operation on low-GWP R-454B.
How Do Refrigerant Regulations Affect Data Center Cooling?
The AIM Act and EPA Section 608 are reshaping refrigerant choices for every data center cooling system that uses a vapor-compression cycle. The AIM Act mandates a phasedown of HFC production and consumption in the United States, pushing operators toward low-GWP alternatives like R-454B. EPA Section 608 prohibits intentional venting of refrigerants and requires certified technicians for all servicing.
As of January 23, 2026, EPA regulations modified refrigerant management requirements for substitute refrigerants including HFCs, while existing requirements for ozone-depleting substances remain in place (Source: EPA, 2026). For practical purposes, this means:
- Any new split system, packaged unit, or chiller installed in a data center environment should use a low-GWP refrigerant. R-454B is the most common drop-in path for residential and light commercial equipment.
- Technicians servicing refrigerant circuits must hold EPA Section 608 certification. The EPA’s Section 608 regulations detail the specific requirements.
- Leak detection and reporting obligations have tightened, especially for systems with charges above 50 pounds.
For a deeper look at how A2L refrigerants like R-454B interact with data center compliance requirements, see our guide to A2L refrigerants and EPA compliance for modular data centers.
For edge and small-footprint data centers, ductless mini splits using R-454B offer a practical, compliant cooling path. A unit like the ACiQ 24000 BTU extreme-cold heat pump provides both cooling and heating on R-454B, keeping the facility compliant with the AIM Act phasedown schedule while operating in ambient temperatures down to -22 degrees F.
How Can You Reduce Data Center Cooling Energy Costs?
The most effective way to reduce cooling energy costs is to raise server inlet temperatures to the upper end of ASHRAE recommended ranges, implement free cooling or economizer hours, and match cooling capacity to actual IT load rather than nameplate ratings. These three steps alone can cut cooling energy by 20-40% without any capital expenditure on new equipment.
Beyond those fundamentals, a cost-reduction roadmap looks like this:
Economizer and Free Cooling Modes
When outdoor air or water temperatures are low enough, economizers bypass compressors entirely. In temperate climates, this can cover thousands of hours per year. Both air-side and water-side economizers are well-documented strategies endorsed by ASHRAE’s data center energy efficiency guidelines.
Variable-Speed Drives and Inverter Compressors
Constant-speed compressors and fans consume the same energy at 30% load as at 90% load. Inverter-driven systems modulate output to match real-time demand. This is why inverter mini splits are a strong fit for edge deployments: they throttle down during low-utilization periods instead of cycling on and off.
Containment and Airflow Optimization
Hot-aisle containment alone can improve cooling efficiency by 10-20% by preventing hot and cold air from mixing. Combine containment with blanking panels, sealed cable penetrations, and variable-speed fan walls for compounding gains.
Monitor PUE and WUE Continuously
You cannot optimize what you do not measure. Power Usage Effectiveness (PUE) remains the industry-standard metric, and tracking it monthly or in real time reveals seasonal patterns and degradation trends. Platforms like Schneider EcoStruxure and Vertiv’s Liebert iCOM integrate power metering and cooling telemetry into a single dashboard. The Uptime Institute and The Green Grid both publish benchmarking resources that help operators compare their performance against industry peers.
For a full breakdown of PUE calculation, benchmarks, and improvement strategies, see our complete guide to Power Usage Effectiveness.
What Standards and Codes Govern Data Center Cooling?
ASHRAE TC 9.9, NFPA 75, EPA Section 608, and the AIM Act form the primary regulatory and standards framework for data center cooling in the United States. Compliance is not optional: state and local building codes typically mandate adherence to NFPA 75 for fire protection, and federal law governs refrigerant handling.
Here is a summary of the key standards every data center cooling designer and operator should know:
- ASHRAE TC 9.9 Thermal Guidelines: Defines recommended and allowable temperature and humidity envelopes for data processing environments. Most operators design to the recommended range for a safety buffer.
- NFPA 75: Covers fire protection of IT equipment areas, including requirements for HVAC shutdown on smoke detection, fire suppression compatibility, and equipment room construction.
- EPA Section 608: Governs refrigerant recovery, recycling, and reclamation. Requires technician certification and prohibits knowing release of refrigerants during equipment maintenance or disposal.
- AIM Act: Directs the EPA to phase down HFC production and consumption. Affects the choice of refrigerant in all new CRAC, CRAH, chiller, and split system installations.
- Uptime Institute Tier Classification: While not a code, Tier requirements (I through IV) dictate cooling redundancy levels. A Tier III facility, for example, requires concurrent maintainability, meaning you must be able to service any cooling component without taking the data center offline.
- Open Compute Project (OCP): Publishes open-source hardware designs including rack and cooling specifications that influence how liquid cooling manifolds and connections are standardized.
For facilities using immersion cooling, check local fire marshal interpretations of NFPA 75 regarding dielectric fluids. Some jurisdictions require additional documentation to classify these fluids as non-combustible. The DOE’s data center energy efficiency resources provide supplementary guidance on efficiency targets that align with federal sustainability mandates.
How Do You Size Cooling for Edge and Small Data Centers?
Sizing cooling for edge and small data centers starts with calculating the total IT heat load in watts, adding infrastructure overhead (UPS losses, lighting, people), and then selecting equipment with enough sensible cooling capacity to reject that total load with a safety margin of 15-20%. A common mistake is sizing to nameplate BTU ratings without accounting for the sensible heat ratio.
Here is a simplified sizing workflow:
- Sum the IT load. Add up the maximum draw of all servers, switches, and storage. For planning purposes, use 80% of nameplate if actual measurements are unavailable.
- Add non-IT loads. UPS inefficiency (typically 3-8% of IT load), lighting, and any personnel heat (approximately 400 BTU/hr per person).
- Convert to BTU/hr. Multiply total watts by 3.412 to get BTU/hr.
- Apply a safety margin. Add 15-20% for future growth and measurement uncertainty.
- Select equipment. Match the total BTU/hr requirement to the sensible cooling capacity of the unit, not the total capacity. Data center loads are almost entirely sensible heat; latent capacity is largely irrelevant.
For example, a two-rack edge deployment with 5 kW of IT load plus 500 W of overhead equals 5,500 W, or roughly 18,770 BTU/hr. With a 20% margin, you need about 22,500 BTU/hr of sensible cooling. A Mitsubishi WX-Series 24000 BTU wall-mount unit on R-454B fits this scenario well, with the inverter compressor throttling down during off-peak hours to save energy.
For larger edge rooms in the 10-15 kW IT load range, a 48,000 BTU system provides adequate headroom. When the facility is in a climate with extreme winters, a heat pump configuration ensures the room does not freeze during low-utilization periods, protecting both IT hardware and fire suppression systems.
Always verify that your selected equipment meets local code requirements, especially regarding A2L refrigerant safety classifications and installation clearances.
What Does the Future of Data Center Cooling Look Like?
The future of data center cooling is liquid-first, driven by AI chip power consumption that air alone cannot manage, and shaped by refrigerant regulations that are eliminating high-GWP options on a fixed legislative timeline. Operators who plan for liquid cooling infrastructure now, even if they deploy air cooling today, will avoid costly retrofits within five years.
Liquid cooling accounted for 84% of all data center cooling solution deals in 2025, driven by 139 transactions focused on direct-to-chip and immersion cooling for high-density AI workloads (Source: Net Zero Insights, 2025). That investment signal is hard to ignore.
Several trends are converging:
- Rack densities will keep climbing. With GPUs projected to exceed 4,000 W by 2029, even mid-tier deployments will routinely exceed what air can handle.
- Refrigerant transitions are mandatory. The AIM Act phasedown is not a suggestion. Equipment purchased today with high-GWP refrigerants will face servicing constraints within a decade.
- Water efficiency matters more. Evaporative cooling data center designs are being re-evaluated in water-stressed regions. Closed-loop liquid cooling at less than 0.2 L/kWh represents the efficiency target for new builds.
- Edge deployments multiply. As AI inference moves closer to end users, thousands of small facilities will need cooling solutions that are simple to install, remotely manageable, and code-compliant without a dedicated facilities team on site.
For a comprehensive look at how cooling, power, and compliance intersect in modular builds, read our research concept paper on modular edge data centers.
Browsing options? Explore AC Direct’s full lineup of ductless mini splits, or request a sizing consultation.
Frequently Asked Questions
How does data center cooling work?
Data center cooling works by capturing heat at server components, transporting it via air ducts, chilled water pipes, or refrigerant lines, and rejecting it to the outdoor environment through condensers, cooling towers, or dry coolers. The goal is to maintain server inlet temperatures within ASHRAE-recommended ranges to prevent thermal throttling and hardware failure.
What is the best cooling system for a data center?
The best cooling system depends on rack density and workload. Air cooling works well below 15 kW per rack. Direct-to-chip liquid cooling suits 15-60 kW racks. Immersion cooling handles 40-140 kW and above. Most modern facilities use a hybrid approach, matching the cooling method to each rack’s actual heat output.
What are the differences between liquid cooling and air cooling?
Liquid cooling transfers heat 25 to 3,000 times more effectively than air, supports much higher rack densities, and uses less energy per watt of heat removed. Air cooling is simpler to install and maintain but becomes impractical above 15 kW per rack. Liquid cooling costs more upfront but delivers lower long-term operating expenses.
How can I reduce energy costs associated with data center cooling?
Raise server inlet temperatures to the upper end of ASHRAE recommended ranges, use economizer or free cooling when outdoor conditions allow, install variable-speed inverter-driven equipment, and implement hot-aisle or cold-aisle containment. Continuously monitor Power Usage Effectiveness to track progress and catch efficiency degradation early.
Is immersion cooling safe for servers?
Yes. Immersion cooling uses engineered dielectric fluids that are non-conductive, non-toxic, and biodegradable. These fluids do not cause corrosion and actually protect components from oxidation. Modern immersion systems are designed with leak containment, fluid monitoring, and fire safety measures that meet or exceed NFPA 75 requirements.
How often should data center cooling systems be inspected?
Most operators inspect cooling systems quarterly, with monthly monitoring of key metrics like supply temperature, return temperature, refrigerant pressures, and filter condition. EPA Section 608 requires leak inspections for systems containing regulated refrigerants above certain charge thresholds. Preventive maintenance schedules should follow both manufacturer recommendations and local code requirements.
What causes hot spots in a data center?
Hot spots are typically caused by bypass airflow escaping through unsealed cable cutouts, recirculation of hot exhaust into cold aisles, uneven rack loading, or missing blanking panels. The fix is better airflow management, not simply adding more cooling capacity. Temperature mapping or computational fluid dynamics modeling can pinpoint problem areas quickly.
Do data centers need to be kept extremely cold?
No. The idea that data centers must run at near-freezing temperatures is outdated. Modern hardware operates reliably within ASHRAE-recommended inlet temperature ranges up to 27 degrees C (80.6 degrees F). Over-cooling wastes energy and increases operating costs without improving reliability. Most efficient facilities run warmer than legacy designs.