The world’s data centers now consume more electricity than the entire United Kingdom. As AI workloads push power densities to unprecedented levels, operators are running out of conventional cooling options, and phase change materials are stepping into the gap.
The statistics are staggering. Global data center power consumption exceeded 500 terawatt-hours in 2025, according to International Energy Agency estimates, and is projected to potentially double by 2028 if current AI-driven trends continue. The primary driver is not the servers themselves, but the cooling infrastructure required to keep them operational. In a typical facility, 30 to 40 percent of total power consumption goes to cooling, and in older or less efficient sites, that figure can exceed 50 percent.
The problem is intensifying at an alarming rate. The shift from general-purpose computing to AI and machine learning workloads has created a step-change in power density. A standard rack of traditional servers draws 5 to 10 kilowatts. A rack of modern AI accelerators, the GPUs that train large language models and run inference at scale, draws 40 to 100 kilowatts. Some next-generation configurations push past 120 kW per rack. Conventional computer-room air conditioning, and even raised-floor hot-aisle/cold-aisle arrangements, were never designed for these densities.
The Cooling Crisis: Why Traditional Approaches Are Hitting Their Limits
Data center cooling has evolved through several generations. First-generation facilities used perimeter CRAC units blowing chilled air under a raised floor. Second-generation introduced hot-aisle and cold-aisle containment to prevent air mixing. Third-generation moved to in-row cooling and rear-door heat exchangers. Fourth-generation, now being deployed at scale, uses direct-to-chip liquid cooling and immersion cooling.
Each generation improved efficiency and density capacity, but each also added cost, complexity, and points of failure. And each generation has struggled to keep pace with the relentless increase in compute density. The fundamental problem is thermal physics: removing 100 kilowatts from a single rack requires either enormous volumes of air, which is energy-intensive and physically impractical, or liquid-based systems, which introduce leakage risks and plumbing complexity.
Even with liquid cooling, the heat has to go somewhere. The chillers, cooling towers, and heat rejection systems on the facility side consume enormous amounts of energy, particularly in hot climates where the temperature differential between the server exhaust and ambient air is small. This is where phase change materials enter the picture, not as a replacement for liquid cooling, but as a thermal energy storage and buffering layer that fundamentally changes the cooling demand profile.
PCM in the Data Center: Three Architectural Approaches
1. Rack-Level Thermal Buffering
The most immediate application of PCM in data centers is at the rack level. PCM panels or modules, integrated into the rack structure or positioned between server rows, absorb heat during peak load periods and release it during off-peak hours. This is particularly valuable for facilities with variable workloads, where AI training jobs may run at full power for hours and then idle while models are evaluated or data is prepared.
The PCM essentially decouples the instantaneous server heat generation from the facility cooling load. During a training run, the PCM absorbs excess heat that the active cooling system cannot immediately remove, preventing server inlet temperatures from breaching safe limits. During the subsequent idle period, the facility cooling system catches up, re-solidifying the PCM and resetting the thermal buffer. This allows the facility cooling infrastructure to be sized for average rather than peak load, reducing chiller capacity, pump sizes, and electrical infrastructure costs.
A server rack drawing 80 kW during peak AI training generates roughly 80 kW of heat. If the facility’s cooling system can handle 60 kW at steady state, the PCM must absorb the 20 kW differential. With a PCM latent heat capacity of 200 kJ/kg, a 200-kilogram PCM module can absorb 40,000 kJ, or 11.1 kWh, of excess heat. That translates to roughly 33 minutes of peak-load buffering, long enough to cover most transient workload spikes without triggering thermal throttling.
2. Facility-Scale Thermal Energy Storage
At a larger scale, PCM thermal energy storage systems can be integrated into the facility’s chilled water loop, functioning as a thermal battery. During nighttime hours, when ambient temperatures are lower and electricity rates are cheaper, the facility overcools the PCM storage tank, freezing the material and storing thermal capacity. During daytime peak hours, the PCM melts, absorbing heat from the chilled water loop and reducing the load on electric chillers.
This approach, known as peak shaving, has significant financial implications. In markets with time-of-use electricity pricing, the differential between off-peak and peak rates can exceed 4:1. A data center that can shift 30 percent of its cooling load from peak to off-peak hours can realize substantial operational savings. A 10-megawatt facility in California, where peak rates can reach $0.35 per kWh and off-peak rates drop to $0.08, could save over $500,000 annually through PCM-based load shifting.
PCM thermal storage also provides a critical resilience benefit. In the event of a chiller failure or power disruption, the PCM system provides a thermal buffer that can maintain safe temperatures for 30 to 90 minutes, depending on system sizing, giving backup generators time to start and stabilize. This is far cheaper and more reliable than battery backup for cooling loads, which would require enormous battery banks to achieve the same duration.
3. Chip-Level and Component-Level PCM
At the most granular level, PCMs are being explored for direct integration with high-power components. Micro-encapsulated PCM, applied as a thermal interface material or embedded in heat sink structures, can absorb transient thermal spikes at the chip surface. AI workloads are characterized by bursty power profiles: inference requests arrive in waves, causing GPU temperatures to spike by 10 to 15 degrees in milliseconds before settling.
Conventional cooling responds to these spikes by ramping up fans or pumps, but the response time of mechanical systems is measured in seconds, while thermal transients at the chip level occur in milliseconds. A thin PCM layer, with its near-instantaneous phase transition response, can smooth these micro-transients, keeping chip temperatures stable and reducing thermal stress that contributes to solder fatigue and component failure.
AI racks at 100+ kW density create a problem that goes beyond cooling capacity: thermal spot loading. Even within a single rack, power distribution is not uniform. GPUs in the center of the rack run hotter than those at the edges. Network switches may idle while compute nodes run at full tilt. Conventional cooling treats the rack as a single thermal zone, leading to overcooling of some components and undercooling of others.
PCM, applied at the component or server level, creates distributed thermal buffering that responds to local heat generation rather than average rack temperature. This granular thermal management can reduce the total cooling energy required while improving the reliability of individual components.
Representative Scenario: PCM Integration in a Hyperscale Facility
Consider a representative 50-megawatt AI training facility housing approximately 500 racks of GPU accelerators drawing 80-100 kW each, struggling with thermal throttling during sustained training runs lasting 12-18 hours. This scenario reflects the operating conditions reported by multiple hyperscale operators facing similar challenges across North America.
Installing PCM panels with a melting point of 28 degrees Celsius between rack rows, covering roughly 40 percent of the server hall floor area, would provide approximately 25 minutes of peak-load thermal buffering at full rack power. The PCM panels, using a paraffin-graphite composite with a latent heat capacity of 185 kJ/kg, represent a commercially available configuration.
Thermal modeling of this configuration projects a 20-25 percent reduction in chiller energy consumption, a 30-35 percent reduction in peak cooling demand, and elimination of thermal-throttling events that would otherwise interrupt training runs. The facility’s PUE could drop from approximately 1.42 to 1.31. At this performance level, the PCM system is projected to achieve payback in 14-18 months through energy savings alone, with additional unquantified benefits from improved training run reliability and reduced component thermal stress.
The Economics: When Does PCM Make Sense?
The business case for PCM in data centers hinges on several variables: electricity prices, workload profile, climate, and facility age. PCM thermal storage delivers the highest ROI in facilities that meet one or more of the following criteria:
- High time-of-use electricity rate differentials. Markets like California, the UK, and parts of Australia where peak-to-off-peak ratios exceed 3:1 offer the strongest payback for load-shifting applications.
- Variable or bursty workloads. AI training facilities, financial trading data centers, and gaming/CDN sites with pronounced peak-to-average ratios benefit most from rack-level thermal buffering.
- Hot climates with limited free-cooling hours. Facilities in the Middle East, Southeast Asia, and the southern United States, where ambient temperatures rarely drop low enough for economizer operation, can use PCM to extend the effective free-cooling window.
- Aging facilities approaching thermal limits. Rather than undertaking a costly chiller upgrade or expansion, PCM can add effective cooling capacity within the existing infrastructure footprint.
- Sustainability mandates. Companies with aggressive carbon reduction targets can use PCM to reduce cooling-related emissions without wholesale facility redesigns.
For new construction, the calculus is different. PCM integration adds 5-15 percent to cooling system capital costs but can reduce total facility cost by downsizing chillers, cooling towers, and electrical infrastructure. Lifecycle savings over a 10-15 year facility lifespan typically run into the millions for large facilities.
Challenges and Limitations
PCM is not a silver bullet, and several challenges must be addressed for widespread adoption. The first is space. PCM modules occupy floor space that could otherwise house servers, and in facilities where real estate is measured in dollars per square foot per month, the opportunity cost is real. This is particularly acute in existing facilities retrofitted with PCM panels, where the panels may reduce server hall capacity by 5-10 percent.
The second challenge is fire safety. Paraffin-based PCMs, while effective, are combustible. Data center operators are understandably cautious about introducing flammable materials into facilities filled with expensive electronic equipment. Non-combustible alternatives, including salt hydrates and clay-based composites, exist but typically offer lower latent heat capacity or face issues with corrosion and phase separation. Glacier Coolant is developing PCM formulations designed to meet data center fire safety requirements, leveraging non-flammable base materials and composite structures that maintain high thermal storage density while complying with building fire codes.
The third challenge is long-term reliability. PCMs undergo thousands of phase transitions over their operational life, and material degradation, phase separation, and leakage are legitimate concerns. Quality matters enormously, and not all PCM products on the market meet the reliability standards required for mission-critical data center applications. Operators should demand cycle testing data, with a minimum of 10,000 cycles demonstrated without significant performance degradation.
The Regulatory Landscape: PUE Targets Push Adoption
Government regulations are increasingly forcing the issue. The European Union’s Energy Efficiency Directive, updated in 2024, requires data centers with IT power consumption above 1 MW to publicly report their PUE and meet progressive efficiency targets. Several member states have gone further: Germany’s Energy Efficiency Act mandates a PUE of 1.2 for new facilities by 2028, and Ireland has effectively banned new data center construction in the Dublin region unless the facility meets stringent sustainability criteria.
In the United States, while federal mandates are lighter, state and local regulations are tightening. California’s Title 24 building energy standards now include provisions for data center efficiency, and several utility companies offer substantial rebates for technologies that reduce peak cooling demand, including thermal energy storage.
These regulatory pressures are transforming PCM from an optional efficiency measure to a compliance necessity. Facilities that cannot achieve PUE targets through conventional means are increasingly looking at PCM as a cost-effective path to compliance without wholesale infrastructure replacement.
Looking Forward: The Thermal Storage Data Center
The data center of 2030 will likely look fundamentally different from today’s facilities. Liquid cooling, once a niche technology, is becoming standard for high-density deployments. Renewable energy, particularly solar and wind, is increasingly integrated directly into facility power supplies. And thermal energy storage, using PCM and other technologies, is emerging as the bridge between variable renewable generation and constant compute demand.
The vision is compelling: solar panels generate electricity during the day, powering servers and simultaneously charging PCM thermal storage. At night, the PCM releases stored cooling, reducing the facility’s nighttime grid power draw. The result is a data center that can operate with a smaller grid connection, lower energy costs, and a smaller carbon footprint, all while maintaining the thermal stability that AI workloads demand.
This is not a distant dream. The technologies exist today. The economics are improving with every electricity rate increase and every new AI workload that pushes densities higher. The question is not whether PCM will be adopted in data centers, but how quickly, and by whom. The operators that move first will gain a structural cost advantage that compounds over the life of their facilities. Those that wait may find themselves paying a premium, both in energy costs and in the competitive disadvantage of a facility that cannot support the workloads of tomorrow.