lennox data center solutions blog

Discover the latest in Data Center Cooling technologies, events and industry trends shaping the future of Data Center cooling. Stay informed with expert insights, awards and recognitions, real-world case studies and valuable resources from Lennox Data Center Solutions

AI Is Changing the Heat Equation: What Rising Power and Density Mean for Data Center Cooling

Updated: November 13, 2026

AI is increasing data center power consumption and concentrating more heat in each rack. U.S. data centers could grow from about 4.4% of national electricity use in 2023 to as much as 12% by 2028, while AI rack densities range from about 40–130 kW.  

Air cooling remains valuable, but liquid cooling offers a more direct pathway for removing concentrated heat. The future is likely to be hybrid, with air and liquid operating as one thermal architecture that responds to changing workloads and supports higher density without overbuilding. 

LinkedIn
AI Is Changing the Heat Equation: What Rising Power and Density Mean for Data Center Cooling

AI growth is increasing data center power consumption, concentrating heat at the rack and raising the need for adaptable, integrated cooling.

Artificial intelligence is changing more than the computing equipment installed in data centers. It is changing data center power consumption; how much computing power can be concentrated in each rack and how the resulting heat must be managed. 

More computing requires more electrical power, and almost all power consumed by IT equipment becomes heat. Concentrating that power in fewer racks makes cooling more difficult. 

For operators, the question is no longer simply, “How do we cool the building?” It is becoming, “How do we design power and cooling infrastructure for increasingly dense and dynamic computing environments?” 

AI growth is becoming a power-infrastructure challenge 

According to Lawrence Berkeley National Laboratory’s 2024 United States Data Center Energy Usage Report, U.S. data centers consumed about 176 terawatt-hours of electricity in 2023, or 4.4% of total U.S. electricity use. The report projects consumption could reach approximately 325–580 terawatt-hours by 2028, equal to about 6.7%–12% of U.S. electricity use. 

AI is not the only source of growth, but it is an important driver. Model training requires substantial computing capacity, while inference adds continuing demand as AI use expands. 

The U.S. Department of Energy reports that hyperscale connection requests can range from 300 MW to 1,000 MW or more, with lead times of only one to three years. Meeting that demand can require new substations, transformers, transmission and generation resources. 

Power availability can shape where and when a facility is built. Electricity used by cooling equipment cannot power IT systems, so thermal efficiency affects how well a site uses available power. These dynamics add a whole new dimension to the importance of effective cooling in determining feasibility and computing capacity of data center projects.

Higher power demand becomes concentrated heat 

Rack density, generally expressed in kilowatts per rack, indicates the power consumed by equipment in a rack. Almost all of that power becomes heat that must be removed. 

AI racks can require far more power than traditional enterprise racks because they combine numerous CPUs, GPUs and supporting components. JLL places AI rack densities at approximately 40–130 kW, with future chip-driven requirements potentially reaching about 250 kW per rack. ASHRAE has also discussed a transition from racks around 120 kW toward several hundred kilowatts. 

One-megawatt racks are not typical today. They are emerging research, development and validation targets, but their inclusion in the U.S. Department of Energy’s COOLERCHIPS 1.5 initiative signals where thermal engineering may be heading. 

Current hardware already illustrates the transition. For example, NVIDIA describes the GB200 NVL72 as a liquid-cooled, rack-scale platform integrating 36 Grace CPUs and 72 Blackwell GPUs. Its power, networking and thermal requirements must be considered at both rack and facility level. Processor selection can no longer be separated from electrical distribution, piping, airflow, coolant distribution and heat rejection. 

Why air cooling alone becomes harder at higher densities 

Air cooling remains familiar, flexible and effective. Fans move cool air through servers, containment limits air mixing, and economizers can reduce mechanical refrigeration in favorable outdoor conditions. 

Air cooling is not obsolete, but concentrated heat creates limits. Higher rack density requires more air through confined spaces, increasing fan energy, airflow-path needs and control demands. Reaching the right components can become difficult or inefficient. 

Liquid cooling moves the cooling medium closer to the principal heat source. Because liquid can carry more heat per unit of volume than air, it can transport concentrated thermal loads with much smaller flow volumes. In direct-to-chip cooling systems, cold plates attached to CPUs and GPUs absorb heat and carry it through a liquid loop to the facility heat-rejection system. A liquid cooling data center does not eliminate heat. It creates a more direct pathway for removing it. 

The future is likely to be hybrid, not air versus liquid 

A rack includes more than CPUs and GPUs. Memory, storage, networking equipment, power supplies and other components have different cooling requirements. Some may remain air cooled even when high-power processors use liquid cooling. 

NVIDIA rack-scale systems illustrate the combination: cold plates and liquid manifolds support high-power processors, while some networking and storage components remain air cooled. Across a facility, AI training zones may use direct-to-chip cooling while lower-density areas remain air cooled. Existing sites can add liquid-cooled rows, pods or localized units without converting the entire building. 

The challenge is integration. Air and liquid systems cannot be planned as separate environments. Their temperatures, flow rates, controls and heat-rejection requirements must work together. Even when liquid captures most processor heat, a residual air-cooling load remains. 

Cooling must follow changing AI workloads 

AI workloads add another variable: thermal load changes as training jobs begin, finish or move and inference demand varies. Some tasks may also shift in time or location when operating and latency requirements allow. 

Power consumption and heat generation generally change with computing activity, but cooling demand does not always change instantly or at the same rate. Servers, coolant, pipes, heat exchangers and the data hall retain heat temporarily. The goal is not instantaneous one-to-one matching. It is a stable, controlled thermal response. 

Variable-speed fans and pumps can adjust output, while sensors monitor temperature, pressure and coolant flow. Controls can stage equipment for actual conditions instead of running every component at full output. In hybrid systems, air and liquid must operate as one thermal architecture. 

Prepare for higher density without overbuilding 

No operator can predict every future processor or workload. Designing only for today can limit choices, while designing an entire facility for one-megawatt racks could add unnecessary cost and inefficiency if that density does not materialize. 

The more practical objective is adaptability: modular cooling capacity, scalable liquid distribution, air cooling for lower-density and residual loads, targeted rack- or row-level cooling, variable-speed equipment, coordinated controls and heat-rejection systems designed for multiple operating modes. 

Energy and water use should also be evaluated across the complete cooling chain. “Waterless cooling” can mean no evaporative water consumption, no facility-water loop or no water in one part of a system. These meanings are not interchangeable, so clear boundaries are essential when comparing solutions. 

Key hybrid-design questions

  • How will coolant be distributed to each rack? 

  • How much rack heat will liquid capture, and how will air remove the remainder? 

  • How will both systems respond as AI workloads change? 

  • Can the infrastructure support different and evolving rack densities? 

Where PhaseFlow fits into the larger picture?

AI data centers need thermal systems that respond to rising energy demand, higher rack density and mixed air- and liquid-cooled equipment. 

PhaseFlow can be positioned as part of that integrated thermal-management pathway, rather than as a replacement for every cooling technology. The product discussion should define which part of the cooling chain PhaseFlow supports, how it responds to changing loads, how it interacts with air- and liquid-cooled IT equipment and what its water-use boundary includes. Performance and capacity claims should be supported by validated product data. 

The future of data center cooling will not be determined by one technology or one peak-capacity number. It will depend on how well the system connects computing demand, electrical capacity, rack-level heat capture, controls and facility heat rejection. 

AI is increasing the heat that data centers must manage. More importantly, cooling is becoming part of the infrastructure that determines how far the computing environment can grow.

How AI Is Changing Data Center Power and Cooling