A modern data center is, at its core, a system for converting enormous quantities of electricity into computing. Servers, storage systems and networking equipment consume power to process and move data. Almost immediately, that electrical energy becomes heat, which must then be removed to keep the equipment within safe operating conditions.

That creates an unusual operating equation. More computing does not simply require more electricity for processors. It also increases the amount of heat that cooling equipment must handle, while placing greater demands on switchgear, uninterruptible power supplies, pumps, fans and other supporting infrastructure.

The scale is becoming increasingly significant. Lawrence Berkeley National Laboratory estimates that U.S. data centers consumed 192 terawatt-hours of electricity in 2024, equal to about 4.7 percent of total U.S. electricity consumption. Its June 2026 update uses a bottom-up model based on equipment shipments, per-device electricity use, cooling simulations, facility types and locations. Under its Reference Case, U.S. data-center consumption reaches 649 TWh in 2030, or 11.8 percent of forecast U.S. electricity use. The report also presents a wider uncertainty range rather than treating that forecast as a guaranteed outcome.

For operators, however, the national totals are less important than what happens inside an individual facility. The central question is how much electricity reaches the computing equipment and how much additional energy is required to keep that equipment powered, protected and cool.

The largest fundamental energy requirement is the IT equipment itself.

Servers draw electricity through their processors, memory, accelerators and other components. Storage equipment and network switches add further demand. Artificial intelligence is increasing this load because accelerated computing systems can concentrate far more electrical power into a relatively small physical area.

Uptime Institute's 2026 Global Data Center Survey found that average rack densities continued to increase and that more operators were reporting peak rack densities of 30 kilowatts or more. Importantly, these are peak densities reported by operators, not evidence that every rack in a modern data center consumes 30 kW. Uptime said average modal rack densities were still rising gradually.

Berkeley Lab's latest U.S. model reaches a similar conclusion from a different direction. Its Reference Case estimates that the average annual power draw of installed eight-GPU AI servers rises above 4 kW per server through 2026. The model combines rated power, idle power and utilization rather than assuming that every machine continuously operates at its maximum nameplate rating.

This distinction matters for cost. A data center designed for 100 megawatts of available capacity does not necessarily consume 100 MW every hour. Actual electricity use depends on how much equipment is installed, its utilization, its power characteristics and the facility systems supporting it.

Yet once the IT equipment consumes electricity, the resulting heat has to go somewhere.

Cooling a data center is much more involved than simply installing air conditioners.

Traditional facilities often use fans to move cold air toward server racks and carry hot exhaust air away from them. Computer-room air handlers, chillers, pumps, heat exchangers and cooling towers can all form part of the thermal chain. Some designs make greater use of outdoor conditions or economization to reduce mechanical refrigeration when temperatures allow.

Cooling can therefore represent a significant share of energy consumption, but there is no universal percentage that applies to every data center.

The U.S. Department of Energy has previously noted that cooling can account for as much as 33 to 40 percent of total data-center energy use in some facilities. That figure describes an upper-end cooling burden and should not be interpreted as the current industry average for every site. Cooling requirements vary widely with equipment density, facility design, climate, operating temperatures and cooling technology.

Berkeley Lab's newer national modeling provides a broader measure. It estimates that all facility infrastructure, including cooling and electrical distribution losses, accounted for about 31 percent of U.S. data-center electricity use in 2024, down from 36 percent in 2018. Cooling dominates that infrastructure category, but the figure also includes other non-IT electrical loads, so it is not a cooling percentage on its own.

That distinction is important when evaluating operating costs. A data center's electricity bill covers both productive computing and the infrastructure required to support it.

The industry's best-known efficiency metric is Power Usage Effectiveness, usually abbreviated as PUE.

PUE is calculated by dividing total facility energy use by the energy consumed by IT equipment. A theoretical PUE of 1.0 would mean every unit of electricity entering the data center goes directly to IT equipment. Real facilities have higher values because cooling, electrical conversion, pumps, fans and other infrastructure also consume power.

Uptime Institute's 2026 Global Data Center Survey reported an industry-wide annual average PUE of 1.52. That figure comes from Uptime's long-running survey of data-center operators and gives each participating facility a place in the headline average. Uptime also calculated a capacity-weighted average that gives greater influence to larger facilities. On that basis, PUE was substantially lower at 1.36, reflecting the generally better efficiency of large, newer sites.

The arithmetic shows why small efficiency differences matter at data-center scale.

Consider a facility carrying a constant 1 MW IT load. At a PUE of 1.52, the total facility load would be 1.52 MW. The additional 0.52 MW represents cooling and other infrastructure energy rather than computing itself. Running continuously for a year would require about 13.3 GWh of total electricity.

At a PUE of 1.36, the same 1 MW IT load would require 1.36 MW at the facility level, or about 11.9 GWh annually. The difference is roughly 1.4 GWh per year for every continuously operating megawatt of IT load.

The financial value of that difference depends entirely on the electricity tariff. At the 1.52 PUE level, every one-cent-per-kilowatt-hour change in the volumetric electricity price changes the annual energy cost by roughly $133,000 for each constant 1 MW of IT load. Actual utility bills may also contain demand, capacity, transmission and other charges, so the calculation represents only the energy component.

PUE itself also has limits. It does not measure whether the servers are doing useful work efficiently, nor does it account directly for water consumption. A poorly utilized server can sit inside a highly efficient facility and still waste energy.

Data-center efficiency statistics can appear contradictory when different methodologies are placed side by side.

Berkeley Lab estimates that the national average PUE associated with U.S. data-center electricity use improved from 1.55 in 2018 to 1.45 in 2024. Its model reflects the changing mix of facility types and the movement of computing into highly efficient data centers.

Uptime's 2026 survey, by contrast, reports an annual average PUE of 1.52 but a capacity-weighted figure of 1.36.

These values should not be treated as competing measurements of exactly the same population. Berkeley Lab is producing a modeled estimate of U.S. data-center electricity consumption using assumptions about facility types and equipment. Uptime is compiling reported PUE figures from its international operator survey. Weighting facilities equally produces a different result from weighting them according to their capacity.

Even within Berkeley Lab's model, facilities serving AI equipment can have lower PUE values than the overall national fleet. The laboratory estimates an average PUE of about 1.145 in 2024 for facilities serving AI equipment, with only a modest decline through 2030. This reflects the concentration of AI systems in newer facilities designed around efficient power and cooling infrastructure.

The broader lesson is that PUE needs context. Facility age, size, climate, IT density and methodology can materially change the number.

As rack power rises, moving heat through air becomes increasingly difficult.

Direct liquid cooling brings coolant much closer to the components generating the heat. Cold plates can be attached to processors and accelerators, with liquid carrying heat through pipes to coolant distribution units and ultimately to the facility's heat-rejection system.

Liquids can carry far more heat than air, and pumping liquid can require less energy than moving the enormous volumes of air needed for very dense computing environments. The U.S. Department of Energy notes that direct liquid cooling can reduce both PUE and water usage in some configurations, although the result depends on the complete cooling design.

Adoption is increasing, but liquid cooling has not replaced conventional air cooling across the industry. Uptime Institute's 2025 cooling survey found that 75 percent of 512 responding owner and operator organizations used perimeter air cooling, while 22 percent reported direct liquid cooling. Respondents could select more than one technology, so the percentages describe technology use rather than mutually exclusive market shares.

Uptime's subsequent 2026 Cooling Systems Survey, which included 1,034 industry respondents, continued to focus on expanding direct-liquid-cooling adoption. The shift is being driven particularly by high-density AI systems, but facilities often remain hybrid environments in which conventional servers are air cooled while selected high-density racks use liquids.

Liquid cooling also introduces costs of its own. Operators need pumps, piping, coolant distribution units, controls, leak management and maintenance procedures. Retrofitting an existing air-cooled facility can be more difficult than designing liquid cooling into a new building. Uptime has noted that large-scale deployment has also required the industry to develop more consistent practices around coolant chemistry, maintenance and system architecture.

The economic case therefore depends on more than the cooling system's electricity consumption. Operators have to consider installation costs, usable rack density, reliability, maintenance and the amount of computing capacity that can be supported within a building's power and thermal limits.

Cooling design can also shift costs between electricity and water.

Evaporative cooling can reduce the electricity required for mechanical refrigeration under suitable conditions because evaporation is an efficient way to reject heat. The trade-off is water consumption. Cooling towers lose water through evaporation and also require additional water to control mineral concentrations within the system.

Dry cooling avoids or reduces that direct water requirement but can have different energy characteristics, particularly as outdoor temperatures rise. The best approach therefore depends on climate, water availability, electricity prices and facility design.

Research published by Berkeley Lab in 2025 found that data-center workload water consumption can vary by more than four orders of magnitude across different conditions. The study examined water consumed per workload and identified server efficiency, the water intensity of electricity generation, utilization, cooling technology, facility efficiency and climate among the major determinants. Its conclusion was that there is no single cooling configuration that minimizes water use in every location.

This makes water another part of the operating-cost decision rather than a separate engineering issue. A design that minimizes electricity consumption does not automatically minimize water consumption, and vice versa.

A data center cannot normally optimize solely for the lowest electricity bill.

Customers expect computing systems to remain available even when individual pieces of equipment fail. Facilities therefore use combinations of UPS systems, batteries, backup generators, redundant electrical feeds, duplicated pumps and cooling equipment.

That redundancy adds capital cost and can also create additional operating losses. Electricity passing through power-conversion equipment is not perfectly efficient. Pumps, fans and standby infrastructure require maintenance, and operators may deliberately keep spare cooling and power capacity available rather than running every component at its theoretical maximum efficiency.

The resulting cost is part of the price of resilience.

Uptime's 2026 survey illustrates why operators continue to make that trade-off. Although the frequency of impactful outages has declined over time, the organization found that one in ten reported outages was still classified as serious or severe, while outage costs continued to rise.

The cheapest facility to operate is therefore not necessarily the most economical facility overall. A small efficiency gain can be a poor bargain if it meaningfully increases the risk of downtime.

One of the most important trends in the industry is that data centers are becoming more efficient while simultaneously consuming more electricity.

Berkeley Lab estimates that facility infrastructure represented 36 percent of U.S. data-center electricity consumption in 2018 and 31 percent in 2024. In other words, a greater proportion of electricity is reaching IT equipment rather than support systems. Yet total electricity consumption is rising because the quantity and power demand of computing equipment are increasing much faster.

The laboratory estimates that U.S. data-center electricity use increased 14 percent between 2023 and 2024, driven primarily by growth in the number and rated power of accelerated servers. Its future figures remain scenarios rather than reported results, but the Reference Case shows the same fundamental tension. Improvements in computing efficiency and facility efficiency do not necessarily reduce absolute electricity consumption when demand for computing expands even faster.

That is the central economics of power and cooling inside a modern data center. Operators are continually trying to direct a larger share of every megawatt toward useful computing while reducing the energy needed to distribute power and remove heat.

The latest generation of facilities is becoming better at doing exactly that. But as AI and other workloads push far more computing into each building, efficiency gains are increasingly being used to support greater computing density rather than simply to reduce the electricity bill.