The global data center fleet consumed 683 terawatt-hours of electricity in 2024, according to ABI Research. Cooling systems account for up to 50 percent of that total energy draw. That puts cooling's share somewhere around 340 TWh last year, more than the entire electricity consumption of South Africa, and the IEA projects total data center demand will reach 1,479 TWh by 2030 at roughly 14 percent annual growth. The cooling bill doubles by decade's end.
Most operators cannot tell you, in real time, where those kilowatt-hours go. They know total facility power. They know IT load. They can produce a PUE number that looks respectable on an investor slide. Granular, continuous measurement of thermal performance across the cooling chain, from chiller plant to CRAC unit to cold aisle to server inlet, remains rare across the fleet. Without it, optimization is guesswork dressed in engineering language.
ASHRAE's recommended thermal guidelines for data center equipment specify a server inlet temperature range of 64 to 81 degrees Fahrenheit (18 to 27 degrees Celsius) with relative humidity between 40 and 60 percent. That window defines the operating regime for every cooling system in the industry. Stay inside it and equipment runs within manufacturer specifications. Drift outside and failure rates climb, warranties void, and thermal throttling degrades compute throughput.
The math inside that window is where the money hides. Every degree Celsius you raise the server inlet temperature reduces cooling energy consumption by approximately 4 percent, so moving from 20C to 25C saves roughly 20 percent of cooling power. At scale, across a 50 MW campus, that 20 percent translates to megawatts freed and millions of dollars in annual operating expense removed from the P&L. We covered this dynamic in detail when we reported on the 67 percent energy savings most data centers leave on the table.
Most facilities still cool to 68F or below because operators lack the instrumentation to know, minute by minute, that raising the temperature two degrees will not create a hot spot in row 14 that takes out a storage array. Without dense sensor coverage and real-time analytics, the rational choice is to overcool. Overcooling is the single largest source of wasted energy in the industry, and it persists as a deliberate strategy rather than an oversight because the cost of a thermal excursion that brings down a production cluster dwarfs the cost of running the chillers harder than necessary.
Power Usage Effectiveness remains the standard metric. Total facility energy divided by IT equipment energy. A PUE of 1.3 means 30 percent of your power goes to overhead, mostly cooling. Hyperscalers report PUEs of 1.1 or below. The industry average hovers around 1.55 to 1.6.
The problem with PUE as a management tool is granularity. It tells you the ratio. It does not tell you which CRAC unit is running at 40 percent efficiency because a damper is stuck. It does not flag that your chilled water loop is overcirculating by 15 percent because the flow setpoint was never adjusted after the last capacity expansion. It does not catch the fact that three racks in zone B are pulling supply air from the hot aisle through a gap in the containment curtain. PUE is an annual physical when you need continuous vital signs.
The data center sector emits roughly 330 million tons of CO2 annually, about one percent of global energy-related emissions. That figure will grow proportionally with demand unless cooling efficiency improves faster than capacity scales. The trajectory right now does not inspire confidence. As we reported in our coverage of Big Tech's climate goals colliding with AI data center expansion, the gap between corporate sustainability commitments and operational reality is widening.
Traditional air-based cooling, specifically Computer Room Air Conditioning (CRAC) units and Computer Room Air Handlers (CRAH), has served the industry for decades. These systems work. They are well understood. They are also approaching fundamental limits as rack densities climb past 30 kW toward 50, 80, and eventually 100+ kW per rack driven by GPU-intensive AI workloads.
Air has a volumetric heat capacity of approximately 1.2 kJ per cubic meter per degree Kelvin. Water's is about 4,180 times denser per unit volume at carrying heat. That physics gap explains why air cooling requires massive airflow volumes, raised floors, hot aisle/cold aisle containment, and progressively larger fan arrays as density increases. Evaporative cooling can extend the range by exploiting the latent heat of water evaporation, but it introduces water consumption as a variable, and in drought-prone regions where data centers are increasingly being built, water is becoming a constrained resource.
The ceiling for air cooling in practice is somewhere around 25 to 30 kW per rack before the economics and physics both break. Above that threshold, you need liquid.
Direct-to-chip liquid cooling routes dielectric fluid or treated water through cold plates mounted directly on CPUs and GPUs, transferring heat from silicon to liquid at the source with dramatically higher efficiency than forcing air across a heatsink. Immersion cooling goes further, submerging entire server boards in dielectric fluid that absorbs heat across every component simultaneously.
Both approaches can handle rack densities of 100 kW and above, reduce fan energy to near zero, and operate at higher fluid temperatures that open the door to waste heat capture and reuse. NVIDIA has demonstrated that hot water cooling at 45C inlet temperatures eliminates the need for chillers entirely, collapsing the cooling infrastructure from a multi-stage refrigeration chain to a simple heat exchanger and a dry cooler.
Liquid cooling also introduces measurement requirements that most facilities are nowhere near ready to handle. Flow rates matter at the individual server level. Fluid temperature differentials across cold plates indicate thermal performance in real time. Pressure drops across distribution manifolds signal clogging or air entrainment. Water quality in open-loop systems affects corrosion rates and biofilm growth. Every one of these parameters needs continuous monitoring, and the instrumentation to provide it is lagging far behind the hardware it needs to serve.
Badger Meter recently framed the cooling transition around their BlueEdge monitoring suite, which tracks flow, energy, gas, and water quality across cooling infrastructure. The industry's measurement infrastructure has not kept pace with the thermal complexity of modern deployments, and the operators who close that gap first will be the ones who actually capture the efficiency gains that liquid cooling promises on paper.
Real-time cooling telemetry provides per-rack thermal profiles that allow dynamic setpoint adjustment, early detection of cooling system degradation before it becomes a capacity constraint, and verification that component-level efficiency gains are translating to system-level savings. It also creates the data layer that feeds AI-driven cooling optimization systems, correlating IT workload patterns with cooling demand in ways that static setpoints cannot. Without dense, accurate, continuous measurement, none of these capabilities exist. You are left with periodic audits and spreadsheet models that describe what happened last quarter.
The industry is building liquid cooling capacity at pace. CDUs, rear-door heat exchangers, immersion tanks. The hardware is arriving. What is lagging is the instrumentation layer that makes all of it work at peak efficiency rather than at whatever setpoints the commissioning engineer picked on day one.
Cooling consumes half the energy in a data center, and with demand growing at 14 percent per year that share represents a problem compounding faster than any efficiency program in the industry can offset. The transition from air to liquid cooling is technically sound and already underway, but none of it matters if operators cannot measure thermal performance across their infrastructure in real time, continuously, at the component level.
PUE is a summary statistic. What operators need is a live thermal map of every cooling loop, every rack, every server inlet, and the technology to provide that already exists in the form of flow meters, temperature sensors, pressure transducers, and analytics platforms capable of processing the data in real time. The gap is deployment. Most facilities, particularly in the colocation and enterprise segments, are still running cooling systems with the same instrumentation density they had a decade ago, before rack densities tripled and the stakes of thermal mismanagement became a material line item on the balance sheet.
The operators who instrument first will optimize first. Everyone else will keep overcooling, overspending, and wondering why their PUE number never moves.