Category Archives: Data Centers

Load Capacity and Energy Requirements: What the Data Says Right Now

If you are laying the groundwork for a data center, the design assumptions available two years ago are already out of date. Here is what the current published data says about rack density, cooling limits, and power availability, and which of those numbers actually belong in an early planning conversation. Data center planning has always started with a load number. What has changed is how fast that number moves. A facility scoped against the industry average of a few years ago is now being asked to carry several times the heat it was designed around, and the equipment driving that increase is shipping today rather than appearing on a roadmap. The figures below come from published industry reporting through 2026. They are the ones we see cited most often in early planning discussions, along with the caveats that usually get left out. Year Avg. kW per rack Change 2021 7 kW Baseline 2023 8.5 kW Gradual 2024 12 kW Accelerating 2025 16 kW Accelerating 2026 27 kW Up ~69% in one year Source: AFCOM, State of the Data Center Report 2026. Figures are blended across a membership base still weighted toward traditional enterprise facilities. 01 — The baseline nearly doubled in a single year Average rack density went from 16 kW in 2025 to 27 kW in 2026. For context, the same survey recorded 7 kW in 2021, meaning the average has roughly quadrupled in five years, with the sharpest move happening in the last twelve months. AFCOM reports that 69% of respondents expect density to keep climbing over the next one to three years and say they are already taking steps to support it. A single-year jump of that size is not a trend line you can extrapolate comfortably, and it is not something a design based on 2023 assumptions absorbs quietly. 02 — The market has split, so the average is misleading The important caveat on any average density figure is that it now describes two different markets wearing one name. A conventional colocation floor still runs largely in the 5 to 15 kW range per cabinet. A purpose-built AI hall next door can run at 130 kW. AI training racks in service today are commonly cited between 40 and 132 kW, and next-generation GPU platforms are being specified well beyond that. Reported industry figures also note that only about one in five operators consider themselves prepared to support the 50 to 70 kW racks already common in AI deployments. The practical takeaway for planning: an average is only useful once you know which population it describes, and whether it measures installed racks or design capacity. Those are two very different numbers on the same drawing set. 03 — Air cooling has a physical ceiling, and it is in sight This is the number that matters most to the mechanical scope. Published analysis puts the practical limit of air cooling somewhere between roughly 20 and 40 kW per rack, depending on who is measuring and how the hall is configured. Above that band, the air volume and velocity needed to carry the heat out of a standard rack form factor exceeds what conventional airflow systems can deliver. Note where the 2026 average of 27 kW sits relative to that band. The industry average is now inside the transition zone, not comfortably below it. That is why cooling has stopped being a downstream facilities decision. Crossing the threshold changes the mechanical load, the water strategy, the redundancy topology, and the total site kilowatts that generation and distribution have to serve. In practice most new high-density halls end up mixed: liquid infrastructure for the dense zones, air retained for the lower-density ones in the same building. 04 — Power, not floor space, is the binding constraint Global data center electricity demand is projected at roughly 132 GW in 2026, climbing toward 290 GW by 2030. The International Energy Agency projects global data center electricity consumption roughly doubling from about 485 TWh in 2025 to around 950 TWh by 2030, close to 3% of world electricity demand, with accelerated servers growing far faster than conventional ones. The nearer-term problem is interconnection. In some major U.S. markets, securing new utility capacity now takes three to four years, which can be longer than constructing the facility itself. Time to power has become as decisive as time to build. That reframes the mechanical question. When the electrical allocation is fixed and hard to grow, every kilowatt the cooling system consumes is a kilowatt unavailable to compute. Efficiency stops being an operating-cost line item and becomes a capacity question. 05 — What this changes about early planning Taken together, the current numbers push three questions much earlier in the process than they used to sit. Which density are you designing to, and for what year? A design-capacity figure and an installed-average figure are not interchangeable. Committing to one without naming the year it describes is how a hall ends up undersized at commissioning. Where is your air-to-liquid boundary? Not whether liquid cooling arrives, but which zones cross the threshold, when, and what has to be in place structurally and hydronically for that to be a staged change rather than a rebuild. What is the fixed electrical envelope? If interconnection sets a ceiling years before the racks arrive, cooling efficiency directly determines how much compute fits inside it. None of that is answered by picking equipment. It is answered by settling load and energy assumptions before the drawings are locked, and writing down which published figure each one came from. THE TAKEAWAY The 2026 average sits inside the band where air cooling stops being sufficient, and the frontier is already several times past it. Any load figure carried into design should name the year and the population it came from. Why we track this AirConnect is the national HVAC company inside Connect Service Solutions. Our work is commercial and industrial mechanical service across the country, and this sector is where the assumptions behind that work