Data Center Operating Costs: Power, Cooling, PUE and Staffing

Data center operating costs decide whether a facility makes money after it is built. This guide, part of the Kurums Data Center Economy hub, breaks down the main opex lines, power, cooling, maintenance, staffing, taxes and insurance, explains how PUE turns into dollars, compares electricity prices across US data center markets and shows why, for AI facilities, depreciation now dwarfs every running cost. It follows our analysis of data center construction cost per MW and pairs with our explainer on how data center business models work.

What are the main operating costs of a data center?

The main operating costs are electricity for IT and cooling, maintenance of electrical and mechanical systems, staffing, property taxes, insurance, network connectivity and security. Electricity is usually the largest single line, while hardware depreciation dominates total cost for AI facilities.

Data center operators group running costs into a few buckets. Energy covers the power drawn by servers, storage and networking plus the overhead used by cooling, power conversion losses and lighting. Facility maintenance covers preventive and corrective work on generators, uninterruptible power supplies, switchgear, chillers and fire systems. People covers facility engineers, technicians, security and network operations staff. Fixed charges cover property and equipment taxes, insurance, ground rent and connectivity.

How these buckets compare depends on who owns what. A wholesale landlord typically passes power through to the tenant, so electricity is the tenant’s cost; a retail colocation provider buys power and resells it with a margin; a hyperscaler pays everything itself. Opsio, a cloud services firm, estimates that power accounts for 60–70% of total operating costs, with cooling consuming 30–55% of facility power and averaging about 40%. Those shares fall sharply in a modern, efficient facility, where cooling overhead may be a fraction of that.

For an AI campus, the headline is that running costs are small next to capital costs. Epoch AI’s May 2026 model of a 1 GW hyperscaler facility estimates total annual cost of ownership at about $8.5 billion, of which only about $0.9 billion is opex. The other roughly $7.6 billion is annualised capex, with servers alone about $5 billion a year when depreciated over five years.

How much does electricity cost to run a data center?

A data center’s power bill equals IT load times hours times PUE times price. A 10 MW IT load running all year uses 87,600 MWh at the IT level; at the 2025 industry-average PUE of 1.54 and $70/MWh, that is about $9.4 million a year.

At gigawatt scale, the numbers become a utility-sized budget. Epoch AI’s model puts annual energy cost for a 1 GW AI facility at about $0.6 billion. Bernstein, in a November 2025 analysis, estimated that a gigawatt-scale site would use about $1.3 billion of electricity a year at a price of $0.15/kWh. The difference between those two estimates is mostly the assumed power price, which shows how much location matters.

Power prices vary widely by market and contract type. The US Energy Information Administration’s 2024 state data show average retail prices across all customer classes ranging from 9.79 cents per kWh in Texas to 27.04 cents in California, against a US average of 12.94 cents in its state table. Large data centers usually buy on industrial tariffs, wholesale contracts or long-term power purchase agreements and pay less than these averages, but the ranking of states is a useful guide. The table applies those averages to a 10 MW IT load; the electricity consumption statistics article covers aggregate demand.

Source: US Energy Information Administration, State Electricity Profiles (2024 data, released November 2025); annual bills are Kurums calculations (87,600 MWh of IT energy × PUE × price). All-sector average prices; large industrial buyers typically pay less.

Wholesale markets add volatility. The EIA reported on-peak daily wholesale prices in the PJM Mid-Atlantic hub ranging from $37 to $150 per MWh during August 2026. PJM’s capacity auction for the 2027/28 delivery year cleared at its price cap of $333.44 per MW-day, up from $329.17 for 2026/27, with a forecast load increase of about 5,250 MW attributed largely to large loads. Capacity charges flow through to bills in PJM states, including Virginia, the world’s largest data center market.

What is PUE and what is the industry average in 2026?

Power usage effectiveness (PUE) is total facility energy divided by IT equipment energy. A PUE of 1.0 would mean zero overhead. Uptime Institute’s 2025 Global Data Center Survey puts the industry average at 1.54, while the most efficient hyperscale fleets report about 1.1.

PUE was developed by The Green Grid and has been the industry’s standard efficiency metric for well over a decade. It measures overhead, not productivity: a facility full of idle servers can still post a good PUE. The industry average has improved only slowly. Uptime Institute reported an average of 1.58 in 2023 and noted that the average had stayed in a band of roughly 1.55 to 1.59 since around 2020, as new efficient facilities were offset by an ageing installed base; the 2025 survey figure of 1.54 shows modest progress.

Hyperscalers set the efficiency frontier. Google reports a fleet-wide trailing twelve-month PUE of 1.09 for 2025, and has published fleet figures around 1.10–1.12 since 2016. Large new campuses built for AI often design for PUEs in the low 1.1 to 1.2 range, helped by liquid cooling and economisers. Older enterprise rooms can run near 2.0; the US EPA’s Energy Star programme has cited 2.0 as the average for US data centers in its older guidance.

How much does cooling cost in a data center?

Cooling is the largest source of non-IT overhead, and at the industry-average PUE of 1.54 it accounts for most of the 54% extra energy consumed above the IT load. Moving from air to liquid cooling raises capital cost but can cut cooling energy significantly.

Cooling cost is mostly energy for chillers, pumps, fans and heat rejection, plus water in evaporative designs and maintenance of mechanical plant. Its share depends on climate, design and density. Facilities in cool climates can use outside air for much of the year; those in hot, humid markets need more mechanical cooling. AI racks at 40 kW and above, per JLL, push operators toward direct-to-chip liquid cooling, which removes heat more efficiently than air but requires coolant distribution units, piping and specialised maintenance. We compare the options in our cooling explainer.

On the capital side, JLL estimates liquid cooling adds about 10% to build costs, and Bernstein puts thermal management at about 4% of an AI campus’s total capex. On the operating side, efficient liquid-cooled halls can reach PUEs close to the hyperscale frontier. Cushman & Wakefield reports cooling equipment prices up 32% from December 2021 to early 2026, so replacement and expansion costs have risen too. The trade-off between water and energy use is covered in {L(‘data-center-sustainability-carbon-water-cfe’, ‘our sustainability article’)}.

How much does it cost to maintain a data center?

Routine maintenance covers generators, UPS systems and batteries, switchgear, chillers, fire suppression and building systems. Equinix’s guidance implies recurring maintenance-type capex of about 3% of revenue, while operating maintenance contracts add further costs that scale with the amount of plant installed.

Maintenance spending has two parts. Operating maintenance includes service contracts, testing, fuel for generator runs and consumables. Recurring capital expenditure replaces equipment as it ages: UPS batteries, cooling units, switchgear components and eventually whole generator sets. Equinix, which reports this separately, guided 2026 recurring capex at $290–310 million, about 3% of its roughly $10.3 billion revenue. Datacentres.com’s 2026 cost guide estimates preventive maintenance contracts at roughly $200,000–400,000 a year per 10 MW and says UPS batteries typically need replacing every five to seven years.

Supply chain inflation has pushed maintenance and replacement costs up. Cushman & Wakefield tracked switchgear prices up 60%, generators and motors up 33% and transformers up 29% between December 2021 and early 2026, with lead times for generators of 60–100 weeks. Operators now hold larger spare-parts inventories and sign longer service agreements to avoid being caught without critical equipment.

How many staff does a data center need and what do they cost?

Staffing is a modest share of costs at scale. Datacentres.com estimates 15–25 full-time staff for a 10 MW facility, covering facility engineering, electrical and mechanical technicians, security and network operations. Bernstein describes staffing costs at a gigawatt-scale AI site as negligible.

Data centers are capital-intensive, not labour-intensive. A typical facility runs 24/7 with shifts of critical facility engineers and technicians, a security team and, depending on the business model, remote-hands staff who install and swap equipment for customers. Hyperscalers automate heavily and run large campuses with relatively small permanent teams, while retail colocation sites need more customer-facing staff.

The labour market is tight, especially for electrical and mechanical skills. Turner & Townsend’s 2026-27 survey found that 81% of respondents consider skilled labour in their region constrained, with mechanical, electrical and plumbing trades most in demand. Uptime Institute has written about the risk of operator skill decay as AI tools take over routine tasks. For local communities, the small permanent headcount is a frequent point of debate, covered in our water, land and jobs statistics.

How do taxes, insurance and connectivity affect operating costs?

Property and equipment taxes can be one of the largest non-energy costs, which is why many US states and countries offer data center tax exemptions. Insurance, network connectivity and security add smaller but steady charges.

Property taxes depend on the value of the land, building and, in many US jurisdictions, the equipment inside. Because servers are expensive and replaced often, personal property taxes on IT equipment can be substantial, and many states exempt data center equipment from sales and use tax to attract investment. Incentives vary widely and are a key factor in data center site selection.

Insurance covers property, business interruption and liability; costs rise with the value of the equipment and the concentration of risk in a single campus. Connectivity costs depend on location: carrier-dense metros offer cheap, diverse fibre, while remote campuses may need to fund long-haul fibre routes. Security, including guards, access control and surveillance, is a fixed cost that does not shrink much for smaller sites, which is one reason small facilities have higher costs per kilowatt.

How can operators reduce data center operating costs?

The biggest levers are buying cheaper power, improving PUE, raising utilisation and extending equipment life. On-site generation, long-term power contracts and batteries can lower energy costs, while liquid cooling, economisers and higher operating temperatures cut cooling overhead.

Power procurement comes first. JLL reports that in parts of EMEA, renewables combined with private-wire connections can cut tenant power costs by about 40% compared with grid supply, and that battery costs have fallen below $90/kWh in select deployments, making storage useful for peak shaving. Long-term power purchase agreements fix prices for a decade or more. Efficiency comes second: raising supply air temperatures, using free cooling and deploying liquid cooling can all lower PUE.

Utilisation is the most underrated lever. A facility running at half load spreads fixed costs, staff, maintenance and taxes over half the kilowatt-hours and often runs at worse PUE because cooling systems are oversized for the load. For AI facilities, keeping expensive GPUs busy matters far more than shaving energy: Epoch AI’s sensitivity analysis shows that shortening the useful life of IT equipment from five to three years would raise the annual cost of a 1 GW facility from about $8.5 billion to about $12 billion.

How do operating costs differ for AI data centers?

AI data centers draw more power per rack and run hotter, so energy and cooling bills per square metre are much higher. But because GPUs are so expensive, energy is a smaller share of total cost: Epoch AI puts opex at only about 11% of a 1 GW AI facility’s annual cost.

For a traditional enterprise or cloud facility, power is often the biggest single cost after the building itself. For an AI campus, the economics flip. Epoch AI’s estimate of about $0.9 billion of annual opex against roughly $7.6 billion of annualised capex means a 10% rise in electricity prices adds about $60 million a year, while a one-year cut in GPU useful life can add billions. Bernstein’s estimate of $1.3 billion a year of electricity at $0.15/kWh is larger, but still small next to the hardware bill.

This changes priorities. AI operators care most about time to power, reliability and keeping accelerators fully utilised, and will pay a premium for sites that deliver capacity quickly. Rising electricity demand from AI also affects everyone else on the grid, a debate we cover in Do data centers raise electricity prices?. For the capital side of the AI equation, see our analysis of Big Tech capex.

Frequently Asked Questions