Data Center Operating Costs: Where the Money Actually Goes

The easiest data center operating-cost chart to make is also one of the easiest to get wrong.

Put electricity at 40%, maintenance at 20%, staff at 15%, cooling at another 15%, add a few smaller slices and the graphic looks authoritative. The problem is that there is no universal data center OPEX pie. A 2 MW enterprise facility, a 50 MW hyperscale campus and a colocation operator do not share the same cost structure, accounting boundary or utilization.

Even “cooling cost” is often counted twice. Cooling electricity already appears in the power bill, while chiller maintenance, pumps, water treatment and replacement equipment belong elsewhere. The same problem appears with labor, network services, depreciation and capital replacements.

A more useful operating model starts by asking what causes each cost to move.

The categories overlap in real accounting systems, but the distinction is useful: electricity behaves almost continuously with load; a generator overhaul does not. Staffing may barely change when a small facility goes from 40% to 60% utilized. A major battery replacement can create a large cost in one year and almost none in the next.

Electricity is the first operating-cost number worth calculating

Electricity is unusual because it can be modeled from a small number of physical inputs:

Take a facility carrying a steady 10 MW of IT load at an annual PUE of 1.52, the industry-wide average reported in Uptime Institute's 2026 Global Data Center Survey. Total facility load is approximately 15.2 MW.

Over 8,760 hours, that is about 133.15 GWh per year. At an illustrative effective electricity cost of $0.08/kWh, the annual bill is roughly $10.65 million.

The 8¢ input is deliberately illustrative. The U.S. Energy Information Administration's latest available data through May 2026 puts the national year-to-date average retail price at 8.83¢/kWh for industrial customers and 13.79¢/kWh for commercial customers. Those EIA sector averages are context, not a data center tariff. A large data center may face a negotiated utility structure, transmission charges, riders, demand charges, taxes or other terms that make its effective cost materially different.

That caveat matters because a one-cent change is enormous at scale.

In the 10 MW / PUE 1.52 example, every 1¢/kWh changes annual electricity cost by about $1.33 million. A procurement decision that improves the effective rate from 9¢ to 8¢ can therefore matter more than trimming several smaller operating line items combined.

PUE tells you how much facility overhead rides on top of the IT load

Uptime's August 2026 analysis gives two useful benchmarks. The normal per-facility average PUE is 1.52. Weight facilities by capacity and the average falls to 1.36, reflecting the efficiency advantage of newer, larger sites. Uptime also notes that leading new facilities routinely report 1.3 or lower.

Hold the 10 MW IT load and 8¢ electricity assumption constant. At PUE 1.52, annual electricity cost is about $10.65M. At PUE 1.36, it falls to about $9.53M. The modeled difference is roughly $1.12M per year.

That does not mean “cooling costs $1.12M.” PUE includes cooling, power-conversion losses and other facility loads. It means the more efficient facility uses roughly 14 GWh less electricity per year at the same 10 MW IT load.

This distinction is one reason our PUE cost guide avoids treating PUE as a cooling percentage.

The server load itself can dwarf every efficiency discussion

Improving facility overhead matters. Reducing unnecessary IT energy can matter even more because every watt removed from the IT side also removes some supporting facility load.

Berkeley Lab's June 2026 update to the U.S. Data Center Energy Usage Report estimates that American data centers could consume 521–843 TWh of electricity in 2030, with a reference case of 649 TWh. The model is built bottom-up from planned equipment shipments, device energy use and cooling-system simulations. The scale of that forecast is a reminder that operating cost growth is being driven heavily by the compute itself, not only by inefficient buildings.

A facility team can improve PUE from 1.5 to 1.3 and still see its electricity bill rise if deployed IT MW grows faster.

This sounds obvious when written out. It is less obvious in annual budgets, where an efficiency project can succeed technically while total utility spend still rises sharply because the business added more servers.

Staffing is a fixed cost until suddenly it is not

A data center needs people whether the racks are 40% or 80% full. That gives labor a different economic shape from electricity.

Operators need some combination of facilities technicians, electrical and mechanical expertise, operations management, security, remote-hands capability, monitoring and escalation coverage. Some roles can support multiple buildings or be contracted externally; others need people physically available at the site.

Uptime Institute's 2026 Global Data Center Survey says more than half of respondents report difficulty finding qualified candidates for open positions, while turnover remains a persistent problem. The staffing constraint matters financially because a data center cannot simply assume an unlimited supply of experienced operators at a stable labor rate.

Scale changes the unit economics. A small private facility may still require 24/7 procedures, specialized contractors and on-call expertise, even though those resources are spread across only a few megawatts. A large operator can spread management systems, training, engineering and some support functions across a much bigger installed base.

This is one of the reasons operating cost per MW generally improves with scale even when the total payroll grows.

Maintenance is not one smooth annual percentage

Maintenance budgets are often represented as a percentage of initial construction cost. That can be useful for a very early model. It hides the way facilities actually age.

Generators need service and periodic major work. UPS batteries have replacement cycles. Switchgear, transfer equipment, controls, pumps, chillers, cooling towers, CDUs, fire systems and security systems all have different inspection, maintenance and replacement patterns.

A new facility may enjoy several years with relatively low corrective maintenance while service contracts and warranties absorb part of the risk. An older facility can enter a period where several expensive systems approach end-of-life together.

The result is a budget that looks more like a staircase than a straight line.

Conceptual pattern only. Actual maintenance cycles depend on equipment, manufacturer requirements, operating hours, condition and facility design.

This is why a one-year OPEX comparison can be misleading. One site may look cheaper simply because its next major battery, chiller or controls replacement sits just outside the model horizon.

Replacement capital belongs in the economic model even when accounting calls it CAPEX

Strict accounting definitions can place major equipment replacements outside operating expense. An owner evaluating the economics of a data center still has to fund them.

If the economic question is “what does it cost to keep this facility usable for fifteen years?”, excluding every large periodic replacement because it is technically capital expenditure produces an incomplete answer.

I would keep replacement capital on a separate line rather than mixing it into day-to-day OPEX. That preserves the accounting distinction while making the lifecycle cost visible.

This becomes especially important for facilities adapting to high-density AI. A working building may need new liquid-cooling infrastructure, denser electrical distribution or additional heat rejection long before the shell reaches the end of its useful life.

Water can be a small invoice and a large strategic cost

Water is a good example of why invoice size and business importance are not always the same thing.

In some facilities, the direct price paid for water is modest relative to electricity. But evaporative cooling can consume substantial volumes, and local scarcity, permitting limits or community opposition can constrain the facility even when the commodity cost itself is low.

Uptime's 2026 survey says more than half of operators now track water consumption. Its research repeatedly cautions against global rules because water use depends heavily on climate, cooling technology and local resource conditions. A facility in a water-stressed location has a different operating risk from one with the same consumption in a water-abundant region.

Berkeley Lab reaches a similar conclusion from the workload side. Its research finds workload-level water use can vary by more than four orders of magnitude depending on server efficiency, grid water intensity, utilization, cooling system, infrastructure efficiency and climate. That is another reason a universal “water cost per MW” would be less useful than it appears.

Network and connectivity costs depend on what the data center is for

A compute campus used mainly for large-scale internal training traffic has a different external-network cost structure from a colocation facility built around hundreds of customer cross-connects, carriers and internet exchanges.

Connectivity spend can include carrier circuits, dark fiber, transit, cross-connects, meet-me-room infrastructure, network operations and redundancy across diverse physical routes.

The cost can be commercially significant without scaling neatly with IT MW. Two 10 MW facilities can have similar power bills and radically different network economics because one is connectivity-heavy and the other is not.

This is also why comparing colocation and ownership requires normalizing the boundary. In colo, some connectivity appears as provider invoices. In an owned facility, similar services may sit inside corporate telecom budgets rather than the facility's OPEX ledger.

Security, insurance and compliance do not disappear because they are difficult to benchmark

Physical security can include guards, access-control systems, CCTV, monitoring, visitor processes and perimeter protection. Insurance cost reflects asset value, geography, business interruption exposure and the insurer's view of risk.

Compliance adds another layer for facilities serving regulated customers or working under particular certification frameworks. Audits, testing, documentation and specialist services cost money even when they are not physically visible in the white space.

These categories are exactly where generic percentage breakdowns become dangerous. A hyperscale owner, wholesale provider and enterprise operator may classify or outsource them in completely different ways.

Fuel is normally a resilience cost, not a primary energy source

Backup generators consume comparatively little fuel during normal utility operation, but fuel still creates testing, storage, treatment, delivery and emergency-planning costs.

The financial exposure is asymmetric. Most years may contain only routine testing. A prolonged grid event can consume fuel rapidly while deliveries become operationally critical.

The same principle applies to many resilience costs: they look expensive relative to their normal annual utilization because their purpose is to be available when normal systems fail.

It is a mistake to treat resilience spending as “unused OPEX”

Redundant UPS capacity, backup generators, spare cooling capacity, testing, maintenance and trained staff can all appear inefficient during a normal year.

Uptime Institute's 2026 outage analysis provides the counterweight. 57% of respondents said their most recent major outage cost more than $100,000, and one in five said their most recent impactful outage cost more than $1 million. Power remains the leading cause of impactful outages. Around one in ten respondents still classified their last outage as serious or severe.

That does not prove every additional redundancy investment pays for itself. It proves that the cost side of resilience has to be compared with the economic consequence of failure rather than with zero.

A backup system can operate for years without producing revenue. That does not make it financially useless. Its return appears in incidents that never become outages.

Low utilization makes almost every fixed operating cost look worse

Electricity tracks load reasonably closely. Many other costs do not.

A 20 MW facility using only 8 MW still needs security, monitoring, maintenance, inspections, insurance and enough operational coverage to run safely. Some cooling and electrical systems also become less efficient at low load.

Divide those costs by only 8 MW of productive IT and the operating cost per used MW can be much higher than the original business case anticipated.

This is one reason colocation can outperform ownership for uncertain demand. The provider aggregates utilization risk across multiple customers instead of one enterprise carrying the cost of an underfilled facility alone.

High utilization creates a different problem: no room to move

Operating a facility very close to its practical limits can make the cost per used MW look excellent. It can also make maintenance, growth and failure states harder to manage.

Capacity has to remain available for redundancy and maintenance according to the design. Local power and cooling limits can become binding even when site-wide utilization still looks acceptable.

Our capacity-planning guide uses the term “stranded power” for the opposite problem — capacity that exists in total but cannot be used where the workload needs it. Both underutilization and over-concentration can damage the operating economics.

AI changes the OPEX mix in more than one direction

The obvious change is electricity. Rack-scale AI systems can consume well above 100 kW per rack and training workloads can operate near high power levels for extended periods.

The less obvious changes are maintenance and cooling. Direct liquid cooling introduces CDUs, piping, coolant management, pumps, controls and different maintenance boundaries between facilities and IT teams. Uptime's 2026 research says the industry is still converging on operating practices for these systems.

At the same time, higher density can reduce the number of physical racks required for a given IT MW. That can remove cabinets, shorten some distribution paths and reduce white-space needs.

AI therefore does not simply add one new “liquid cooling OPEX” line. It changes the physical architecture and shifts where costs sit.

Automation may reduce labor pressure before it reduces headcount

Operators are increasingly using analytics and automation for monitoring, alarms, predictive maintenance and capacity management. Uptime's 2026 survey says confidence in AI is highest for lower-risk operational uses such as sensor-data analysis and predictive maintenance.

That does not automatically translate into fewer staff. In a sector already struggling to hire qualified people, the first economic value may be allowing existing teams to manage more infrastructure, detect problems earlier and spend less time on repetitive monitoring.

There is also a risk trade-off. Uptime's outage research continues to identify procedural failures and human error as important contributors to outages, while more automation can introduce new failure modes of its own.

I would therefore model automation savings cautiously: reduced manual workload is easier to defend than an assumption that a new control platform allows a fixed percentage of the operations team to disappear.

A good annual budget separates four different kinds of money

Keeping the four categories visible prevents several accounting tricks from distorting the operating model. A site should not look artificially efficient because battery replacement sits in a capital budget, nor artificially expensive because the IT hardware electricity bill is charged to facilities while another site books it to technology.

The ten-year model should not hold everything flat

A serious lifecycle model needs changing assumptions.

Electricity prices move. IT load ramps. PUE changes as facilities fill and age. Maintenance grows unevenly. Staff costs inflate. Major components reach replacement cycles. New hardware generations can alter rack density and cooling requirements.

Even a simple model improves substantially if it allows different annual values for IT MW, PUE and electricity rate rather than multiplying the first year's OPEX by ten.

The 2026 market makes this especially important because the workload mix is changing quickly. Berkeley Lab's new 2030 electricity forecast ranges from 521 to 843 TWh precisely because reasonable assumptions about AI equipment shipments, utilization and power behavior produce very different futures.

The number I would distrust most is “OPEX per MW” without utilization

A facility designed for 50 MW but operating at 20 MW and a facility designed for 20 MW and operating at 20 MW can both be described as “20 MW data centers” in casual conversation.

Their cost structures are not the same. The first may be carrying the staff, maintenance and fixed infrastructure of a much larger build while waiting for demand. The second may be close to full and have little room for expansion.

When someone quotes annual operating cost per MW, I would want to know whether the denominator is design MW, available IT MW, contracted MW or actual average IT load. Changing that denominator can alter the apparent unit cost without changing one dollar of expense.

So where does the money actually go?

In a large, well-utilized facility, electricity can become the dominant recurring cash cost simply because the physical energy flow is enormous. Our 10 MW example already produces more than $10 million per year at 8¢/kWh and PUE 1.52.

Around that electricity bill sits the machinery and organization required to keep the energy flowing safely: facilities staff, maintenance contracts, replacement equipment, cooling consumables, security, network services, insurance, monitoring, compliance and resilience.

The relative size of those categories changes with scale, age, architecture, location and accounting. That is why this article does not end with a universal pie chart.

The better operating-cost model has no mystery percentage. It starts with the actual IT load and electricity contract, measures facility overhead, then adds the people and equipment required to keep that load available over time. Finally, it includes the irregular costs — replacements, upgrades and failures — that annual averages are especially good at hiding. That is where the money actually goes.