Data Center Capex: Build Costs and Time-to-Power

Introduction

Data centers are not a buy-and-hold business; they are a capital-deployment business, and the two numbers that govern them are the cost to build a megawatt of capacity and the time it takes to energize it. Get both right and a developer earns a wide spread between development cost and stabilized value. Get either wrong, by overpaying to build or by stranding capital in a project that cannot get power, and the economics collapse. This is why a credible view of any data center platform starts not with its in-place rents but with its cost stack and its development pipeline. A standard facility runs roughly $10 million to $12 million per megawatt of critical load, and an AI-optimized facility can be double that.

The defining feature of the cost stack is that the building is the small part. Most of the capital goes into the electrical and mechanical systems that deliver and condition power, not into the shell. Across recent estimates, electrical systems alone account for roughly 40% to 50% of total cost, mechanical systems another 15% to 20%, and the powered shell, the land and base building, only about 15% to 20%.

Critical Load (Cost per MW)

The IT power a facility can actually deliver to computing equipment, measured in megawatts, and the denominator against which development cost is quoted. Stating cost per megawatt of critical load, rather than per square foot, normalizes for the fact that a data center's value is its deliverable power, not its floor area.

The headline unit is simple: take the all-in development cost and divide it by the critical megawatts the facility can deliver. This is the number every cost comparison and pipeline estimate ultimately reduces to, which is why a banker quotes a build in dollars per megawatt before any other metric.

The headline figures show how fast costs have risen and how much AI raises them. Across nineteen United States markets, the cost to develop one megawatt of critical load averaged about $11.7 million in 2025, ranging from roughly $9.3 million in San Antonio to $15 million in Reno. On a square-foot basis, construction costs climbed from about $183 in 2020 to over $1,000 by the end of 2025. AI-optimized facilities, with their liquid cooling and far higher power densities, can push the all-in figure to $20 million per megawatt or more.

Land, often assumed to be the main cost in real estate, is a modest line item: average data center land ran about $244,000 per acre in 2024, though demand for large parcels has pushed prices for 50-plus-acre sites up sharply. The point for a banker is that data center development is an electrical and mechanical engineering project wrapped in a real estate envelope, which is why it behaves so differently from the hyperscale, wholesale, and retail segments' traditional-property comparables.

Why Costs Have Surged

Two forces have driven the cost surge. The first is sheer demand for specialized equipment: lead times for transformers, switchgear, and generators have stretched to a year or more, and that scarcity feeds straight into price. The second is the density of AI hardware, which demands far more power distribution and cooling per square foot than any prior generation. Together they explain why a square foot that cost under $200 to build in 2020 can cost over $1,000 today, and why cost discipline has quietly become a real competitive differentiator between platforms.

The Equipment Bottleneck Behind the Timeline

The scarcity is most acute in the electrical gear that makes up the largest slice of the budget, and it has become a binding constraint in its own right. Lead times for large power transformers have stretched to roughly 128 weeks, about two and a half years, with generator step-up units running near 144 weeks and some large orders quoted as far out as five years. Prices moved with the scarcity: transformer costs have risen on the order of 77% for power units and as much as 95% for distribution units since 2019. Because a site cannot energize without this equipment, the supply chain now gates delivery as tightly as the interconnection queue does, and the two delays compound rather than overlap. Industry analysts estimate that 30% to 50% of the capacity planned for 2026 will slip to 2028 as a result. For a developer, this turns procurement into a strategic function: ordering long-lead equipment years ahead, sometimes before a site is fully permitted, and treating a secured transformer slot almost the way it treats a secured megawatt of power. A platform that locked in equipment orders early holds an advantage that does not show up in its rent roll but is worth as much as a queue position.

Because returns come from continuously building, the leading platforms run enormous, sustained capital programs. Equinix spent roughly $3.4 billion of total capital expenditure in 2024, and under its "Build Bolder" plan the company expects to step that up to as much as $5 billion every year through 2029, on the order of $20 billion to $25 billion over the five years. That is not a one-time build; it is a treadmill, and the platforms that can fund it cheaply win.

Cost by Platform and Geography

The cost to build also varies by platform and geography in ways that reveal the underlying economics. Digital Realty's United States development pipeline implies an average build cost near $9.5 million per megawatt, while its European pipeline runs closer to $14 million on higher construction and energy costs, and Equinix's multi-region wholesale pipeline sits around $11.5 million per megawatt. Those figures, and the pace at which each platform can deploy them, are what the market is really valuing when it prices the data center REITs.

Cost is only half the equation. The other half is time-to-power, the years it takes to actually energize a site, which the power constraint article covers in depth. The reason time matters so much is that the entire return thesis is a development spread: build at a cost, lease at a rent, and capture the gap between the development yield and the cap rate at which the stabilized asset trades.

If a developer builds at an 8% development yield and the stabilized asset trades at a 6% cap rate, the spread between them is the value created, and it can be substantial on a multi-hundred-million-dollar facility.

But that spread is only earned if the project gets built and energized on schedule. Every year of delay in securing power pushes out the rent, raises the carrying cost of the capital already sunk, and erodes the spread. Time-to-power is therefore not a side issue; it is the variable that determines whether the development spread is real or theoretical, and it feeds directly into the valuation of any data center.

The other side of that risk is absorption: whether the capacity is spoken for before it energizes. The single best signal is the pre-lease rate, the share of a new build's capacity already committed under signed leases at the point of delivery.

A facility delivering fully pre-leased earns its development yield from day one, while a speculative build that energizes empty carries its cost basis with no income until tenants sign, which is why lenders and equity partners watch the pre-lease rate as closely as the cost per megawatt.

The capex math reframes the central question from "what does this asset yield" to "how much capital can this platform deploy, at what cost per megawatt, how fast can it energize, and what spread does that create."

Development Yield

A project's stabilized net operating income divided by its total development cost, expressed as a percentage. It is the developer's return on the cost to build, and the gap between it and the cap rate at which the finished asset trades (the development spread) is the value a developer creates by building rather than buying. A platform with a low cost basis, secured power, and a cheap cost of capital can compound value by building; one without those advantages cannot, no matter how strong demand is. The relationship between development yield and the prevailing cap rate environment is the lens through which the whole sector should be read.

Strip the sector down to its economics and a data center is a capital-deployment machine, not a building: power and cooling systems, not the shell, consume the budget; AI roughly doubles the cost per megawatt; and the leading platforms spend billions a year because the return comes from building at a development yield and capturing the spread to the cap rate, not from collecting rent on what already exists. The variable that decides whether that spread is real or theoretical is time, both the years to energize and the years to procure the equipment, which is why a platform's edge ultimately rests on three things the income statement does not show: a low cost basis, secured power, and a cheap cost of capital to fund the treadmill.