How a data center gets built.

A modern data center is one of the most demanding construction projects in the world: a building that draws the power of a small city, rejects megawatts of heat continuously, and is expected to never — ever — go dark. The building itself takes 18–24 months to put up. The full development cycle around it commonly takes 3 to 5 years, and longer in power-constrained markets. Here is how each stage works, in order.

STAGE 01Site selection & power procurement

Every data center project begins with the same question: where can we get power? Modern campuses request 50–500 MW from the grid — a load comparable to a mid-sized city — and grid capacity, not land, is the scarce resource. A developer's site-selection screen weighs power availability and cost first, then fiber routes, water, natural-hazard risk, tax incentives, and proximity to network hubs.

Securing that power means entering the utility's interconnection process. The utility runs a system impact study — typically 6–18 months, paid for by the developer — to determine what grid upgrades the new load requires. In constrained markets like Northern Virginia, the queue for large new interconnections runs 3–7 years. This is why sophisticated developers negotiate with utilities before they close on land, and why "powered land" — sites with secured grid capacity — trades at a large premium over ordinary industrial acreage. A site without power is just dirt.

STAGE 02Entitlement & permitting

In parallel with the utility, the developer must win the legal right to build. Entitlement covers zoning (or rezoning) the parcel for data center use, site-plan approval, environmental review, and the local political process that comes with all of it. Data centers bring tax revenue with very few school-age children — a fiscal profile counties historically loved — but noise, diesel generators, water use, and visual impact have made approvals genuinely contested in many jurisdictions.

Beyond zoning, the project needs building permits, air-quality permits for the generator fleet (dozens of large diesel engines are a regulated emissions source), stormwater and water-use approvals, and often FAA review for tall utility structures. Entitlement risk is one of the first things an acquirer diligences: a fully entitled, powered site can be worth multiples of the same land without approvals.

STAGE 03Design

With power and entitlement moving, engineering begins. The design team sets the parameters that will govern the asset for decades: total IT capacity in megawatts, rack density (traditional enterprise racks draw 8–17 kW; AI training racks now demand 80–150 kW+ and force liquid cooling), the redundancy topology (N+1 shared spares versus 2N fully duplicated systems), and the cooling architecture — air-cooled, chilled-water, or direct-to-chip liquid.

The single most important drawing is the one-line diagram: the schematic showing the entire power path from utility point of delivery, through the substation and medium-voltage switchgear, to generators, UPS systems, PDUs, and finally the racks. Every box on that diagram is capital, lead time, and a failure mode. Designs are targeted to the tenant: a hyperscaler build-to-suit follows the tenant's own specification, while a speculative colocation building is designed for flexibility.

STAGE 04Long-lead procurement — the 2–4 year transformer problem

Here is the fact that most changes data center schedules today: large substation transformers are running 2–4 year lead times in 2025–2026, driven by surging demand from data centers and grid electrification. Medium-voltage switchgear has stretched to 40–52 weeks. Backup generators — typically 2.5–3 MW diesel units, ordered by the dozen — are the second-longest item after transformers.

This is why equipment procurement is now a competitive weapon. Large operators reserve manufacturing slots years ahead, buy positions in other developers' orders, and sometimes acquire whole projects primarily for the equipment already on order.

STAGE 05Construction

Construction proper runs roughly 18–24 months for a large facility and follows a deliberate sequence:

• Sitework and civil — grading, utilities, stormwater, foundations. Data halls carry enormous floor loads and vibration-sensitive equipment, so the concrete package is substantial.
• Shell — structural steel or precast concrete walls and roof. A "powered shell" (building complete, utility power delivered, interiors unfinished) is itself a leasable product.
• Electrical and mechanical fit-out — the heart of the job and the majority of the budget. Substation energization, generator placement, switchgear lineups, UPS rooms, chilled-water plant, piping, and air handling. Electrical and mechanical trades routinely account for well over half of total project cost.
• White space fit-out — the data halls themselves: busway or whips to the racks, containment aisles, fire suppression, security, and monitoring systems.

Large campuses are built in phases: the first data hall is commissioned and generating revenue while later halls are still in construction — which keeps capital efficient but means construction, commissioning, and live operations coexist on one site for years.

STAGE 06Commissioning — Levels 1 through 5

A data center is not finished when construction ends; it's finished when it has proven it cannot be killed. Commissioning (Cx) is the staged testing regime that provides that proof, and tenants — especially hyperscalers — will not accept space without it.

The Level 5 "pull-the-plug" test is the graduation exam: utility failure, generator start and transfer (the ATS completes the swap in roughly 10 seconds while UPS batteries carry the load for the 5–15 minute window), cooling continuity, and controls all demonstrated at once, witnessed and documented. Commissioning findings are punch-listed and retested until the facility passes clean.

STAGE 07Operations — the next 20 years

Handover to operations is where the building starts earning. Tenants move in through structured onboarding; critical-facilities teams run the plant 24/7 against SLAs that promise near-perfect uptime ("five nines" — 99.999% — allows about five minutes of downtime a year). Every maintenance action follows written procedures (MOPs), because most outages are caused by human error, not equipment.

Operations is also where design choices become money. PUE — total facility power divided by IT load — measures efficiency: modern hyperscale facilities achieve 1.08–1.2, while the industry average sits near 1.5. On a 100 MW campus, that gap is millions of dollars a year in power cost, flowing straight to operating margin for the next two decades. If the terms in this article are new to you, the data center glossary defines all of them with the numbers operators actually use.