The headlines describe data center investment in numbers that are hard to make sense of. Microsoft's Mount Pleasant, Wisconsin commitment: $3.3 billion. AWS's Mississippi expansion: $10 billion at announcement, since scaled up toward $25 billion. Meta's Louisiana campus: $10 billion. Stargate (the OpenAI-Oracle-SoftBank joint venture): $500 billion across the announced rollout. The major hyperscalers' combined capital spending is on track to clear half a trillion dollars in 2026 alone.

These numbers are not exaggerated. They are also not what most people imagine. The cost breakdown of a hyperscale campus has very little to do with the building itself, very little to do with the land, and very much to do with electrical infrastructure that has supply chain bottlenecks the industry cannot engineer around in the timeframes investors want.

Understanding where the money actually goes is one of the more useful exercises for anyone trying to think clearly about this buildout, because it explains why the investment is concentrated in particular places, why timelines are what they are, and why "just build it somewhere else" is not a real option in most of the proposed alternatives.

The Building Is the Cheap Part

A 100-megawatt data center building in central Arizona, the kind currently being built across the West Valley and East Valley, costs roughly $400 million to $700 million for the structure itself and the basic mechanical, electrical, and plumbing rough-in. Per megawatt of IT capacity, that works out to $4 million to $7 million per megawatt for the shell and interior fit-out. A typical hyperscale campus has eight to twelve such buildings phased over several years.

Building shells are large and they look impressive, but the construction itself uses well-understood techniques. Tilt-up concrete walls. Steel structure. A ridiculously oversized HVAC and electrical core. Raised floors or slab construction depending on cooling architecture. The contractors who build these are competent, the timelines are predictable, and the costs are not the constraint. A general contractor can deliver the shell on schedule almost regardless of where you choose to build it. Land cost (typically $50,000 to $200,000 per acre for industrial land in the Phoenix metro) is also not the constraint. A 200-acre campus with permitting and entitlements is a single-digit-percentage line item on the project budget.

The expensive part of a data center is the infrastructure that connects it to the grid and runs inside it. That's where the multi-billion dollar numbers come from, and that's where the supply chain problems live.

Power Delivery: The Real Cost Center

Building a 100-megawatt data center requires bringing 100 megawatts of utility power to the property line. That is not a trivial amount of electricity. For comparison, 100 megawatts is roughly the peak demand of 75,000 to 100,000 typical Phoenix homes. Delivering that to a single industrial property requires a dedicated substation, dedicated transmission, and often a dedicated upgrade to the regional grid.

A new utility-scale substation purpose-built for a data center campus costs roughly $50 million to $150 million depending on size and complexity. Transmission line extensions to reach the substation can add another $20 million to $100 million. Grid upgrades that the utility requires before agreeing to interconnect, the so-called "system impact study" mitigation, can run another $50 million to $200 million. Total electrical infrastructure to get power to the property line, before a single watt is consumed inside the building, is regularly in the $200 million to $500 million range per campus.

That money largely goes to the utility, the engineering firms, and the equipment vendors. Substantial portions of it benefit the broader grid. The substation that serves the data center also serves the surrounding industrial area. The transmission line that extends to it opens future capacity. The grid upgrades that the utility requires get built faster than they would have otherwise. None of this changes the fact that it shows up on the data center developer's capital plan as a hundreds-of-millions-of-dollars line item.

Inside the property line, the cost compounds. Switchgear lineups, the high-voltage equipment that distributes power within the campus, run $30 million to $80 million per building. Medium-voltage transformers, around 50 of them per building at varying sizes, total another $20 million to $40 million. Uninterruptible power supply (UPS) systems, typically rotary or static units sized for full IT load, cost $50 million to $100 million per building. Generators, typically 20 to 60 diesel units per building at 2 to 3 megawatts each, cost $80 million to $200 million per building including paralleling switchgear and fuel infrastructure.

Add it up and the electrical infrastructure inside one 100-megawatt building runs $300 million to $500 million. Multiply across a multi-building campus and the electrical capex alone is in the multi-billion dollar range.

The Lead Time Problem Nobody Has Solved

Money is not the constraint. The constraint is supply chain.

Large power transformers are the most acute bottleneck. The kind required for utility-scale substation work and large data center campuses are manufactured by a small number of global suppliers. Lead times for new large transformers now run from 80 to 210 weeks, roughly two to four years. Several vendors have temporarily stopped accepting new orders in certain transformer categories because they cannot promise delivery within reasonable windows.

The transformer shortage isn't a temporary disruption. It's a structural mismatch between the buildout pace the AI economy is demanding and the manufacturing capacity that exists for the equipment that buildout requires. Building a new transformer manufacturing line is itself a multi-year, multi-hundred-million-dollar capital project. Several manufacturers are expanding capacity, but the new capacity won't come online fast enough to clear the backlog.

Switchgear is in similar shape, with the largest lineups stretching to 18 to 24 months. Generator lead times have stretched from around 6 months pre-pandemic to roughly 12 to 15 months currently. Specialty cooling equipment, particularly for liquid cooling deployments, has 12 to 18 month lead times. Even commodity items like medium-voltage cable have seen 6 to 12 month delays during the worst of the recent supply chain stress.

The implication for siting is that the projects that are getting built are the projects that placed equipment orders 18 to 36 months ago. The projects that haven't placed orders yet are looking at 2028 or 2029 commissioning dates. The places that already have working hyperscale campuses (Phoenix being one of them) have operating engineers who know the equipment, contractors who know the configurations, and supplier relationships that get prioritized when allocations are tight. That advantage doesn't transfer easily to greenfield markets.

The IT Equipment Multiplier

The capex numbers above are for the building, the infrastructure to power it, and the systems that distribute power inside it. None of those numbers include the actual servers, GPUs, and networking equipment that go into the building once it's commissioned.

The IT equipment cost is the largest single line item for an AI-focused hyperscale buildout. Nvidia GB200 server racks, fully populated, cost roughly $3 million to $4 million per rack. A 100-megawatt building filled with GB200-class equipment can hold 700 to 1,000 of these racks, putting IT capex at $2 billion to $4 billion per building.

Across an eight-building campus, IT equipment alone can total $16 billion to $32 billion. That is the AI workload money. It depreciates faster than the building does, gets refreshed every 3 to 5 years, and represents the actual product the data center is designed to host. The campus itself, the buildings and electrical infrastructure, are the platform. The chips are the value.

This is part of why the announced investment numbers have grown so large so fast. The infrastructure costs were always significant, but they were on the order of single-digit billions for a major campus. The IT equipment costs, in the AI era, are an order of magnitude larger and refreshed continuously. A campus like Stargate's Abilene site or Microsoft's Mount Pleasant complex is properly understood as ongoing capex of several billion dollars per year for the operational life of the facility, not a one-time construction cost.

The Labor Cost Buried in the Schedule

The construction labor on a hyperscale campus is a major line item, but it gets reported under "construction costs" rather than broken out separately. Total construction labor on a single 100-megawatt building runs $80 million to $150 million across the 18 to 30 month construction period. That includes electrical, mechanical, controls, sheet metal, ironwork, civil, framing, roofing, fire protection, and general carpentry. Across a multi-building campus phased over several years, total construction labor reaches into the high hundreds of millions to over a billion dollars per campus.

This labor cost is mostly captured by Phoenix-area workers. Contractors hire locally where they can, both because of cost and because the local trade unions have the workers. Out-of-state crews are pulled in for specialty work, but the bulk is local. The economic activity those wages support, in housing, retail, services, gas stations, restaurants, schools, is part of why the regional economic impact of the buildout is significantly larger than the operational employment number suggests.

The Bottom Line on Cost

The $10 billion campus is not $10 billion of land or $10 billion of building. It's roughly $1 billion of land, site work, and shell construction, $2 to $4 billion of electrical and mechanical infrastructure, $2 to $5 billion of IT equipment in the first deployment wave, and ongoing capex that totals more than the initial build over the campus's operating life. The supply chain for the electrical infrastructure is the binding constraint. The buildings are the easy part. The chips are the expensive part. And the campus is, viewed correctly, a multi-decade industrial commitment that nobody enters lightly because the dollars don't allow for casual decisions.

Understanding the breakdown changes how you read the headlines. A $10 billion announcement is not a check that gets written once. It is a 5 to 10 year commitment with most of the spending happening in years 2 through 7, most of the labor happening locally, and most of the equipment ordering happening before ground is broken because the lead times don't allow anything else.

That is the actual capital structure of what's getting built across Phoenix right now. The number is large. The mechanics are real. And the supply chain bottlenecks mean that the industry cannot pivot to alternative locations on short notice, even if the local conversations get harder. Where the projects can land at the speed the economy requires, that is where they are landing. Phoenix is one of perhaps three or four metros in the country where the answer to "can you actually deliver this in time" is yes.

Sources