AI Infrastructure as an Investment: How to Evaluate Data Center Opportunities

A framework for judging an AI infrastructure asset the way an operator would, before the projections take over.

AI infrastructure is one of the largest capital build-outs of the decade. The returns will not come from AI adoption. They come from the physical fundamentals underneath the compute.

Capital is moving into AI infrastructure faster than most sectors can absorb it. McKinsey estimates data center capital outlays will reach roughly $7 trillion by 2030, with about $5.2 trillion tied specifically to AI workloads. That scale of spend pulls in money that treats every asset the same and prices the sector on the theme rather than the fundamentals.

The assets that perform will be the ones underwritten on how they actually operate, not on the assumption that demand covers every mistake. This is a framework for evaluating an AI infrastructure opportunity the way the people who build and run these facilities would, one factor at a time.

Why AI Infrastructure Is Not a Bet on AI Adoption

It is tempting to underwrite an AI data center as a proxy for AI growth. That framing hides the risk. The demand for compute can be real and rising while a specific asset still underperforms, because the asset earns on power delivered, hardware utilized, and contracts signed, not on the size of the market around it. A facility with the wrong power position or the wrong tenant will struggle in the strongest market. The theme sets the tailwind. The fundamentals set the return.

Power Availability Decides the Deal Before Anything Else

Power is the first underwriting question, not the last. An AI campus with 500 megawatts of design capacity and no firm power agreement is a land parcel with a diagram. The practical test is simple: is the power contracted, and when does it energize? A firm power purchase agreement with a defined date is an asset. A letter of intent from a utility is a hope, and the gap between the two is where most AI infrastructure timelines slip. Before any other number matters, confirm that the electricity behind the deal is real and dated.

Location Is a Power and Latency Decision, Not a Real Estate One

The value of a site comes from what connects to it, not its square footage. Two variables dominate. The first is access to energized, affordable power. The second is proximity to the workloads it will serve. Training clusters can sit far from population centers because latency matters less, which lets them chase cheap power in remote markets. Inference capacity has to sit near users, which pulls it toward more expensive, more constrained metro grids. A site that looks cheap on a per-acre basis can carry a multi-year grid delay that erases the discount. Judge location by power cost, grid queue position, water access for cooling, and network latency to the demand, in that order, and treat cheap land with no power as what it is.

Tenant Quality and Contract Structure Set the Cash Flows

The lease is the asset. A hyperscaler on a long triple-net lease produces a different risk profile than a venture-funded startup on a two-year term, even at the same headline rent. Underwrite the tenant’s ability to pay through a full model cycle, not just their current funding round. Look at contract length against the hardware refresh cycle, escalators against power cost inflation, and what happens to the space if the tenant fails. In a market this young, tenant concentration is a real risk. One anchor that represents most of the revenue is a single point of failure wearing a marquee logo.

Capacity and Utilization Separate a Productive Asset From an Idle One

Nameplate capacity is a marketing number. Utilization is the one that pays. A facility rated for 100 megawatts that runs at 40 percent earns like a 40-megawatt facility while carrying the cost base of a 100-megawatt one. Ask how power is actually drawn across the day, how much of the design capacity is contracted rather than speculative, and whether the electrical and cooling systems can sustain high-density racks at full load. Modern AI racks now push past 100 kilowatts, and facilities built for 5 to 10 kilowatt racks cannot host them without a retrofit that changes the return math. The difference between rated and real is the difference between the pitch and the performance.

Financing Structure Changes the Return More Than Rent Does

How the asset is financed moves the equity return more than a few dollars of rent per kilowatt. AI hardware depreciates on a faster cycle than the building around it, which makes capital structure a central question rather than a back-office detail. Off-balance-sheet financing, asset-backed structures on the hardware, and separating the real estate from the equipment each change the risk and who carries the depreciation. McKinsey’s breakdown of the projected AI build-out puts about 25 percent of the capital, roughly $1.3 trillion, into the power and energization layer alone. Investors who understand where capital sits in the stack, and who bears the hardware risk, will price these deals more accurately than those looking only at yield on cost.

Time-to-Power Is the Clock Every Asset Runs On

Time-to-power is now the defining constraint on AI infrastructure returns. According to RMI, the average time from an initial US grid interconnection request to commercial operation reached nearly five years in 2024, up from under two years in 2008. More than 2.2 terawatts of generation and storage sit in interconnection queues, and only 19 percent of projects that requested connection between 2000 and 2019 had reached operation by the end of 2024. Every month of delay is a month of carrying cost against no revenue. When you evaluate a deal, the energization date is the number that decides when the asset starts earning. Treat any project without a credible path to energized power as pre-development, and price it that way.

What to Confirm Before Underwriting an Opportunity

Due diligence for an AI infrastructure asset comes down to a short set of questions, and a weak answer to any one points to where the return will leak. The first is whether the power is contracted and when it energizes, stated as a date rather than a plan. The second is the real time-to-power given the site’s queue position. The third is who the tenant is, how long the lease runs, and whether their credit holds through a full cycle rather than a single round. The fourth is how much capacity is contracted versus speculative, and what utilization the model actually assumes. The fifth is whether the facility can sustain racks above 100 kilowatts at full load, or whether a retrofit is hiding in the numbers. The last is execution: whether the operator has delivered megawatts before, or only planned them. A team that controls hardware sourcing, financing, colocation, and operations across the chain deploys in weeks, while a group assembling those pieces deal by deal loses quarters to sequencing.

This is the vantage point Vertical Data works from, coordinating hardware, financing, and energized colocation as one stack so an asset moves from diagram to deployed without stalling in the gaps between vendors. For an investor, the same coordination is what separates an AI infrastructure asset that earns on schedule from one that waits.