Data Center Real Estate: The AI Demand Surge

Introduction

For most of their history, data centers sat in a quiet corner of industrial real estate: low-rise boxes near fiber and substations, leased to telecom carriers and corporate IT departments, and underwritten like specialized warehouses. That framing collapsed over the past three years. The hyperscale cloud and artificial-intelligence build-out has turned the asset class into the most capital-hungry corner of real estate. Alphabet, Amazon, Microsoft, and Meta alone committed more than $350 billion of capital expenditure in 2025 and have guided toward roughly $700 billion combined in 2026, nearly double the prior year, with Amazon pointing to about $200 billion, Alphabet to $175 billion to $185 billion, Meta to $125 billion to $145 billion, and Microsoft to $110 billion to $120 billion on its own. Vacancy across North American markets has fallen to a record-low 1.4%, and tenants now sign leases for capacity that will not be delivered for two or three years. The binding constraint is no longer land, capital, or even tenant demand. It is power, and the multi-year queue to secure it.

A decade ago, a real estate banker could treat data centers as a footnote to the industrial coverage group: a niche with a handful of specialist owners and a valuation framework borrowed from warehouses. The numbers no longer permit that. Aggregate data center capex rose 57% in 2025 according to Dell'Oro Group, and total spending on data center infrastructure is forecast to surpass $1 trillion in 2026. To put that in perspective, the four largest hyperscalers' combined annual capital budget now exceeds the entire yearly transaction volume of most traditional commercial property types in the United States.

Broken out by company, the four largest hyperscalers' 2026 capital plans show how concentrated the spend has become:

The reclassification matters because the operating model is genuinely different from anything else in real estate. A warehouse is a shell that protects goods from weather; a data center is a precision environment that delivers conditioned power and cooling to racks of computing equipment at extremely high density, with redundancy engineered so that a single utility outage does not take the building offline. The real estate is the smaller part of the capital stack. Most of the dollars go into electrical and mechanical systems: transformers, switchgear, uninterruptible power supplies, generators, chillers, and increasingly liquid-cooling distribution. That is why a data center cannot simply be folded into the industrial real estate supercycle thesis even though both ride a logistics-and-technology tailwind. The unit of value is power delivered, not square feet leased.

Hyperscale Data Center

A very large facility, typically tens to hundreds of megawatts of IT capacity, built to serve the cloud and AI platforms of a single large operator or a small set of them. Capacity is measured and leased in megawatts of critical IT load rather than in square feet, because power, not floor area, is the scarce input.

This shift in the unit of account is the first thing a generalist banker needs to internalize. When a hyperscaler signs a lease, the headline term is a megawatt figure and a rate per kilowatt per month, not a rent per square foot. Capacity, supply, and absorption across the sector are all reported in megawatts and gigawatts. A market is "tight" when its available power is exhausted, regardless of how much vacant land surrounds it.

Because the unit of value is power rather than floor area, the basic occupancy metric is expressed in megawatts too. Capacity utilization measures how much of a facility's built critical IT load is actually leased, and in this cycle it sits near the ceiling almost everywhere:

A market running at record-low vacancy is simply one where this ratio has pushed toward one across its operators, which is why a banker reads a near-zero vacancy figure as a signal that built power is fully spoken for and new supply has to come from energizing fresh capacity rather than backfilling empty shells.

The contrast with a conventional warehouse is sharp across nearly every underwriting dimension, which is why the two genuinely cannot share a valuation template:

Most of the dollars in that capital stack flow to systems a warehouse never contains, and a meaningful share of a stabilized facility's replacement cost can sit in equipment that may need refreshing within a single hold period. That is why the development and operating discipline of a data center owner looks more like an infrastructure business than a traditional landlord, and why the public owners describe themselves as platforms rather than property companies.

The demand surge is not simply more of the same cloud workloads. AI has changed the physical requirements of the building. Training a large model concentrates thousands of high-end accelerators into tightly clustered racks that draw far more power and generate far more heat per cabinet than the general-purpose servers that filled data centers a decade ago. Rack densities that were once measured in single-digit kilowatts now run to tens of kilowatts and, for the densest AI clusters, beyond, forcing a move from air cooling toward liquid cooling and reshaping the mechanical design of the entire facility.

AI represented roughly a quarter of all data center workloads in 2025, with model training driving most of that incremental demand. The composition is expected to shift around 2027, when inference (running trained models to serve user queries) is widely projected to overtake training as the dominant AI requirement. That transition matters for real estate because the two workloads have different geographic logic.

Training Versus Inference and Why Location Logic Diverges

Training is latency-tolerant and power-hungry. A training cluster can sit far from population centers, wherever large blocks of cheap power and land are available, because a few extra milliseconds of network distance do not affect a multi-week training run. Inference is latency-sensitive: serving a chatbot response or a recommendation needs to happen close to the user. As inference grows, demand spreads from a handful of mega-campuses in remote, power-rich locations toward a denser network of facilities near metropolitan populations. For bankers, that bifurcation explains why both giant gigawatt greenfield campuses and smaller edge-adjacent facilities are being financed at the same time, and why the hyperscale, wholesale, and retail colocation segments are all expanding rather than one cannibalizing the others.

The Cooling Transition and What It Costs

The density jump has forced a physical redesign of the building. For decades, data centers rejected heat with air: computer-room air handlers pushing cold air under raised floors and through hot and cold aisles. AI clusters generate too much heat per rack for air alone to manage economically, which has pushed operators toward direct-to-chip and immersion liquid cooling. That transition is expensive and it is not optional for the densest deployments, so the cost of the mechanical fit-out per megawatt has climbed even as operators have raced to add capacity. For an owner, the design choice is also a commercial one: a facility plumbed for liquid cooling can serve the densest and most valuable AI tenants, while an older air-cooled building may be limited to lower-density workloads.

The other consequence of AI density is that obsolescence risk has changed shape. A well-located warehouse stays useful for decades. A data center's shell and power infrastructure are long-lived, but the cooling and electrical fit-out tuned to a given generation of chips can require substantial reinvestment as densities climb. Underwriting therefore has to separate the durable elements (land, power interconnection, base building, and structural capacity for cooling) from the elements that may need refreshing across a hold period. A banker who treats the entire asset as a single long-lived box will mis-price both the upside and the reinvestment drag.

If there is one idea that defines data center real estate in this cycle, it is that the scarce resource has migrated from land and capital to electricity. Goldman Sachs has estimated that AI will drive a 165% increase in data center power demand by 2030, and the sector is forecast to add roughly 97 gigawatts of capacity between 2025 and 2030, nearly doubling its size in five years. United States data center electricity consumption is projected to climb from about 4.4% of national usage in 2023 to somewhere between 6.7% and 12.0% by 2028. Generating that power, and physically connecting it to the grid, has become the gating item for new supply.

Nowhere illustrates this better than Northern Virginia, the largest data center market in the world. The region reached more than 1,100 megawatts of net absorption in 2025 and over a gigawatt of new capacity delivered, expanding to roughly 4,040 megawatts of total inventory. Data centers consumed an estimated 25.6% of Virginia's electricity as of 2023, a share that could approach 46% by 2030. The local utility, Dominion Energy, saw its contractual commitments to data center customers nearly double from 21 gigawatts to 40 gigawatts in 2024 alone. Available powered shell space in the region fell below 5% by early 2025, and the practical question for a new project is no longer whether land is available but how many years it will take to energize.

That constraint reframes how capital competes. In a normal real estate market, capital chases the best buildings. In data centers today, capital chases the best access to power, and the developers who win are those that locked up interconnection agreements, land, and utility relationships years before. The link to the broader power complex is direct enough that data center coverage now overlaps materially with the work covered in the energy investment banking group, where generation, transmission, and power-purchase economics are the core subject matter.

The demand for data center capacity arrives in a stack of tenant types, and understanding who sits where is essential to reading the market. At the top are the hyperscalers, the public cloud and AI platforms that take down the largest blocks of capacity, often hundreds of megawatts at a time and frequently on a build-to-suit basis. Below them sit enterprises and AI-native companies leasing wholesale capacity, and beneath that a retail colocation layer of smaller deployments and interconnection-driven customers.

The pace at which those tenants take down capacity is captured by absorption, the headline demand-pace metric the sector watches each quarter. Like everything else here it is denominated in power, not space:

In a supply-constrained market absorption is effectively a read on how much new power got energized and immediately spoken for, which is why a record absorption figure alongside record-low vacancy tells a banker the building boom is running flat out yet still failing to outpace demand.

The defining market fact is scarcity. North American primary-market vacancy at 1.4% is the lowest on record, and by some measures tighter still, with JLL putting availability near 1% for a second consecutive year and effectively zero for large contiguous blocks of powered capacity. The scarcity has held through a building boom: primary markets absorbed a record 2,497.6 megawatts of capacity in 2025, far above the prior 1,809.5-megawatt record set in 2024, yet construction still could not outrun demand. Absorption has concentrated in a handful of markets where power and land could still be assembled at scale, with Northern Virginia leading and a second tier of Sun Belt and Midwest markets growing quickly behind it.

Net Absorption

The net change in occupied capacity over a period, measured here in megawatts of critical IT load rather than square feet. Positive net absorption means tenants took down more capacity than they vacated, and in a supply-constrained market it is effectively a read on how much new power got delivered and leased.

The leading markets and their 2025 momentum give a sense of how concentrated the demand has become:

Northern Virginia alone now holds several times the capacity of all United States secondary markets combined, and wholesale colocation, particularly for hyperscaler and AI tenants, accounts for the majority of new leasing volume. Equinix, the largest retail-and-interconnection operator, reported record bookings across its small, medium, and large deal categories in 2025 and delivered record capacity, including more than 90 megawatts through its xScale hyperscale unit; that unit has now leased roughly 430 megawatts across more than twenty operational facilities in over a dozen metros. Digital Realty, the largest wholesale operator, reported a substantial leasing backlog that gives clear revenue visibility into 2026 and beyond. Together the two anchor a United States colocation market projected to exceed $72 billion by 2030, up from roughly $40 billion in 2025.

The Gigawatt Campus and the Deals Defining the Cycle

The clearest evidence of how far the unit of account has scaled is the arrival of the gigawatt campus. Where a marquee data center deal once meant tens of megawatts, the defining 2025 and 2026 projects are measured in gigawatts and tens of billions of dollars, and they are increasingly financed as standalone infrastructure ventures rather than ordinary leases. The Stargate joint venture, announced in January 2025 by OpenAI, Oracle, and SoftBank, set out to invest up to $500 billion in AI data centers, with its flagship campus in Abilene, Texas designed to house more than 450,000 Nvidia GB200 accelerators. Oracle separately signed a five-year, roughly $300 billion compute agreement with OpenAI beginning in 2027, one of the largest contracted commitments in the sector's history.

The hyperscalers are matching that scale on their own balance sheets and through partners, and the roster of gigawatt-class projects now under way captures the absolute size of the build-out:

Meta's Hyperion sits on a site several times the size of Central Park, and Amazon's Rainier is dedicated entirely to Anthropic, showing how a single tenant relationship now anchors a multi-gigawatt campus. For a banker, the lesson is that the underwriting unit is no longer a building or even a lease but a multi-year, multi-billion-dollar capacity commitment, and that the line between a real estate financing and a corporate infrastructure financing has effectively dissolved at the top of the market.

Why Pre-Leasing Has Become the Norm

When vacancy sits near zero and delivery timelines run years, tenants cannot wait for completed buildings. They pre-lease capacity that is still in the design or construction phase, and in many cases reserve power that has not yet been energized. That pre-leasing is what allows developers to finance projects and what gives the public REITs their forward visibility, but it also concentrates risk: a market's reported absorption can be dominated by a handful of enormous commitments, and the tenant credit of the hyperscalers becomes the load-bearing assumption in every underwriting model. A 200-megawatt lease is only as good as the counterparty's willingness and ability to pay rent across a fifteen-year term.

The flip side of record demand is genuine concentration risk. A small number of tenants drive a large share of leasing, a small number of markets hold a disproportionate share of inventory, and a single utility's interconnection capacity can gate an entire region. Bankers underwriting the sector have to hold the demand euphoria and the concentration risk in view at the same time.

The Global Map: FLAP-D and the European Power Squeeze

The same dynamic is reshaping Europe, where the dominant cluster is known by the acronym FLAP-D: Frankfurt, London, Amsterdam, Paris, and Dublin. Combined live capacity across those five markets more than doubled from roughly 1.8 gigawatts in 2019 to about 3.6 gigawatts by 2025, and roughly 70% of new European supply continues to land in them, with London and Frankfurt alone expected to account for around half of total European capacity. Colocation vacancy across FLAP-D fell from a high near 17% in 2021 to a record-low 6.3% by the end of 2025, and the European data center market is projected to roughly double from about $47 billion in 2024 to $97 billion by 2030.

The European story is, if anything, a more acute version of the power squeeze. Amsterdam and Dublin have at various points imposed moratoria or restrictions on new connections because data centers were absorbing too large a share of local grid capacity, and that scarcity is pushing growth toward emerging markets in Spain, Italy, Greece, and the Nordics where power and land remain available. For a banker working cross-border mandates, the lesson is that the binding constraint travels even when the regulatory regime does not: every major market is rationing power, and the winners everywhere are the platforms that secured interconnection early. Asia-Pacific markets such as Tokyo, Singapore, and emerging hubs in Malaysia and India add a third leg, with Singapore's own historical moratorium on new capacity a textbook example of power and sustainability limits forcing demand into neighboring geographies.

Scarcity has done what scarcity always does: it has pushed up pricing. Rental rates have risen across primary markets as available powered land has dwindled, with the steepest increases in build-to-suit projects where rising costs of capital, land, construction, and equipment all flow through to the rent. For the first time in the asset class's history, landlords in the tightest markets hold meaningful pricing power against even the largest tenants. The hard numbers bear this out: average asking rates for a 250-to-500-kilowatt requirement in Northern Virginia rose 6.5% year-over-year to about $195.94 per kilowatt per month at the end of 2025, a fourth consecutive annual increase, with primary-market rates for that size band now pushing past $200. Land has repriced in step, with site costs for recent and pending data center transactions in Northern Virginia and the Northeast exceeding $8 million per acre.

On the capital side, stabilized data centers trade at cap rates that have stayed relatively firm despite the broader rate environment, with institutional-quality, well-leased facilities generally pricing on the order of 100 to 150 basis points above the ten-year Treasury yield. Investor sentiment has cooled slightly at the margin, with a larger share of investors expecting cap rates to drift up than in the prior year, but demand for quality assets has continued to outstrip supply, which has kept pricing resilient. The mechanics of how those cap rates are set, and why a long lease to a hyperscaler compresses them, follow the same logic explained in the broader cap rate framework, adjusted for the credit and lease-duration profile specific to this sector.

That capital intensity changes the question a banker asks. The relevant comparison is not "what does this building yield" but "how fast can this platform deploy capital into new capacity at an attractive development spread, and how is it funding that deployment." The full build-cost and time-to-power math behind those development spreads is laid out in the data center capex article. A platform that can fund a multi-year pipeline cheaply and energize it quickly is worth far more than the same square footage in the hands of an owner who cannot.

The data center surge is a capital-formation story as much as a property story. The volume of capital required to build out the AI infrastructure pipeline far exceeds what the existing public REITs and private platforms can fund from retained cash flow, which means the sector has become a magnet for new equity, new debt, joint ventures, and entirely new vehicles. The financing structures are themselves becoming a banking product: Meta's Hyperion campus is funded through a roughly $27 billion joint venture with Blue Owl Capital that keeps most of the build off Meta's own balance sheet, a private-credit-meets-real-estate template that the rest of the sector is now studying, while Stargate pools sovereign, strategic, and private capital into a single development vehicle of unprecedented size. The arrival of fresh public-market vehicles, including Blackstone's digital infrastructure trust, signals that sponsors see room to bring more capacity public, and the public data center pipeline is widening accordingly.

The deal flow that results spans the full investment banking product set:

The same AI-capex cycle that is forcing software and private equity to reprice their assets, a dynamic visible in the broader AI valuation reset, is on the real estate side a once-in-a-generation expansion of a property type into a core institutional allocation. The capital required exceeds what any single source can supply, which is exactly why the sector has drawn sovereign wealth funds, pension plans, infrastructure funds, and dedicated digital-infrastructure vehicles into the same deals.

Data centers have become a distinct asset class with their own vocabulary, their own scarce input, and their own capital dynamics, sitting at the intersection of real estate, energy, and technology. A candidate who can explain why power is the binding constraint, why capacity is measured in megawatts, why pre-leasing concentrates tenant-credit risk, and why the returns come from development rather than passive ownership has already demonstrated more sophistication than the headline "AI is driving data center demand" that everyone repeats. The sections that follow build out each piece of that picture, from the public REIT landscape to the lease structures, the tenant universe, and a full valuation walkthrough.