4. The Moment Servers Became an Asset Class

The Day the Cornfield Disappeared

In the spring of 2021, a farmer in the American Midwest heard from a neighbor that the parcel next to his had sold. The buyer wasn’t a grain company, or a neighboring farm operator. It was an unfamiliar real estate development entity, and behind that entity stood one of the most famous tech companies in the world. A few months later, where the cornfield had been stood a concrete box the size of dozens of football fields combined, windowless, with almost no people in sight. Inside, fans spun and cables hummed around the clock while tens of thousands of servers computed answers to the questions we ask.

Around the same time, similar scenes were playing out on the other side of the planet. In Asia, a wave of new site development swept through places like Johor Bahru and the outskirts of Jakarta, which offered both grid capacity and proximity to subsea-cable landing points, while in Europe the opposite story was unfolding in places like the outskirts of Dublin, already dense with data centers, where new grid interconnection had effectively frozen. Land that was once a field, an empty lot, or an already-saturated industrial park either turns into a data center overnight, or finds itself stuck, wanting to but unable to. The name for this two-track drama isn’t simply “IT infrastructure buildout.” It is the redrawing of the real-estate asset-class map itself.

From “Appendage” to “Asset Class”

As recently as 2021, the data center was an odd creature in the world of commercial real estate. Rather than being treated as an independent asset class alongside office, retail, industrial, and residential, it was handled as an appendage tacked onto the telecom/IT sector, typically stuffed into a parenthetical, as in “industrial real estate and data centers.”

Over the past five years, that status has flipped entirely. The fact that global real estate advisory firms have begun publishing separate, annual market-outlook reports devoted solely to data centers is itself evidence of the shift.¹ A commercial real estate advisory firm standing up a dedicated research unit and an annual publication for a single asset class is a signal that the class is no longer a sideline: it has been recognized as an independent object of capital allocation. Just as office and logistics did before it, the data center now has its own line on the institutional investor’s allocation table.

As of 2026, the top 14 data center operators worldwide are estimated to be spending roughly $750 billion a year in combined capital expenditure.² Capital moving at this scale can no longer be called “equipment purchasing.” You buy land, secure permits, bring in power, and put up buildings — following the same grammar as a conventional real estate development project, just at a different scale and speed.

Annual capex among America’s six hyperscalers — Microsoft, Meta, Amazon, Alphabet, Oracle, and Apple — is projected to grow roughly sixfold from 2022 to 2026, approaching $700 billion.³ The top five companies spent roughly $256 billion in 2024, an estimated $443 billion in 2025, and an estimated $602 billion in 2026.³ Roughly three-quarters of this is going to AI infrastructure, not just chips and servers, but the buildings to house them and the generation and transmission assets to power them. Real estate, energy, and manufacturing are effectively fusing into a single industry.

When an asset class’s identity changes, so do the rules of the game played over it. An office building’s value used to be set by tenant creditworthiness and location accessibility. A data center’s value is now set by three entirely different variables: power, cooling water, and distance to the telecom backbone. If the maxim of twentieth-century real estate was “location, location, location,” the maxim of twenty-first-century data-center real estate has become “wherever the power surplus is.”

Five Years From Training to Inference: Demand Changes Character

Layered onto all this is the fact that AI demand itself isn’t static. As of 2025, AI accounted for roughly a quarter of total data-center workload, and most of that was still training AI models.⁴ Training demand tends to cluster in a small number of enormous clusters. You pile a huge amount of compute in one place and run it nonstop for weeks or months, so the field of candidate sites narrows to a handful of mega-campuses. A training cluster doesn’t need to talk to a user in real time, so as long as power and cooling water are secured, it can sit in the remotest corner of the earth.

Starting around 2027, “inference” demand — users asking AI questions in real time and getting answers back — is expected to overtake training demand.⁴ Inference is a completely different animal. It needs to minimize response latency to the user, so it has to be deployed close to users, as nodes distributed across many regions. If you ask a chatbot something and the answer takes three seconds too long, that service has failed. In other words, the data-center real estate map of the next five years is entering a phase of reorganization — from “a few mega-campuses” to “many regionally distributed nodes.” This isn’t simply capacity expansion; it’s a redesign of location strategy itself. Data centers are repeating, at a much faster pace, a trajectory similar to what logistics centers went through a generation ago, evolving from a single urban warehouse into a regional fulfillment network.

This inflection point hands developers a double assignment. Right now, betting on training-era mega-campuses is what wins the large leases that capital is chasing, but it also means shouldering the risk that, five years from now, that oversized asset becomes a stranded asset that’s simply “too big and too remote.” Conversely, a developer that starts putting a small foot into regionally distributed nodes today looks modest in scale right now, but stands to enjoy the first-mover advantage of an already-built network once the inference era hits full stride.

The Bottleneck Moved: From Money to Electricity

For a real estate developer, the scariest question was always “can I raise the money?” In the data-center game, that question is no longer the scariest one. As recently as 2021, the development bottleneck was financing and chip supply. By 2025–2026, the bottleneck had shifted, unmistakably, to power.⁵

In some regions, a new data center may wait up to four years to connect to the local grid.⁵ Capital is abundant, but there isn’t enough electricity to build: an ironic bind. Developers are cutting through this bottleneck by bypassing the existing grid altogether and securing their own generation, in a strategy known as BYOP (Bring-Your-Own-Power).⁵ Data-center developers have effectively started doubling as power producers.

In March 2024, Amazon Web Services (AWS) acquired a data-center campus adjacent to Talen Energy’s Susquehanna nuclear plant in Pennsylvania, and entered into a long-term supply agreement directly with the generator for up to 960 megawatts of power.⁶ Just as twentieth-century industrial parks clustered next to ports, twenty-first-century AI industrial parks have begun clustering next to power plants. Energy supply agreements have become mandatory attachments to real estate contracts. This is also a qualitative shift from the 2021–2023 pattern, when Big Tech mostly bought virtual power purchase agreements (vPPAs, essentially renewable energy certificates), toward directly owning generation assets or drawing power directly, behind the meter.⁶

The rise of rural areas as the new stage for development, where land is cheap, power plants and transmission lines are close, and zoning review is simpler, follows the same logic.⁷ Knowing that a single large data center can consume as much power as a decent-sized town explains why developers head for farmland to avoid urban permitting fights and community opposition.⁷ The competition for rural land near renewable energy plants is playing out not just in America’s Corn Belt but across Europe too. Every single question we type into our phones is now capable of triggering a boardroom decision, somewhere on earth, about whether to spin up new generating capacity.

A World of 2% Vacancy: Two Parallel Real Estate Markets

During this same period, downtown office markets were still struggling with vacancy in the wake of remote work. In the same cities (sometimes even the same suburban industrial parks), the data-center market was moving in exactly the opposite direction. The average vacancy rate across major U.S. data-center markets fell below 2% in 2025, its lowest level in at least 12 years.⁸ In Europe, supply has failed to keep pace with demand, and vacancy is projected to fall to 6.5% by 2026.⁸

In real estate, vacancy below 2% signals an extreme seller’s market: one where the landlord sets the terms, not the tenant. While office buildings sit emptying out, somewhere a cornfield is being filled in with a server-packed concrete box. Real estate is no longer a single market; it has splintered into multiple parallel worlds running on different cycles. Watching the same country, the same capital markets, live through recession on one side and one of the hottest booms in history on the other, simultaneously, would have been hard to imagine even five years ago.

How the Capital Gathers: A Financing Structure That Grew Sixfold in Five Years

Financing this much capital-intensive development required the funding structure to evolve too. Over the past five years, four distinct streams of capital have poured into the data-center asset class — publicly listed REITs offering a tax pass-through, large-scale private infrastructure fund inflows, project-finance debt, and hyperscalers’ own direct lease commitments, forming not a triangle but a full four-sided formation.⁹

Each type of capital has a different character. Listed REITs draw public money chasing dividends and promise stable rental income, but they can’t escape the quarterly earnings pressure that comes with being publicly traded. Private infrastructure funds have a far longer investment horizon and can absorb entire power-generation assets, but haven’t yet produced a confident answer to the question of how they’ll exit.⁹ Hyperscalers’ direct leases (commitments running 10 or 15 years) give developers a cash flow arguably safer than a bank loan, but in exchange, the developer takes on concentration risk, depending on a small handful of tenants for its entire revenue. Unlike a traditional office building mixing five or ten tenants, a single data center often has exactly one.

Poorvu’s insight that a REIT is “two things at once” — a real estate deal and a Wall Street product — applies here as well.¹⁰ On the surface, a data-center REIT is a stable, dividend-paying real estate product; underneath, it’s a far more complex infrastructure asset, sometimes owning power plants outright, sometimes shouldering long-term power contracts.

Some regulatory questions remain unresolved. There’s no clear standard for whether nuclear or natural-gas generation assets can be housed inside a REIT structure.⁹ For private-fund investors, “who exits to, and how” has emerged as a new core risk. Sale to a listed REIT, or an IPO of the platform itself, are the exit scenarios most often floated, but few proven track records exist yet.⁹ A sector that a handful of specialist REITs quietly ran as recently as 2021 has, by 2026, fully transformed into a competitive asset class that mainstream institutional capital is racing to enter.

Redrawing the Game Diamond

Let’s take a moment to overlay the lens running through this whole book onto the data center. Poorvu drew the real estate game as a diamond, four variables (assets, capital markets, players, and the external environment) connected by arrows to one another.¹¹ Set this new game board, the data center, on top of that diamond, and it becomes clear that all four corners now wear a completely different face than they did five years ago.

Assets are no longer priced by location, square footage, and finish quality. Power capacity (measured in megawatts), position in the grid-interconnection queue, and cooling method have become the new appraisal line items. Capital markets now see private infrastructure funds and hyperscalers’ own bond issuance sitting alongside traditional bank debt and public equity: capital that’s far larger, far more patient, and concentrated in far fewer hands than office development ever attracted. The player deck has been reshuffled entirely. In place of the traditional local developer, a new lineup has emerged: the hyperscaler (effectively the world’s largest tenant, and often a co-developer as well), specialist data-center REITs, power generators, and the real estate advisory firms brokering all of it. What stands out in particular is that the tenant, the hyperscaler, has started investing in power assets itself, taking on a dual role as both developer and tenant, a part rarely seen in Poorvu’s traditional game. The center of gravity in the external environment has shifted too, away from tax policy and demographic trends and toward grid regulation, local political battles over generation permitting, and judgment calls about the sustainability of AI demand itself, all of which now decide who wins and loses this game.

One principle Poorvu emphasized still holds: the four card decks keep pushing and pulling on one another endlessly. Grid bottlenecks (external environment) give rise to self-generation strategies (a redefinition of the asset), which in turn draws in private infrastructure funds willing to take on power assets too (capital markets), which pushes new players — hyperscalers doubling as power producers — to center stage. The diamond’s skeleton is unchanged, but what fills it is an entirely different game from five years ago.

AI’s Real Demand Isn’t the Cloud

Let’s return here to the core message of this chapter. We tend to think of AI as a software problem, an algorithm problem, a cloud-subscription problem. The experience of asking a chatbot a question and getting an answer feels highly virtual, intangible. But behind that experience sits infrastructure that is intensely physical and tangible: land, concrete, wiring, cooling water, and power plants.

What the past five years of data-center real estate history shows is that AI’s real bottleneck isn’t algorithmic sophistication — it’s the electricity to run it and the real estate to house it. The race to build the world’s smartest AI model is converging into a race to see who can secure land next to a power plant, and win grid-interconnection approval, faster. If the winner of the late-twentieth-century real estate game was whoever spotted the best location first, the winner of the AI-era real estate game is whoever spots the surplus power first.

And the next phase of this game has already begun. As demand shifts from training-centric to inference-centric, the data-center real estate map of the next five years will be redrawn not as a handful of mega-campuses but as countless distributed nodes scattered across the globe. Before that map is finished, the next chapter looks deeper into where, exactly, this new asset class will actually get built — the geography of a new scarcity made of electricity, water, and land.

---

Rule of the Game

What AI needs isn’t the cloud — it’s land and electricity.

In five years, the data center has been promoted from “IT appendage” to independent asset class. Driving that promotion: hyperscalers’ roughly $750 billion in annual capex, and the inflection point where demand shifts from training-era mega-campuses to inference-era distributed nodes. The game’s bottleneck moved from capital to power, and the funding structure has evolved into a new four-way formation mixing REITs, private infrastructure funds, and direct leases. On Poorvu’s game diamond, the emergence of a new role — the tenant (the hyperscaler) doubling as its own developer and power producer — is the most fundamental change these five years have produced.

---

Footnotes

1. JLL, “2026 Global Data Center Market Outlook”; CBRE, “Global Data Center Trends 2026” — the fact that CRE advisory firms have begun issuing annual reports dedicated to data centers is itself cited as a signal of the asset class’s independence. ↩

2. Ropes & Gray, “Data Center Investment in 2026: AI Demand, Power Constraints, and Private Equity Trends”; HB Capital RE, “Data Centers CRE 2026: The $700B Industrial Adjacency” — estimate of roughly $750 billion in combined annual capex among the top 14 data-center operators. ↩

3. CreditSights, “Technology: Hyperscaler Capex 2026 Estimates”; Introl, “Hyperscaler CapEx Hits $600B in 2026”; Yahoo Finance, “Meta, Microsoft, Amazon, and Alphabet are about to spend a shocking amount of money to dominate the AI era” — roughly sixfold growth in U. S. six-hyperscaler capex 2022–2026, with year-by-year estimates for 2024–2026. ↩ ↩²

4. BloombergNEF, “AI Data Center Build Advances at Full Speed: Five Things to Know”; JLL / Data Center Frontier, “JLL’s 2026 Global Data Center Outlook” — 2025 AI workload share and the projected training-to-inference demand shift (around 2027). ↩ ↩²

5. The AI Consulting Network, “AI Data Center Power Crisis: CRE Site Selection 2026”; EnkiAI, “Hyperscaler AI & Data Center Energy 2026” — grid-interconnection wait times (up to four years), the BYOP (Bring-Your-Own-Power) strategy. ↩ ↩² ↩³

6. FTI Consulting, “Power, Renewables & Energy: 2025 M&A Review, 2026 Outlook” — the shift from vPPAs to behind-the-meter power procurement; AWS’s acquisition of the campus adjacent to Talen Energy’s Susquehanna nuclear plant and the up-to-960MW supply agreement (March 2024, per SEC filings). ↩ ↩²

7. LandApp, “Why Data Centers are Driving Rural Land Demand”; LightBox, “A Growing Demand for Land: Site Selection for Data Centers Insights”; American Farm Bureau Federation, “Balancing Data Center Growth with American Agriculture” — the preference for rural sites and the tendency to bypass urban permitting. ↩ ↩²

8. Ropes & Gray, “Data Center Investment in 2026”; CBRE, “European Data Centres Outlook 2026” — sub-2% vacancy in major U. S. markets (2025, lowest in 12 years); projected 6.5% European vacancy in 2026. ↩ ↩²

9. Ropes & Gray, “Data Center Investment in 2026”; Angel Investors Network, “Data Center REITs: 39–45% Returns on AI Infrastructure” — REIT/private infrastructure fund/project-finance funding structures, regulatory uncertainty around housing power assets in REITs, exit-strategy issues. ↩ ↩² ↩³ ↩⁴

10. William J. Poorvu & Jeffrey L. Cruikshank, The Real Estate Game (1999) — reinterpretation of the frame describing REITs as “a real estate deal and a Wall Street product” (paraphrased, not a direct quotation). ↩

11. William J. Poorvu & Jeffrey L. Cruikshank, The Real Estate Game (1999), the “game diamond” frame (assets–capital markets–players–external environment) — reconstructed in this chapter and applied to the data-center asset class. ↩