Data Centers Are the New Warehouses: The Hottest Trade in Real Estate

Every Cycle Gets a Darling

Each real estate cycle takes one type of property and turns it into the business everyone should own. In the 2010s that was the warehouse. Online retail needed logistics space close to every major city. Vacancies fell to 3 percent by 2022. Rentals piled up for a decade and a category that investors used to dismiss as boring boxes became the best-performing property type of the time. The 2020s have delivered thatcrowns the data center the specialized industrial building that houses the servers behind cloud computing and now artificial intelligence. Same pattern. A technological change creates a structural demand for a physical shell and the shell changes price. What's different this time is the size of the numbers behind it and I would say what the building actually is under the concrete and the security fence. More on that later

Seven Hundred Billion Dollars of Demand

Demand is dominated by hyperscalers the handful of companies running cloud platforms on a global scale. The six largest Microsoft Amazon Alphabet Meta Oracle and Apple are projected to spend roughly $700 billion on capital spending in 2026. That's nearly six times what the same group spent in 2022 and most of the increase is AI infrastructure: buildings power equipment and the chips inside them. Training and running large AI models requires density.that older facilities simply can't offer so almost all of that demand falls on new construction. The year 2026 is on track to set another leasing record. Tenants now routinely sign up for buildings years before they exist a practice the industry calls pre-leasing.When your client commits capital before you have poured concrete you know who has the negotiating power

Vacancy Barely Exists

National data center vacancy in the United States is hovering around 2 percent a record low. Northern Virginia the largest data center market on the planet is effectively 0.3 percent full. Atlanta is running close to 1 percent. For comparison a healthy office market was previously considered to have 10 percent vacancy putting office vacancy at about five times the data center rate even in a good year.Office prices continue to tighten. Rents are at record highs and continuing to rise in every major market and one detail shows how much this has changed: Large tenants who buy 10 megawatts or more used to negotiate volume discounts the same way a wholesale buyer gets a better unit price almost anywhere else. Today those same tenants pay a premium for scale because guaranteed capacity is worth more to them than a lower overall rate. When your customers get moreImportant companies stop asking for a discount you are no longer in a normal market

The Building Is a Shell Around a Power Contract

This is the fact that took me the longest to internalize: data center leasing has no price for space. The industry doesn't measure deals in square feet. It sizes them in megawatts of energy capacity because electricity not square footage is the scarce input. Ask a data center broker how big a deal is and the answer will be in megawatts before it gets to square feet. Leasing a warehouse is a bet on location and access to a highway interchange. Leasing a data center behaves more like a long-term power purchase agreement that comes bundled with a building

That distinction changes what actually limits supply. For a warehouse developer the binding constraint is land near the right highway exit and land you can typically buy if you pay enough for it. For a data center developer the binding constraint is position in interconnection queue that is your place in line for a utility to connect your site to the grid at the capacity you need. A grid connection for a large campus can take 24 to 48 months to secure longer when new transmission lines or new generation must first be built and that queue not the zoning not the land itself is now the industry's real bottleneck. Of the roughly 16 gigawatts of U.S. capacity planned for2026 about two-thirds had not begun construction as of this spring and analysts expect that $150 to $200 billion of planned spending will slip through 2027 and 2028 for exactly this reason: the power was not yet available. The shift toward giant AI campuses of 500 megawatts and larger has pushed construction schedules into multi-year territory and converted land that already has secured power making theindustry calls motorized land in the most valuable input of the business

There is a second consequence of pricing the deal based on energy rather than space and I think it is the one that is underweight. A warehouse ages the same way a building ages: the roof needs to be replaced the loading docks need to be resurfaced the roof needs to be replaced again. A data center ages the same way a laptop ages. Chip density continues to increase and cooling requirements continue to change from ordinary air cooling to liquid cooling and an engineered facility.For 2020 racks it can fall off the top tier long before its concrete needs actual repair. This asset's obsolescence clock runs on the tenant's hardware upgrade cycle not the landlord's maintenance schedule. This is a truly unusual risk for a real estate investor because typically the building sets the pace of deterioration and in this case the client's chips do

How the Landlords Get Paid

Two business models divide the industry and each monetizes the power constraint a little differently. hyperscale or the wholesale model leases an entire building or campus to a single giant tenant in a long-term triple net lease meaning the tenant not the landlord pays taxes insurance and maintenance in addition to rent. A fifteen-year lease to Microsoft behaves much more like a corporate bond than a building and investors value it accordingly with lower yields than most other commercial real estate. placement This model run by companies like Equinix rents smaller spaces to many tenants at a time and charges separately for the interconnection between them. Its management is more operationally complex and more diversified since no single tenant's decision can empty the building overnight. Public investors come to the sector mainly through REITs such as Equinix and Digital Realty while private giants and infrastructure funds own much of what remains. The biggest profits from the business are foundin development: build at a high return on cost then watch how the market values the finished and leased asset at a much lower stabilized return. That differential between the two returns is where the sector's fortunes are really made and it's worth analyzing with real numbers

A Worked Example: Pricing a Lease by the Kilowatt

Let me nail down that performance with round illustrative numbers not an actual project. Let's say a developer builds a 50-megawatt data center campus meaning that 50 megawatts of critical IT capacity is actually delivered to the servers once power and cooling losses are subtracted. Construction cost $12 million per megawatt a reasonable illustrative figure given how capital-intensive this build has become. Total cost:50 megawatts for $12 million per megawatt or $600 million

Now price the lease. Data center rent is priced per kilowatt per month and one megawatt is equal to 1,000 kilowatts so 50 megawatts is 50,000 kilowatts. Suppose the market rent is $100 per kilowatt per month. The monthly rent is 50,000 kilowatts times $100 or$5,000,000 per month. Multiply that by 12 months and the annual rent is $60,000,000. Because this is a triple net lease the tenant already covers the taxes insurance and maintenance so I treat that $60 million as close to the owner's net operating income with no further deductions

Performance over cost is the annual revenue divided by the total cost: 60 million divided by 600 million is 0.10 or 10 percent. That's the return the developer gets on the money he actually spent to build the thing

Now let's assume that the market once the building is stabilized and fully leased values a bond-like asset like this at a 6 percent yield consistent with the source point that these long-term leases are priced closer to a bond than typical real estate. The value equals income divided by the yield: 60 million divided by 0.06 is $1,000,000,000 or $1 billion.The developer spent $600 million to build something that the market will pay $1 billion for. That's $400 million in value created about 67 percent more than it cost to build

One more number worth considering: simple payback cost divided by annual income completely ignoring financing and taxes. 600 million divided by 60 million is exactly 10 years. This is not the same as the developer's actual yield since it ignores the eventual sale but it is a quick check of how long the building has to continue collecting rent before it has paid for itself on cash flow alone. Each number in this example is illustrative. Actual projects vary depending on themarket tenant credit and the amount of energy cost that is passed on. But the relationship between the three numbers construction cost per megawatt rent per kilowatt and the gap between return on cost and stabilized return is the real driver of the business

Case Study: Amazon, Talen Energy, and the Nuclear Plant Next Door

The clearest example I know of how the powered earth really decides who wins in this business is the deal Amazon signed with Talen Energy around the Susquehanna nuclear plant in Pennsylvania. Amazon purchased a data center campus located directly next to the Talen nuclear station and arranged to draw power directly from the plant itself rather than waiting in the normal queue to connect to the regional grid. The logic is simple once you see it: Instead of spending years securing a new grid interconnection from scratch buya site that is already on top of a huge reliable and already allowed generation and skip the line completely

The agreement did not go smoothly. The regional grid operator PJM Interconnection and several utilities filed objections with the Federal Energy Regulatory Commission concerned that allowing a large private customer to draw power directly from a plant that also serves the public grid could increase costs or reliability risk for everyone else on the system. Regulators rejected parts of the original agreement and it had to be revised in a more limited way. I don't think the exact regulatory exchange is the important part here. The important thing is that one ofThe best-capitalized companies in the world with roughly the same access to land and capital as any other company has ever had chose to go directly to a nuclear plant rather than fight a normal interconnection queue and even then ran into a political and regulatory fight over network access. If Amazon can't simply get around the power constraint that indicates the constraint is real

The Bear Case: Where This Model Breaks

The risks here are real and deserve more than a sentence of coverage added at the end. The concentration of demand is extreme. A handful of tenants manage the majority of leasing in the entire sector so the fate of the sector is tied to the capital budgets of six companies. Every capital spending boom in history has eventually been surpassed including the warehouse boom to which this sector continues to be compared where industrial vacancy roughly doubled its 2022 lows onceThe wave of new supply has finally landed. There's no reason to assume that data centers are exempt from that pattern just because the current numbers seem extreme

The technology risk is the one I take most seriously because it flies in the face of the bull case in a way that warehouse history has never had to contend with. Increasing chip power densities and the move toward liquid cooling continue to change what an optimal building actually looks like so a facility designed for 2020 workloads can surpass the upper tier long before its concrete needs real repair. A warehouse built in 2010 still moves boxes perfectly today. It's entirely possible that adata center built in 2010 may no longer be competitive for cutting-edge AI training work whatever its physical condition. And the speculative margin is growing: land purchased without assured power capacity built before any signed lease in a rush to claim sites with power before someone else does. Some of those people will not find a tenant at the assumed rental price or will not find a tenant at all

Here's the honest version of the other side not just a list of risks. The strongest bearish argument is not that demand for AI is slowing. It's that the hyperscalers themselves stop wanting to rent the shell. If a company like Microsoft or Amazon decides that the value lies solely in owning the power and chips and that the building itself is an asset that it can build directly once it has secured the land and a grid connection then the owner's role is reduced to financing and managing theconstruction a much lower margin position than a bond-like fifteen-year lease implies today. I don't think that's the base case because these companies clearly like to keep capital off their own balance sheets when a REIT or infrastructure fund will hold the real estate at a lower required return. But it's the scenario that would make me wrong the most and it deserves to be taken seriously rather than dismissed

How I Actually Think About a Data Center Trade

If someone handed me the presentation for a data center REIT and asked me to underwrite it in an hour I wouldn't start with line-of-trade funds like I would with an ordinary REIT. I would start by asking what the tenant mix is like because a building fully leased to one or two hyperscalers on fifteen-year triple net terms is a completely different risk than a colocation building with fifty smaller tenants on shorter terms. The first one is close to a bonuscorporate.The second is closer to a diversified operating business and I would model it that way

So I would ask about power not space. How much of the portfolio's capacity is on secure interconnected sites and how much is on land that still needs a grid connection a few years from now? I would treat land with electric power in the pipeline as the real growth option in the stock and undeveloped land without power as nearly worthless until proven otherwise no matter how big the acreage number appears on a presentation. My honest read is that many retail investors still value thesecompanies the same way they would evaluate an office owner in terms of square footage and location and that is the wrong perspective for this asset class

The other thing I really do and it took me a while to start doing it consistently is to separate the permanent portfolio from the development process in my head as two different bets rather than one company. The standing and leased buildings are the bonded part. I judge them by the credit of the tenant and the lease term and I don't lose much sleep over the technological risk in a building that is already fully leased for the next decade. The pipeline is the capital-like part where both the technological risk andinterconnection risk and where I specifically want to know if the power is insured before I give the company credit for the megawatts it says it plans to deliver. Putting those two together into a single number which is what an overall gigawatt-in-development figure tends to do is one of the easiest ways to overpay for this sector. I'd rather have less and know exactly what bucket each megawatt falls into

The Bottom Line

Data centers are the real estate sector that defines this cycle: near zero vacancies record rents and roughly $700 billion in annual capital spending at hyperscale. But the asset underlying the commerce isn't actually a building. It's an energy contract sized in megawatts controlled by an interconnection queue that now lasts years instead of months and exposed to an obsolescence clock set by the tenant's chips instead of the owner's roof. The numbers crunched above show whereprofit really comes from: a gap between the return you get from building the thing and the return the market pays once it's rented and stabilized. The story of Amazon and Talen shows that even a company with nearly unlimited capital can't simply buy its way around a power constraint. And the bear case is no footnote: concentration of demand a technology cycle that can outpace building and a hyperscaler's decision to internalize the entire stack are all real ways this trade stops working. ForAs an industry scholar the two numbers that matter most remain hyperscaler capex and interconnect queue because this is the only type of property where the owner's real product is electricity not space