Data Center REITs: The Real Estate Play Hiding Inside the AI Boom

Here is a trade that fell right through the crack between two crowds. The real estate people spent 2025 arguing about office vacancy and cap rates and where mortgage rates were going. The tech people spent it arguing about model benchmarks and whether the AI capex was a bubble. Almost nobody was standing in the middle, looking at the companies that are, quite literally, the landlords of the AI boom.

That middle is where data center REITs live. And in 2026 the middle won. The two public pure-plays, Equinix and Digital Realty, plus their storage-and-records cousin Iron Mountain, put up total returns of roughly 39%, 41%, and 45% over the twelve months ending April 20, 2026. The Vanguard Real Estate ETF, the thing most people think of as “real estate,” returned 12.87% over the past year. The data center names are up about 36% year-to-date while the rest of the sector barely moved.

I studied computer engineering, so my instinct with anything labeled “AI trade” is to ask what the actual physical bottleneck is. Not the narrative. The bottleneck. And once you trace the AI stack down from the chatbot to the model to the GPU to the rack to the building to the substation feeding it, you land on a boring, unglamorous, deeply un-sexy asset: powered floor space. That is what these REITs sell. There is no software moat, no model to leapfrog. They own the one thing every AI company needs and none of them can conjure out of thin air.

What these companies actually rent out

A data center REIT owns the building, the power feeds, the cooling, and the security, then leases it to tenants who bring their own servers. Because it is structured as a real estate investment trust, it passes the rental income through to shareholders. The tenants are the names you already know: cloud providers, the hyperscalers, and increasingly the AI labs renting capacity to run inference at scale.

Equinix and Digital Realty are not the same business, though, and if you are going to own one you should know which bet you are making. Equinix runs interconnection colocation. Think dense, high-value facilities where thousands of networks physically cross-connect to each other. That is a premium, sticky, high-margin model. Its adjusted EBITDA margin hit 51% in Q1 2026. Digital Realty leans the other way, toward large hyperscale build-to-suit capacity, whole buildings leased to a single cloud giant. Lower margin per square foot, but enormous scale and the ability to swallow a 200 megawatt deal in one signature.

The Q1 2026 numbers are the whole story

Guidance raises and record bookings are where the AI demand stops being a story and starts being cash flow. Equinix reported Q1 2026 revenue of $2.444 billion, up 10% year over year, with net income up 21% to $415 million. The number that matters most for a REIT is adjusted funds from operations, the cash-flow measure that actually backs the dividend, and Equinix crossed $1 billion in a quarter for the first time ever, landing at $1.07 billion. It booked a record $378 million in first-quarter annualized gross bookings and raised full-year guidance to revenue of $10.14 to $10.24 billion, adjusted EBITDA up to $5.25 billion, and AFFO of $4.20 to $4.28 billion.

Then management said the quiet part into the microphone: AI-driven demand accounted for 60% of its largest deals in the quarter. And crucially, they pointed at inference, not training, as the binding driver. That distinction is the whole investment thesis, and I will come back to it.

Digital Realty's quarter was arguably louder. It signed record leasing generating $707 million of annualized GAAP rental revenue at 100% share, its second-highest quarter ever and nearly 70% above the next-highest quarter on record. Its single largest lease was a 200 megawatt AI inference deal with an AA-rated hyperscaler in Charlotte, the biggest in the company's history. A record 21% of its smallest 0-to-1 megawatt bookings were AI-oriented too, which tells you the demand is not just a handful of whale deals. It raised 2026 core FFO per share guidance by $0.10 to $8.10, implying about 9% growth over 2025, and ended the quarter with a signed-but-not-commenced backlog of $1.8 billion in annualized rent at 100% share.

Vacancy is the tell

If you want to know whether a landlord has pricing power, don't read the earnings call, look at vacancy. Empty space means tenants have options and rents drift down. North American data center vacancy sat at a record-low roughly 1% for a second straight year in JLL's year-end 2025 report. Northern Virginia, the biggest data center market on Earth, was at 0.3%. Atlanta at 1%. That is not a healthy market, that is a market on fire.

The wild part is what happened alongside it. CBRE reported that North American inventory grew 33% in Q1 2026. Supply is being added at a blistering pace, and vacancy still won't budge, because the space is pre-leased before the concrete cures. When you can build a third more capacity in a quarter and still can't create an empty room, you are not looking at a normal real estate cycle. You are looking at demand that structurally outruns supply.

The real scarce asset is power, not property

Here is where the computer engineer in me gets interested, because the bottleneck is not what most investors think it is. It is not square footage. It is not even the GPUs at this point. It is electricity, and the wires to deliver it.

Worldwide data center power demand is projected to rise about 27% in 2026 to 132 gigawatts, up from 104 GW in 2025. The IEA's base case has global data center electricity consumption roughly doubling to about 945 terawatt-hours by 2030, just under 3% of all electricity on the planet. Goldman Sachs sees US data center power demand surging from about 31 GW in 2025 to roughly 66 GW by 2027. Gartner has global consumption growing 26% in 2026 alone. Pick your source, the arrow points the same way, steeply up.

And you cannot summon a gigawatt. New high-voltage transmission or generation can stretch grid interconnection timelines to 24, 36, or even 48-plus months. That is the moat, and it is an accidental one. A REIT that already holds a site with power secured is sitting on something a competitor cannot replicate for three or four years no matter how much capital they throw at it. This is why 64% of North American capacity under construction has moved to frontier power markets like West Texas, Tennessee, Wisconsin, and Ohio. The industry is chasing electrons, not real estate.

Inference is the durable part

Back to that Equinix comment about inference over training, because it is the most important thing management said all quarter.

Training a model is a spiky, one-time event. You spin up an enormous cluster, burn through it for weeks or months, ship the model, and the demand ends. If AI capex were only about training, you'd be right to worry about a cliff once the frontier labs slow their pace. Inference is different. Inference is what happens every single time somebody sends a prompt, generates an image, or lets an agent run. It scales with usage, and usage only goes one direction as these tools get embedded into everything.

So when both Equinix and Digital Realty flag inference as the driver, with Digital Realty's largest lease in history explicitly inference-oriented, they are telling you the demand is shifting from the lumpy, cyclical part of AI to the steady, recurring part. For a landlord signing 10 and 15-year leases, that is exactly the demand profile you want. It is the difference between renting to a film crew for a summer and renting to a business that needs the building forever.

The smart money noticed too

You are not early to this, and I would rather be honest about that. Blackstone launched a new REIT designed for pure-play exposure to AI-boom data centers. When the largest alternative asset manager on the planet builds a dedicated vehicle to crowd into the exact same asset class as the public names, that is confirmation and a warning in the same breath. Confirmation that the thesis is real. A warning that a lot of capital is chasing the same scarce powered land, which eventually compresses returns.

As of mid-2026 the two public names carry very different price tags. Equinix sits near a $102.5 billion market cap, a share price around $1,051, and a dividend yield of roughly 1.9% to 2.0%. Digital Realty is near $63.6 billion, around $178 a share, yielding closer to 2.7%. Neither is cheap on a yield basis. You are not buying these for the income. You are buying the AFFO growth and the pricing power that a 1% vacancy market hands its landlords.

How I'd actually think about it

A few honest takes, because the hedge-everything version of this is useless.

• This is an infrastructure bet, not a real estate bet. Don't value these like office or retail REITs. The right comparison is a toll road on AI usage. The moat is power and the interconnect queue, not the building.
• Inference is the thesis, watch it. If the AI demand narrative ever wobbles, the question to ask is whether inference volume is still climbing, not whether the labs are still training. Inference is the recurring revenue. Training is the spike.
• The tenant quality is the underrated part. These leases are with Microsoft, Oracle, and AA-rated hyperscalers on long terms. That is investment-grade contracted cash flow. Very different risk than a landlord with a hundred small tenants who can vanish in a downturn.
• Power is the risk and the moat at once. The same grid constraint that protects incumbents can also cap their growth. If a REIT can't energize its pipeline, that backlog just sits there. Track how much of their capacity actually has power secured.
• Valuation leaves no room for error. Up 36% in a year prices in a lot of good news. The demand is real, but so is the multiple. Size the position accordingly.

The intersection is the whole point

The reason this trade sat there for the taking is that it required holding two ideas at once. You had to understand REITs well enough to read an AFFO line and a lease backlog, and you had to understand the AI stack well enough to know that inference demand is structural and that power is the real bottleneck. Real estate investors mostly had the first half. Tech investors mostly had the second. The people who had both got to buy the landlords of the entire AI boom while everyone else argued about the tenants.

That is usually where the interesting money is. Not in the obvious thing everyone is looking at, but in the boring physical layer underneath it that has to exist for the obvious thing to work. GPUs get the headlines. The buildings they sit in, and the substations feeding those buildings, quietly collect the rent.