Data Centre REITs: The Landlords of the AI Build

Marketing Communication. Capital at risk. For Professional Investors Only. Please read fund legal documentation before making any final investment decisions.

Every AI model runs on chips that need a building, a grid connection and constant cooling.¹ Data centre REITs own and operate those buildings, and lease the capacity to the companies developing and deploying AI, including the hyperscalers themselves.²

Key characteristics:

• They rent space and power facilities, rather than the compute itself. Tenants install and own their servers and chips; the operator provides the building, electricity, cooling and connectivity.³
• Hyperscalers are major tenants, not competitors. Over 40% of current hyperscale capacity is leased from DC operators, with the rest being self-built.⁴
• Power is one of the scarce inputs. An operator with the power capacity can deliver years faster than a tenant building from scratch.⁵
• Contracts lengths vary. Contracts span from 1-5 years for colocation tenants, to 15-20- years for the largest recent hyperscale lease contracts with fixed annual escalators, many of which are signed before construction ends. This supports more predictable, stable cash flows.⁶
• They pay income. REIT rules require most taxable income to be distributed, a rare feature of AI infrastructure companies. ⁷

What a Data Centre REIT Does

"Data centre REIT" is largely a tax and legal label.⁸ Functionally, these companies are data centre operators: they develop, manage and maintain the facilities, in addition to owning them.⁹ The listed universe includes US REITs such as Equinix and Digital Realty, as well as operators in the US, Asia and Australia that are not always classed as REITs but run the same core business.¹⁰

That broadly consists of owning the buildings that house computing equipment and renting out the space, power, cooling and connectivity that the equipment needs to generate compute capacity. ¹¹

This split defines the business: the landlord collects rent while tenants replace their chips several times over the life of a lease. Which chip or AI model wins does not change who pays the rent.

Who the tenants are

Much of the AI news flow tracks hyperscaler capex going towards construction of their own data centres: Amazon, Alphabet, Meta and Microsoft have a combined guidance of $720-745bn for 2026.¹² Yet the hyperscalers do not spend only on their own buildings, renting over 40% of their data centre capacity from data centre operators.¹³ The largest spenders in AI are among the largest tenants.

The largest tech companies rent around 40% of their data centre capacity

Hyperscalers are the most visible tenants, but every other company using AI ends up in the same buildings, typically indirectly, without necessarily signing a lease with them:

1. A bank, retailer, start-up or AI lab rents computing power from a cloud provider or specialist AI cloud (such as Azure or AWS).¹⁴
2. The cloud provider owns the servers and houses them either in its own data centres or in space and power leased from a data centre operator.¹⁵
3. The operator owns the building and secures electricity from the utility.¹⁶

Some companies also lease space directly: banks, telecoms carriers, large enterprises and AI developers place their own equipment in operators' buildings, next to the clouds and networks they connect to.¹⁷

Whether demand starts with a hyperscaler, an AI start-up or a bank, a large share of it ends up in data centre operators’ buildings. The operator therefore sits beneath the AI stack and is paid by the companies that run on it.

Why tenants lease rather than build

The scarce but essential inputs are grid connection, cooling and land.¹⁸ Securing a grid connection can take years, and lead times are increasing. The average duration from interconnection request to commercial operation went from just over 2 years in 2005 to over 5 years in 2025.¹⁹

Consequently, a facility that is already connected to the grid can deliver capacity far sooner than a tenant starting from scratch.²⁰

The market shows how tight capacity is. Vacancy across primary US markets stood at a record-low 1.4% in H1 2026, 0.2% in Northern Virginia, and 80% of capacity under construction was already pre-leased.²²

US data centre vacancy has fallen to record lows

Even the largest builders use landlord-style structures. Microsoft has said it is classifying more future data centre leases as operating leases.²³ Meta's Hyperion campus in Louisiana sits in a ~$30bn vehicle in which Blue Owl holds 80%, financed largely by debt and leased back to Meta over a horizon of around 20 years.²⁴

Only the largest technology companies and well-funded private capital can build at gigawatt scale. The listed operators' edge is existing powered capacity, scale, and at the retail end, dense networks of interconnected customers.

Long contracts, visible cash flows

Hyperscale leases typically run 10-20 years with fixed annual rent increases, and many are signed before the building is finished. Recent transactions show the pattern.

Hyperscale leases run 15 to 20 years

Terms already on Digital Realty's books reflect the same structure: 13.7 years remaining on the lease with a "Leading AI Chip Maker" and 10.8 years with Oracle.³⁰

Long leases provide:

• Contracted revenue for a decade or more, often with investment-grade counterparties. These are typically on a triple net lease structure, in which the tenant pays base rent plus most property costs, commonly taxes, insurance, maintenance and utilities.³¹
• Built-in growth from annual escalators, before any increase in market rents.
• Pre-leasing that reduces development risk: capital is committed against a signed tenant.

However, they do not remove:

• Concentration. A handful of hyperscalers account for a large share of wholesale revenue.
• Renewal exposure. Not every lease is long: Digital Realty's top 20 tenants have a weighted average of 6.3 years remaining,³² and retail colocation contracts typically run 1-5 years.
• Counterparty risk. Leases to newer AI cloud providers carry more credit risk than leases to the largest technology companies.

On the other hand, data centre operators with near-term lease expirations can benefit from repricing opportunities. Digital Realty, for example, has approximately 30% of its portfolio (by annualised base rent) expiring through to 2027.³³ The company has upgraded its rental rate growth on renewal leases (cash) from 6-8% as of February 2026, to 9-11% as of July 2026, signalling growing pricing power.³⁴ This is consistent with the annual increases in average asking rental rates across primary U.S. markets since 2021.³⁵

US data centre rents have risen for four consecutive years

The role in the AI landscape

AI spending flows through three layers: the chips and servers that compute, the networks that connect them, and the facilities that house and power them.³⁶ Data centre operators own the third layer, and they are paid on a different clock from the first.³⁷

Digital Realty's signed-but-not-yet-commenced leases illustrate the lag. Of $1.4 billion in annualised rent at its share, about $635 million was scheduled to commence in H2 2026, $480 million in 2027 and $312 million in 2028 and beyond.³⁹

The two layers are complements, not substitutes. Every new AI cluster needs a powered building, and the tenant typically renews or replaces its chips several times within one lease.⁴⁰

Income in an AI-exposed portfolio

US REITs must distribute at least 90% of their taxable income to shareholders to keep their tax status.⁴¹ Data centre REITs therefore pay regular income, which is somewhat uncommon among AI beneficiaries: most chipmakers and AI platforms pay little or no dividend.⁴²

Payout policy reflects a trade-off. Equinix, for example, retains around half its cash flow to fund development; higher-payout REITs rely more on debt and new equity to grow, which ties their growth to the cost of capital.⁴³

Data centre REITs' dividend yields

This document is not intended to be, or does not constitute, investment research as defined by the Financial Conduct Authority.