How Data Center Financing Actually Works

Introduction

A hyperscale data center is a warehouse full of extremely expensive computers, and yet the way it gets paid for looks almost nothing like the way a technology company gets funded. It looks like a toll road. Roughly 100 GW of new capacity is expected to come online between 2026 and 2030, close to doubling the global installed base, and JLL's 2026 outlook puts the all-in price tag at around $3 trillion once you count both the buildings and the chips inside them. Morgan Stanley and Moody's landed on a similar number in early 2026. Nobody writes that check out of retained earnings.

So the money arrives in layers, and each layer has its own lender, its own security package, and its own tolerance for risk. Land is bought with equity and, increasingly, with dedicated land-cost facilities. Construction is funded with bank debt that prices completion risk explicitly. Once the building is finished, leased, and throwing off steady cash, the whole thing gets refinanced into the bond markets at a fraction of the original cost of capital. That progression, from speculative dirt to investment-grade paper, is the entire game.

This post walks the capital stack stage by stage: what gets funded when, who lends at each point, and why the tenant's credit rating matters more than the concrete. If you want the macro picture of where all this money is coming from and what it means for credit markets, our post on how the AI boom is financed covers the demand side. This one covers the plumbing.

Before walking the stages, it helps to see the destinations. A stabilized data center can be financed through four broadly distinct channels, and the choice depends on portfolio size, how many assets the sponsor wants to pledge, whether the sponsor wants the debt on its own balance sheet, and how much structural flexibility it needs. The figures below reflect market conditions as of mid 2026.

Those average deal sizes come from Barclays estimates cited in Structured Finance Association research published in July 2026. The channels are not mutually exclusive. A single sponsor might use a land-cost facility to assemble a site, a construction loan to build it, a private credit bridge to carry it through lease-up, and an ABS master trust to term it out. Each step is a refinancing of the last, at a lower cost, once a specific risk has been retired.

The instinct among students is to file data centers under technology. Credit markets file them under real assets, and that classification drives every structuring decision that follows.

Contracted cash flow replaces product risk

A software business is underwritten on growth, churn, and margin expansion. None of that applies here. A leased data center produces a defined stream of rent from a named counterparty under a contract that typically runs 15 to 20 years with annual escalators of 2% to 3%. There is no product to get wrong and no market share to lose. What is being financed is a contracted revenue stream attached to a physical asset with a long useful life, which is exactly the profile that project finance and infrastructure lending were built for.

What that changes in the capital structure

Infrastructure classification has concrete consequences. It supports higher leverage than a corporate credit of comparable size, because the cash flow is contracted rather than cyclical. It supports longer tenors, because the asset life is measured in decades. It attracts a different investor base, principally insurers and pension funds hunting for long-duration, investment-grade-rated paper that matches their liabilities. And it makes the asset securitizable, because contracted, predictable cash flows are precisely what a securitization structure needs in order to isolate risk and tranche it. A pre-revenue AI startup could never be securitized. A leased campus with an A-rated tenant can be, and routinely is.

Everything before the first tenant moves in is development risk, and development risk is expensive to finance. This is the part of the lifecycle where equity does most of the work and where the newest financing tools have appeared.

Site control is really power control

Land is the cheap part. The scarce input is an electrical interconnection, and in the constrained markets that matter (Northern Virginia, Dallas, Phoenix, parts of the Nordics) the queue to connect to the grid can run years. A site with signed power is worth a multiple of an identical site without it, which is why developers now buy land primarily as a proxy for the interconnection rights attached to it. Underwriting a development-stage data center therefore starts with a power diligence question, not a real estate one: what is the megawatt allocation, when does it energize, and what happens to the economics if it slips.

Land-cost facilities and pre-construction capital

Historically this stage was pure equity. That has changed. As of mid 2026, outside capital arrives much earlier in the lifecycle through land-cost facilities, which are loans secured by land-banked sites held in development vehicles, and through hybrid structures collateralized by a contracted asset base. Sponsors use them to hold more sites for longer without tying up all their equity, which matters when the constraint on growth is how many powered sites you can control. These are higher-risk, higher-priced instruments than anything downstream, and they are almost entirely a private credit product.

Construction debt and completion risk

Construction loans fund the vertical build and are drawn in tranches against certified milestones. Lenders here are pricing something securitization investors never touch: the risk that the building is late, over budget, or never finishes. Typical protections include completion guarantees from the sponsor, fixed-price contracts with the general contractor, and reserves for cost overruns. Pricing sits well wide of the eventual permanent debt, which is exactly why sponsors are in a hurry to reach stabilization and refinance. The gap between construction pricing and stabilized pricing is the developer's reward for taking build risk.

The moment an asset moves from "under construction" to "stabilized" is the single most valuable event in the lifecycle, because it changes which investors are allowed to buy the paper.

What "stabilized" actually means

Stabilization is not a vague adjective. Rating agencies and bond investors apply a specific test, and both the ABS and CMBS markets have overwhelmingly financed assets that pass it. KBRA's research on the sector notes that these deals have generally been backed by built and stabilized cash-flowing assets with little or no remaining construction or lease-up risk. In practice that means the facility is complete and commissioned, capacity is contracted under executed leases, rent has commenced, and operations are running at a level the underwriter can project forward with confidence.

Stabilized Asset

A completed, leased, and cash-generating property that no longer carries meaningful construction or lease-up risk. In data center finance, stabilization is the threshold that lets an asset move out of expensive development and construction debt and into securitization or long-term investment-grade financing.

Once an asset clears that bar, the refinancing follows almost mechanically. Construction debt is repaid with cheaper permanent debt, the sponsor recycles equity into the next project, and the asset joins a pool that can be tapped again later. This is the flywheel that lets platforms grow far faster than their own equity base would allow, and it is why the industry's financing calendar tends to cluster around commissioning dates.

Fifteen-year triple net leases and why tenant credit is the asset

The lease is what makes any of this financeable. Hyperscale capacity is typically let on 15-year triple net leases, commonly with extension options to investment-grade tenants, frequently rated A or better, with the parent company guaranteeing the payments and annual rent escalators built in. Operating expenses, property taxes, and insurance pass through to the tenant, which produces unusually clean net operating income for the landlord.

Triple Net Lease

A lease under which the tenant pays property taxes, insurance, and maintenance in addition to rent, leaving the landlord with near-pure rental income. Triple net structures dominate hyperscale data center leasing because they make the landlord's cash flow highly predictable and therefore easy to finance.

That structure means the credit being underwritten is not the data center operator's. It is the tenant's. A speculative-grade developer can raise investment-grade-rated debt because the cash flow behind that debt comes from a company rated several notches higher. The pattern shows up repeatedly in public filings: Cipher Mining disclosed a 15-year triple net lease with Amazon in which the parent fully guarantees the lease cash flows, filed with the SEC in 2026.

With a stabilized asset and a creditworthy tenant, the sponsor can finally access the deep, cheap end of the market. There are four main ways to get there, and they coexist rather than compete.

Data center ABS and the master trust

Data center ABS has existed since 2018 and is now a mainstream corner of the structured market. JPMorgan projected in early 2026 that data center securitization issuance would run $30 billion to $40 billion annually in 2026 and 2027, representing roughly 7% to 10% of combined ABS and CMBS issuance, up from about $27 billion in 2025.

Data Center ABS

Asset-backed securities issued against the lease revenue of stabilized data centers, usually through a master trust that lets the sponsor add qualifying facilities and issue further notes over time. The bonds are typically fixed rate and repaid on an anticipated repayment date rather than a hard maturity.

The master trust format is the structural signature of the asset class. Instead of financing one building, the sponsor establishes a trust into which eligible assets can be contributed over time, then issues successive series of notes against the growing pool. That creates a scalable funding platform rather than a one-off loan, which suits operators building dozens of facilities. KBRA published a global rating methodology for the sector in January 2026 built around net cash flow, simulated tenant defaults and lease rollover, and stress tests on the repayment structure.

Anticipated Repayment Date

The date on which securitized notes are expected, but not legally required, to be repaid. If the notes remain outstanding past that date, the structure usually imposes penalty interest and sweeps excess cash to accelerate repayment, which pushes the issuer to refinance without triggering a hard default.

Single-borrower CMBS and the 144A investor base

CMBS entered the sector around 2021 and takes a different shape. Where ABS pools many assets in a revolving trust, single-borrower CMBS typically finances one campus, one asset, or a discrete portfolio. CMBS deals also skew floating rate, while data center ABS skews fixed. Deal size differs accordingly: Averages also depend on the window you measure. KBRA's since-inception tally of the market as of mid-2025 counted 75 ABS transactions totaling $34.5 billion at an average of roughly $460 million, below the recent-vintage figure above because early deals were smaller, against 13 CMBS deals totaling $14.2 billion at an average near $1.09 billion, published in its comparison research.

Almost all of this paper is sold under Rule 144A, meaning it is privately placed with qualified institutional buyers rather than registered for public sale. That route is faster and less disclosure-intensive, and the buyer base (insurers, pension funds, and asset managers hunting long-duration investment-grade product) does not need retail protections. The practical consequence for a junior banker is that much of the deal flow in this sector never appears in public offering statistics, which is one reason estimates of total AI-related issuance vary so widely.

Structured credit shows up well beyond the securitization desk: Work through debt, valuation, and capital-structure questions with worked answers, start practicing interview questions for free and find the gaps before an interviewer does.

Why private credit moved in

Private credit went from a bit player to a central financier of this sector in about three years. Outstanding private credit loans to AI-related companies grew from close to zero to more than $200 billion by early 2026, and the Bank for International Settlements has projected that figure could reach $300 billion to $600 billion by 2030. Morgan Stanley expects private credit to supply roughly $800 billion of data center financing of the $1.5 trillion external financing gap it projects through 2028, led by asset-based finance.

The reason is speed and shape. Data center deals frequently need capital before an asset is financeable in the public markets: during lease-up, across a portfolio at different stages, or against collateral that no rating agency has a published methodology for yet. Private lenders can underwrite a bespoke structure in weeks and hold the whole thing, which banks constrained by capital rules and syndication risk often cannot. That flexibility is the same dynamic driving the broader expansion of private credit and direct lending across the market.

Hyperscaler bonds and off-balance-sheet joint ventures

The cheapest money in the system belongs to the hyperscalers themselves. Alphabet, Amazon, Meta, and Oracle issued about $93 billion of investment-grade debt in 2025, roughly 6% of all US investment-grade corporate bond issuance that year, and Morgan Stanley forecast $250 billion to $300 billion from hyperscalers and related joint ventures in 2026. The gap between what these issuers pay and what a speculative-grade developer pays is the whole reason the investment grade versus high yield distinction dominates so much of DCM.

Increasingly, though, the hyperscalers would rather not carry the debt at all. The template is the Meta joint venture with funds managed by Blue Owl Capital to develop the Hyperion campus in Louisiana, announced by Meta in 2025. Blue Owl funds took 80% of the venture and Meta retained 20%, with roughly $27 billion of debt raised at the venture rather than at Meta. Meta gets the capacity through a lease, the asset managers get long-dated infrastructure exposure, and the borrowing sits outside Meta's reported balance sheet.

Debt structures come up in every technical round: Download our comprehensive 160-page PDF, access the IB Interview Guide covering capital markets, credit analysis, and deal structures end to end.

Alongside the real estate sits a second, newer credit market financing the equipment that goes inside the buildings. It behaves differently enough that it deserves separate treatment.

Lending against the chips

Specialist compute providers fund GPU purchases with debt secured by the chips themselves plus the customer contracts they support. CoreWeave's GPU-collateralized borrowing grew from $2.3 billion in 2023 to a $7.5 billion facility in 2024 and then an $8.5 billion facility that carried the first investment-grade rating on a GPU-backed financing. Valor Equity Partners arranged roughly $20 billion of equity and debt in 2025 against Nvidia processors. The rating logic mirrors the real estate side exactly: the borrower is speculative grade, the customer whose contract underpins the loan is not, and the structure imports the stronger credit.

The duration mismatch nobody has solved

This is where GPU lending diverges sharply from building finance. A data center shell has a useful life of 20 to 30 years. A GPU generation has a commercially competitive life of roughly five to six years on the most generous public benchmark, the depreciation schedules hyperscalers themselves use, and critics argue the real figure is shorter if performance keeps improving at the current pace. Lending five-year money against an asset with a seven-year economic life and no established secondary market is a fundamentally different proposition from lending against a leased building.

Capital is abundant. Electricity is not, and that inversion is the defining feature of the sector as of mid 2026. The International Energy Agency projects data center electricity demand will more than double by 2030 to around 945 TWh, roughly Japan's current annual consumption, and warns that around 20% of planned projects could face delays connecting to the grid without major transmission investment, in its analysis of energy and AI.

For financiers this converts an engineering problem into a credit problem. Delivery schedules are set by interconnection dates rather than construction timelines, so a financing model that assumes revenue starts on completion may be wrong by years. Sponsors respond by contracting power directly through long-term purchase agreements, on-site generation, and in some cases nuclear supply deals, each of which is itself a financing needing its own lenders. The result is that data center coverage now overlaps heavily with power and utilities coverage, and a credit analyst who cannot read an interconnection agreement is missing the variable that determines whether the deal works.

The market's growth has been fast enough that the risk conversation is still catching up. Four concerns dominate diligence discussions.

Technology obsolescence and residual value

Data centers built for one generation of hardware may not suit the next. Power density per rack has risen sharply, and cooling requirements have shifted from air to liquid in high-density deployments, which can leave older facilities functionally mismatched to current demand. Rating agencies describe this as a capital-intensive, technically complex sector with genuine obsolescence risk, where a state-of-the-art facility can become a stranded asset. Because securitizations rely on a terminal value at the end of the note's life, obsolescence feeds directly into stressed capitalization rates and therefore into how much debt an asset can carry.

Concentration, in every direction

The other three worries are all versions of the same problem. Tenant concentration is extreme, since a handful of hyperscalers account for the majority of contracted capacity, so a single tenant's credit deterioration would hit many deals at once. Geographic concentration clusters assets in a few power-rich markets exposed to the same grid and the same local politics. Refinancing concentration follows from the anticipated repayment date convention: a large share of outstanding paper is expected to refinance in a narrow window, which is comfortable in an open market and much less so in a closed one. Investors used to diversified pools should note that a data center securitization is often closer to a single-name credit exposure wearing a structured wrapper, a point worth keeping in mind alongside conventional real estate valuation metrics when comparing the sector to traditional property.

Data center financing is unusual in that it touches almost every product and coverage group at once, which is why it has become such a common interview topic. The work splits roughly like this:

The honest summary is that data center financing works beautifully as long as three things hold: the tenants stay creditworthy, the power arrives on schedule, and the refinancing window stays open. Each has held so far, and the structures built around them have become some of the most sophisticated in credit markets. What makes the sector worth understanding is not the size of the numbers but the mechanics underneath them, because the same logic (contracted cash flow, credit substitution, staged refinancing) is how toll roads, pipelines, and power plants have been funded for decades. Data centers are simply the newest and fastest-growing member of that family, and the bankers who can explain the stack stage by stage will find themselves useful across more desks than they expected.