How Are Data Centers Financed? REITs, ABS, Project Finance and Private Credit

How data centers are financed has become one of the most important questions in global credit markets, because the AI build-out is now too large for any company to fund from cash flow alone. This guide, part of the Kurums Data Center Economy hub, explains each layer of the data center capital stack, from construction loans and asset-backed securities to REITs, joint ventures, private credit and the off-balance-sheet structures used by hyperscalers, with 2025–2026 examples. It builds on our analysis of the AI capex supercycle and our guide to data center construction costs.

How are data centers financed?

Data centers are financed through a capital stack that typically combines developer or sponsor equity, construction loans during the build, long-term debt or securitisation once leased, and, for the largest projects, joint ventures with sovereign wealth, pension or private credit funds. Hyperscalers also fund campuses directly from cash flow and bonds.

The financing follows the life cycle of the asset. In the development phase, a sponsor buys powered land and starts construction with equity and a construction loan, ideally after signing a pre-lease. In the stabilised phase, once a creditworthy tenant is paying rent, the asset can be refinanced with cheaper long-term debt such as commercial mortgage-backed securities (CMBS), asset-backed securities (ABS) or corporate bonds. At the portfolio level, owners recycle capital by selling stakes to joint-venture partners or listing as real estate investment trusts (REITs).

The scale of the build-out has stretched every layer. JLL estimates that data center spending could approach $3 trillion over the next five years including tenant fit-out, creating about $1.2 trillion of real estate value and requiring about $870 billion of new debt. JLL’s view is that debt markets will remain open to credible developers, but the cost and structure of that debt depends heavily on who the tenant is. For an overview of who builds and owns the facilities, see our guide to data center business models.

Source: JLL 2026 Global Data Center Outlook; Equinix Q2 2026 results; CoreWeave Q2 2026 disclosures; Oracle Q4 FY2026 release; Wikipedia summaries of Stargate and Blue Owl reporting.

How do construction loans for data centers work?

A construction loan funds the build in stages as work is completed, usually alongside sponsor equity. Lenders focus on whether the project has a signed lease with a creditworthy tenant, a firm power connection and a credible contractor, because those factors determine whether the loan can be refinanced on completion.

Pre-leased projects are the easiest to finance. When a hyperscaler or investment-grade tenant has signed a long lease before construction, banks can lend against the expected rental income. Speculative projects, built without a tenant, need more equity and pay higher spreads. The 2025 financing of OpenAI’s Stargate site in Abilene, Texas, shows the pattern: JPMorgan Chase agreed in May 2025 to lend $2.3 billion for the campus, developed by a joint venture of Blue Owl Capital, Crusoe and Primary Digital Infrastructure and leased to Oracle.

Construction risk is rising. JLL reports that 57% of projects saw schedule delays in 2025, and Cushman & Wakefield says pad-mounted transformer lead times run 68–113 weeks. Every month of delay adds interest while the borrower earns no rent, a point JLL’s Matthew Giddings made in October 2026 when he warned that financing costs keep accumulating while projects wait for power. Lenders increasingly ask for evidence of equipment slots and grid connection dates, issues covered in our article on the power bottleneck.

What are data center ABS and CMBS?

Data center asset-backed securities (ABS) and commercial mortgage-backed securities (CMBS) are bonds secured by the rents from stabilised, leased facilities. They let owners refinance construction debt at lower cost once buildings are full, and JLL says issuance could reach $50 billion as early as 2026.

In a typical ABS, an owner places leased facilities into a bankruptcy-remote issuer that sells rated notes repaid from contracted rents; CMBS are secured by mortgages on the properties. Both appeal to insurers and other long-term investors.

Securitisation has grown with the market because stabilised data centers with hyperscale tenants look like long-term infrastructure: long leases, high renewal rates and creditworthy counterparties. The risks are concentration (a pool may depend on one or two tenants), technological obsolescence if a facility cannot support future density, and refinancing risk when notes mature. JLL’s estimate of $50 billion of ABS and CMBS, possibly as soon as 2026, would make data centers one of the largest esoteric asset classes. Investors comparing vehicles can read our guide to investing in data centers.

How do data center REITs raise capital?

Data center REITs such as Equinix and Digital Realty fund growth with a mix of unsecured bonds, new equity, retained cash flow and joint ventures. Because REITs must distribute most taxable income as dividends, they rely on capital markets to fund development.

Equinix is a case study in REIT funding. It guided 2026 capex at $5–6 billion and $5–7 billion a year for 2027–2029. In the first half of 2026, free cash flow was negative $560 million after capex, while it paid $1.03 billion in dividends; it raised about $2.42 billion of net proceeds from senior notes, taking senior notes outstanding to $19.7 billion and total debt principal to about $22.2 billion. It also funds hyperscale-scale buildings through its xScale joint ventures, including one with Singapore’s GIC and the Canada Pension Plan Investment Board, keeping much of that capital off its own balance sheet.

Digital Realty follows a similar playbook. Its second-quarter 2026 materials showed about 1.4 GW under construction at a total cost of roughly $20 billion, about 63% pre-leased, with an expected stabilised yield of around 11.5%, and it reported a record $2.5 billion backlog of signed leases in Q3 2026. That yield, compared with its cost of capital, is the spread that makes development worthwhile. We compare the two REITs and their peers in our ranking of colocation providers.

How do joint ventures fund data centers?

In a joint venture, an operator or hyperscaler partners with a capital provider, typically a sovereign wealth fund, pension fund or infrastructure investor, to share the cost of building or owning data centers. The operator brings development skills and tenants; the partner brings patient equity.

Joint ventures let operators grow faster than their own balance sheets allow and earn fees for development and management. Equinix’s xScale programme, Digital Realty’s fund and JV structures, and many private platforms backed by infrastructure investors all follow this pattern.

The largest deals now involve consortia. The AI Infrastructure Partnership, whose backers include BlackRock’s Global Infrastructure Partners, MGX, Microsoft and Nvidia, agreed in 2025 to buy Aligned Data Centers, a deal we cover in our M&A guide. Stargate itself is a joint venture: OpenAI, SoftBank, Oracle and MGX, with SoftBank and OpenAI reported by The Information to have committed $19 billion each, and the remaining funding expected from limited partners and debt. Private equity platforms are profiled in our guide to private equity data center empires.

What is an off-balance-sheet data center SPV, and how did Meta use one for Hyperion?

An off-balance-sheet SPV is a separate company that owns and finances a data center and leases it to the user. If structured accordingly, the user records a lease rather than the full asset and debt. Meta used a joint venture with funds managed by Blue Owl Capital, valued at about $27 billion, for its Hyperion campus.

Hyperion, in Richland Parish, Louisiana, is Meta’s largest AI campus. Meta first described it in December 2024 as a $10 billion-plus investment; by the time it announced the Blue Owl joint venture in October 2025, the project was valued at about $27 billion. The campus is designed to draw up to 5 GW, and the utility Entergy plans at least 3.8 GW of generation to serve it. In structures like this, outside investors fund most of the capital through equity and debt, the hyperscaler develops and leases the facility, and lease payments service the financing.

The appeal is that the hyperscaler shares the cost of a multi-decade asset without issuing all of the debt itself, while keeping control of design and operations. Critics argue that such SPVs obscure true leverage, because lease and guarantee obligations can be as binding as debt. Meta’s 10-Q showed $454.5 billion of non-cancelable contractual commitments at 30 September 2026, about $121 billion due in 2027, and Moody’s reportedly estimated about $662 billion of signed but not yet commenced data center leases across the hyperscalers. Analysts increasingly add these commitments to reported debt when assessing hyperscaler balance sheets, a theme in our AI bubble debate.

How does private credit finance AI data centers and GPU clouds?

Private credit funds lend directly to developers and GPU cloud operators, often secured by GPUs and by long-term customer contracts. These loans are faster and more flexible than bank or bond financing but more expensive, and they let neoclouds grow far beyond their equity base.

CoreWeave pioneered GPU-backed borrowing. In August 2023 it raised a $2.3 billion facility led by Magnetar Capital and Blackstone, collateralised by Nvidia H100 GPUs, the first time that hardware had been used as loan collateral. By 30 June 2026 its total debt had reached about $35.1 billion, of which about $31.4 billion was recourse and $3.7 billion non-recourse debt structured against hyperscaler contracts. In February 2026 it sought $8.5 billion of new financing using contracts with Meta as collateral, and in April 2026 it priced an upsized $3.5 billion convertible note offering. Management says its weighted average cost of debt fell by nearly 300 basis points over the past year.

The cost is visible in its income statement. CoreWeave’s net interest expense was $640 million in Q2 2026, up from $267 million a year earlier, and it guided Q3 interest of $860–940 million, against 2026 capex guidance of $35–39 billion. Lenders accept the risk because the debt is sized against contracted revenue: CoreWeave’s backlog was about $104 billion at mid-2026. The flip side is concentration, since Microsoft accounted for about 67% of its 2025 revenue. Futuriom has flagged problems in the private credit sector as one of the risks to AI data center projects.

How do hyperscalers finance their own data centers?

Hyperscalers mostly fund data centers from operating cash flow, supplemented by bonds, finance and operating leases, and joint ventures. As 2026 capex approaches or exceeds cash flow, they are leaning more on leases and external partners.

The four largest US hyperscalers generate enormous cash flow, but 2026 capex of roughly $700 billion is testing it. Amazon’s trailing twelve-month free cash flow was an outflow of $42.4 billion at Q3 2026, and Meta’s Q3 capex of $34.9 billion exceeded operating cash flow of $32.8 billion. Leases help: Microsoft said in 2026 that more future data center leases would be classified as operating leases, which are not counted in capex, and it extended building useful lives from 15 to 25 years.

Oracle is the most debt-dependent of the group. Its free cash flow was negative $23.7 billion for the four quarters to May 2026 and negative $5.4 billion in its fiscal first quarter of 2027, when capex reached $28.5 billion. Oracle said it expects about $40 billion of debt and equity financing in fiscal 2027 and relies heavily on leasing campuses built and financed by developers. Its ability to keep borrowing is closely watched by credit investors, as discussed in our hyperscaler comparison.

What are the risks in data center financing?

The main risks are tenant concentration, construction and power delays, refinancing at higher rates, technological obsolescence and the possibility that AI demand falls short of the capacity being financed. Circular deals between chipmakers, AI labs and cloud providers add a layer of correlation risk.

Real estate lenders ask whether a building will still be leasable when the first lease ends, since 15 kW halls may not suit 100 kW AI racks. GPU lenders face faster obsolescence: if newer chips make existing GPUs uneconomic before loans are repaid, collateral values fall.

Concentration is the cross-cutting risk. Many financings ultimately depend on the same handful of tenants: Microsoft, Amazon, Google, Meta, Oracle and the leading AI labs. If one of them slows spending, the effect would ripple through developers, private credit funds, ABS investors and equipment suppliers at once. Supporters counter that these are among the most creditworthy companies in the world and that backlogs remain at record levels. We present both sides in Is there an AI data center bubble? and track country-level funding in our US investment guide.

Frequently Asked Questions