A data centre lease signed this year carries risks its predecessors did not. The chips inside the building lose value faster than the debt against them runs off, the tenant is often one of a handful of companies, and the power connection can take longer to secure than the shell takes to build. None of that is hidden. It is disclosed in filings, priced by rating agencies and argued about in public by the people who buy the debt. This piece describes those structures; it is not investment advice.
Scale is what moved financing from a footnote to a board item. The International Energy Agency reported on 16 April 2026 that electricity demand from data centres rose 17% in 2025, with AI-focused data centres growing faster still, against 3% growth in global electricity demand, and that capital expenditure at five large technology companies passed $400 billion in 2025 and was set to rise a further 75% in 2026. Moody's Ratings, in commentary reported in July 2026, projected capital spending across Microsoft, Amazon, Alphabet, Meta, Oracle and CoreWeave at $785 billion in 2026 and approaching $1 trillion in 2027, with direct debt across the six at about $460 billion and off-balance-sheet data centre lease commitments at $1.2 trillion, which it treats as debt-equivalent. Those are two different measures, borrowing on the balance sheet and lease obligations off it, and a single headline debt figure usually shows only one of them.
The instruments in use
The build-out draws on corporate bonds, project finance placed with private credit, loans secured on the computing hardware itself, vendor support from equipment suppliers, and long-dated leases. Each puts the risk with a different party, which matters more than the headline amount.
Project finance and private credit
Meta's Hyperion campus in Richland Parish, Louisiana, is the reference case. In October 2025 Meta and Blue Owl Capital set up a joint venture in which Blue Owl's funds hold 80% and Meta 20%. According to reporting on the deal, the vehicle raised about $27 billion of debt, rated A+ by S&P and fully amortising to 2049, plus about $2.5 billion of equity, with Morgan Stanley as sole bookrunner and PIMCO as anchor investor; the same reporting called it the largest private capital transaction ever completed. The shape is familiar project finance: a special purpose vehicle, rated debt and a long maturity, secured on a specific asset and its cash flows and sold to institutional buyers, rather than the operating company borrowing directly. As with private credit in other wrappers, the vehicle changes who holds the risk and how it is rated, not the dependence on the tenant's rent.
Oracle has borrowed more directly. Its fiscal 2026 results, published on 10 June 2026, showed capital expenditure of $55.7 billion against operating cash flow of $32.0 billion, free cash flow of negative $23.7 billion, and borrowings of $129.5 billion at 31 May 2026. The company said it expects to raise about $40 billion of debt and equity in fiscal 2027, including a previously announced $20 billion at-the-market equity programme. In November 2025, Octus reported that banks were preparing about $56 billion of debt, roughly $38 billion of term loans and $18 billion of bonds, for Stargate-related campuses in Wisconsin, Texas and New Mexico being developed for Oracle by Vantage Data Centers and STACK Infrastructure. Oracle's remaining performance obligations reached $638 billion at the end of the year, up $85 billion on the previous quarter, and the company said most of the increase came from large AI contracts where the customer prepaid for the GPUs or supplied them itself. Remaining performance obligations are contracted future revenue, not cash, and do not retire debt on their own.
GPU-backed loans
CoreWeave has financed much of its fleet through delayed draw term loans secured on the infrastructure and the customer contracts attached to it. The facility it announced on 31 March 2026 is an $8.5 billion non-recourse delayed draw term loan, with about $7.5 billion available initially, secured by substantially all assets of a dedicated subsidiary, CoreWeave Compute Acquisition Co. VIII. It has a floating tranche at SOFR plus 2.25% and a fixed tranche at about 5.9%, and matures in March 2032. Moody's rated it A3 and DBRS A (low). MUFG and Morgan Stanley structured it, Goldman Sachs and JPMorgan were lead arrangers, and Blackstone Credit & Insurance anchored it. CoreWeave described it as the first investment-grade rated financing secured by high-performance computing infrastructure and an associated customer contract. The company's 10-K shows total debt of $21.4 billion at 31 December 2025, against $7.9 billion a year earlier. Here the hardware, not the company's general balance sheet, is the collateral, which is why the structure draws the closest scrutiny.
Vendor financing
Equipment suppliers are carrying part of the load. CoreWeave's 10-K lists OEM and software licence financing arrangements with $4.2 billion of principal outstanding at the end of 2025, up from $1.2 billion a year earlier. Nvidia has gone further. According to its 17 August 2026 SEC filing, as reported by Data Center Knowledge, it is providing residual value guarantees of up to $105 billion on leased infrastructure at an OpenAI campus in Ohio built and owned by SB Energy. The guarantees are not cash paid upfront. If specified default or insolvency events occur, Nvidia may have to cover the shortfall between a guaranteed minimum value under the leases and what SB Energy recovers by re-leasing or selling the assets, and OpenAI has agreed to reimburse Nvidia for anything it actually pays. The initial phase covers 4.25 gigawatts of IT capacity under a 20-year lease, with an option for a further 3.75 gigawatts, and Nvidia is separately investing $1.5 billion in SB Energy. No trade credit changes hands, but in substance the supplier is underwriting the resale value of the asset its own products fill.
What the collateral is worth
The concern the official sector keeps returning to is two clocks running at different speeds. The Hyperion debt amortises to 2049; the chips inside such buildings are replaced far sooner. CoreWeave depreciates its data centre computing equipment over six years, having extended the estimate from five years from January 2023, and its 10-K warns that changes to those useful-life estimates, or an inability to redeploy equipment beyond its contracted life, could significantly affect its results. At the European Central Bank's Sintra forum on 30 June 2026, Tobias Adrian, director of the IMF's Monetary and Capital Markets Department, argued, according to contemporaneous reporting, that the larger risk from AI is not equity valuations but hyperscalers "borrowing long to buy short-lived assets". The same reporting cited Dealogic data showing that Amazon, Alphabet, Meta, Microsoft and Oracle issued $159 billion of corporate bonds in the first five months of 2026, more than their total borrowing over the previous five years. A Bank for International Settlements bulletin published on 7 January 2026 by Iñaki Aldasoro, Sebastian Doerr and Daniel Rees judged that macroeconomic and financial stability risks from the AI boom appear moderate, while noting that investment needs will push financing from operating cash flow towards debt, with private credit playing a rapidly increasing role, and that equity prices have run far ahead of debt market pricing.
Whether a GPU is good collateral depends on what it fetches once the first tenant leaves. A residual value guarantee sized in the tens of billions shows the supplier is willing to stand behind that value. It has not yet been tested at this scale, because the current generation of large leases has not run its course.
Who is on the other end of the lease
The other half of the credit is who pays the rent. CoreWeave's 10-K, filed on 2 March 2026, states that "a substantial portion of our revenue is driven by a limited number of our customers, and the loss of, or a significant reduction in, spending from one or a few of our top customers would adversely affect our business, operating results, financial condition, and prospects." The same filing puts Microsoft at about 67% of 2025 revenue and the top two customers at about 77% of 2024 revenue. A loan secured on a customer contract is only as good as that customer's willingness and ability to keep paying, so a corporate credit question sits on top of the asset-value question.
Power is a financing constraint
A data centre earns nothing without power, and connection has become the slower half of the build. A May 2026 legal update from Eckert Seamans noted that PJM's queue backlog exceeded 300 gigawatts of projects during its reform process, that the size of required network upgrades is one of the most significant drivers of withdrawals, and that utilities often charge several million dollars per project for interconnection studies and engineering. At federal level, the Department of Energy directed the Federal Energy Regulatory Commission in October 2025 to consider rules for connecting large loads such as data centres, through an advance notice of proposed rulemaking. On 18 June 2026 FERC chose not to issue a nationwide rule. It issued show cause orders to PJM, MISO, SPP, CAISO, ISO New England and NYISO, requiring each to justify its existing tariff or propose changes, with responses due on 17 August 2026 and the rulemaking docket left open. Those proceedings are under way; no new large-load tariff has yet been approved under them. Operators are also contracting for power years ahead: the IEA's April 2026 figures put conditional offtake agreements between data centre operators and small modular reactor projects at 45 gigawatts, up from 25 gigawatts at the end of 2024, for reactors that are not yet built.
Proof of Talk's Paris programme put this framing on stage in June 2026. Chris Miglino, billed as chief executive of Axe Compute, a role the company announced on 9 February 2026, gave the keynote "The Compute Capital Stack: A View of AI's Global Infrastructure Buildout" on the Taostats Stage on 2 June 2026.
The crypto-adjacent version
Some bitcoin miners are converting energised sites into AI hosting capacity, on the view that grid connections and power infrastructure, not mining hardware, are the scarce asset. The risk there sits in whether announced contracts become delivered, paying capacity on schedule, and it is covered in the miners' pivot to AI data centres. It is a variant of the same collateral and counterparty question, not a separate one.
Where this leaves an allocator
The instruments are conventional: project finance, secured loans, vendor support, leases and bonds. What is new is where they are pointed: at hardware whose economic life is contested, at a small set of tenants whose concentration is disclosed in their own filings, and at sites whose power connections are still moving through queues and regulatory proceedings. Rating agencies have priced some of this; CoreWeave's facility shows investment-grade debt can be built around hardware and a customer contract. Whether the wider market has priced the gap between chip life and bond maturity that the IMF highlighted in June remains, as of September 2026, an open question.