Tech groups shift $120bn of AI data centre debt off balance sheets
Creative financing helps insulate Big Tech while binding Wall Street to a future boom or bust
Tech companies have moved more than $120bn of data centre spending off their balance sheets using special purpose vehicles funded by Wall Street investors, adding to concerns about the financial risks of their huge bet on artificial intelligence.
Meta, Elon Musk’s xAI, Oracle and data centre operator CoreWeave have led the way on complex financing deals to shield their companies from the large borrowing needed to build AI data centres.
Financial institutions including Pimco, BlackRock, Apollo, Blue Owl Capital and US banks such as JPMorgan have supplied at least $120bn in debt and equity for these tech groups’ computing infrastructure, according to a Financial Times analysis.
That money is channelled through special purpose holding companies known as SPVs. The rush of financings, which do not show up on the tech companies’ balance sheets, may be obscuring the risks that these groups are running — and who will be on the hook if AI demand disappoints.
SPV structures also increase the danger that financial stress for AI operators in the future could cascade across Wall Street in unpredictable ways.
“Eighteen months ago this would have been unfathomable, and fast forward to today, it’s very much the norm,” said a senior executive at one of the large financing institutions about the tens of billions of dollars flowing into SPVs to fund data centres.
“The tech industry can access meaningfully more capital than any other because of the credit profile,” he added.
Silicon Valley giants have traditionally generated a lot of cash and had little debt, giving these companies excellent credit ratings and high confidence from investors.
The race to secure computing power for advanced AI has pushed tech groups to borrow more than ever before, however. Tapping private capital funding through off-balance sheet structures protects companies’ credit ratings and flatters their financial metrics.
Meta in October completed the largest private credit data centre deal, a $30bn agreement for its proposed Hyperion facility in Louisiana that created an SPV called Beignet Investor with New York financing firm Blue Owl Capital.
The SPV raised $30bn, including about $27bn of loans from Pimco, BlackRock, Apollo and others, as well as $3bn in equity from Blue Owl.
The deal meant Meta could in effect borrow $30bn without any of the debt appearing on its balance sheet. This made it easier to raise a further $30bn in the corporate bond market a few weeks later.
Oracle has led the way in structuring large debt deals through third parties to pay for its enormous commitments to lease data centre power to OpenAI.
Larry Ellison’s tech group has partnered with builders and financiers such as Crusoe, Blue Owl Capital, Vantage and Related Digital to build numerous data centres that will ultimately each be owned by SPVs.
Its off-balance sheet financing deals include about $13bn invested by Blue Owl and JPMorgan, including $10bn of debt, into an SPV that owns its OpenAI facility in Abilene, Texas; a $38bn debt package to pay for two data centres in Texas and Wisconsin; and an $18bn loan for a site in New Mexico.
In each case, Oracle has agreed to lease the facilities from the SPVs. In the event of a default, lenders would have recourse over the assets — the data centre, the land it sits on and the chips that power it — and not the companies that manage the sites.
Raising off-balance sheet debt via an SPV has become more popular as the amount of capital needed to fund AI infrastructure has skyrocketed, stretching tech companies’ cash reserves. Morgan Stanley estimated that $1.5tn of external financing was needed to fund tech companies’ AI plans.
In many cases, investors in these data centre deals have been convinced that the financial risk ultimately still lies with the tech company leasing the site, should demand for AI services fall, resulting in a hit to the value of these huge computing facilities.
In the case of Beignet Investor, Meta owns 20 per cent of the SPV and has given a “residual value guarantee” to the other investors. This means that the social media group would have to repay the SPV investors if the value of the data centre drops below a certain level by the end of the lease and Meta decides not to renew.
Musk’s AI start-up xAI is pursuing a similar structure as part of a $20bn fundraise, including as much as $12.5bn in debt. The SPV will use the money to buy Nvidia graphics processing units and lease them to xAI.
CoreWeave said in March that it had created an SPV to fulfil a $11.9bn contract to supply computing power to OpenAI, which would “incur indebtedness to finance its obligations”. In July, it borrowed $2.6bn to fund its OpenAI contracts.
Private capital investors are keen to get in on the AI boom, boosting demand for these novel structures. Tech companies had borrowed about $450bn from private funds as of early 2025, $100bn more than over the previous 12 months, according to UBS.
This year, about $125bn flowed into “project finance” deals — long-term financing of infrastructure projects — such as the Meta and Blue Owl transaction, UBS said.
Data centre construction has become largely reliant on deep-pocketed private credit markets, a rapidly inflating $1.7tn industry that has itself prompted concerns due to steep rises in asset valuations, illiquidity and concentration of borrowers.
“There is risky lending and underlying credit risk built up in private credit already,” said one banker close to data centre financing deals. “This creates a very interesting set-up for the next several years, as you have two material risks to the outlook becoming more intertwined.”
The risk in these structures largely depends on how widespread they become. If multiple AI companies are using SPVs, stress can spread through the private credit funds behind them simultaneously with little transparency.
The AI data centre boom also largely relies on a small group of clients. OpenAI alone has made more than $1.4tn in long-term computing commitments across most of the sector’s big players.
Lenders across multiple different data centres could therefore be exposed to the same risks if one tenant falters. They also face uncertainty around access to power, AI regulation or technological shifts that make the current generation of AI hardware obsolete.
Not all of the big “hyperscaler” data centre companies have joined the trend. Google, Microsoft and Amazon — which all had large, established data centre businesses before the AI boom — have continued to finance construction using cash.
While Google and Amazon have recently tapped bond investors to raise more debt directly, the three companies have not yet disclosed any significant SPV financing.
Wall Street also is pushing into more obscure structures around data centre transactions.
A number of tech bankers said they had even seen securitisation deals on AI debt in recent months, where lenders pool loans and sell slices of them, known as asset-backed securities, to investors. Two bankers estimated these deals currently numbered in the single-digit billions of dollars.
These deals spread the risk of the data centre loans to a much wider pool of investors, including asset managers and pension funds.
Matthew Mish, UBS’s head of public and private credit strategy, said most investors “feel that actually it’s a good thing you ultimately end up with hyperscaler risk” given these companies’ strong balance sheets and credit profiles.
But Mish added SPV financings still “add outstanding liabilities” for tech companies, meaning the “overall credit quality for the hyperscaler would be worse than what’s currently being modelled”.
