Big Tech’s $1.65 Trillion Hidden AI Debt Reaches Pension Funds

Five of the biggest U.S. tech companies have built $1.65 trillion in AI obligations that never touch their balance sheets, Japan’s Nikkei reported on July 20. The figure, pulled from footnotes in recent filings from Alphabet, Microsoft, Amazon, Meta and Oracle, is eight times what it was four years ago and now exceeds the five companies’ combined $1.35 trillion in officially recorded debt.

Almost none of that $1.65 trillion stays contained inside those five companies. Chunks of it have already been packaged into private-credit funds and bonds now sitting inside pension systems, insurers and asset managers, which means a slowdown in AI demand would not stop at a tech company’s income statement.

The Beignet Workaround

Meta’s own invisible debt reached roughly $420 billion, the highest of the five and 2.8 times its official debt, according to Nikkei. A slice of that tab runs through a joint venture Meta struck with funds managed by Blue Owl Capital to build its Hyperion data center campus in rural Louisiana, a project whose total development cost tops $50 billion.

The mechanics are simple once you see them. Blue Owl, a private-credit investment firm, owns 80% of Hyperion through a special-purpose vehicle nicknamed Beignet. Meta holds the remaining 20% and leases the finished facility back on a long-term basis. The vehicle raised $30 billion to build it, including roughly $27 billion in loans, so the debt sits on Beignet’s books instead of Meta’s.

Meta still carries the downside. It reportedly agreed to guarantee losses in full if the joint venture unravels, an obligation that shows up only in a footnote, not on the balance sheet investors actually see.

Oracle Leases Its Way Around Its Own Balance Sheet

Oracle’s invisible debt climbed to $273.3 billion, a 30-fold jump in four years, and most of it traces back to leases feeding the Stargate data center buildout it shares with OpenAI, the ChatGPT maker.

One piece of that chain runs through developer Related Digital, which is building a compute campus for OpenAI with Oracle as anchor tenant. The project’s price tag has jumped to $16 billion, up from $10 billion just last fall, and the debt sits with Related Digital’s own special-purpose vehicle rather than with Oracle. Oracle’s lease payments are what service those bonds, and Oracle’s debt-to-equity ratio already sits near 400% before any of that off-book leasing is counted.

The chain has two links that both have to hold. OpenAI has to keep paying Oracle, and Oracle has to keep paying the developer regardless of what OpenAI does. A legal client alert on emerging litigation risks in AI data center financing tallies several similar Oracle-linked vehicles: roughly $13 billion from Blue Owl and JPMorgan into an SPV for Oracle’s Abilene, Texas facility built for OpenAI, a $38 billion debt package for sites in Texas and Wisconsin, and an $18 billion loan for a New Mexico campus. Those three deals alone approach $69 billion, a fraction of Oracle’s $273.3 billion invisible-debt total.

Pension Funds Bought Into Blue Owl’s Data Centers

The Beignet structure does not stop with Meta and Blue Owl. Retirement systems have money riding on it too.

The California State Teachers’ Retirement System and the Pennsylvania State Employees’ Retirement System both hold stakes in Blue Owl Capital funds that back these data center deals, according to a report from the Alliance for Affordable Energy on Meta’s Hyperion lease. That same report notes Meta can exit its Louisiana lease as early as 2033, which could leave local ratepayers covering gas plants, transmission lines and substations built to serve the project.

Insurers are exposed by a similar route. A note published in March and cited by CNBC found that opaque data center financing carries second-order litigation risk for downstream pension funds and insurers that later discover how concentrated their exposure really is. The adviser, identified in that report only as Rana, told CNBC that lease-value disputes between AI data center landlords and tenants are already underway, regardless of whether the broader market ever cracks.

• CalSTRS – holds a stake in Blue Owl Capital funds financing Meta’s Hyperion campus
• Pennsylvania SERS – also invested in the Blue Owl funds behind the same Louisiana project
• BlackRock bondholders – purchased more than $3 billion of bonds issued by the Beignet joint venture
• Insurers and asset managers – flagged by industry analysts as carrying concentrated, hard-to-see exposure through private-credit funds

None of those institutions set out to bet on GPU depreciation curves. They bought into credit funds and bonds, and the AI buildout came along with them.

A Bondholder Lawsuit Already Names Oracle

The exposure has already reached a courtroom. On January 14, 2026, a proposed class of bondholders sued Oracle in New York state court in a case captioned Ohio Carpenters’ Pension Plan v. Oracle, according to the same legal client alert cited above. The suit is an early sign that the off-book leasing structures big tech has leaned on are starting to draw the kind of scrutiny usually reserved for on-balance-sheet debt.

Wall Street Has Seen This Financing Trick Before

The pattern has a precedent, and it is not a flattering one. During the dot-com bubble, telecom equipment makers Nortel and Lucent lent money to their own customers to manufacture the appearance of revenue growth. When those customers defaulted, the loans turned worthless and helped push Nortel into one of the largest corporate bankruptcies in Canadian history.

Even the executives at the center of today’s buildout have acknowledged the risk out loud.

Misspending a couple of hundred billion dollars would be very unfortunate.

Those are Mark Zuckerberg’s own words, cited in Quinn Emanuel’s litigation client alert. Sam Altman, OpenAI’s chief executive, has gone further, conceding that “some investors are likely to lose a lot of money.”

Can GPUs Outlast the Debt Built Around Them?

Not on current lease terms, according to the industry source Nikkei quoted, and that mismatch is the mechanical risk underneath all of this financing.

“A commercial real estate lease retains asset value, but AI data centers rely on semiconductors that depreciate rapidly due to fast-paced technological advances,” the tech industry source told Nikkei. “If utilization rates drop due to weak AI demand, big tech firms will face massive asset impairment charges.”

The scale involved makes that mismatch matter more each quarter. Hyperscalers are now spending an estimated 45% to 57% of revenue on capital expenditure, a ratio previously seen mostly in utilities and telecoms, and Morgan Stanley and JPMorgan estimate the sector will need to issue roughly $1.5 trillion in new debt over the next three years. Investor Michael Burry and other analysts have separately argued that hyperscalers’ depreciation schedules understate true wear on AI hardware by close to $176 billion between 2026 and 2028, since GPUs are typically written down over five to six years despite an economic life closer to two or three.

Read against each other, the two ends of that table are the whole argument. One side says the obligations are real and growing fast. The other says the contracted demand behind them is real too.

Big Tech Insists the Backlog Is Real

Big Tech’s own answer to all of this is the backlog number, not a denial of the debt. Matt Garman, chief executive of Amazon Web Services (AWS, Amazon’s cloud division), put it plainly: “Current AI infrastructure investments are not speculative.”

The $1.45 trillion combined backlog across Microsoft, Amazon and Alphabet is the evidence he is pointing to. It is also, for now, the only thing standing between $1.65 trillion in footnoted obligations and the impairment charges that industry sources warn would follow if AI demand comes in soft. Oracle’s bondholder lawsuit, filed in January, suggests some investors are no longer willing to wait and see which number wins.