Big Tech buries billions in AI debt as bubble fears mount
Meta is paying roughly $6.5bn (£4.82bn) in extra financing costs to keep $27bn of AI infrastructure borrowing off its balance sheet, a costly accounting choice that captures the mood in Big Tech’s race to build the pipes of AI without spooking investors.
The arrangement, known as special purpose vehicle financing (SPV), allows an external entity to raise debt, construct the data centre, and lease it back to the tech group.
On paper, Meta books lease payments rather than traditional borrowing, but really, it has committed to decades of payments tied to huge computing facilities.
The structure was used for Meta’s $30bn data centre project in Louisiana, which was financed largely through private credit heavyweights like Blue Owl Capital, Pimco, BlackRock and Apollo.
Meta owns around 20 per cent of the vehicle and has offered a residual value guarantee, meaning it could be required to compensate investors if the project’s value falls below agreed levels at the end of the lease.
Oracle, too, has pushed tens of billions of dollars of AI data centre investments in similar ways, including a $38bn package tied to its partnership with OpenAI.
Elon Musk’s xAI has raised $20bn via a comparable structure, with debt secured against Nvidia chips.
And in some cases, Nvidia has even invested equity in customers that then use it to buy its hardware, a circular flow of capital that keeps revenue ticking while the chip giant’s liabilities sit elsewhere.
The accounting is legal and disclosed, but it is unfolding against a backdrop of eye-watering AI forecasts and a surge in borrowing across the sector.
Morgan Stanley estimates hyperscalers could raise $400bn in corporate bonds in 2026 alone to fund AI expansion.
JPMorgan has calculated that AI and data centre firms now account for 14.5 per cent of its $10tn investment-grade bond index, which is about $1.5tn in debt exposure.
UBS says roughly $450bn has flowed from private capital into tech infrastructure as of early 2025.
For market watchers, that scale and complexity are stirring memories of past tech bubbles.
AJ Bell’s Russ Mould points to Richard Bookstaber’s study of past market crises. “He argued that leverage, complexity and opacity help to fuel bubbles,” he told CityAM.
“The use of special purpose vehicles and off-balance sheet structures to fund enormous AI capital investment will bring back bad memories for experienced investors.”
He also adds that while these structures comply with accounting rules, “more debt and more complexity mean more risk”, particularly if returns do not match the spending.
Strong balance sheets
But this doesn’t yet look like a rerun of the late-1990s telecom crash, and the biggest US tech firms are still sitting on huge piles of cash.
Among the hyperscalers, only Oracle and Apple currently carry more long-term debt than cash and short-term investments.
Nvidia’s debt-to-capital ratio stands at 8.3 per cent; Alphabet’s at 10.3 per cent; Meta’s at 27.9 per cent.
Oracle’s is far higher at 83.9 per cent, though still investment grade, albeit on negative watch.
Matt Britzman, senior equity analyst at Hargreaves Lansdown, told CityAM: “Among the four largest public market AI investors – Amazon, Alphabet, Meta and Microsoft – total calendar-year 2026 capex is forecast at over $600bn, so it’s not like these companies are trying to hide their ambitions”.
He adds that the combined operating cash flow across the group is expected to approach $700bn in 2026.
“Off balance sheet arrangements also look modest in scale relative to the enormous cash flows that big tech are pulling in, which reduces concerns about hidden leverage.”
“Demand for compute remains extremely strong, and cloud giants are still seeing rental demand for six-year-old A100 chips”, Britzman added.
The question, then, is less about solvency today and more about future durability.
Gartner forecasts global AI spending will hit $2.52tn in 2026, up 44 per cent year on year.
By 2030, it expects AI to completely dominate IT budgets. But credit agencies have flagged that some of the sector’s biggest customers, including OpenAI, are not expected to turn profitable until later in the decade.
These data centres are financed on the basis of 20-year demand assumptions.
And if that demand translates into projected revenue growth, these structures will look prescient. On the other hand, if it does not, the risk will sit in private credit vehicles and long-term leases.