Hyperscaler AI debt: $662B in hidden data centre liabilities rattles credit markets

Not long ago, a credit default swap on a Big Tech name was a sleepy trade — low volume, tight spreads, the sort of instrument a bank held for regulatory capital reasons and never thought about. That changed sharply in the first half of 2026.

“The volumes we’re seeing in single-name CDS on these names are multiples of what they were 18 months ago,” Matt Mandell, BofA’s head of US single-name CDS, told Bloomberg in late May. Traders are buying protection on Amazon, Microsoft, Alphabet, Meta and Oracle at a pace that has reshaped the investment-grade derivatives desk. The reason is straightforward: the five largest cloud providers are projected to spend $750 billion in capital expenditure this year — 38 per cent of their combined revenue, according to S&P Global — and an increasing share of that is being financed with debt.

The debt itself is not the problem. What has begun to worry credit markets is the structure of the borrowing, the scale of what is not on balance sheets, and the growing entanglement between AI infrastructure finance and the private credit system.

The derivatives signal

Bloomberg reported in its Credit Weekly column on 23 May that hedge funds are selling CDS protection on hyperscaler names in volume, attracted by premiums that look like easy money against companies with trillion-dollar market caps. On the other side, banks that underwrite the bond issuance are buying protection to hedge their exposure. The notional value of credit derivatives referencing hyperscaler debt has grown sharply in 2026, though precise figures remain opaque.

Barclays analysts warned on 21 May that Big Tech’s AI debt binge is testing the capacity of the high-grade bond market to absorb it. The concern is not default risk in the conventional sense — these are investment-grade borrowers — but rather a liquidity mismatch: if AI revenue growth disappoints and the market reprices the sector simultaneously, the unwind could be disorderly.

Oracle’s debt load is the canary

If the other hyperscalers are raising eyebrows, Oracle is raising alarm bells. The company carries $160 billion in outstanding liabilities, of which roughly $133 billion is tied to its AI infrastructure build-out, according to JPMorgan estimates cited by Fast Company. Its debt-to-equity ratio sits at 415 per cent, compared with under 80 per cent for Amazon, Microsoft, Alphabet and Meta.

In April, private credit firm Blue Owl Capital pulled out of a US$10 billion data centre financing deal with Oracle, prompting Apollo and Pimco to step in as replacement lenders. A class-action lawsuit filed by Oracle bondholders alleges the company used special-purpose vehicles to obscure the true scale of its AI-related commitments — Fast Company’s Francis Northwood reported that Oracle had $66 billion in SPV-linked data centre obligations as of early 2026.

If your software business is in healthcare, the fund classifies it as healthcare exposure. The software exposure is meaningfully higher than it looks.
— Robert Dodd, Raymond James analyst, on private credit fund misclassification, as quoted by Fast Company

Oracle is one company. But the SPV structure it used is common across the sector, and the bond market is beginning to ask which hyperscaler is next.

The off-balance-sheet problem

In February, Moody’s Ratings flagged that the five largest US hyperscalers had accumulated $662 billion in off-balance-sheet data centre lease commitments — obligations that do not appear as debt on corporate balance sheets but represent legally binding future payments. That figure has almost certainly grown since.

These commitments take several forms. One is the residual value guarantee, or RVG: Meta, for example, structured a $28 billion RVG — known internally as Project Hyperion — through a joint venture with a data centre operator. Meta guarantees the operator a minimum return and retains an option to buy the assets at a pre-agreed price. The obligation is real; the accounting treatment keeps it off the books.

Researchers at the Dallas Federal Reserve published a paper in February examining how AI-related debt issuance affects interest rate markets through what they call “duration supply” — the sheer volume of long-dated fixed-income paper that hyperscaler borrowing is injecting into the system. Their core finding: the AI capital cycle is now large enough to move the yield curve.

S&P Global’s ratings arm summarised the mood in a note to clients: “The margin for error is narrowing.”

Private credit and the shadow banking link

The private credit market, estimated at $3 trillion globally, has become the primary lender for AI infrastructure beyond what the public bond market will absorb. More than $450 billion had flowed into technology through private credit channels by late 2025, according to Fast Company, and the figure has kept climbing.

This creates a dangerous interlock. Private credit funds are less transparent than banks, face lighter capital requirements, and — as Raymond James analyst Robert Dodd’s observation above makes plain — frequently classify their exposure in ways that obscure concentration risk. A fund holding $2 billion in loans to data centre operators may report its exposure as “real estate” or “infrastructure” rather than “technology” or “AI.”

When you see one cockroach, there are probably more.
— Jamie Dimon, JPMorgan CEO, on the private credit selloff, as quoted by Fast Company

The cockroach Dimon was referring to was the SaaSpocalypse — the sharp selloff in software-company debt that rippled through private credit portfolios in late 2025 and early 2026. The concern is that a similar repricing of AI infrastructure debt, triggered by disappointing AI revenue growth or a rate shock, would be larger and harder to contain.

What it means for cloud buyers

For enterprise customers — including the Australian banks, government agencies and ASX-listed companies that have migrated core workloads to AWS, Azure and Google Cloud — the hyperscaler debt story is not an abstraction.

PIMCO published an analysis in early 2026 arguing that while hyperscaler balance sheets remain strong by conventional metrics, the off-balance-sheet commitments and private credit interconnections introduce “micro risks” — pricing pressure, capacity allocation shifts, or changes in service terms — that enterprise procurement teams are not currently modelling. If debt servicing costs rise and capex budgets are reassessed, the cloud services that enterprises depend on could see margin compression passed through in contract renewals or reduced investment in non-AI infrastructure.

Australian enterprises are disproportionately exposed. The three major cloud providers — AWS, Azure and Google Cloud — hold roughly 75 per cent of the Australian cloud market, according to the ACCC’s most recent digital platform services inquiry. Concentration risk on debt-heavy providers is not yet a boardroom topic. The credit markets suggest it should be.

The big ‘if’

Within the private sector, the question is will AI create the productivity acceleration to overcome the profligate spending we’re currently engaged in. The world needs a savior, and the hope is that AI is the savior we need.
— Ken Griffin, Citadel CEO, speaking at Davos in January 2026, as quoted by Fast Company

Griffin’s framing captures the asymmetry at the heart of the hyperscaler credit story. If AI revenue materialises at the scale the hyperscalers are betting on — if OpenAI’s reported $1.4 trillion in multiyear infrastructure commitments produces the returns its backers expect — then $750 billion in annual capex will look prescient, and the debt will be serviced comfortably from cash flow.

If it does not, the credit derivatives market has already begun to price the alternative. Enterprise technology buyers, investors and regulators are only starting to ask the question the CDS desks asked 18 months ago: what happens if the AI capex cycle turns before the AI revenue cycle has really begun?

S&P Global’s credit analysts put it bluntly: “The margin for error is narrowing.” For Australian enterprises whose operations run on hyperscaler infrastructure, the margin is narrowing too.