AI data centers are being built on debt Wall Street barely understands

Alphabet, Microsoft, Amazon, Meta, and Oracle are carrying roughly $1.65 trillion in AI infrastructure debt that does not appear on their balance sheets — an eightfold increase in four years, and now larger than the group's combined on-balance-sheet debt of about $1.35 trillion. This isn't a rounding error in a footnote. It's the financial structure underneath the entire AI buildout, and it's built to be largely invisible to the people who buy these companies' stock.

The mechanism is the special-purpose vehicle, or SPV. A hyperscaler doesn't borrow the money directly and put a data center on its own books. Instead, it sets up (or partners into) a separate legal entity that owns the data center, borrows the construction money from private credit funds, and signs a long-term lease or compute-purchase agreement with the parent company. Because the hyperscaler typically owns a minority stake — Meta holds 20% of its Hyperion SPV with Blue Owl Capital, for instance — the debt doesn't consolidate onto the parent's balance sheet under standard accounting rules.

The scale is no longer a curiosity

Meta's off-balance-sheet obligations now stand at roughly $420 billion — 2.8 times its reported debt. The Hyperion structure alone routes about $27 billion of project financing through an entity Meta doesn't fully own on paper, even though Meta is the only real customer for the compute it produces.

Oracle has gone further. The company has arranged a roughly $13 billion SPV with Blue Owl and JPMorgan to own the OpenAI facility in Abilene, Texas ($10 billion of that as debt), a $38 billion debt package for two more data centers in Texas and Wisconsin, and an $18 billion loan for a New Mexico site. Oracle's total obligations have grown more than thirtyfold in four years, reaching $273.3 billion as of the end of May 2026.

What happens when a bet like this goes wrong

Oracle is the live case study. S&P downgraded Oracle's credit rating to BBB- on July 9 — one notch above junk — citing AI infrastructure spending that is outpacing revenue. AI-related business made up about 27% of Oracle's fiscal 2026 revenue, and S&P expects that to climb toward 60% by 2028. The agency is forecasting a free operating cash flow deficit of nearly $42 billion for fiscal 2027.

Oracle's debt-to-equity ratio now sits at roughly 500%, against about 50% at Amazon and 30% at Microsoft. And Oracle is far more customer-concentrated than its peers: OpenAI accounts for roughly half of Oracle's $638 billion in remaining performance obligations — the revenue Oracle is contractually owed but hasn't yet recognized. If OpenAI's own economics wobble, so does a very large share of the contracted revenue backing Oracle's debt load.

Why this structure exists at all

None of this is illegal or even particularly novel — SPV financing for large infrastructure projects (pipelines, toll roads, power plants) has existed for decades. What's different is the speed and the counterparty risk. Traditional infrastructure SPVs finance assets with 30-to-50-year useful lives against contracted revenue from utilities or governments. AI data center SPVs are financing assets whose core hardware — GPUs — depreciates on a 3-to-5-year cycle, against revenue contracts from companies (OpenAI, Anthropic, and the hyperscalers' own AI divisions) whose own profitability is still unproven at scale.

Private credit funds — Blackstone, Blue Owl, Apollo, Pimco, and BlackRock chief among them — have become the primary lenders here precisely because banks, constrained by post-2008 capital rules, are less willing to hold this exposure directly. Morgan Stanley projects private credit will supply an additional $800 billion in data center financing over the next two years. That's $800 billion of exposure landing on entities with far less regulatory oversight and far less public disclosure than a bank balance sheet.

The read-through for the rest of the industry

This financing structure is now directly connected to stories we've been tracking all week: the Pentagon's reported $5 billion loan discussion with Fluidstack to secure data center supply chain capacity, and Microsoft's rumored 38GW data center buildout target for 2032. Power availability and capital availability are now the two hard constraints on AI scaling — not chip supply, not model architecture. Both constraints are being solved with debt structures that keep growing faster than the revenue meant to service them.

What to watch

Three signals matter more than the headline capex numbers. First, watch credit rating actions on the hyperscalers with the highest customer concentration — Oracle's downgrade is a template, not an outlier, if OpenAI or Anthropic revenue growth disappoints. Second, watch private credit fund disclosures for how much AI infrastructure debt they're carrying relative to their total book — concentration risk in the lenders is just as real as concentration risk in the borrowers. Third, watch whether any hyperscaler starts consolidating these SPVs back onto its balance sheet voluntarily; that would be the clearest signal yet that the off-balance-sheet structure is becoming a liability rather than a convenience.