All week the AI story has been about capability and geopolitics — models breaking into servers, governments accusing each other of theft, five-billion-dollar chip deals. Underneath it runs a less glamorous question that got louder this week: who is actually paying for all of this, and where is the bill being recorded?

A Nikkei analysis of the financial footnotes at Alphabet, Microsoft, Amazon, Meta, and Oracle put a number on it: roughly $1.65 trillion in off-balance-sheet obligations tied to AI infrastructure — data-center leases, GPU supply commitments, and special-purpose vehicles. That is more than the $1.35 trillion those same companies carry as formal, on-balance-sheet debt.

What "off the books" actually means

Start with the important caveat, because it's the whole story: this is legal. The obligations are disclosed — in the footnotes — and the accounting is GAAP-compliant and signed off by auditors. Nobody is committing fraud.

What they're doing is structural. Instead of borrowing directly to build a data center, a tech company partners with or creates a special-purpose vehicle — a separate legal entity that raises the debt and owns the asset. The tech company then signs a long-term lease or capacity contract to use it. The SPV's borrowing sits on the SPV's books, not the tech company's, even though the tech company is effectively on the hook for the payments.

The concrete deals are large. In Abilene, Texas, a roughly $10 billion vehicle — backed by Blue Owl and JPMorgan-led debt — owns an OpenAI facility that Oracle leases long-term; a separate $38 billion package funds data centers in Texas and Wisconsin; an $18 billion loan backs a site in New Mexico. Multiply that pattern across five hyperscalers and you get to $1.65 trillion.

Why anyone should care

The Enron comparison is doing a lot of work in the coverage, and it needs handling with care. Enron used special-purpose vehicles too — but, as analyst Gil Luria put it, "Enron's crime wasn't having special purpose vehicles. Enron's crime was hiding them." Big Tech isn't hiding these; they're in the disclosures. The worry is subtler: off-balance-sheet obligations are harder to see at a glance, they understate how leveraged these companies really are to the AI build-out, and when a data center goes live, its lease can roll onto the balance sheet all at once.

This isn't hypothetical. S&P has already downgraded Oracle's credit outlook, explicitly citing stretched leverage from AI commitments that include something like $260 billion in future lease obligations. Moody's and the Bank for International Settlements have flagged the broader trend. The first credit consequences of the AI capex boom are landing.

The through-line

Read alongside this week's other money story — AMD putting up to $5 billion into Anthropic in exchange for tens of billions in chip orders — a picture forms. The financing of the AI build-out is getting genuinely exotic: circular deals between chipmakers and labs, off-balance-sheet vehicles between hyperscalers and lenders, capital moving in loops that keep demand looking robust.

None of this means a crash is coming. Legal, disclosed, GAAP-compliant financing is not a scandal. But it is a reason to read the footnotes. The headline numbers on AI spending are large; the numbers just underneath them are larger, and they're the ones that decide whether this boom is built on solid ground or on a structure that only holds while everything keeps going up.