Big Tech's hidden AI debt hits $1.65 trillion, Nikkei finds

A quietly assembled number from Japan's business press is doing more to reframe the AI capex story this week than any earnings call. Five US tech giants, Alphabet, Amazon, Meta, Microsoft, and Oracle, are carrying roughly $1.65 trillion in 'hidden' debt tied to AI infrastructure, according to a Nikkei Asia study reported by Tom's Hardware. That figure is roughly 122% of the $1.35 trillion these same companies disclose on their balance sheets, and it has grown eightfold over four years.

The mechanics are the interesting part. Most of the obligation comes from long-term data center lease agreements and future purchase commitments for servers, graphics accelerators, and other compute hardware that has yet to be delivered. Because those liabilities live in accounting footnotes attached to SEC filings rather than on the balance sheet itself, they are legal under current US accounting rules but easy to miss for anyone reading only the headline numbers. The clearest illustration, described by TheNextWeb, is Meta's Hyperion data center in Louisiana, where Meta and Blue Owl Capital placed equity into a separate structure that assumed $27 billion in debt. Meta's total off-book number is estimated near $420 billion, close to three times what it disclosed in its FY 2026 financial statement.

Oracle is the more acute case. Its off-balance-sheet exposure sits at $273.3 billion, a roughly 30-fold increase over four years, and is largely tied to long-term leasing agreements related to the Stargate AI data center project across Texas, New Mexico, Michigan, and Wisconsin. S&P has already downgraded Oracle's short-term rating from A-2 to A-3 and long-term from BBB to BBB-, just above junk. Morgan Stanley and Moody's have also raised concerns about the broader reporting practice.

The honest caveat is that this is one analyst house's methodology, not an audited restatement, and the disclosures themselves are legal under current US rules. What the reporting doesn't give you is a revenue-versus-obligation schedule, i.e. whether AI revenue arrives fast enough to service the leases. If it doesn't, the first layer exposed isn't the hyperscalers; it's the private credit sponsors quietly funding the buildout behind them.