Off-balance-sheet debt at Alphabet, Microsoft, Amazon, Meta, and Oracle has grown roughly eightfold since 2022, reaching an estimated $1.65 trillion, according to a new Nikkei analysis. That figure now exceeds the five companies’ combined on-balance-sheet debt of roughly $1.35 trillion, meaning the liabilities investors can’t see in a standard debt-to-equity ratio are larger than the ones they can.
The gap comes from how the AI buildout is being financed. Instead of borrowing directly and putting new debt on the balance sheet, the hyperscalers are increasingly locking in long-term data center leases, GPU supply commitments, and joint ventures with private credit funds — structures that lock in future cash outflows without showing up as formal liabilities today.
Meta is the starkest example. Its off-balance-sheet debt is estimated at around $420 billion, nearly three times its transparent debt load. Much of that runs through vehicles like its $27 billion private-credit joint venture with Blue Owl, used to fund the Hyperion data center campus in Louisiana. Oracle’s hidden debt has grown even faster in relative terms: an estimated $273 billion as of the end of May, more than thirty times what it was four years ago, funded partly through project-finance packages tied to AI sites in Texas and Wisconsin. Oracle’s debt load, at roughly 2.5 times sales, was enough to prompt S&P Global to downgrade the company to its lowest investment-grade tier.
Alphabet, long seen as the hyperscaler with the cleanest balance sheet, has moved in the same direction. Its most recent quarterly filing discloses $40.7 billion in future funding commitments to off-balance-sheet vehicles, including a roughly $30 billion equity derivative structured as AI infrastructure financing, and a new joint venture with Blackstone that shifts a portion of its capex onto outside capital.
Why It’s Called Shadow Borrowing
The Bank for International Settlements has its own term for this: shadow borrowing. Its concern is straightforward — a company can raise the external capital it needs to fund a data center without its formal debt figures ever reflecting the obligation. That makes leverage harder for investors, rating agencies, and regulators to see in real time. Moody’s has flagged a related risk in the underlying contracts themselves: many of the leases involved are pre-operational, meaning the company is committed to paying for capacity before it generates any revenue from it.
Morgan Stanley’s own estimate of the broader AI financing picture puts total off-balance-sheet exposure across the industry at roughly $1.8 trillion, made up of nearly $1 trillion in purchase commitments, more than $800 billion in leases that haven’t yet begun, and about $110 billion in accounts-payable financing. By the bank’s numbers, hyperscaler leverage ratios have climbed from 0.9x to 1.8x, and annual AI-related debt issuance is on pace to exceed $570 billion a year.
The Circularity Problem
Nikkei’s framing draws a direct historical parallel: an industry structure where capital circulates among Nvidia, the hyperscalers, data center operators, and AI startups, each funding the next link in the chain. That’s the same basic shape as the early-2000s telecom buildout, where equipment makers financed the internet companies buying their gear — a structure that worked exactly until demand growth stopped outrunning the debt taken on to build ahead of it. Nikkei’s caution is that if AI infrastructure supply keeps expanding faster than monetized demand, the mismatch shows up first in the parts of the balance sheet that were never on the balance sheet to begin with.