Big Tech's $1.65 Trillion Hidden AI Debt Problem

Five of the world's most valuable companies have quietly built up $1.65 trillion in hidden AI debt that never appears on a standard balance sheet. It now dwarfs the borrowing investors can actually see. The accounting is entirely legal, and that is exactly what should worry you.

By Blumefield | July 22, 2026

The liabilities that never reach the ledger

The largest technology companies on earth have discovered a way to spend like never before without frightening the people who lend them money. According to a new analysis from Nikkei, the combined hidden AI debt sitting off the balance sheets of Alphabet, Microsoft, Amazon, Meta and Oracle has reached roughly $1.65 trillion, a figure that has grown almost eightfold since 2022. What makes the number remarkable is not just its size but its invisibility. It is already larger than the roughly $1.35 trillion of conventional debt those five firms report in plain view.

That gap matters because the ratios most investors rely on, debt to equity, debt to EBITDA, interest coverage, are all built from the visible number. When the bulk of a company's future obligations live somewhere the standard ratios cannot reach, the picture of financial health those ratios paint is flattering by design. The hidden AI debt does not represent fraud. It represents the outer edge of what accounting rules permit, and the hyperscalers have driven straight to that edge.

The scale of the underlying spending explains the urgency. Capital expenditure across the five companies is on course to hit about $725 billion in 2026, nearly five times the $148 billion they spent in 2023. Almost none of that buildout is funded from a single quarter of cash flow. It is financed, leased, and promised forward, and a growing share of those promises never lands on the ledger.

How the leases disappear

The mechanics of this hidden AI debt are mundane, which is part of why they have gone largely unexamined. Under the current lease accounting standards overseen by the Financial Accounting Standards Board, a lease becomes a reported liability only when it commences, meaning when the asset is actually available for the tenant to use. A data center that has been designed, contracted and signed but not yet energized sits in a kind of accounting waiting room. The obligation is real and often runs for well over a decade, yet it appears only in the footnotes until the day the keys change hands.

Roughly $662 billion of signed data center leases currently sit in that waiting room across the industry. Morgan Stanley, examining the wider financing landscape, puts total off-balance-sheet exposure closer to $1.8 trillion once purchase commitments and supplier financing are included. Its breakdown counts nearly $1 trillion in purchase commitments for chips and capacity, more than $800 billion in leases that have not yet begun, and about $110 billion in accounts-payable financing. Read the accounting framework and none of this violates it.

The comparison critics keep reaching for is Enron, which used off-balance-sheet vehicles to keep debt out of view before its collapse. The parallel is imperfect. The hyperscalers are profitable, cash-rich and disclosing these commitments in their filings. But the instinct behind the structure, keeping the scariest liabilities where the headline ratios cannot see them, is uncomfortably familiar.

Oracle becomes the warning label

If the industry has a canary, it is Oracle. On July 9, S&P Global Ratings cut the company to BBB-, one notch above junk, citing the cash demands of its accelerating AI infrastructure business. Oracle's hidden AI debt had swelled to about $273 billion by the end of May, more than thirty times its level four years earlier. The agency added some $260 billion of further lease commitments expected to begin between 2027 and 2029, on top of unconditional purchase obligations.

The rating agency flagged a mismatch that should trouble anyone watching the sector. Oracle has locked itself into data center leases running fifteen to nineteen years, while many of its largest customer contracts, including its landmark deal with OpenAI, run closer to five. OpenAI alone accounts for roughly half of Oracle's $638 billion remaining performance obligation backlog, and OpenAI is a private company that has never turned a profit. S&P projects an Oracle free operating cash flow deficit approaching $42 billion in fiscal 2027.

Investors can read the terms in Oracle's own disclosures. The company's funding runway, the time it can sustain current spending before needing materially more capital, is estimated at under two years, far shorter than any of its peers. Oracle is not the whole story of hidden AI debt, but it is the clearest illustration of how quickly the structure can turn from clever to precarious.

Meta, Alphabet and the private-credit machine

Meta carries the largest hidden AI debt load in absolute terms, with off-balance-sheet obligations of around $420 billion, roughly 2.8 times the debt on its books. A signature example is its $27 billion joint venture with private-credit firm Blue Owl, structured to fund the sprawling Hyperion data center campus in Louisiana. The arrangement channels tens of billions into construction while keeping the associated borrowing at arm's length from Meta's own balance sheet. The details sit in Meta's investor disclosures for those willing to dig.

Alphabet has taken a subtler route, disclosing about $40.7 billion in future funding commitments to off-balance-sheet vehicles, including a roughly $30 billion equity derivative dressed as infrastructure financing. These are not back-alley instruments. They are engineered by some of the most sophisticated finance teams in the world, precisely to raise enormous sums without denting the credit metrics that keep borrowing cheap.

The bond market is starting to notice. Gross issuance tied to AI buildouts in 2026 has already surpassed all of 2025, with roughly $152 billion sold in the first four and a half months and perhaps $300 billion expected across the year. As covered in earlier Blumefield reporting on the AI debt wave, credit investors are beginning to demand more for the privilege, a shift that quietly raises the cost of the whole enterprise.

What breaks if the AI revenue is late

For now, the visible metrics look reassuring. Excluding Oracle, debt to EBITDA ratios remain modest, with Microsoft and Alphabet near 0.6 times, Meta at 0.8, and Amazon at 1.3. These are not stretched balance sheets by any conventional reading. The bet embedded in the hidden AI debt is straightforward. Demand for computing power will keep rising, the data centers will fill, and the revenue will arrive in time to service obligations that in some cases stretch two decades out.

The danger is timing. Leases and chip commitments are fixed and long. AI revenue, by contrast, is still largely a forecast, and the customer contracts underpinning it are often short. If enterprise adoption slows, if a cheaper model undercuts pricing, or if a single anchor tenant stumbles, the cash outflows do not pause to wait for the story to recover. That asymmetry is the real risk hiding inside the footnotes.

None of this means the AI infrastructure boom is a mirage. The compute is being used, the largest players are genuinely profitable, and off-balance-sheet financing is a legitimate tool. But $1.65 trillion is not a rounding error, and its location outside the headline numbers means the market is pricing these companies on a partial view. The most important lesson from every previous cycle is simple. When an entire industry agrees to stare at the same reassuring number, the risk is usually sitting somewhere else. Readers tracking this story can follow the continuing analysis at Blumefield.

The next real test arrives with earnings season, when investors will finally get to weigh the promises buried in the footnotes against the revenue that shows up in the results.