For the first few years of the AI boom, the bulls had an unanswerable argument: this time the spending is real, and it is being paid for in cash. The richest companies in history were building data centers out of their own profits. No bubble, no leverage, no problem. That argument is now out of date. The boom has entered a more expensive phase, and the money funding it has changed.
The numbers are staggering. Eight major technology companies — Google, Meta, Microsoft, Amazon, Nvidia, Oracle, Broadcom, and SpaceX — face an estimated $3.6 trillion in off-balance-sheet commitments, according to an analysis based on SEC filings and a Morgan Stanley Research framework. That figure covers future chip purchases, data center leases, financial guarantees, and investment agreements. About $2.03 trillion of it is purchase commitments, and about $1.25 trillion is leases that have not yet commenced. Alphabet alone accounts for roughly $871 billion, followed by Microsoft at about $558 billion, Nvidia at about $426 billion, Oracle at about $322 billion, and Amazon at about $286 billion.
Look closer and the pattern sharpens. Oracle disclosed $288 billion in additional leases not yet commenced plus $34.15 billion in unconditional purchase obligations in its August 2026 filing. Nvidia reported $366 billion in scheduled future commitments, plus $56 billion in additional commitments and $108.5 billion in maximum guarantee exposure. Meta disclosed $454.52 billion in noncancelable contractual commitments in its filing for the quarter ended September 30, 2026, mostly for cloud capacity, servers, networking, energy, and hardware. These are not press-release promises. They are contractual obligations, filed with the SEC.
None of this proves a crash is coming, and it is important to say so plainly. The demand for AI computing is real, the customers signing these contracts are real, and the companies involved are among the most profitable in history. Debt used to build something people want is not the same as debt used to speculate on something nobody wants. The dot-com comparison, beloved of skeptics, has real limits: the fiber laid in 1999 eventually got used, just not by the companies that paid for it.
But the financing has changed in ways that should make even an optimist pay attention. When the cash runs low relative to the ambition, the gap gets filled with bonds, bank loans, private credit, and leasing structures designed to sit off the balance sheet — and with a web of deals among Nvidia, OpenAI, Oracle, and CoreWeave in which money appears to chase itself in a circle. The Bank of England and the IMF have both begun treating AI-related borrowing as a financial-stability issue rather than merely a technology-sector development. That is not a prediction of disaster. It is a sign that the people whose job is to worry about the financial system have started worrying.
The honest position is neither panic nor complacency. The AI buildout may pay off handsomely, and the companies building it may be right that demand will grow into the capacity. But the bill for that capacity is now measured in trillions, it is increasingly borrowed rather than earned, and it will come due on a schedule that does not care whether the AI revolution arrives on time. The technology story was the fun part. The debt story is the part that decides who survives it.
