Capital expenditure is the number everyone watches. For the past two years the AI trade has been priced off it: how much did the hyperscalers spend on servers, chips and buildings last quarter, and did they raise the guide. Roughly $600 billion across the group over the past reported year — a figure large enough to carry its own news cycle.
A Wall Street Journal review of the footnotes in the most recent securities filings of nine of the biggest technology companies found something about five times that size sitting a few pages further back. Around $3 trillion of commitments, overwhelmingly AI-related, that do not show up as debt on the balance sheet. Roughly $1.2 trillion of it is leases on facilities that have not begun operating. The other $1.9 trillion is contracted purchases of chips, equipment, energy and services.
The names reviewed include Alphabet, Amazon, Microsoft, Meta, Oracle, Nvidia, Broadcom, AMD and SpaceX. None of this is hidden in the sense of concealed — it is disclosed, in the notes, exactly where accounting rules say it belongs. It is hidden in the sense that almost nobody reads it, and no headline metric captures it.
Why the number doesn’t count as debt
Two mechanics do most of the work. A lease on a data centre that has not yet been handed over does not become a balance-sheet liability until the asset is available for use — so a signed 15- or 20-year commitment on a building still under construction lives in the footnotes. And a purchase obligation for future chips, power or compute capacity is a contractual promise to pay, not borrowed money, so it never enters the debt line at all.
Both are real cash out the door on a fixed schedule. Neither is optional in any ordinary sense. And the scale is now roughly triple what these companies owe under their recognised leases and long-term borrowings combined.
The individual disclosures show how fast the shift is happening. Alphabet reported $811 billion of purchase and contractual obligations, against $332 billion three months earlier — roughly 2.4 times higher in a single quarter. Meta carries $347 billion of future lease commitments. Oracle’s off-balance-sheet lease commitments have grown roughly 30-fold in four years.
Our take: The bear case on AI capex has always been “what if the demand doesn’t show up.” It has been easy to wave off, because capex is discretionary — a hyperscaler that sees demand soften simply stops buying, and the spending line falls with it. Off-balance-sheet commitments break that assumption. A 20-year lease on a building that isn’t finished and a take-or-pay contract for compute running into the 2030s do not flex with demand. They are the part of the build-out that keeps spending money after the enthusiasm stops. That is a different risk profile from the one the market has been underwriting, and it is why the footnotes now matter more than the guide.
What this does and doesn’t tell you
It does not say these companies are overextended. Alphabet, Microsoft, Amazon and Meta generate enormous operating cash flow, and a contracted purchase of chips you fully intend to use is not a distressed obligation. Committing early is also how you secure scarce supply — and as SanDisk’s investor day made plain, signed multi-year volume is exactly what the supply side is now selling.
What it does say is that the reported numbers have stopped describing the exposure. If you are sizing the AI build-out from capex and long-term debt, you are looking at roughly a quarter of it. Two of the more striking recent stories in this space — Broadcom’s $370 billion financing backstop and Nvidia’s shrinking guarantee to OpenAI — were both about obligations that never appeared as debt either. That is not an anomaly. It is the structure of how this build-out is being financed.
What to watch
- The quarterly delta in purchase obligations. Alphabet’s move from $332 billion to $811 billion in one quarter is invisible in the capex line and enormous in the footnote. Read the change, not the level.
- Take-or-pay language. The difference between “we intend to buy” and “we owe regardless” is where the actual risk sits, and it is only legible in the contract terms disclosed in the notes.
- Lease commencement dates. The $1.2 trillion of not-yet-operating leases converts into recognised balance-sheet liabilities as buildings come online — a mechanical, scheduled increase in reported obligations over the next several years, independent of any new signing.
- Whether auditors or regulators push for prominence. Nothing about the current disclosure is improper. But when a category grows to triple the recognised balance-sheet obligations, pressure to surface it usually follows.
- Cost discipline everywhere else. Oracle’s pattern — cutting headcount while raising capex — is what servicing fixed, long-dated commitments looks like from the inside.
The AI trade has been argued for two years over a number that turns out to be the smaller one. The bigger one has been sitting in the filings the whole time.