The AI Buildout's Real Number Isn't on Anyone's Balance Sheet
Six companies have signed roughly $1.5 trillion in purchase commitments for compute, chips, and power — with another $1.5 trillion in leases beside it. Neither figure is debt, and neither shows up where investors are trained to look.
A Financial Times analysis published this month puts the combined purchase commitments of Alphabet, Microsoft, Amazon, Nvidia, Oracle, and Meta at close to $1.5 trillion — obligations tied largely to computing infrastructure, chips, data-center capacity, and energy.
Separately, Goldman Sachs has identified roughly another $1.5 trillion in lease commitments across the AI data-center buildout.
Neither number is capital expenditure. Neither is debt. And that is precisely why they matter.
What a purchase commitment is
Capex is money already spent. Debt is money already borrowed, sitting on the balance sheet where every screen and every credit model can see it.
A purchase commitment is neither. It is a contractual agreement to spend in a future period — a signed obligation to buy a specified quantity of GPUs, or take a specified quantity of power, or occupy a specified quantity of colocation capacity, at prices and volumes fixed in advance. It appears in the notes to the financial statements, not on the balance sheet proper, and it does not carry a coupon, a maturity ladder, or a credit rating.
For years this was an unremarkable disclosure. Every large company has some contractual future spend — supplier agreements, real-estate leases, minimum volume commitments. It was footnote material because the magnitude was footnote-sized.
At $1.5 trillion across six companies, it is not footnote material. It is comparable in scale to the sovereign debt of a mid-sized economy, and it is larger than the capex figures that anchor essentially all coverage of the AI buildout.
Why the capex number understates it
The standard framing of the AI arms race is a capex race: Alphabet raised to $205 billion, Microsoft held flat, Meta's free cash flow went to zero. Those are the numbers that move stocks on earnings night.
But capex measures what a company spent in the quarter that ended. Purchase commitments measure what it has already promised to spend in the quarters that haven't started. The first is history. The second is the forward curve — and right now the forward curve is dramatically steeper than the trailing print suggests.
Alphabet's purchase commitments rose sharply between the first and second quarter as it locked in long-term infrastructure and energy agreements. That is the mechanism in miniature: the company signs multi-year supply for power and compute capacity, and the obligation lands in the notes long before a dollar of it appears as capex.
The practical effect is that an investor reading only capex is reading a lagging indicator of a commitment that was made quarters earlier.
The comparison to debt is imperfect, and imperfect in an uncomfortable direction
Purchase commitments are not debt. They are not borrowed money, they do not accelerate on covenant breach, and they can sometimes be renegotiated with a counterparty who would rather keep the customer than enforce the contract.
But they share the property that matters most: they are fixed future cash obligations, largely insensitive to whether the revenue that was supposed to service them arrives.
Debt has a well-developed apparatus for exactly this problem — ratings, covenants, disclosure standards, stress-testing conventions, an entire analyst discipline built around leverage ratios. Purchase commitments have almost none of it. There is no standard metric for commitment coverage. There is no convention for what an appropriate ratio of committed future spend to operating cash flow looks like for a hyperscaler. There is no agency assigning a grade.
The obligations are real. The analytical infrastructure for pricing them is not built yet.
What the leases add
The $1.5 trillion in lease commitments identified by Goldman sits alongside, not inside, the purchase figure. Leases have somewhat better accounting treatment post-ASC 842 — operating leases now capitalize onto the balance sheet as right-of-use assets and corresponding liabilities — but the AI data-center buildout has produced structures that stretch what the standard was designed to capture: capacity agreements with neocloud providers, power purchase agreements with independent generators, and colocation contracts whose classification depends on details buried in the terms.
Add the two and the AI buildout has roughly $3 trillion of contractual future obligations that are not what most people mean when they say "debt," across a handful of companies.
The counterargument, stated fairly
These are the most cash-generative businesses in the history of public markets. Combined operating cash flow across the six companies runs into the hundreds of billions annually. A $3 trillion obligation spread across five to ten years, serviced from that base, is not obviously imprudent — it is a company with an enormous income statement pre-buying scarce inputs at fixed prices, which is what you do when you expect the inputs to get scarcer.
If AI revenue materializes at anything close to the projected trajectory, these commitments will look like brilliant forward-buying of capacity nobody else could get. Locking in power in 2026 at 2026 prices, in a market where the constraint is increasingly electricity rather than silicon, is defensible on its own terms.
The commitments are also, importantly, mostly to each other and to their suppliers — chip makers, foundries, equipment vendors, utilities — which means the counterparty risk is distributed through a supply chain that has every incentive to keep delivering.
The read
The number is not a scandal. Nothing here is hidden; it is all disclosed, in the notes, where the accounting standards say it belongs.
The issue is that the center of gravity of the AI trade has moved into a disclosure format the market has not learned to read. Investors have spent two years arguing about capex — a backward-looking number — while twice that magnitude accumulated in a forward-looking one that generates no headlines because it does not sit on a line anyone screens for.
The question worth asking about any of these six companies is no longer what they spent last quarter. It is what they have already agreed to spend through 2030, and what happens to that obligation if the revenue curve flattens a year earlier than the contracts assume.
That answer exists. It is in the notes.