$662 Billion of Data Centre Leases Sit Off Big Tech's Balance Sheets

Moody's Ratings analysed the year-end 2025 disclosures of the five largest US hyperscalers and found $969 billion in total future data centre lease commitments. Of that, $662 billion covers leases that have not yet commenced and therefore sit off the balance sheet entirely. That unrecorded figure equals 113% of the same five companies' adjusted debt.

The numbers, precisely

In February 2026, Moody's Ratings published an analysis of the financial disclosures of Amazon, Meta, Alphabet, Microsoft and Oracle. It found $969 billion in total undiscounted future lease commitments at year-end 2025.

More than two thirds of that, $662 billion, relates to leases that have not yet commenced. Under generally accepted accounting principles, those are not recognised as current liabilities. They do not appear on the balance sheet at all.

The comparison Moody's drew is the one that made this a story. The unrecorded $662 billion equals 113% of the same five companies' most recent adjusted on-balance-sheet debt. Their invisible commitments exceed everything their credit profiles currently show.

Why the accounting rule allows this

Lease accounting recognises a liability when the lessee obtains the right to use the asset. That is the commencement date, not the signature date.

The logic is defensible. Until the landlord has delivered the property, the tenant has not received anything. Recognising a liability for services not yet provided would misrepresent the current position.

For an office lease starting in three months, the gap is immaterial. For a data centre purpose-built over several years, the gap between signature and commencement can span most of a decade. Terms then run for up to twenty-five years afterwards.

What the ratings agency actually said

Moody's analyst David Gonzales was careful about the framing when speaking to Fortune. His point was that the companies had not avoided a liability through structuring. They had not yet received the services that trigger it. They will.

That is a materially different claim from concealment, and it is worth repeating because most coverage flattened it. The disclosure exists. It sits in footnotes rather than in the headline debt figure, which is a visibility problem rather than a reporting failure.

Watching it move inside one company

The aggregate figure is abstract. One company's own filings make it concrete.

Alphabet disclosed in its second quarter 2025 filing that it had entered leases primarily related to data centres that had not yet commenced. Future lease payments came to $23.9 billion, not recorded on its consolidated balance sheet. It stated those leases would commence between 2025 and 2031. Noncancelable terms ran from one to twenty-five years.

One quarter later, that figure had risen to $42.6 billion, including a purchase option considered reasonably certain to be exercised.

That is one company across three months. Nothing about the disclosure was improper and nothing about it was prominent.

When it arrives, and what happens

These obligations do not disappear. They are recognised as each lease commences, which produces three mechanical effects over the coming years.

Three effects, all automatic

Reported adjusted debt rises without any new borrowing. A company that issues no bonds and signs no new agreements will still show higher leverage. The previously signed leases simply begin.

Lease-related cash outflows increase. Rent begins being paid. That reduces free cash flow at exactly the point in the cycle when capital expenditure is also elevated.

And credit metrics move. Moody's stated it foresees a material increase in adjusted debt and lease-related cash outflows for these companies in coming years. It expects this to be mitigated by earnings growth, while acknowledging significant uncertainty about the growth and profitability of the AI market.

The clause that matters

That last clause is the whole question. The obligations are certain. The earnings expected to absorb them are not, which is the subject of the analysis of where AI returns have actually been measured.

How big this is in context

Against the other numbers in this cycle

Moody's also projected that combined capital expenditure across six major companies would reach $785 billion in 2026 and approach $1 trillion in 2027. The lease commitments sit alongside that, not inside it.

Both figures grew between the February analysis and the mid-year updates. That tells you lease financing is scaling alongside the build rather than substituting for it.

The comparison worth holding is the second bar against the fourth. Recorded commitments sit at $307 billion. Unrecorded commitments sit at more than twice that. An analyst modelling these companies from the balance sheet alone is working with the smaller of the two numbers and has no indication in the primary statements that a larger one exists.

Why lease rather than buy

A reasonable question sits underneath all of this. These are among the most cash-generative companies on earth. Why lease capacity at all rather than build and own it?

Three reasons, and they are commercial rather than accounting-driven. Speed is the first. A specialist developer already holds land, power agreements and permits. That delivers capacity years faster than an in-house programme can.

Flexibility is the second. Leasing preserves the option to walk away at the end of a term. A region may turn out to be the wrong place, or a hardware generation may make a facility uneconomic. Ownership removes that option.

Capital allocation is the third. Money not tied up in property is available for accelerators, which are the scarce input. Given the choice between owning a building and owning the chips inside it, operators have chosen the chips.

The accounting treatment is a consequence of those decisions rather than the motive for them. That distinction matters, because it separates a visibility problem from an allegation of engineering, and only the first is supported by the evidence.

The case that this is a non-story

Three arguments, and they carry real weight.

The disclosure exists. Every figure quoted in this post comes from company filings or a ratings agency reading them. Nothing was uncovered and nothing was leaked. Anyone willing to read a footnote could have found it.

The companies remain highly creditworthy. Moody's is not arguing these businesses are overleveraged today. Amazon, Alphabet, Meta, Microsoft and Oracle continue to hold some of the strongest investment-grade credit profiles in the market.

And commitments are not the same as risk. A lease obligation matched by contracted revenue from customers using that capacity is an ordinary arrangement. The question is whether the revenue arrives, not whether the lease exists.

The honest summary is that this is an earnings and leverage visibility issue rather than a solvency one. It matters because it changes how these balance sheets should be read, not because it suggests anyone cannot pay.

What to watch

The first row is the practical one. It updates quarterly, it requires no subscription, and it moved 78% in a single quarter at one company without generating a single headline.

The fourth row is the slower signal and the more interesting one. If operators begin structuring shorter leases, or shifting toward ownership, that would indicate management has revised its own view of how long this capacity earns. Structure changes before commentary does.

One caution on all four. None of these signals has a threshold. They tell you the direction of travel and they do not tell you when anything happens, which is a limitation of reading disclosure rather than a flaw in the disclosure itself.

Frequently asked questions

Where to start this week

One footnote, twenty minutes.

Open the most recent quarterly filing of any large cloud provider and search for the lease commitments note. Find the figure for leases not yet commenced. Then open the filing from four quarters earlier and find the same number.

Whatever the rate of change is, you now know something about that company's forward obligations that its headline debt figure does not tell you. It is public, it is audited, and almost nobody reads it.

Repeat it once a quarter. Four data points will tell you whether the lease programme is still accelerating, and that is a better read on management's own confidence than any earnings call commentary will give you.

Related reading