Hyperscaler Off-Balance-Sheet Leases: $1.15 Trillion Not Yet on the Books

Microsoft, Meta, Amazon, Alphabet, Oracle and CoreWeave have signed $1.15 trillion of leases, mostly for data centers, that haven't started yet and don't count as debt. That's more than twice the $517.1 billion they owe, and triple what they had signed a year earlier. Microsoft's $329.1 billion is eight times its debt.

For the four hyperscalers alone, the total is $830.5 billion, 2.3 times their debt. Lease figures are the future payments each company discloses in the notes to its latest quarterly or annual filing, compared with the same filing a year earlier; totals use unrounded figures. Debt is bonds, bank loans and commercial paper, as in our hyperscaler debt tracker.

Signed-but-not-started leases tripled in a year

A year earlierLatest filing

100200300$400B0

$92.7B$329.1B

MicrosoftA year earlier$92.7BLatest filing$329.1B

$99.8B$288.0B

OracleA year earlier$99.8BLatest filing$288.0B

$52.6B$279.0B

MetaA year earlier$52.6BLatest filing$279.0B

$68.9B$137.2B

AmazonA year earlier$68.9BLatest filing$137.2B

$23.9B$85.2B

AlphabetA year earlier$23.9BLatest filing$85.2B

$34.0B$35.5B

CoreWeaveA year earlier$34.0BLatest filing$35.5B

MicrosoftOracleMetaAmazonAlphabetCoreWeave

What is a lease that hasn't started?

It's a signed contract to rent a building or equipment that isn't ready yet. Under US accounting rules a lease only goes on the balance sheet when the company gets the keys. Until then, the promise sits in a footnote. When the data center opens, the company records a lease liability, which works like debt, and an asset of the same size.

So these figures are, in effect, future debt. Moody's describes such commitments as "yet to be on the balance sheet" and may add them to its own measure of a company's debt, Fortune reported. Renting lets a hyperscaler add capacity without spending its own cash up front or borrowing in its own name. The developer, often backed by private credit, builds the data center; the tech company pays rent for 15 to 20 years.

Two caveats:

• The figures are total future payments, not today's value. A dollar of rent due in year 15 counts as a full dollar here. The lease liability booked when the building opens will be smaller, because it's discounted.
• They aren't capex. Rented capacity sits outside the capex guidance the market watches, so the true cost of the AI build is bigger than capex alone.

Which hyperscaler has the most off-balance-sheet debt?

Microsoft, by a wide margin. Its latest annual report lists $329.1 billion of leases, "primarily for datacenters", that haven't started, up from $92.7 billion a year earlier. The leases run 1 to 20 years and start over the next seven years. Microsoft adds that some of the deals still depend on "certain contractual conditions being met".

Microsoft borrows little: its debt is $40.3 billion. But it rents much of its capacity from data center developers, so its leases already on the balance sheet are $88.5 billion, more than twice its debt. The $329.1 billion still to come will add to that as each building opens.

Microsoft's rent promises dwarf its debt

Debt$40.3bn

Leases already on the balance sheet$88.5bn

Leases signed, not yet started$329.1bn

Leases on the balance sheet are the present value of operating and finance leases. Not-yet-started leases are undiscounted future payments, so the two aren't directly comparable.

How big are Oracle's lease commitments?

$288 billion, up from $99.8 billion a year earlier and from the $248 billion that drew headlines three quarters ago. Oracle's latest quarterly filing says the leases are "substantially all related to data center arrangements", start within about three years and run 15 to 19 years.

That number matters more for Oracle than for anyone else, because Oracle already owes $125.3 billion and spent $28.7 billion more than its operations brought in over the last 12 months. Spread evenly over 15 to 19 years, $288 billion is roughly $15 billion to $19 billion of rent a year once the buildings are open. Oracle's operating cash flow over the last 12 months was $46.9 billion. More on its build in Oracle's capex plans.

How fast are Meta's leases growing?

Faster than anyone's. Meta's not-yet-started leases rose 5.3 times in a year, from $52.6 billion to $279.0 billion. The month after the quarter ended, it signed another $68 billion of data center leases starting over the next two years, which aren't in the table. Meta also guarantees the value of buildings in joint ventures it doesn't own outright. The details are in Meta's $279 billion of off-balance-sheet leases.

Amazon, Alphabet and CoreWeave

Amazon is the only one of the four that sets out when the payments fall due. Of its $137.2 billion, about $93.5 billion is due more than four years out, so the bill builds slowly. Amazon also has the most leases already on its books: $109.8 billion of lease liabilities, 83% of its $132.5 billion of debt.

Alphabet owns more of its data centers, so its leases are smaller: $85.2 billion not yet started, less than its $100.2 billion of debt, plus a separate $5.8 billion short-term lease signed in the last month of the quarter. It is still growing fast, 3.6 times in a year.

CoreWeave barely moved, from $34.0 billion to $35.5 billion. That isn't a slowdown. Its leases are starting: lease liabilities on its balance sheet rose from $3.4 billion to $16.5 billion over the year, as signed sites turned into running data centers. The same shift will reach the others as their buildings open. CoreWeave's own bonds now yield over 15%.

Can they afford the rent?

One way to size it: how many years of operating cash flow would cover the leases not yet started? For Alphabet, less than six months. For Oracle and CoreWeave, more than five years.

Oracle's unstarted leases equal six years of its operating cash flow

Oracle6.1 years

CoreWeave5.1 years

Meta2.1 years

Microsoft1.8 years

Amazon0.9 years

Alphabet0.5 years

Leases signed but not yet started divided by cash from operations over the last 12 months.

The rent isn't due at once; it's spread over 15 to 20 years. For the four hyperscalers, the annual cost looks manageable against cash flow today. For Oracle and CoreWeave, it stacks on top of debt they're already struggling to finance.

What it means for shareholders

Three things change when these leases start.

1. Debt-like liabilities rise. Each new building adds a lease liability. Net cash, the cushion the debt tracker measures, will look worse even if no new bonds are sold.
2. Costs hit earnings. Rent (or, for finance leases, depreciation and interest) shows up in operating expenses, on top of the depreciation from capex. The depreciation lag gets longer.
3. The commitment can't be cut quickly. Capex can be slowed next quarter. A signed 18-year lease can't. If AI demand cools, the rent is still due.

Our read

The capex numbers understate the AI build. Microsoft and Meta have each signed rent promises bigger than a year of capex (Microsoft spent $115.9 billion over the last 12 months, Meta $89.3 billion), and they sit in a footnote, not in the debt line. For the four hyperscalers, the cash flow is there to pay, so this is a quality-of-earnings issue more than a solvency one: margins will carry these costs for two decades.

Oracle is different. Its unstarted leases are worth six years of operating cash flow, on top of $125.3 billion of debt and negative free cash flow. That's where a demand slowdown would hurt first. Watch whether each quarter's not-yet-started figure keeps climbing, and whether lease liabilities start jumping as buildings open.

Frequently asked questions

Do hyperscalers have off-balance-sheet debt?

Not debt in the legal sense, but they have signed leases that will become debt-like liabilities when the buildings open: $830.5 billion across Microsoft, Meta, Amazon and Alphabet, and $1.15 trillion including Oracle and CoreWeave.

Why does Microsoft rent so many data centers?

Renting from developers gets capacity online faster and spreads the cost over the lease term instead of paying up front. Microsoft's leases already on its balance sheet are more than twice its debt.

What is Oracle's $248 billion lease figure?

It was the not-yet-started lease total in an earlier Oracle quarterly filing, three quarters before its latest one. The figure has since grown to $288 billion.

Are these leases included in capex guidance?

No. Capex covers what a company buys. Leased data centers show up later as lease liabilities and rent, outside reported capex.

Limits of this data

• Lease figures are undiscounted totals; the liabilities eventually booked will be lower.
• Microsoft says some of its leases depend on conditions still to be met; others may also be renegotiated or cancelled.
• The filing dates differ: Oracle's latest quarter ends two months after the others.
• Meta's $68 billion signed after the quarter and Alphabet's $5.8 billion short-term lease are excluded from the totals.
• Purchase commitments for chips, servers and cloud capacity are separate and not included.

How we built this

Leases not yet commenced come from the lease or commitments note in each company's filing with the SEC: Microsoft's Form 10-K for the year ended June 30, 2026 and June 30, 2025; Alphabet's 10-Qs for June 30, 2026 and June 30, 2025; Amazon's for June 30, 2026 and June 30, 2025 (the "Leases not yet commenced" row of its commitments table); Meta's for June 30, 2026 and June 30, 2025; Oracle's for August 31, 2026 and August 31, 2025; and CoreWeave's for June 30, 2026 and June 30, 2025. Oracle's $248 billion figure is from its 10-Q for November 30, 2025, as reported at the time. Lease liabilities on the balance sheet are the present value of operating and finance leases (operating only for Meta). Debt and 12-month operating cash flow are the figures in our hyperscaler debt tracker. We update this page after each quarterly report.

Go deeper with member research

• Issue 03: The AI Depreciation Clock: when capex, construction in progress and these unstarted leases reach the income statement, company by company.
• Issue 05: NVIDIA Is Building the Credit Market for AI: the private credit and lease structures that fund the data centers behind these numbers.