The $1.09T Hidden Lease Burden Behind AI's Data Center Boom

$1.09 trillion in future lease payments, signed and committed, and not a single dollar on the balance sheet yet.

That is what Microsoft, Meta, Oracle, Amazon and Alphabet disclosed in their latest filings, compiled by Reuters on August 4. The leases are for data centers that do not exist yet. The buildings are not built, the chips are not installed, and under standard accounting rules the obligations do not count as debt until the facilities open. So the largest single capital commitment in the history of corporate technology sits, for now, in the footnotes.

The contracts behind the number

A hyperscaler lease signed in 2024 does not hit the income statement as a liability. It appears only when the facility is ready for use. Until then, the terms live in the notes to the financial statements: the dollar amount, the start dates, the renewal options, the take-or-pay provisions.

Technically, this is not evasion. The accounting rule, broadly consistent across US GAAP and international standards, says a lease becomes a liability when the lessor makes the asset available. Before that, the company has a contractual commitment, disclosed, but not a balance-sheet obligation. Moody's reached this exact conclusion in its February analysis of the five hyperscalers: the treatment does not hide an existing liability, because the service has not been delivered.

What the treatment does do is push the true financial exposure somewhere an investor has to go looking for it.

What is growing

The pipeline is compounding, and it is compounding fast.

Oracle disclosed $248 billion in upcoming lease agreements at the end of November 2025. That was already a 148 percent increase from the end of August of that year. By late March 2026, the figure had climbed again, past $261 billion. Reuters now puts its uncommenced commitments at $260 billion, with the leases expected to begin between fiscal 2027 and fiscal 2029.

The four other hyperscalers moved in lockstep. Microsoft carried the largest disclosed pipeline at $329.1 billion, against $88.52 billion of recognized lease liabilities. Meta reported $278.99 billion, then added $68 billion in July. The Bank for International Settlements, in its March 2026 Quarterly Review, described the mechanics precisely: hyperscalers substitute upfront capital expenditure with multi-year operating expenses, financing the build-out through dedicated vehicles backed by private credit funds, and keep most of the associated debt off their own balance sheets.

What is falling

The share of AI data center obligations that actually shows up on balance sheets.

Recognized lease liabilities sit at roughly $285 billion across the five companies. Total known pipeline now runs past $1.09 trillion and, with Meta's July signings, toward $1.16 trillion. Recognized debt is the smaller and shrinking fraction of the total. Every new committed lease widens the gap.

The same pattern applies to reported leverage. Oracle's borrowings equaled about 4.4 times trailing earnings before interest, taxes, depreciation and amortization (EBITDA) at the end of May. Fold in the recognized operating and finance lease liabilities, and the ratio rises to about 5.7 times. Report the uncommenced leases, and the number goes higher still.

The new structures

This is not traditional bank financing. The Bank for International Settlements calls it shadow borrowing: obligations that are economically debt-like but sit outside corporate balance sheets.

The mechanics run through special purpose vehicles. A consortium of sponsors capitalizes a vehicle, which develops the data center. The vehicle raises debt through private placements. The hyperscaler takes a minority equity stake, signs a long-term operating lease or a capacity offtake agreement, and often guarantees part of the borrowing. The debt itself is held by private credit funds and insurers, sometimes with investment-grade features supported by the hyperscaler's guarantees.

Alphabet disclosed $28.4 billion in credit-derivative backstops at March 31, plus $9 billion in financial guarantees, and had agreed to up to $33.3 billion of further support, subject to final terms. Meta's Louisiana data center venture carries a different shape: it owns 20 percent, has an initial lease commitment of about $12.31 billion, and provided residual-value guarantees with an aggregate threshold around $28 billion. Oracle guaranteed up to $3.3 billion of one lessor's borrowing.

The consequence is a new linkage between hyperscalers and non-bank credit markets. Banks support these vehicles with funding lines. When the vehicles face refinancing pressure, or when private credit appetite turns, the shock travels straight back to the credit risks the market has already started to price: credit default swap spreads on the lower-rated hyperscalers have risen since early 2025, per the BIS.

Where the gaps concentrate

Concentration risk is not evenly distributed.

Oracle's position deserves scrutiny out of proportion to its market cap. Its $260 billion in uncommenced commitments run nearly seven times the recognized lease liabilities, and the leases extend 15 to 19 years, with starts spread across fiscal 2027 to 2029.

It has said so plainly in its own filings: the duration, renewal terms and pricing of its data center leases may not align with customer contracts, leaving the company exposed if customers do not renew or cannot perform. That is a candid disclosure of what could become the single largest corporate lease mismatch this decade.

S&P Global Ratings has already adjusted its view, incorporating the $260 billion into its adjusted-debt forecast and projecting leverage around 4.4 in fiscal 2027.

Why investors should watch the footnote discipline

Underestimate the gap, and the ratio analysis in a normal credit model misses the build-out's true financing need.

The scale of uncommenced commitments is staggering relative to history. Market-level data quantifies the step change: analysts at Axis Intelligence, citing Moody's Ratings, put the five hyperscalers' total undiscounted future data center lease commitments at $969 billion at the end of 2025, of which $662 billion sat off reported balance sheets, equal to 113 percent of the same companies' combined on-balance-sheet debt. Moody's observation that the treatment does not understate an existing liability is technically true and financially small comfort: the structures understate economic risk, which is exactly what Moody's said it may adjust for.

A second-order effect follows. As the lease obligations grow, rating agencies look at adjusted debt figures, not headline debt. The day they re-rate a hyperscaler on adjusted metrics, the reported financial ratios lose half their meaning, and the financing costs rise accordingly.

The commitments are not hidden. They are disclosed, in notes and footnotes, in the language of accounting. What changed this summer is that the scale stopped fitting in the footnotes. $1.09 trillion is a number that rearranges how the AI build-out gets financed, and a private investor evaluating a hyperscaler position now has to read the notes, not just the balance sheet.

The economics of the build-out have an explicit consumer, and that consumer is the AI revenue cycle itself. If token demand keeps compounding, the facilities underpin the next phase of cloud growth. If it stalls, companies keep paying for costly, long-lived capacity that is difficult to shed. That is the whole trade, in two sentences.

As we wrote in August, six asset managers are building a $500 billion market for Nvidia compute financing, and every one of those structures now sits atop the same lease economics this article describes. The financing layer grows; the underlying obligations grow with it.