The $250B Footnote: Off-Balance-Sheet Compute Commitments

In Microsoft's Form 10-K for the fiscal year ended 30 June 2025, one sentence in the notes discloses leases "primarily for datacenters, that had not yet commenced of $92.7 billion," with terms running to 2031.¹ The commitments table adds $109,953 million of purchase commitments and $32,149 million of construction commitments.¹ None of these three figures is recorded as a liability on the balance sheet. By the company's next quarterly report, for the period ended 31 December 2025, the leases-not-yet-commenced line alone had grown to $155.1 billion.² Taken together, Microsoft had by the end of 2025 on the order of a quarter of a trillion dollars of contracted future spending that the balance sheet does not carry.

This is not where a reader expects to find a company's largest forward obligations. Microsoft's balance sheet showed long-term debt of $40,152 million at 30 June 2025.¹ The contracted commitments disclosed in the footnotes are several times that figure. The comparison has to be made carefully: the commitments are a cumulative, multi-year stream of future outflows, while the debt is a single point-in-time balance, so the two are not like for like, and Microsoft generates more than enough operating cash flow to service what it has signed. The point that survives the caveat is this. The recognized debt is what standard leverage ratios measure, and the contracted forward spending that those ratios do not capture is now much larger than the debt that they do. The thesis of this piece is that a large and rising share of the AI buildout's forward commitment sits in exactly this position, contracted and economically real but unrecognized as a liability, and that the same contract which reads as a celebrated backlog on the seller's disclosures is an undisclosed forward obligation when the buyer is private. The layer examined here sits between private project debt and issued bonds: the contractual commitments that are economically debt-like but never recognized as liabilities.

What the Footnote Says

Start with the anchor company, because Microsoft files the cleanest version of the disclosure. The $92.7 billion of leases not yet commenced reported at 30 June 2025 is, in the company's own words, leases "primarily for datacenters" that "will commence between fiscal year 2026 and fiscal year 2031 with lease terms of 1 year to 20 years."¹ A signed lease does not produce a right-of-use asset or a lease liability until it commences, which is when the lessor makes the asset available for use. Until then, however binding the signature, the obligation is disclosed in a footnote and recognized nowhere. By the quarter ended 31 December 2025 that line had grown to $155.1 billion, a rise of more than half in six months, the whole of it outside the balance sheet throughout.²

A lease that has not commenced is not pure leverage, and the piece should say so. When it commences it brings a right-of-use asset onto the balance sheet alongside the liability, so it is not equivalent to unsecured borrowing. What makes it matter is the other property: it is a fixed, multi-year payment stream the company is already committed to, and the company itself states the terms are largely non-cancelable. The concern is the rigidity and the timing of the cash calls, not a claim of hidden insolvency.

Alongside the leases, the FY2025 contractual-obligations table reports $109,953 million of purchase commitments and $32,149 million of construction commitments.¹ These are not borrowings; they are promises to pay for goods, services, and facilities not yet received. Add the leases not yet commenced to the purchase and construction commitments and the off-balance-sheet total is roughly $235 billion at 30 June 2025, crossing $250 billion in the second half of the year as the lease line alone reaches $155.1 billion. The figure in the title is not a single printed number. It is what the footnotes sum to once the reader adds the obligations the balance sheet leaves out, and it is worth stating plainly that this is a reconstruction from several footnote lines rather than one disclosed total.

The Same Number, Read From the Other Side

Remaining performance obligations, or RPO, are a seller's disclosure: the aggregate price of contracts a company has signed but not yet fulfilled, reported under the revenue-recognition standard as a measure of booked future business. When a cloud provider's RPO rises, the market reads it as a backlog, a runway of contracted revenue, and treats it as a positive signal.

Oracle is the clearest case. On 10 December 2025 the company reported remaining performance obligations of $523 billion, up 438 percent year over year, with the chief financial officer attributing the jump to "new commitments from Meta, NVIDIA, and others."³ Three months later, on 10 March 2026, the figure was $553 billion, and Oracle stated that "most of the increase in RPO in Q3 related to large scale AI contracts."⁴ CoreWeave, the GPU-cloud operator examined in the first piece of this series, reported remaining performance obligations of $60.7 billion at 31 December 2025, up from $15.1 billion a year earlier, against fiscal-2025 revenue of $5.1 billion and a net loss of $1.2 billion.⁵ In each case the backlog is presented, correctly, as evidence of demand.

The inversion is to read the same contract from the buyer's side, and it has one clean form and one qualified form. The clean form is the private buyer. CoreWeave's 10-K discloses that OpenAI "has committed to pay us up to approximately $6.5 billion through May 31, 2031."⁵ On CoreWeave's disclosures that commitment is part of a $60.7 billion backlog presented as an asset; on OpenAI's side it is a multi-year, largely fixed payment obligation that appears nowhere at all, because OpenAI is a private company that files no public statements. The larger reported arrangement between OpenAI and Oracle, a compute purchase of roughly $300 billion over about five years effective 2027, first reported by the Wall Street Journal and not confirmed in itemized form by either company, is the same structure at far greater scale: the seller books the demand inside its RPO and the private buyer's obligation is invisible.⁶

The qualified form is the public buyer, and here the word "hidden" has to be dropped. When the buyer is itself a public company, the obligation is disclosed, in its own purchase-commitments footnote, just not on the balance sheet and not in headline leverage. So the obligation is not concealed; it is recorded as a recognized liability by neither side and shown in full by neither side's primary statements. That is a real asymmetry, but a narrower one than "nobody can see it."

A second qualification is required, because the argument fails without it. A seller's RPO and a buyer's purchase-obligation footnote describe the same underlying contract from opposite sides, but they are not the same accounting object and they do not reconcile dollar for dollar. RPO is measured under the revenue standard, with its own treatment of enforceability, variable consideration, and contracts shorter than a year; the buyer's disclosure follows a different rule for unconditional purchase obligations.⁷ Either side may include amounts the other excludes, for instance where a customer prepays or supplies its own hardware. Oracle in fact says the latter is happening: most of its Q3 RPO increase, it states, relates to contracts where "Oracle does not expect to have to raise any incremental funds" because the equipment "is either funded upfront via customer prepayments" or "the customer buys the GPUs and supplies them to Oracle."⁴ That detail cuts against treating Oracle's RPO as a proxy for any single counterparty's hidden liability, and it is a reason the inversion is an argument about economic substance, that a binding contract is a liability-like obligation for the buyer and a revenue runway for the seller while being recognized as debt by neither, rather than a claim that the two figures mirror each other to the dollar.

Why It Stays Off the Balance Sheet

None of this is a loophole peculiar to AI, and the piece is weaker if it pretends otherwise. The accounting is orthodox. A liability is recognized when a past event creates a present obligation. A signed promise to buy compute in the future is an executory contract: neither side has performed, the seller has not delivered the capacity and the buyer has not paid, so the contract is disclosed rather than recognized.⁷ The lease version works the same way: under the lease standard a lessee recognizes a right-of-use asset and a lease liability only at commencement, so a signed but not-yet-commenced lease sits in a footnote however large it is.⁸

The disclosure regime was built for exactly this kind of obligation. The standard requiring firms to disclose unconditional purchase obligations was written in 1981, aimed at take-or-pay and throughput contracts, where a buyer owes a fixed stream of payments whether or not it uses the capacity.⁷ That is the category these AI commitments are disclosed under, and the category is rigid by definition. A caveat belongs here, though, because it is where the bear case overreaches. The specific cancellation terms of the largest compute purchase agreements, the OpenAI-Oracle arrangement above all, are mostly not disclosed, so their rigidity cannot be confirmed from the filings; what can be confirmed is that the leases not yet commenced are stated by the companies to carry non-cancelable terms, and that purchase obligations as a disclosure category are the fixed-or-minimum kind. The honest position is that the lease commitments are demonstrably rigid and the headline compute-purchase commitments are reported as firm but with terms the reader cannot fully check.

Two consequences follow. First, because none of these commitments is a recognized liability, they do not enter the leverage ratios in their reported form. Net debt to EBITDA and debt to equity key off recognized, interest-bearing debt, so a company can carry hundreds of billions of contracted future outflows and still post a low reported leverage ratio.⁹ This is true of the headline ratios; it is less true of the rating agencies, which already make adjustments for leases and, to varying degrees, for other fixed commitments, so the obligations are not invisible to a careful credit analyst even if they are absent from the screen-level number. Second, the obligations have become harder to assemble. The standardized contractual-obligations table that once forced purchase obligations and other commitments into a single comparable grid was eliminated in the 2020 modernization of the disclosure rules; companies now describe their "material cash requirements" in narrative form, scattered across the commitments footnote and the management discussion.¹⁰ The information is still there. It is no longer laid out so a reader can compare one company to the next at a glance.

Not Just Microsoft

Microsoft is the cleanest filer, not an outlier, though the figures across companies are not strictly comparable and should not be summed as if they were one line item.

Amazon's FY2025 10-K reports, in its commitments note, $96,373 million of leases not yet commenced, $84,772 million of unconditional purchase obligations described as "not reflected on the consolidated balance sheets," and $18,868 million of other commitments, against recognized long-term debt of $65,648 million.¹¹ Alphabet reports $149.1 billion of purchase commitments and other contractual obligations, but the composition matters: the figure spans "energy take-or-pay contracts, licenses ... and technical infrastructure and inventory orders," and $113.0 billion of it is short-term, which is closer to ordinary near-term procurement than to multi-year rigid commitment.¹² The cleaner data-center number in Alphabet's filing is its leases not yet commenced, roughly $58.5 billion, which the company states carry "non-cancelable lease terms primarily between one and 25 years."¹² Meta reports $131.05 billion of contractual commitments, "mostly related to third-party cloud capacity arrangements and our continued investments in servers and network infrastructure, data centers," with $30.63 billion due in 2026, plus approximately $103.77 billion of leases not yet commenced, against long-term debt of $58.74 billion.¹³

The pattern is consistent without the figures being identical. Each of these companies discloses data-center and cloud-capacity commitments, recorded as footnote items and as leases not yet commenced, that are large relative to the debt their balance sheets carry. None of the four breaks out a discrete "AI compute" line, so the honest reading is that the AI buildout is the stated driver of the data-center and cloud-capacity commitments, not that every dollar of disclosed purchase obligation is AI compute. Meta is the most explicit about the driver, tying its commitments to "third-party cloud capacity arrangements" and data centers.

The Counter-Case

Three objections to this framing deserve a hearing, because parts of all three are correct.

One reading says that footnoted purchase obligations are ordinary. Every large company carries them: airlines disclose aircraft orders, retailers disclose inventory and lease commitments, energy firms disclose take-or-pay pipelines, and the treatment is decades old.⁷ That is true. What it does not address is magnitude and concentration: the AI case differs in degree rather than in accounting kind, and the degree is large enough that the disclosed commitments now exceed recognized debt by several times.

A more substantive objection is that the commitments are funded operationally and matched to productive assets. The Federal Reserve Bank of Dallas, in a note dated 10 February 2026, observed that a large share of the buildout, on the order of $500 billion to $600 billion since 2023, "appears to have been internally funded by hyperscalers from retained earnings," and that the underlying assets have economics that "do not vary with interest rates over the course of a business cycle."¹⁴ These are cash-generative firms buying capacity against real demand, not speculative borrowers. The same note, however, records that the firms "have begun turning more recently to public and private debt markets," with issuance "large and long in duration," alongside "structured off-balance-sheet borrowing from private lenders," so even the benign reading documents the shift toward debt and off-balance-sheet structure.¹⁴

The strongest version of the bull case points at the credit market itself: spreads on AI-related issuers do not price distress, which is some evidence that sophisticated lenders regard the obligations as manageable.¹⁵ That is a fair point about the present, and the kind of signal that lags. The countervailing fact is that some of the borrowing is being routed off the balance sheet on purpose. Reporting in early February 2026 described a roughly $30 billion Meta data-center financing in Louisiana in which the debt sits with a separate vehicle rather than on Meta's balance sheet, one of a growing set of special-purpose structures keeping the borrowing off the parent.¹⁶ When obligations are deliberately placed where the reported leverage ratios do not reach them, the absence of alarm in those ratios carries less information.

What to Watch

The honest framing of the current position is that the accounting is correct and the commitments are visible to anyone who reads the footnotes; what is understated is the reported leverage picture, which leaves these commitments out by construction, and what is genuinely unknown is the cancellability of the largest compute deals. Three things, with concrete markers, will indicate how the condition evolves.

The first is the gap between leases not yet commenced and recognized lease liabilities. As signed leases commence, they move from footnote disclosure onto the balance sheet as matched assets and liabilities, so recognized leverage can rise sharply without a single new contract being signed. Watch whether the leases-not-yet-commenced line keeps growing, which signals fresh forward commitment, or falls as a wave of commencements lands on the balance sheet.

The second is the conversion of seller RPO into revenue. Oracle's $553 billion and CoreWeave's $60.7 billion are backlogs that must be delivered and paid for on schedule. Watch the ratio of RPO to realized revenue, and whether the backlog keeps growing faster than the revenue behind it, which is the same warning the second piece in this series drew from the gap between committed capacity and end demand.

The third is disclosure of the terms themselves. The off-balance-sheet commitment is benign in proportion to how cancellable it is and how reliably the buyer can pay. Watch for any disclosure of the cancellation provisions on the largest compute commitments, for movement of these obligations into dedicated off-balance-sheet vehicles, and for the first impairment or onerous-contract charge taken against a compute commitment, which would be the point at which a footnote obligation finally reached the income statement.

Personal View

The body above reads the footnotes. What follows is what I make of them.

What strikes me about this layer is not its size but how little of the thing that actually matters is disclosed. The variable that decides whether these commitments are benign or not is their cancellability, and for the largest compute deals that is precisely the number nobody publishes. Oracle reports a $553 billion backlog; it does not report what happens to that backlog if a counterparty wants out, or cannot pay. Until those terms are visible, my view is that anyone claiming certainty here, bull or bear, is reading more into the footnotes than the footnotes contain. The most important figure in this story is the one we do not have.

The concern I will state more firmly is about where these obligations land. For Microsoft, Amazon, Alphabet and Meta, a quarter of a trillion dollars of contracted forward spending is real but absorbable; these are among the most cash-generative companies in the world, and when a lease commences or a commitment converts, they can carry it. What I find harder to dismiss is that the same rigid, footnote-only obligations sit, on the other side of these contracts, with the counterparties least able to demonstrate they can pay: the private model labs and the neoclouds examined earlier in this series. That is exactly where the obligation is most invisible, because a private buyer files nothing, and least cushioned, because the revenue behind it is real but unprofitable. The risk I take seriously is not that a hyperscaler is quietly overlevered. It is that the weakest node in the chain has signed the most rigid promises, and that the structure is built so we would read about it in a footnote, if at all, only after it mattered.

In my view that is the right place to keep watching, and the right amount of caution to hold while watching it, no more and no less.

This is part 3 of a four-part series on AI infrastructure financing. Part 1 examined the SPV layer where the clusters are funded; Part 2 traced the vendor-financing loop that drives the demand these footnotes capture. Part 4 takes up the on-balance-sheet credit picture the footnotes ultimately feed into.

END · 03