How Meta Keeps $39B of Debt Off Its Balance Sheet

Meta is on track to spend between $130B and $145B on AI infrastructure in 2026 alone. Yet a reader scanning the face of its balance sheet would find no trace of the $27.3B Louisiana bond, nor the $12.3B El Paso bond, listed as a Meta liability. The debt exists. It just lives somewhere else.

That gap is not a rounding error. With roughly $278.99B in uncommenced lease obligations and a $28B residual value guarantee sitting in the footnotes of Meta’s SEC filings rather than on the balance sheet itself, the question of what Meta actually owes has become anything but academic. This is the space that sophisticated credit analysts and law firms are now examining line by line.

After reading this, you will know how the accounting actually works, whether the “concealment” framing survives scrutiny, and what the practical risk implications are for anyone trying to gauge Meta’s true leverage. Treat it as a tool for reading company filings more critically, not just a Meta story.

The two deals that put $39 billion of debt somewhere else

The pattern is easier to see once you look at both structures side by side. Meta’s Louisiana campus, code-named Hyperion, sits inside a special purpose vehicle called Beignet Investor, LLC. Its El Paso project sits inside Sopaipilla Investor, LLC. Both use the same ownership split and the same functional logic.

In Beignet, Blue Owl holds 80% and Meta holds 20%. In Sopaipilla, BlackRock-managed funds hold 80% and Meta again holds 20%. That 20% figure is not incidental, and it is the reason both structures behave the way they do.

Legally, Meta is a minority equity holder in a bankruptcy-remote vehicle it does not control. Economically, Meta is the sole revenue source for both, funding each entity entirely through long-term lease commitments. The legal form and the economic substance point in opposite directions.

The Beignet bond, described as the largest corporate bond in history against a Hyperion development cost exceeding $50B, carries terms worth reading closely:

• 6.581% fixed coupon, quarterly pay
• Priced at approximately 225 bps over Treasuries
• Rated A+ by S&P Global, the only agency to rate it
• Morgan Stanley as sole bookrunner
• PIMCO as anchor investor, with BlackRock and Prudential also participating

The Sopaipilla notes launched on 23-24 July 2026, arranged by JPMorgan and Morgan Stanley. Combined, the two structures carry roughly $39.6B in SPV debt. None of it appears as a Meta liability, and the 20% stake is the accounting hinge that keeps it there.

How the accounting rules make this possible

The gap you just saw is not a trick played around the rules. It is a product of the rules working exactly as written. Three mechanics do the work.

The equity-method treatment

Because Meta holds sub-50% stakes in both SPVs, it uses equity-method accounting. That means it records a single line item representing its ownership interest rather than consolidating the full assets and liabilities of each vehicle.

This is the direct mechanism that keeps $39.6B off the balance sheet. Consolidation would drag every dollar of SPV debt onto Meta’s books. A 20% equity stake does not.

Why “not probable” carries so much weight

Under US Generally Accepted Accounting Principles (GAAP), a guarantee becomes a recorded liability only when payment is judged probable. Meta’s position, stated in both its 2025 10-K and its Q2 2026 10-Q, is that “payments are not probable and therefore no liability has been recorded” against its $28B residual value guarantee.

That single word, “probable”, is a GAAP threshold, not an editorial preference. If AI infrastructure values were to fall significantly, that probability assessment could change, and a large liability could appear on Meta’s balance sheet with little warning for investors who were not reading the footnotes.

ASC 842 and the uncommenced lease cliff

Under lease accounting standard ASC 842, a lease liability is recognised only when the lease commences. Meta’s Hyperion rent begins 1 June 2029 and its El Paso lease begins in 2028, so those obligations have not yet commenced.

The FASB’s ASC 842 lease standard requires that a right-of-use asset and corresponding lease liability be recognised only at commencement, which is the precise mechanism allowing hundreds of billions in signed but unstarted obligations to remain outside the headline balance sheet figures that most screening tools consume.

That is why approximately $278.99B in uncommenced lease obligations as of 30 June 2026 sits in footnotes rather than on the face of the balance sheet, alongside the $131.05B in non-cancelable contractual commitments disclosed in the 2025 10-K. The figure is not hidden. It is simply excluded from the headline leverage ratios that most investors and stock screeners rely on.

Law firm Quinn Emanuel put the mechanism plainly in a client alert dated 13 March 2026.

Combining short initial lease terms with large residual value guarantees can cause reported liabilities to materially understate a company’s true obligations.

The takeaway is that Meta is operating within the rules, not around them. That is precisely what makes the gap between reported debt and total economic exposure so hard for standard analytical tools to catch, and it tells you where to look in any company’s filings, not just Meta’s.

The concealment debate: what the critics argue, and what the defenders say

This is where the interpretations split, and neither side is easy to dismiss.

The critics’ case

The core argument is functional. Because Meta is the sole revenue source for both SPVs, critics contend it is economically the borrower regardless of the legal wrapper, with the 20% stake serving as an accounting device rather than a genuine sharing of ownership.

The residual value guarantee sharpens the point. Quinn Emanuel flagged that in a downside scenario, RVGs and long-dated leases could transmit large losses back to Meta even though the SPV legally owns the assets. This is the “shadow leverage” concern: obligations that are economically debt-like yet excluded from debt/EBITDA metrics.

The defenders’ case

The counter-argument rests on three points. First, these obligations are disclosed in SEC filings and footnotes, so the information is public rather than concealed. Second, S&P evaluated the full structure when it assigned the A+ rating. Third, off-balance-sheet SPVs have become the default structure for large AI campuses across the sector, not an isolated Meta manoeuvre.

The scale of the pattern supports that last point. A Morgan Stanley analyst note reported on 25 August 2026 quantified it.

Across the five major hyperscalers, undiscounted commitments exceed $2.7T, roughly three years of combined operating cash flow.

Context matters on the numbers, too. Some media coverage has cited a $420B off-balance-sheet figure for Meta, but that number is not independently confirmed in Meta’s own filings; the verifiable components are roughly $278.99B in uncommenced leases and $28B in RVG commitments. Oracle, for comparison, carries only around $3.3B in comparable guarantee obligations, which shows the guarantee structure of primary concern is concentrated at Meta and Alphabet rather than spread evenly across all five.

The real question is not whether Meta is hiding debt. It is whether footnote disclosure is enough when the figures involved are large enough to change your view of the company’s risk profile.

What Meta’s capex trajectory reveals about its structural transformation

Step back from the accounting and a bigger shift comes into focus. Meta has quietly crossed from an asset-light software company into an infrastructure-heavy operator, and most investors have not recalibrated their mental model.

The capex numbers tell that story on their own. Full-year 2025 capital expenditure, including finance lease principal payments, came in at $72.22B. The 2026 guidance nearly doubles that, and the SPV structures exist partly because even this direct number is manageable only with off-balance-sheet augmentation.

The guidance climbed in steps through the year:

1. 28 January 2026: initial 2026 range of $115-135B
2. 29 April 2026: raised to $125-145B on higher component costs and added construction
3. 29 July 2026: lower end lifted to $130B, giving a current range of $130-145B

Reuters framed the top of that range as roughly double the prior year’s spend, almost entirely AI infrastructure.

The broader AI infrastructure investment thesis rests on more than $2.3 trillion in legally contracted, undelivered backlog across the four major hyperscalers, a figure that reframes the capex debate from speculative to demand-confirmed, while leaving the execution and leverage questions that Meta’s footnotes expose firmly unresolved.

The revenue mismatch problem

Here is the structural catch. That spending is being funded primarily by Meta’s legacy advertising revenue, not by AI-generated revenue, which has not yet scaled to match the commitments being made.

The obligations are long-duration and fixed. The revenue upside that is meant to justify them remains speculative. That asymmetry is the core risk in the whole arrangement.

When you adjust hyperscaler leverage ratios to include uncommenced leases and contractual commitments, they become broadly comparable to a traditional data centre operator like Equinix in its pre-AI era. If you are valuing Meta on software-company multiples and asset-light assumptions, the capex trajectory and adjusted leverage suggest you are pricing a materially different business than the one that now exists.

When you adjust hyperscaler leverage ratios to include uncommenced leases and contractual commitments, they become broadly comparable to a traditional data centre operator like Equinix in its pre-AI era, and the valuation implications of that adjusted leverage are explored in detail through a scenario-based framework that separates the Family of Apps earnings profile from the infrastructure obligations now sitting alongside it.

The downside scenario that footnotes are written to survive

It is worth sitting with the specific mechanism by which a footnote commitment becomes a balance-sheet liability, because the risk here is concrete and conditional rather than abstract.

The transmission runs through the RVG. If AI infrastructure values fall or demand disappoints, the “not probable” assessment on the $28B guarantee could be reassessed. Recognition of a large liability could then follow, with limited prior-period warning in the headline metrics investors watch.

The specific pressure points are worth naming:

• The RVG reassessment trigger, moving $28B from footnote to recorded liability
• The uncommenced lease cliff, as roughly $279B in obligations commences
• The $14.7B contingent purchase commitment treated as a financial guarantee
• The loss of pivot optionality that asset-light models once preserved

That last point matters most. In prior technology cycles, Meta could walk away from failed initiatives at low cost. Multi-decade leases, SPV guarantees, and hundreds of billions in uncommenced obligations substantially eliminate that flexibility, making the AI buildout an effectively irreversible strategic direction.

Quinn Emanuel described the flow of risk directly.

In a downside scenario, residual value guarantees and long-dated leases could transmit large losses back to Meta, even though the SPV legally owns the assets.

Remember what the A+ rating on the Hyperion bonds actually reflects: the creditworthiness of Meta as tenant, not a judgment that risk to Meta itself is low. The structure is bankruptcy-remote, so SPV creditors like PIMCO and BlackRock’s investors are protected. Meta’s lease and guarantee obligations are what remain exposed. The design protects the bondholders, not Meta equity, precisely because Meta was built to be the creditworthy anchor all along.

Reading Meta’s filings with the full picture in view

Pull the three layers together and a usable posture emerges. The deal mechanics show how the debt stays off the balance sheet. The accounting framework shows why that treatment is technically compliant. The risk asymmetry shows where the exposure actually sits if AI demand underperforms.

None of that requires investigative work, because the figures are already disclosed. The challenge is analytical, not investigative. When reading any hyperscaler’s filings, three questions do most of the work:

• What is the uncommenced lease obligation buried in the footnotes?
• What guarantees are disclosed but not recognised as liabilities?
• How does adjusted leverage, inclusive of those obligations, compare to a traditional infrastructure operator?

The Morgan Stanley figure of $2.7T in aggregate hyperscaler commitments means this framework applies well beyond Meta. Any of the five major players presents structurally similar footnote dynamics, and Meta’s adjusted position looks more like an infrastructure operator than a software business.

The practical read is not that Meta is fraudulent. It is that applying software-company shortcuts to a company that has structurally transformed will produce systematically wrong risk assessments. Investors who read the footnotes and adjust their models are working with a materially more complete picture than those relying on headline metrics.

For investors wanting to build the analytical habits that make footnote scrutiny routine across any company, our dedicated guide to reading earnings filings covers non-GAAP manipulation, cash flow divergence signals, and the metrics that quietly disappear from releases when underlying performance deteriorates.

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions.

Past performance does not guarantee future results, and the downside scenarios described here are speculative and subject to change based on market developments and company performance.