Off the Books: How AI's Build-Out Quietly Became a Financing Risk

A weak signal in artificial intelligence: hyperscalers carry strong balance sheets, but the AI build-out is increasingly funded through special-purpose vehicles, private credit and residual-value-guaranteed leases GAAP does not capture, and four of the world's major regulators began naming the concentration in 2026.

The consensus is that hyperscalers are paying for the AI build-out from cash: record capex, strong balance sheets, no obvious leverage problem. The 2026 evidence inverts that. A growing share has migrated off balance sheet into special-purpose vehicles, private credit and residual-value-guaranteed leases that published debt figures do not reflect. The IMF, FSB, ECB and Bank of England all began flagging this opaque leverage layer between February and May 2026. A disappointing AI-revenue print would no longer be only a tech-equity correction; it would land on insurers, pension funds and a private-credit book several times the visible bank exposure.

Signal Identification

This is a structural shift in how capital reaches AI infrastructure, not a slowdown in the technology. Leases not yet commenced, SPV-issued private-credit notes and residual-value guarantees sit outside GAAP debt yet behave like leverage, and the holders are no longer mostly banks.

What's Changing

The off-balance-sheet layer is now quantifiable. Moody's Ratings counts $969 billion in total undiscounted future data-centre lease commitments at the top five US hyperscalers at end-2025, of which $662 billion is for leases not yet commenced and so off balance sheet under GAAP, equivalent to 113% of the five firms' adjusted debt. AI hardware has a useful life of just four to six years, pushing firms toward shorter leases backstopped by residual-value guarantees recorded as no liability, including Meta's $28 billion guarantee (Fortune, February 2026).

The funding mix is shifting in parallel. The FSB estimates global private credit at $1.5-2.0 trillion at end-2024, concentrated in technology, with AI firms turning to private lenders to fund data centres; direct bank exposures of around $220 billion of drawn and undrawn credit lines understate the linkage and commercial estimates run to $270-500 billion (Financial Stability Board, May 2026). The IMF reaches the same point, flagging circular financing along the AI value chain (International Monetary Fund, April 2026).

The ECB is explicit about who absorbs the loss. It estimates US private credit at about $1.4 trillion at end-2024, comparable to the $1.5 trillion US subprime market, and modelled a stress scenario in which an AI-revenue disappointment cascades through private-credit holdings: pension funds lose 5-6% of total assets, insurers around 4%, with bank losses contained at no more than 1.3% of equity (European Central Bank, May 2026). The Bank of England's FPC identified vulnerabilities in risky credit markets including private credit as worsening, and asked for further work on AI risks (Bank of England, April 2026).

Disruption Pathway

The pathway runs in three stages. First, financing migration: capex outruns operating cash flow, so the marginal dollar is raised through SPV-issued private-credit notes, lease commitments not yet commenced, and residual-value guarantees on short-life equipment. The IMF's circular-financing language and Moody's $662 billion off-balance-sheet count are the same phenomenon viewed differently. Second, holder transformation: that financing accumulates in private-credit funds whose LPs are mostly insurers and pension funds, with banks now the warehouse rather than the holder. Third, transmission: an AI-revenue disappointment or shortened hardware life impairs the lease and the loan that funded it; the loss lands first on the private-credit holder, then on its insurer or pension-fund LP.

Stress concentrates at three points: technology-cluster concentration (a sector shock is not diversified away), valuation opacity (private ratings and infrequent mark-to-market delay loss recognition), and hardware obsolescence (a model shift stranding a GPU vintage calls the guarantee).

Why This Matters

For boards, investors and insurance and pension trustees, the assumption that needs revising is that AI capex is funded from hyperscaler operating cash flow on lightly leveraged balance sheets. On the evidence, a material share is off the visible balance sheet and sits in private-credit holdings larger than the visible bank exposure. That changes the loss-allocation tree if AI revenue disappoints: insurers and pension funds, not banks, take the first hit, and four authorities have said so in their 2026 outputs.

Decision-action posture for this signal: Prepare. The structure is documented and four authorities are monitoring, but no trigger event has yet materialised; institutions should map their indirect AI-infrastructure exposure through private-credit allocations and lease counterparties before the next cycle, while regulators continue to gather data.

Implications

The sources point to a durable rewiring of who finances and who absorbs AI-infrastructure risk. The inflection window is 2026-2029, defined by whether exposures keep growing before regulators tighten disclosure and capital treatment. Winners map indirect AI exposure across private-credit and insurance allocations and price residual-value and lease-tail risk into counterparty diligence. Losers treat hyperscaler reported debt as the full picture and discover the leverage layer only when a revenue miss prompts the first markdown.

Early Indicators to Monitor

• An IMF, FSB, ECB or BoE follow-up proposing disclosure or capital treatment for AI-linked data-centre leases or SPV-issued private credit.
• A hyperscaler restatement bringing off-balance-sheet lease commitments into reported debt, or an equivalent rating-agency adjustment.
• A private-credit fund with AI-infrastructure exposure marks down its book, or a life insurer or pension fund discloses an AI-linked loss.
• A residual-value guarantee called on a stranded GPU vintage.
• FASB or IASB opens a project on lease commitments not yet commenced or technology residual-value guarantees.

Disconfirming Signals

• Hyperscaler AI revenue prints above consensus across two consecutive quarters.
• Private-credit growth in technology slows, with origination shifting back to bank balance sheets.
• The IMF, FSB, ECB or BoE downgrades AI-linked private credit to routine monitoring.
• Hardware useful-life assumptions extend rather than shorten.
• Insurers and pension funds cut private-credit allocations, reducing the loss-absorption layer.

Strategic Questions

• Should boards ask for a memo on indirect AI-infrastructure exposure through private-credit and lease counterparties?
• At what point do regulators move from naming the off-balance-sheet layer to requiring its consolidation?
• Do insurer and pension trustees treat the ECB scenario as a planning input or wait for a trigger?

Keywords

AI infrastructure financing; hyperscaler capex; data centre leases; off balance sheet; residual value guarantees; private credit; circular financing; financial stability; insurer exposure; pension fund exposure; GAAP; IMF GFSR; FSB; ECB FSR; Bank of England FPC

Bibliography

Source tiers: Tier 1, regulators and intergovernmental bodies. Tier 2, think-tanks, consultancies and rating agencies. Tier 3, quality journalism. Tier 4, vendor and practitioner sources.

• Tier 1 Global Financial Stability Report briefing. International Monetary Fund. April 2026.
• Tier 1 Private credit vulnerabilities. Financial Stability Board. May 2026.
• Tier 1 Financial Policy Committee Record. Bank of England. April 2026.
• Tier 1 Financial Stability Review. European Central Bank. May 2026.
• Tier 2 Moody's flags $662 billion data-centre lease risk. Fortune (reporting Moody's Ratings). February 2026.
• Tier 3 Tokenmaxxing is over. Fortune. May 2026.
• Tier 3 Starbucks retired its AI agent. Fortune. May 2026.

Analyst inferences and editorial framing

Claim-fidelity self-disclosure. The $969 billion, $662 billion, 113%, four-to-six-year and $28 billion figures are faithful summary of Moody's Ratings via Fortune (February 2026); the chart's on-balance-sheet bar is implied from the 113% ratio and labelled directional. The $1.5-2.0 trillion, $220 billion and $270-500 billion figures are faithful summary of the FSB (May 2026). The $1.4 trillion, $1.5 trillion subprime comparison and the 5-6%, 4% and 1.3% loss figures are faithful summary of the ECB (May 2026). The circular-financing framing is faithful summary of the IMF (April 2026), the BoE framing of the FPC Record (April 2026). The opaque-leverage-layer characterisation and three-stage pathway are analyst synthesis. The Starbucks and tokenmaxxing references are directional Tier 3 corroboration, not anchors.