Every structure described here is legal, permitted under current accounting rules, and disclosed by the companies in footnotes and commitment tables. This is not investment advice. The question is not whether the numbers exist — it is where they live, and what happens to the metrics built on top of the place they don’t.
AI Infrastructure Financing — Key Figures (via FT, Sep 2026)
Big Tech AI infrastructure exposure held off balance sheets via guarantees and SPVs, as reported by the FT
$43.8B
Alphabet’s data-centre guarantees as reported — up from $16.9B over six months, per FT reporting
~$30B
Meta’s reported SPV raise — ~$27B loans, ~$3B equity — with the vehicle owning the data centre and leasing it back
>$120B
Data-centre financing debt moved to investors via SPVs across multiple companies, per FT
Note: The ~$300B (off-balance-sheet exposure) and a separate ~$3T figure (sector-wide commitments) are different scopes and are not combined here. All figures via FT reporting; not independently verified.
The Financial Times reported on September 20, 2026 that Big Tech companies hold roughly $300 billion of AI infrastructure exposure off their balance sheets through guarantees and special-purpose vehicles. To be precise about what that figure covers: it refers to off-balance-sheet exposure at the company level, not the separate and much larger figure of roughly $3 trillion in sector-wide commitments — those are different scopes and the FT does not combine them, and neither does this piece.
The Alphabet case is the FT’s clearest illustration of scale. Alphabet’s data-centre guarantees are reported to have risen from $16.9 billion to $43.8 billion over six months, with less than 2% of that figure reaching the balance sheet. The obligation grew by a large multiple in half a year; the measured quantity — the number that flows automatically into every ratio, screen, and covenant test built on top of balance-sheet figures — barely moved.
Meta’s reported structure shows the mechanism in its most explicit form. A special-purpose vehicle raised roughly $30 billion — approximately $27 billion in loans from lenders including Pimco, BlackRock and Apollo, alongside roughly $3 billion in equity from Blue Owl. The vehicle owns the data centre. Meta leases it. The borrowing therefore sits in the vehicle, not on Meta’s balance sheet. Companies including Meta, xAI, Oracle and CoreWeave are reported to have used such vehicles to move more than $120 billion of data-centre financing debt to investors; no individual share of that total is assigned to any of them here, and no return, fee or term is attributed to any lender.
The key insight: Disclosed and measured are not the same thing — and only one of them propagates automatically through the financial systems built on top of it. Both sets of numbers are public. Only one of them shows up by default in every leverage screen, covenant test, and credit model without a human having to go looking for it first.
Both numbers were published. Only one of them enters a leverage ratio without somebody choosing to put it there.
The Structural Read
The most important thing to understand about this story is that it is not about concealment. The companies published the figures that made the FT’s reporting possible. The issue is architectural, and it operates at the level of how financial information propagates once it is recorded.
A number on the balance sheet enters every derived metric automatically. Leverage ratios, screening filters, credit models, covenant tests — they all consume it without anybody deciding to include it. A number in a footnote enters those same calculations only where a human reads the footnote, understands what the structure means, forms a judgment about how much of it to count, and manually adds it back. Both are disclosure. Only one is measurement in the system-level sense: it arrives at its destinations by default rather than by deliberate retrieval.
The Alphabet illustration makes the arithmetic visible without requiring any alarm. Someone relying on a balance-sheet-derived ratio rather than the commitment footnote would have seen a company whose measurable obligations were roughly stable across a period when the reported guarantee figure nearly tripled. They would have been reading a correct number. It would have been answering a different question from the one they thought they were asking.
Product Overhang Doctrine — Applied to Capital Structure
“Capability builds invisibly until it surfaces all at once.” The same logic applies here — not to product features, but to financial obligations. Commitments accumulate in footnotes, lease schedules, and SPV books. The balance sheet registers almost none of the change. The overhang is real. It simply hasn’t arrived at the place most people are looking.
The Metastructure is worth describing neutrally, because neutrality is actually more clarifying than verdict. When a separate entity owns an asset and leases it to you, what you carry is the right to use the asset and an obligation to pay rent — not the asset itself and not the debt that purchased it. This is a decades-old financing technique used across shipping, aviation, property, and energy. It is a real transfer of real economic risk to parties who are compensated to hold it. Nothing here calls it a trick, a loophole, or a dodge.
What connects this particular financing wave to everything else happening in AI infrastructure is the nature of the underlying asset. These are not finished data centres; many of the commitments are contracted ahead of capacity being built. The build itself is the thing that everything depends on — construction timelines, hardware delivery, power procurement. And when you look at a single construction project from both sides of the ledger, something worth noticing appears: a commitment to pay for capacity and a guarantee of somebody else’s borrowing against that same capacity are two distinct obligations arising from one project. Neither is necessarily visible in the place a reader looks first. This piece does not estimate any combined exposure, add any figures, or suggest stress of any kind. The observation is only about where obligations from a single project come to rest — and how many different sets of books they can rest in without any of them being wrong.
Three Implications
IMPLICATION 1 — FOOTNOTE LITERACY BECOMES A COMPETITIVE SKILL
Analysts and capital allocators who read commitment tables and SPV disclosures in full will see a materially different picture of AI infrastructure exposure than those who stop at balance-sheet-derived metrics. Both groups are reading public information. The gap in the picture they construct is a function of retrieval effort, not availability. As the scale of off-balance-sheet AI financing grows, that gap widens — and so does the edge held by those who close it manually.
IMPLICATION 2 — SPV STRUCTURES RESHAPE WHERE AI CREDIT RISK CONCENTRATES
When an SPV owns the data centre and a hyperscaler leases it, the credit exposure migrates from the hyperscaler’s balance sheet to the vehicle’s lender syndicate. Institutional capital — asset managers, private credit funds — becomes the effective infrastructure creditor. The hyperscaler’s reported leverage stays contained; the risk doesn’t disappear, it relocates. Understanding where it relocates, and under what conditions it might return, is the relevant structural question for anyone modelling AI infrastructure at scale.
IMPLICATION 3 — ACCOUNTING STANDARDS FACE A STRUCTURAL LAG
Every structure described in the FT’s reporting is permitted under current rules. That is not a loophole; it reflects the fact that accounting standards are written for the asset classes and financing patterns that existed when they were drafted. AI data-centre guarantees and pre-construction SPVs are new enough in their current scale that the standards haven’t caught up. Regulators and standard-setters will eventually close that lag — but the financing has already been done. The question for the next cycle is whether the rules evolve before the next wave of commitments is contracted.
Business Engineer Framework
The Map of AI — Where Infrastructure Financing Sits in the Stack
The FT’s reporting lands squarely in the capital and infrastructure layers of the AI stack — the layer beneath models, chips, and products that most coverage ignores. The Map of AI traces all nine layers, from compute procurement to application distribution, and shows how power and margin accumulate differently depending on where in the stack a company operates and how it finances the build. Understanding that map is the prerequisite for understanding why the financing structures described here exist, who benefits from them, and where the obligations actually rest.
The FT’s reported $300 billion figure is not a scandal — it is a systems property. Every obligation it describes is legal, disclosed, and findable by anyone willing to read past the balance sheet. What it reveals is that the AI infrastructure buildout has grown large enough, fast enough, that a substantial share of its financial weight now lives in the part of the reporting stack that requires deliberate human retrieval rather than automatic propagation. The companies published the numbers. The question is which systems, and which analysts, are actually reading them.
Not investment advice. All structures described are permitted under current accounting rules and disclosed by the companies concerned. Figures reach this piece through FT reporting and subsequent coverage and have not been independently verified.
This is not investment advice. Every structure described above is permitted under current accounting rules and is disclosed by the companies concerned in footnotes and commitment tables. Nothing above alleges concealment, fraud, impropriety or wrongdoing by any company or any person, and nothing above says that any company is hiding anything. The figure of roughly $300 billion refers to exposure held off balance sheets; the separate figure of roughly $3 trillion refers to sector-wide commitments. These are different scopes and are not combined above. All figures reach this piece through Financial Times reporting and subsequent coverage rather than independent verification. No individual share of the reported $120 billion is assigned to any named company, and no return, fee or term is assigned to any named lender. No ratio is computed above and no total debt, credit rating or cost of capital is stated. Sale-and-leaseback and special-purpose-vehicle structures are long-established across shipping, aviation, property and energy. Nothing above forecasts default or stress, and nothing is predicted.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
