AI capex: the $300B obligation off the books

The table buyers use to size a vendor commitment no longer holds the fastest-growing obligation. According to FT reporting, Big Tech has moved roughly $300 billion of AI infrastructure exposure off balance sheet through residual value guarantees and special purpose vehicles: Alphabet data center lease guarantees rose from $16.9 billion to $43.8 billion in six months, with under 2 percent on the books. Our judgment: this is not an accounting compliance story, it is a measurement story. When the fastest-growing obligation lives in the notes, anyone underwriting a vendor on balance-sheet ratios is deciding with a number that no longer covers the exposure.

1. The facts: guarantees and SPVs move the obligation out of the statements

Per FT reporting, with secondary sources noting they could not independently verify the figures, Alphabet data center lease guarantees climbed from $16.9 billion to $43.8 billion in six months, with under 2 percent entering the balance sheet. It is the cleanest case on record: the obligation multiplied while the number read by every ratio, screen and covenant test barely moved.

Meta shows the mechanism end to end. Per FT reporting, the roughly $50 billion Hyperion project in Louisiana is held through the Delaware SPV Beignet Investor, funded 80 percent by Blue Owl and 20 percent by Meta. About $28 billion of residual value guarantees back roughly $27 billion of Pimco, BlackRock and Apollo debt. The SPV owns the data center and Meta leases it back, so the borrowing stays on the SPV books while Meta carries guarantees and lease commitments. Sources: AI Weekly coverage of the FT report and FourWeekMBA structural breakdown.

It is not a single-company pattern. Meta, xAI, Oracle and CoreWeave are reported to have moved more than $120 billion of data center financing debt to investors through such vehicles, without a company-level split. Broadcom, Nvidia through the SoftBank and OpenAI development arrangement, Oracle and Amazon are said to use variants of the same structure. Source: Business Post follow-up.

Two citation rules matter here. First, the roughly $300 billion is company-level off-balance-sheet exposure from guarantees and SPVs, while the roughly $3 trillion figure widely quoted in the industry is an industry-wide commitment total; FT does not merge them, so they cannot be added or swapped. Second, these structures are legal under current accounting rules and are disclosed in notes and commitment tables. Reading this as fraud is a misreading. The accurate statement is that the disclosure location distorts the metric. The original FT investigation sits behind a paywall (FT original, HTTP 403 on our check, anti-bot and paywall).

2. Our judgment: the question is not whether the obligation exists, but where it lives

We read this as an ordinary pattern. Wherever a structural gap opens between self-reported scope and verifiable fact, wrong metrics grow in it. In the model safety layer, vendor-claimed boundaries and measured overreach are two faces of the same problem. In the capital layer, vendor-claimed commitment scale and actual obligation are two faces of the same problem. The regularity holds: the fastest-growing obligations are the least verifiable ones.

For buyers the implication is concrete. What you are underwriting in AI infrastructure is not a technology stack but a balance sheet that keeps adding leverage. Once compute, racks and power are locked in multi-year commitments, supplier credit and concentration become your operating risk, and that risk has been moved into the notes.

Our call: moving AI vendor diligence from statement ratios to an obligation ledger is the cheapest upgrade available in enterprise AI procurement this year. It needs no new tooling, only a few lines in the review checklist.

3. The one decision this changes for each buyer

4. From ratios to an obligation ledger

Six items go straight into a vendor review checklist. All come from disclosures and contract text you can request today:

5. Conclusion and this month action

Our judgment: the real scale of AI infrastructure commitments is not on the statements, so vendor assessment built on statements has systematically understated how fast obligations grow. Buyers do not need more data. They need the right input location, which means pulling notes and commitment tables into the assessment.

The smallest useful move this month: at your next AI vendor review, make unfunded and off-statement obligations a standing agenda item, ask for guarantee and SPV detail, and put the request in the minutes. Then answer one question: if the obligation ledger holds a guarantee you have never seen, what exactly was your risk assessment built on?

References: FT investigation (original, paywalled, HTTP 403 on our check): ft.com/content/7f11afae-c4e3-4054-a65b-873f3647f563 | AI Weekly coverage (Alphabet $16.9B to $43.8B, Meta Hyperion and Beignet, $28B residual value guarantee): aiweekly.co alert | FourWeekMBA breakdown ($300B versus $3T scope split, $120B-plus of SPV debt): fourweekmba.com breakdown | Business Post follow-up: businesspost.ie report. All specific figures are attributed to FT reporting; the secondary sources state they could not independently verify them. This article does not merge the $300B off-balance-sheet exposure with the $3T industry commitment total.