AI Governance & Digital Sovereignty  ·  27 September 2026

The Arithmetic Was Already Public

Michael Burry's forensic Substack breakdown puts a number on $3tn in undisclosed AI hyperscaler debt, the same off-balance-sheet mechanism this publication reported on in August.
By Alan Wright  ·  The Haunted Lighthouse Limited  ·  Peel, Isle of Man

On 27 September, the Sunday Telegraph ran a piece by Chris Price on Michael Burry, the investor Warren Buffett nicknamed “Cassandra” for calling the 2008 subprime crisis years before the market did. Fittingly, Burry's own Substack is called Cassandra Unchained; the material the Telegraph is reporting on secondhand sits there under the series title “The Heretic's Guide to AI's Stars”. Part IV, “The Big 5 Hyperscalers & the Missing $3 Trillion” (19 September), established the headline figure: over $3tn in AI hyperscaler commitments that do not appear on any balance sheet, drawn from footnotes to the companies' own 10-K and 10-Q filings. Part V, “Capital Cycle IQ & Forensic Files on the Big 5 Hyperscalers” (24 September, abridged version 26 September), supplies the forensic detail underneath it.

Readers of this publication will recognise the shape of that number. It is the same mechanism we described in “Why Frontier AI Companies Need, Not Want, Your Cash” (7 September), and the same depreciation and revenue-timing argument behind the reset wall in “Not Inevitable, But Sooner Than You're Prepared For”. Both traced back to the SEC's ABS disclosure changes we covered in August, when Meta's Hyperion vehicle and Goldman's roughly $1tn estimate of leasing commitments first put a number on the practice. Burry's series is the same filing-level evidence, aggregated across five hyperscalers, with a bigger number and a household name attached.

The Numbers

Burry's own figures: nearly $1.2tn in uncommenced lease commitments and more than $1.5tn in purchase commitments across the five companies, rising above $3tn once special-purpose vehicles, guarantees and contingent backstops are added in. He frames the risk as a duration mismatch; data centres take three to five years to build and lock the hyperscalers into leases running thirteen to twenty years, while the GPUs and TPUs going inside them see meaningful power and cooling changes every twelve to eighteen months. Firms risk being left holding bespoke, purpose-built concrete for hardware that's already a generation behind; Burry cites Satya Nadella's own past comments about not wanting Microsoft tied to infrastructure built around a single hardware generation as, in effect, a hyperscaler CEO conceding the point. More than $400bn of that sits in construction in progress assets across the five companies, which stay classified as construction in progress and so accrue zero GAAP depreciation until the projects enter service, while losing real economic value the whole time they sit idle; extending useful-life assumptions on that hardware doesn't remove the cost, it just defers when it hits the income statement. That >$400bn figure is independently confirmed by TheStreet's own reporting on Burry's post. Burry has also estimated, in an argument first made last November and reiterated in August, that the same extended-life accounting could overstate Oracle's earnings by roughly 26.9% and Meta's by roughly 20.8% by 2028, a combined shortfall of around $176bn across 2026 to 2028.

A Live Example

Burry also names circular financing directly, cloud providers funding AI labs that return the capital straight back as cloud spend, obscuring how much of the demand is genuine end-user demand rather than money moving in a loop. That argument has a live example sitting inside the same story: Project Jupiter, the New Mexico data centre at the centre of this week's Oracle news below, is being built for Oracle under the $400bn Stargate deal it signed with OpenAI and SoftBank last year. The same circularity “Why Frontier AI Companies Need, Not Want, Your Cash” described in September, cloud providers and AI labs financing each other's revenue, is the financing structure behind the specific building now at the centre of the accounting argument, not a separate case.

The Accounting Trick

The sharper detail from the Telegraph's own reporting is what Joachim Klement, Panmure Liberum's head of strategy, economics and ESG, said about Microsoft: shifting future data centre leases from finance leases to operating leases drops the projected capex figure from $190bn to $175bn without the company changing its spending by a single dollar. That is the Theatre Pulldown gap in its purest accounting form; the institutional claim is “reduced capital commitment”, the operational reality is identical spending under a different label.

What Oracle Isn't Calling an Exit

Oracle's force majeure notice on Project Jupiter is the sharpest new fact in the reporting, and it is worth being precise about what is and isn't confirmed. Bloomberg reported on 24 September that Oracle sent the notice to Stack Infrastructure, the Blue Owl Capital portfolio company building the 2.45 GW Doña Ana County campus, seeking to defer payments should the site miss its 2028 start date; the project has been dogged by local opposition and permitting setbacks, including a six-month delay to the Energy Transfer gas pipeline meant to fuel it. Both sides immediately pushed back on the “exit” framing: Oracle called force majeure notices “commonplace” and said the project “remains on our planned schedule”, while Blue Owl said the notice “does not change the financial commitments to this multiyear project”. Take that pushback for what it's worth, but weigh it against numbers neither side disputes: Oracle's $18bn in loans tied to the project were quoted at 89 to 91 cents on the dollar the same week, a gauge of Oracle's credit risk hit a record high, and Oracle's shares fell 4.4% on the day, part of a 29% slide for the year. Companies can call a filing routine. Bond markets pricing that filing at a discount are a harder thing to talk down.

One Correction

The Telegraph's own reporting contains a small inconsistency worth noting: its prose names Apple as one of the five companies behind the spending, while its accompanying chart lists Amazon instead. Burry's Substack settles it: the five are Microsoft, Amazon, Oracle, Meta and Alphabet; Apple isn't in his analysis at all. The chart, sourced straight from the companies' own filings, got it right; the prose didn't. A small, concrete example of why checking against the primary document beats trusting the write-up about the write-up.

Whose Book This Serves

Worth stating plainly rather than burying: Burry is not a disinterested forensic accountant here. He has disclosed short positions against Oracle (at roughly $145 a share) and Nebius, both citing the same off-balance-sheet leverage argument, alongside separate shorts on Micron and others. That doesn't make the filing-footnote arithmetic wrong; the lease and purchase-commitment totals are the companies' own disclosed numbers, not his estimate. But a reader is owed the fact that the person doing the forensic accounting also holds a trading position that benefits if the market agrees with him.

None of this is a victory lap. Burry himself made the same claim about priority back in August, posting on X that people “could have heard it first, months ago” about the same $3tn figure. Everyone in this space wants credit for calling it early; the more useful question is whether the underlying filings support the number, not who said it in which order. The Fed's own spring 2026 survey, already covered here, found half of market contacts flagging AI as a possible shock to financial stability, up from 9% a year earlier. Set against that: cloud revenue held up across the hyperscalers last quarter, capex guidance for next year is still climbing toward $1tn, and the Magnificent Seven were still posting gains the same week Burry's post was circulating. That gap, strong reported growth on one side and balance-sheet forensics on the other, is the actual story, and it doesn't resolve just because a well-known short-seller put a number on it. Burry attaching an investor-facing figure to a mechanism doesn't change what the filings said; it confirms that the filings said what we already reported they said.

So when the same bosses who signed these leases start talking publicly about restraint, is that a change of heart, or the story being prepared for when the numbers force one anyway?


Sources


Questions about this analysis, or interested in working with The Haunted Lighthouse?
consultancy@haunted.lighthouse.co.im

The Sovereign Auditor covers digital sovereignty, cybersecurity governance, and data protection policy, with particular focus on Isle of Man jurisdiction and Crown Dependency issues.

Support independent analysis. Subscribe directly, or scan on your phone.

Payments via PayPal. Credentials delivered by email. No Substack. No Stripe. No middlemen.