The Guarantee Has No Date

Microsoft had signed $329.1 billion of leases, mostly for data centers, that had not begun by June 30, 2026, according to its annual report. Oracle reported $288 billion on August 31, 2026 in its quarterly report, to start by fiscal 2029. At June 30, 2026, Meta reported $278.99 billion, Amazon $137.2 billion and Alphabet $85.2 billion. The five add up to $1,118.5 billion, by DeepStack's sum. Moody's put the same line at $662 billion at the end of 2025, equal to 113% of the five companies' adjusted debt, Fortune reported on February 25, 2026. Meta's share then was $103.77 billion, by its annual report for 2025.

The published totals look irreconcilable until they stand in a row. Moody's and DeepStack count the same instrument, leases signed and not yet begun, at two dates, and the count rose 69% in six to eight months. The Bear Cave relayed on September 27, 2026 Michael Burry's tally of nearly $1.2 trillion of such leases plus more than $1.5 trillion of purchase commitments, a wider instrument; Burry wrote on September 19 that vehicles, guarantees and backstops could lift the total above $3 trillion. UBS Asset Management reached about $3.5 trillion on September 29 from Bloomberg data of August 2026 by changing the base: $808 billion of reported debt, plus $1.25 trillion of leases not yet commenced, about $2 trillion of service and production commitments and $86 billion of backstops, less about $620 billion of cash.

Part seventeen asks how much of the AI buildout is debt the balance sheets do not show yet, and who carries it if a tenant walks away. Parts one, three, eight and eleven followed the financing without setting the totals side by side. DeepStack read 396 sources by 111 voices between June 15 and October 11, 2026, and verified 167 disagreements, 57 of them contradictions or refutations.

The lease is already debt

Burry did the arithmetic for Microsoft on September 24, 2026. Discounted at the 4.5% rate Microsoft files, he wrote, the signed book adds about $240 billion to the $88.5 billion of lease liabilities already carried, about $330 billion of lease debt against $40 billion of bonds, so the company cannot simply slow down if demand changes. Microsoft's annual report for the year to June 30, 2026 confirms the $88.5 billion and starts the $329.1 billion between fiscal 2027 and fiscal 2033.

Steve Eisman of The Real Eisman Playbook went after the structures on September 25, 2026. Meta, he wrote, financed a $27 billion Louisiana campus through Beignet Investor LLC, kept 20%, bears delays and overruns and promised to rent it for 20 years, yet the debt stays off its books. Such designs, he argued, spare a company the rating agencies and work until they don't. He read the critical audit matter Ernst & Young wrote on the venture as the sign of an auditor unsure the deal belonged off the balance sheet.

The documents narrow his reading. Meta's announcement of October 21, 2025 gives the other 80% to funds managed by Blue Owl Capital; PIMCO, which Eisman counted among the owners, bought debt. Meta's 10-Q for June 30, 2026 starts the venture's leases in 2029 on four-year terms renewable to 20 years. In Meta's annual report for 2025, Ernst & Young called auditing Meta's judgment of who directs the venture "especially challenging," which records a hard call and finds no error in it.

The cash flows cover it

Gaetano and Burry drew opposite lessons from Meta's own history. Gaetano pointed on September 15, 2026 to 2023, no AI case by his own account, when Meta's capital spendingCash a company spends on long-lived assets: data centers, chips, networks, power. More → fell and its free cash flowOperating cash flow minus capital spending: the cash left after the company has paid for its own buildout. More → rose from $18.4 billion to $43 billion. Burry cited on September 19 Meta's 2022 charges for abandoned offices and data centers: a purpose-built data center lacks an office's sublease market, so capacity nobody needs gets written down. What separates 2023 from 2026 is a signature. A capital budget can be cut; a signed lease is a promise about someone else's building.

Harley Bassman, on Erik Townsend's MacroVoices on September 17, 2026, said most hyperscalersOne of the largest cloud and platform companies that build data centers at continental scale. More → own cash businesses that can carry their debt; shareholders may suffer, but short of a breakup the debt gets repaid. Harris Associates Oakmark wrote in its third-quarter 2026 letter that it holds bonds of Meta, Alphabet, Amazon and Oracle for their wider spreadsThe extra yield a borrower pays above a Treasury of the same maturity, in basis points. More → and lower leverage than the investment-grade market.

The filings give this side its best facts, and they concern timing. Amazon's 10-Q for June 30, 2026 schedules $93.5 billion of its $137.2 billion for after 2030. Moody's analyst David Gonzales told Fortune in February 2026 that the $662 billion was "yet to be on the balance sheet" rather than missing. The agency itself was warier: it said "the accounting liability is unlikely to reflect certain plausible future scenarios" and that it may make a nonstandard adjustment to its debt measure as lease dates draw nearer. Oguz Erkan of Capitalist Letters noted on August 23, 2026 that Meta can delay or cancel its own builds, though not its pay-or-take contracts with neoclouds.

The cash is a closer call than the timing. YCG Investments wrote in its second-quarter 2026 letter that Microsoft and Amazon, which once paid for the buildout from free cash flow, now spend enough to push it below zero and are turning to debt. The filings split the two. Microsoft's annual report shows $182.9 billion of operating cash in the year to June 30, 2026 against $115.9 billion of additions to property and equipment, leaving $67.0 billion. Amazon's 10-Q shows $161.4 billion of operating cash and $173.0 billion of property and equipment purchases in the 12 months to June 30, 2026, a gap of $11.6 billion before a single new lease starts. Alphabet's 10-Q puts its first half of 2026 at $84.9 billion of operating cash and $80.6 billion of capital spending.

DeepStack's reading is that a signed lease is debt with a start date. It enters the balance sheet when it commences, as late as 2036 for Meta, and at least one agency counts it sooner: S&P Global Ratings, cutting Oracle to BBB- on July 9, 2026, added to the debt in its forecast the $260 billion of leases Oracle then expected to start between fiscal 2027 and 2029. Moody's has said only that it may adjust. The lease's risk is timing, and the risk is concentrated: Oracle's $288 billion is 4.6 times the $63.2 billion of undiscounted payments left on the leases it carried on August 31, 2026. On timing, the cash side has the better of it. On cash, its case holds at Microsoft and leans on borrowing at Amazon.

Who holds the guarantee

On June 30, 2026, Meta put its maximum loss on the Hyperion venture at $46.03 billion, a figure that counts its equity, its funding commitments and its leases from the venture as well as its guarantees, against $2.92 billion carried as an investment. The guarantees are residual value guarantees with a threshold of about $28 billion: if Meta ends or does not renew a lease and other conditions are met, it pays any shortfall between the campus's value and the threshold. The 10-Q calls such payments "not probable, and therefore no liability has been recorded to date."

Burry read the structure on September 24, 2026 as a tell: lenders would not carry the residual risk without Meta's protection, so the parties closest to the asset doubt its value. Haim Zaltzman of Latham & Watkins, on Bloomberg Intelligence's Tech Disruptors on September 29, read the same tool the other way: deals this large and this fast probably need guarantees and credit support to clear. JunkBondInvestor priced the gap on September 20: debt backed by such a guarantee typically pays 100 to 150 basis pointsOne hundredth of a percentage point: 25 basis points is 0.25%. More → over the guarantor's own bonds. On July 26, JunkBondInvestor had tracked the Hyperion paper itself, issued at par, later above 109 and near 95 by then. It trades off the tenant's curve, JunkBondInvestor wrote, so it cheapens when the tenant's own bonds do, even with nothing changed at the project.

Alphabet's 10-Q for June 30, 2026 reports $43.8 billion of credit derivatives backstopping other parties' data center payments, up from $16.94 billion on December 31, 2025, with an estimated $24.1 billion more awaiting final terms. The liability recorded on them rose from $69 million to $815 million over the same six months, from 0.41% of notional to 1.86% by DeepStack's division. Exposure rose two and a half times and the liability nearly twelvefold. The filing adds that terms run up to 15 years and that on a default Alphabet may take over the leases.

Nvidia wrote the largest promise. Its 8-K of August 17, 2026 describes residual value guarantees, capped at $105 billion, on leases at a campus in Pike County, Ohio, with OpenAI as tenant and buildings expected ready from 2028. OpenAI will reimburse anything Nvidia pays, and the obligation ends if OpenAI earns a satisfactory credit rating. That is the cash side's best fact on guarantees, and it rests on the tenant whose default would trigger them.

DeepStack's reading is that a guarantee is debt with no date. A lease reaches the balance sheet on a known day; a guarantee has no such day. A lessee's residual value guaranteeA promise to make up the difference if an asset is worth less than expected at the end of a lease. More →, like Meta's, is booked only once payment turns probable, and a backstop marked at fair value, like Alphabet's, is carried at a fraction of its notional. The guarantee does not take risk off the company that writes it. It keeps the first loss with the guarantor, up to the cap, while little is recorded: nothing for Meta on Hyperion's residual value, 1.86% of notional for Alphabet on other parties' data center payments. Lenders hold what lies past the cap, and until then they hold the tenant's credit in the price of their paper. Eisman's Enron comparison fits the consolidation judgment only; nothing in these filings suggests a misstatement.

The weak link pays first

Kris of Potential Multibaggers reported on September 28, 2026 that Oracle had sent a force majeure notice on Project Jupiter, its New Mexico campus tied to Stargate, and hoped it would not prove the first crack. Eisman said on September 25 that the notice went to Blue Owl, with the project's debt near 90 cents. Quanta 72 added on September 27 that the notice lets Oracle delay payments if power delays push the campus past its 2028 start; Oracle has missed no payment, and Blue Owl calls its commitments unchanged.

Bassman had answered the premise on September 17, 2026: the hyperscalers, not labs such as Anthropic, do the borrowing, so a lab's failure would be an equity problem rather than a bond problem. Oracle is the hyperscaler that answer fits least. Kakashii noted on September 24 that Oracle entered the AI era already owing tens of billions of dollars.

Leyla Kunimoto of Accredited Insight traced the Jupiter money on October 5, 2026: about $18 billion of construction debt and $3 billion of Blue Owl equity. Oracle is not the borrower, she wrote, but its rent is the loan's main source of cash once the lease begins. JunkBondInvestor drew the line on September 20 between borrowers with a strategic counterparty such as Meta or Alphabet, who can pay up, and a developer carrying an unhedged floating-rate construction loan and no revenue before 2027. Set beside Eisman on Beignet, the same structure reads as evasion to one and as strength to the other, and both readings hold, for different tenants.

Oracle's own cash shows the strain. In the quarter to August 31, 2026, its earnings release reported $23.1 billion of operating cash, of which $11.4 billion, or 49.2%, came from customer prepayments with a significant financing component. Capital spending was $28.5 billion in the same quarter; without the prepayments, operating cash covered 41% of it.

The tests ahead

Oracle reports first, on December 14, 2026 by its investor relations page, and the annual reports follow in February 2027; every test reads a period still open on the day this story appears. The first, read on December 14, is whether customer prepayments with a significant financing component exceed 40% of Oracle's operating cash for the six months to November 30, from 49.2% in the quarter to August 31, 2026.

On February 16, 2027, two tests read the year-end filings. One is whether the leases not yet commenced that Alphabet, Amazon and Meta print for December 31, 2026, with Microsoft's figure for the same date and Oracle's for November 30, sum to more than $1.2 trillion, undiscounted. The base is $1,118.5 billion at mid-2026, and Meta's 10-Q for June 30 already reports about $68 billion of leases signed in July. The other is whether the liability on Alphabet's data center credit derivatives exceeds 1.86% of their gross notional at December 31, 2026, from 0.41% at the end of 2025 and 1.86% at June 30, 2026. The last waits for Meta's annual report for 2026, read on February 26, 2027: whether it consolidates Hyperion or records a liability above zero for the guarantees, neither of which it did at the end of 2025 or at June 30, 2026.

What would prove DeepStack wrong is a year-end sum of the five lease lines no higher than the $1,118.5 billion of mid-2026, Alphabet's ratio at 1.86% or less at December 31, 2026, or Oracle's prepayments at 40% or less of its six-month operating cash on December 14, 2026. A year-end sum above $1,118.5 billion but not above $1.2 trillion would lose DeepStack's bet on that test without flipping the Call, because the pile would still have grown.

The lease will be counted when it starts. Nobody has set a date to count the guarantee.