The Deal That Set the Template: Eighty Percent of a Data Centre Belongs to a Fund
On 21 October 2025 Meta announced a joint venture with funds managed by Blue Owl Capital to develop and own Hyperion, its data centre campus in Richland Parish, Louisiana. The release stated the terms with unusual precision: the Blue Owl funds hold 80 percent of the venture and Meta 20 percent; the partners fund their pro rata shares of "approximately $27 billion in total development costs"; the funds contributed about $7 billion in cash at closing; Meta received a one-time distribution of about $3 billion; and Meta provides construction management and property management for a campus it will operate and no longer owns [1]. Days earlier the venture's financing vehicle, Beignet Investor LLC, had priced $27.3 billion of 6.581 percent senior secured bonds due May 2049 in a private placement, at 225 basis points over Treasuries, rated A+, with Morgan Stanley as sole underwriter; IFR, which gave the transaction its financing-package award, and Bloomberg reported that PIMCO took about $18 billion of the issue [2]. The rating sits one notch below Meta's own, on the strength of a lease and an undertaking to make the bondholders whole should Meta exit at one of the four-year intervals the lease allows, and the debt sits on the venture's balance sheet rather than Meta's [2].
The structure deserves the attention it has received, though for a reason the commentary tends to pass over. A company with one of the largest cash flows in corporate history chose to finance its flagship asset through a vehicle it does not consolidate, at 6.58 percent for a claim that rests in the end on its own lease payments, and the market absorbed $27 billion of that paper in a single placement. The premium is the price of keeping the asset and its debt off the sponsor's balance sheet, and the willingness of long-dated capital to pay for a 23-year claim on a single-tenant building is the first fact about this cycle. The second fact is who supplied the capital. PIMCO manages the retirement savings, insurance reserves and pension assets of people who have never heard of Richland Parish. The risk of the AI build-out did not disappear when it left Meta's balance sheet. It moved.
A Century Bond for a Two-Year Product Cycle
Alphabet's turn came in February. On 13 February 2026 the company closed concurrent public offerings of $20 billion of dollar-denominated senior notes and £5.5 billion of sterling-denominated senior notes, in tranches running from 3.700 percent notes due 2029 to 5.750 percent notes due 2066 in dollars and, in sterling, from 4.125 percent notes due 2029 to "£1,000,000,000 aggregate principal amount of 6.125% notes due 2126" [3]. The form of global note filed with the Securities and Exchange Commission describes the instrument without ornament: a £1 billion security, "redeemable prior to February 13, 2126 at the Company's option" [4]. Nine days earlier the company's fourth-quarter results had told investors why: "our 2026 CapEx investments are anticipated to be in the range of $175 to $185 billion," against $91.4 billion of purchases of property and equipment in 2025 [5]. A hundred-year bond financing a doubling of capital spending on chips and servers whose replacement cycles are measured in years is a mismatch of durations that only an issuer of Alphabet's standing can place, and the fact that it was placed, at 6.125 percent, in a currency not its own, is a reading of how far the search for long-dated yield has travelled.
The Leases That Are Not Yet Liabilities
The bonds are the visible part. The larger part sits in a footnote that the accounting standard allows to stay off the balance sheet until a building is handed over. Under ASC 842 a lease is recognised when it commences; a lease signed for a data centre still under construction appears only as a commitment, in a note to the accounts, and the five largest buyers of computing capacity now disclose commitments of a size without precedent. Microsoft's annual report for the year to 30 June 2026 records "additional leases, primarily for datacenters, that had not yet commenced of $329.1 billion," with terms of one to twenty years and commencement dates running to fiscal 2033 [6]. Meta's second-quarter report puts its figure at approximately $278.99 billion, "consisting of data centers, colocations, and certain network infrastructure," with terms of up to thirty years, and adds that in July 2026 it signed further data-centre leases of approximately $68 billion [7]. Amazon's commitments table carries a line, "Leases not yet commenced," of $137.2 billion, of which $93.5 billion falls after 2030 [8]. Alphabet discloses $85.2 billion of future payments on leases "primarily related to data centers, that have not yet commenced," with terms of up to 26 years [9]. Oracle, whose fiscal year ends in May, reported $260 billion at 31 May 2026 and $288 billion three months later, "substantially all related to data center arrangements," for terms of fifteen to nineteen years, including one lease on which it has guaranteed up to $3.3 billion of the landlord's own borrowing [10].
Add the five footnotes and the figure is about $1.12 trillion of lease commitments that appear on no balance sheet, at 30 June for four of the companies and 31 August for Oracle, before Meta's July signings. For scale, the lease liabilities the same companies do carry are a fraction of that: Microsoft's operating and finance lease liabilities total about $88.5 billion, Oracle's about $43.8 billion and Alphabet's about $20.6 billion [6][9][10]. The off-balance-sheet boom is therefore not a metaphor. By the companies' own disclosures, more than a trillion dollars of future rent has been promised to landlords whose buildings are not yet built, and whose own financing, as Hyperion shows, has been sold to the holders of long-dated savings. The tenants are among the most creditworthy companies in the world, and the question this essay asks is not whether they will pay. The question is who stands between them and the people whose money built the buildings, and on what terms.
Private Credit Found Its Borrower
The lenders who have grown fastest into that gap are not banks. The Bank for International Settlements' September 2026 Quarterly Review traced the shift: private credit outstanding reached almost $2.5 trillion in 2025, and technology companies, which borrowed $22 billion from private lenders in 2010, or 22 percent of the market, borrowed more than $1 trillion in 2025, or 44 percent of it [11]. A BIS Bulletin published in July looked inside the business development companies through which retail and institutional investors reach that market and found about $115 billion lent to software firms, roughly a fifth of all BDC lending and more than 80 percent of their technology portfolios, with a conclusion the authors stated carefully: the revenue risk that generative AI poses to those borrowers "has not affected these loans yet, and neither BDCs nor their equity investors have priced software exposure differently" [12]. The Federal Reserve Bank of Boston, reading some 890,000 loan observations across 168 BDCs, found that the share of loans paying interest in kind rather than in cash rose from about 6 percent in early 2022 to about 10 percent in early 2026, "a 67 percent increase over three years," and that the rise had spread across nearly every industry [13].
The case that tested those prices arrived in October, and it concerned a company whose business is to rent the computing capacity the hyperscalers have not yet built for themselves. CoreWeave's report for the quarter to 30 June 2026 shows $35.6 billion of debt principal, of which $7.5 billion falls due within twelve months; quarterly revenue of $2.575 billion; net interest expense of $640 million, up from $267 million a year earlier; and a net loss of $626 million [14]. Its contracted backlog stood at approximately $104 billion [15], and in September it added $4.2 billion of 2.875 percent convertible notes due 2033 to the stack [16]. Its secured facilities carry effective rates of 9 to 15 percent; its unsecured 9.625 percent notes due 2032, issued in June, were trading, according to Reuters and Barron's, at yields near 14.8 percent in the second week of October, as the shares fell 15 percent in a session [14][17]. The same week Gartner put more than $3.5 trillion of AI infrastructure spending at risk of impairment by the end of 2030 and expected more than 80 percent of the independent "neocloud" providers to have ceased independent operation by then [18]. A borrower whose secured lenders charge 9 percent and whose unsecured bonds yield 15 percent is a borrower on whom two prices have been set, and the distance between them is the market's estimate of what the first lien is worth when the second one is not.
The Insurance Balance Sheet as the Terminal Buyer
Follow the chain one step further and it ends, more often than not, on the balance sheet of a life insurer. The Congressional Research Service reported in September that the number of US insurers owned by private-equity firms rose from about 25 in 2017 to 139 in 2024, that they held about $700 billion of cash and invested assets, 7.8 percent of the industry's total, and that they are twice as likely as their peers to hold assets originated by their own parent, to rely on private ratings and to cede reserves to affiliated reinsurers offshore [19]. The Federal Reserve, which added private credit to the Financial Accounts of the United States for the first time in September, published the holders a month later: private credit vehicles now manage more than $2.6 trillion and have lent about $1.5 trillion to US businesses, roughly 6 percent of all nonfinancial business debt, and of the shares in private debt funds, defined-benefit pension funds hold just over 40 percent, life insurers 14 percent, and households and nonprofits 23 percent [20]. The three largest holders of the asset class that financed the build-out are, in other words, the pension, the policy and the household itself.
The reserves behind those policies have been moving offshore at the same time. A BIS paper of October 2025 on the transformation of the life industry records that US life insurers had ceded $2.1 trillion of reserves to reinsurers by 2023, up from $500 billion in 2017 and equal to a quarter of their assets, with offshore reinsurers taking 40 percent of the ceded risk against 14 percent six years earlier; it notes that Bermuda's insurance sector exceeds 150 times the island's GDP, and it records the 2023 failure of 777 Re, "a PE-linked reinsurer" domiciled there [21]. The Financial Stability Oversight Council's 2025 Annual Report carries the series forward: reserves ceded rose "by over 70 percent since 2019, rising to $2.4 trillion in 2024 from $1.4 trillion," the offshore portion "more than doubled and exceeded $1.1 trillion," "nearly 70 percent of life and annuity reserves ceded offshore go to affiliated reinsurers," and insurers with asset-manager or private-equity sponsors account for 46 percent of what is ceded to those affiliates [22]. The Council's own description of the arrangement is the one this essay would use: it "may introduce additional opacity and complexity." An annuity sold in Ohio, reinsured in Hamilton by an affiliate of the asset manager that also originated the private loan that financed the data centre, is a chain with four links and one set of eyes on it.
The Mark-to-Market Nobody Sees
Insurers carry most of their bonds at amortised cost, so the price of a long-duration promise appears in a disclosure table long before it reaches an income statement. Japan offers the clearest current reading of that table, because its four largest life insurers publish it every quarter. At 30 June 2026 the unrealised losses on their domestic bond holdings stood, by each company's own report, at ¥6.28 trillion for Nippon Life, ¥4.04 trillion for Dai-ichi Life, ¥2.53 trillion for Sumitomo Life and ¥2.28 trillion for Meiji Yasuda, a combined ¥15.13 trillion, up from ¥14.08 trillion three months earlier [23][24][25][26]. The losses are unrealised because the bonds are held to match reserves and the companies intend to keep them; the point is not that any of the four is impaired but that one quarter of rising yields added roughly a trillion yen to a loss no policyholder will see in a statement. The private loans that have replaced such bonds on American insurers' balance sheets have no quoted price at all. Their value is a model's output, reviewed by an auditor, under a solvency regime that in Bermuda, as the BIS observes, allows more latitude than the prescriptive rules of the United States [21]. The Japanese table is the honest version of a mark that elsewhere is not taken.
What the Referee Says
The International Monetary Fund's Global Financial Stability Report of October 2026, released in Bangkok on 13 October, names the first of its three salient risks in a sentence: weaker or less durable AI investment and profitability "could lead to a correction in AI-related asset prices," and "Corrections could be amplified by rising leverage and concentrated exposures" [27]. Its catalogue of vulnerabilities reads like an index of this essay: "Liquidity mismatches among insurers and pension funds have increased"; "Rising leverage and contingent liquidity risks in private credit could amplify stress"; and "Globally, hedge funds now hold about $13 trillion in assets, more than three times their 2013 level" [27]. The Fund does not forecast the correction. It observes that the leverage has been built in places that report no daily price, and that the institutions holding it have promised a liquidity they may be unable to deliver on the day it is asked for. A reader of The Liquefaction of Forty Trillion will recognise the shape: the obligations are long, the funding is shorter than it looks, and the gap is carried by whoever cannot leave.
Who Holds the Bag: The Household on the Liability Side
The household's exposure to this cycle is asymmetric, and the asymmetry is the argument. On the upside sits the equity: the Federal Reserve's Distributional Financial Accounts for the second quarter of 2026 put 88.1 percent of corporate equities and mutual fund shares in the hands of the top 10 percent of households and 0.6 percent in the hands of the bottom half [28]. The capital gains of the AI build-out have accrued, to the extent they have reached households at all, to one household in ten. On the downside sit the liabilities. Fifty-six percent of families own stock directly or through a fund, down from 58 percent in 2022 in a decline the Fed's Survey of Consumer Finances attributes to "families outside the top 10 percent of the income distribution," while about 65 percent participate in a retirement plan [29]; the retirement plan holds the pension that holds 40 percent of the private debt fund shares; the household directly holds another 23 percent [20]; and the policy or annuity is backed by reserves that have, in rising proportion, been reinsured offshore by an affiliate. The Two-Speed Household measured the distance between the family that lives on assets and the family that lives on wages. This cycle widens it from both ends: the gains concentrate in the first family's brokerage account while the risk is distributed into the second family's pension, policy and power bill.
The power bill is not a figure of speech. On 3 August 2026 the Governor of Texas directed the state's utility commission and grid operator to audit every data centre in the interconnection queue before any is allowed to proceed, stating that ERCOT "is currently considering approximately over 474 gigawatts of requests to connect to the Texas grid," more than five times the system's record peak demand, that "Approximately 90 percent of the new power requests are data centers," and that any project failing the audit "will be denied connection" [30]. The ratepayer is the last holder in the chain, and the one with no fund document to read.
The comparison with 2008 is made often and should be made carefully. After that crisis Congress required the sponsor of a securitisation to keep "not less than 5 percent" of the credit risk it sold, the rule now codified at 17 CFR 246.4 [31]. The structures of this cycle are not securitisations and the rule does not reach them: a joint venture that owns a building, a private loan held in a fund, a reserve reinsured by an affiliate. Nothing requires the sponsor of the Hyperion venture to hold 5 percent of its bonds, the manager of a BDC to hold 5 percent of its loans, or the asset manager behind an insurer to retain 5 percent of what it cedes to its own reinsurer. Retention, where it exists, is a matter of the sponsor's choosing. That is the structural difference from 2008, and not a reassuring one.
The SAVI Reading: The Charter Over the Footnote
The SAVI Capital Model takes no view on whether the AI build-out will earn its cost of capital. It takes a view on where the household sits in the structure when the answer arrives. Every arrangement described above places the household on the liability side: as the pensioner whose fund bought the venture's bonds, the policyholder whose reserves were ceded to the affiliate, the saver whose BDC shares cannot be redeemed in the quarter they are needed, the ratepayer whose grid is spoken for. The model's first tenet, equitable profit sharing, moves the household to the other side of the ledger: in a company built on the model, half of net profit is paid in cash to the people who work in it, so that the return on the capital a workforce helps to produce reaches that workforce as income rather than as the residual value of a lease it will never see. The third tenet answers the opacity the Council describes: a structure whose economics must be read in a footnote, a reinsurance treaty and a private mark fails a standard of stewardship that requires the terms to be legible to the people who bear them. Alitheia's tokenised structures, with ownership and distributions recorded on a shared ledger, are transparency by construction rather than by disclosure, and the firm has argued in Alitheia and the Intelligence Alpha that the technology is a means to that end and never the thesis.
The Growth Equity Thesis and Capital After the Keynesian Model set out why the firm expects the next cycle of capital formation to be judged on the distribution of its proceeds rather than the size of its commitments. The off-balance-sheet boom is the test case. More than a trillion dollars of promised rent, a reinsurance system that has moved $1.1 trillion of reserves offshore, a lending market whose technology book grew from $127 billion in 2019 to more than $1 trillion in 2025, and an equity market whose gains belong to a tenth of households: these are the terms on which the present cycle has been raised. The white paper argues [32], and the firm commits to write into the legal terms of every fund it raises, that the terms can be different, and that a charter putting the household on the profit side of production is not a cost borne by investors but the condition on which the consent of the people who do the work is kept. The footnote records what the sponsor chose. The charter records what the model requires.