The most consequential question in the artificial intelligence capital expenditure cycle is no longer how much the largest technology companies are spending. That figure is now disclosed, debated, and capitalized into prices with a precision that would have been unimaginable two years ago. The more consequential question is where the obligations that finance the spending are recorded—and, increasingly, the answer is that a growing share of them are recorded nowhere on the face of the consolidated balance sheet.
That is not, in itself, an accusation. United States generally accepted accounting principles permit a reporting entity to fund the use of an asset without owning the asset, and to participate in a financing vehicle without consolidating the vehicle, provided a specific and demanding set of conditions is satisfied. The structures now being deployed to fund data centers are, by the public record, the product of careful work by sophisticated counsel, accountants, and rating agencies, and they are disclosed in the manner the standards require. Buxton Helmsley does not allege that any issuer has misapplied the consolidation, lease, or guarantee literature, and this article identifies no company by name for the simple reason that the forensic question is structural rather than particular. The question is whether the institutional investor, the limited partner allocator, and the audit committee member have been given enough information—and have learned to ask for enough information—to understand the economic exposure that sits behind a reported balance sheet that, by design, does not show it.
The distance between the cash that leaves a company and the liability that the company reports is precisely where forensic accounting lives. In the present cycle, that distance has widened faster than the disclosure framework built to measure it. What follows is intended to assist the institutional reader in closing it.
The dimensions of the build-out are now a matter of record. Combined capital expenditure across the five largest United States cloud and artificial intelligence infrastructure providers rose from approximately $162 billion in 2022 to roughly $448 billion in 2025, growing at an average annual rate on the order of seventy percent since the second quarter of 2023, according to research that reconstructs the figures from the companies' own filings.¹ For 2026, the four largest of those companies are, on their own guidance, expected to increase capital expenditure by more than sixty percent above the historic levels reached in 2025, with the four collectively approaching seven hundred billion dollars in a single year.² These are capital intensity ratios—capital expenditure measured against revenue—that resemble those of regulated utilities and heavy industry rather than those of the asset-light software businesses these issuers were a decade ago.
For most of the cycle, the spending was financed in the manner least likely to trouble a credit analyst: out of the operating cash flow of a small number of exceptionally profitable companies. That is no longer sufficient. By the estimate of one major investment bank, meeting projected data-center demand will require capital expenditure on the order of $2.9 trillion globally between 2025 and 2028, of which roughly $1.4 trillion can be funded from corporate cash flow and the remaining $1.5 trillion must be raised externally.³ The same research identifies asset-based private credit as the single largest external channel, with an estimated contribution of approximately $800 billion, and the balance distributed across investment-grade corporate debt, securitized credit, bank lending, and the equity of sovereign and private vehicles.³
The shift is visible in the issuance data. Of roughly $950 billion of debt issuance in 2025, approximately $170 billion went to data-center-related loans, an increase of about fifty-seven percent from the prior year; in the United States bond market alone, technology issuers raised approximately $157 billion through late September of 2025, up roughly seventy percent year over year.⁴ ⁵ Issuance of data-center asset-backed and commercial-mortgage-backed securities exceeded $20 billion in 2025, eclipsing the full-year 2024 total of $11.6 billion, and the same bank expected $250 billion to $300 billion of issuance in 2026 from the hyperscalers and the joint ventures organized alongside them.⁶ ⁴ Convertible issuance, a structure that finances today's construction by selling tomorrow's dilution, reached a twenty-four-year high of roughly $167 billion globally in 2025, lifted materially by the sector.⁴
The point of reciting these figures is not to suggest that debt financing is improper. It is to establish a single proposition on which everything that follows depends: the financing of the artificial intelligence build-out has moved, in the space of roughly eighteen months, from the internally generated cash flow of the issuers to the external capital markets and, within those markets, increasingly to channels and structures whose obligations do not necessarily appear on the issuer's own consolidated balance sheet. Much of the issuance just described is recorded on the issuers' own balance sheets, as conventional debt should be; the concern of this article is the growing portion that is not. The regulators charged with watching the system have noticed. The Bank of England's Financial Policy Committee, which had observed that artificial intelligence infrastructure investment was to that point financed largely by the cash flows and equity of large, profitable technology companies, warned that debt financing of the sector was increasing quickly and in ways that included a material contribution from private markets, and that if such financing increased as projected, the financial-stability consequences of any artificial-intelligence-driven fall in asset prices would increase with it.⁷ The International Monetary Fund, for its part, identified the growing role of nonbank financial intermediaries—and their importance as intermediaries in private credit specifically—as a vulnerability capable of transmitting stress to the core banking system.⁸ The structures examined below are the precise mechanism by which that transmission would occur.
For a generation, the phrase "off-balance-sheet financing" referred most commonly to the operating lease. An issuer that needed the use of a building, an aircraft, or a fleet of servers could, under the lease accounting that prevailed through 2018, structure its contract as an operating lease and disclose the resulting payment obligations only in the footnotes, leaving the balance sheet itself unburdened by either an asset or a liability. The practice was lawful, widespread, and, in the judgment of standard setters, corrosive of comparability, because two issuers with economically identical commitments could present materially different balance sheets depending on how their contracts were drafted.
That door is now substantially closed. Accounting Standards Codification Topic 842, the lease standard the Financial Accounting Standards Board issued in 2016 and that took effect for public companies in fiscal years beginning after December 15, 2018, requires a lessee to recognize a right-of-use asset and a corresponding lease liability on the balance sheet for virtually every lease with a term exceeding twelve months, whether the lease is classified as finance or operating.⁹ A lease is classified as a finance lease if it meets any of five criteria—among them that the present value of the lease payments, together with any residual value guaranteed by the lessee, equals or exceeds substantially all of the fair value of the underlying asset, or that the asset is so specialized that it is expected to have no alternative use to the lessor at the end of the term—and is otherwise classified as operating; but in either case the obligation now appears on the face of the balance sheet.⁹ The standard's stated purpose was to address precisely the off-balance-sheet financing that operating-lease accounting had permitted.
The institutional reader should understand what that accomplished, and what it did not. It accomplished the capitalization of the ordinary lease. It did not, and could not, reach an arrangement structured so that there is no lease to capitalize on the issuer's books at all—because the asset is owned by a separate entity that the issuer does not consolidate, because the issuer's use of the asset is documented as something other than a lease of sufficient term, or because the issuer's interest in the financing entity is accounted for under the equity method rather than consolidated line by line. In other words, the closing of the operating-lease door did not end off-balance-sheet financing; it relocated the question from the lease standard to the consolidation standard. The frontier of off-balance-sheet financing in the artificial intelligence cycle is no longer the lease. It is the entity.
The structure that has come to define the present cycle can be described in the abstract, because its essential features recur across transactions. An operating technology company that requires an enormous, purpose-built data-center campus does not build it on its own balance sheet and does not enter a conventional long-term lease for it. Instead, a separate legal entity—a joint venture or special-purpose vehicle—is formed to develop and own the campus. The operating company takes a minority equity interest in that entity, frequently on the order of twenty percent, while funds managed by an asset manager hold the majority, frequently on the order of eighty percent. The vehicle raises the bulk of the construction cost as debt, increasingly placed with private-credit funds and the insurance balance sheets behind them. The operating company serves as the developer, the operator, and the tenant of the completed facility, and it frequently provides the vehicle with credit support—commonly a residual value guarantee covering a substantial portion of the early operating life of the campus—that protects the vehicle's lenders against the risk that the operating company will not renew its use of the asset.
In a transaction of this kind reported in October 2025, an approximately $27 billion vehicle was formed to develop a single campus, with the operating technology company retaining a roughly twenty percent interest and funds managed by an asset manager holding the remaining roughly eighty percent; the operating company was to act as developer, operator, and tenant, and it provided the venture a residual value guarantee covering the first sixteen years of operation.¹⁰ ¹¹ Reported as the largest private-debt financing on record, the vehicle's senior debt reportedly carried a single investment-grade rating and priced at a yield of approximately 6.58 percent—a level closer to high-yield than to the operating company's own cost of funds, which is itself a disclosure about where the market believes the risk resides.¹² Because the financing entity, rather than the operating company, is the borrower, the associated debt does not appear on the operating company's consolidated balance sheet.¹¹ The structure has been openly described as a potential template for the industry, and variants have proliferated: in at least one reported case, an artificial intelligence developer pursued a comparable vehicle of roughly $20 billion in which it would rent the underlying chips rather than own them.¹¹
It is essential to state plainly what is and is not unusual here. There is nothing unusual, and nothing improper, about financing a long-lived asset with debt raised against that asset's contracted cash flows; that is the ordinary logic of project finance, and it has built power plants, pipelines, and toll roads for a century. What is new is the application of that logic, at this scale and velocity, to the core productive infrastructure of the most valuable companies in the public equity markets—and the consequence that the reported balance sheets of those companies understate, sometimes by tens of billions of dollars per transaction, the capital deployed to build the infrastructure they design and operate, and the economic exposure they retain. The operating company in the structure described above does not own the campus, but it designed it, it runs it, it occupies it, and it has guaranteed a substantial portion of its residual value. Whether that constellation of rights and obligations amounts to control for accounting purposes is not a matter of intuition. It is a matter of a specific standard, and that standard is where the institutional reader must go next.
The governing literature is Accounting Standards Codification Topic 810, Consolidation. When an entity is a "variable interest entity"—broadly, an entity whose equity holders lack a controlling financial interest through their equity, or whose equity is insufficient to finance its activities without additional subordinated support—the question of who consolidates it is not answered by counting votes or shares.¹³ It is answered by identifying the "primary beneficiary," and a reporting entity is the primary beneficiary, and must consolidate the vehicle, if it has both of two characteristics of a controlling financial interest.¹⁴
The first characteristic is power: the power to direct the activities of the entity that most significantly affect the entity's economic performance. The second is economics: the obligation to absorb losses of the entity, or the right to receive benefits from the entity, that could potentially be significant to the entity.¹⁴ A reporting entity that has both consolidates; a reporting entity that has only one does not. Critically, the analysis is qualitative, not arithmetic. A twenty percent equity interest does not answer the question, in either direction. An entity holding a minority equity stake can be the primary beneficiary if it holds the power to direct the venture's most significant activities and bears or receives economics that could be significant; an entity holding a majority stake can fail to be the primary beneficiary if it lacks that power. The codification specifically directs increased skepticism toward arrangements in which a party's economic interest is disproportionate to its stated power, precisely because such arrangements can be engineered to place power and economics in nominally different hands.¹⁵
Set the abstract test against the recurring structure. The operating company designs the campus, operates it, and is its sole tenant; that is a strong indication of power over the activities that most significantly affect the vehicle's economic performance, because the value of a purpose-built, single-tenant data center is overwhelmingly a function of that tenant's occupancy and the operator's management of it. The operating company also provides a residual value guarantee covering a substantial early period; that is an obligation to absorb losses, and potentially a significant one. Whether the structure nonetheless avoids consolidation turns on carefully constructed features—the allocation of decision rights to the majority equity holder, the precise scope and probability of the guarantee, the term and renewal mechanics of the tenancy—that determine, under ASC 810, whether power and significant economics in fact reside with the operating company. The institutional reader is not equipped, from the outside, to redo that analysis. The institutional reader is, however, entitled to demand the disclosure that would permit an informed judgment of whether the conclusion is robust or merely defensible.
This is not a hypothetical concern that auditors and regulators have failed to anticipate. The Securities and Exchange Commission's staff has, in its comment-letter practice, repeatedly required registrants to substantiate exactly these conclusions—to explain why an entity is or is not a variable interest entity, why the registrant is or is not its primary beneficiary, and, in one instructive line of comment, why a registrant accounting for an eighty percent interest under the equity method had concluded it was not required to consolidate.¹⁶ The existence of that line of inquiry is itself informative. It tells the institutional investor that the consolidation conclusion is contestable, that reasonable professionals can and do disagree about it, and that the conclusion is therefore exactly the kind of estimate—like the useful life of a server—that should be read as a position rather than received as a fact.
ASC 810 also imposes a disclosure regime that does not disappear when consolidation is avoided. A reporting entity that holds a variable interest in a variable interest entity it does not consolidate must still disclose the nature of its involvement, the nature and carrying amount of the assets and liabilities recognized in connection with that involvement, and its maximum exposure to loss as a result of it.¹³ That maximum-exposure-to-loss disclosure is, in the structures under discussion, among the most important numbers an institutional investor can obtain, because it is the figure that translates an off-balance-sheet arrangement back into the language of the balance sheet. It is also, in practice, among the least consistently presented, and an investor who does not look for it will not be handed it.
The residual value guarantee that recurs in these structures is not merely a fact bearing on the consolidation analysis. Where it falls within the scope of Accounting Standards Codification Topic 460, Guarantees, it is an obligation in its own right. ASC 460 requires a guarantor, at the inception of certain guarantees, to recognize a liability generally measured at the fair value of the guarantee, and—whether or not initial recognition is required for a particular guarantee—to disclose the nature of the guarantee, its term, the events that would require performance, and, importantly, the maximum potential amount of future payments the guarantor could be required to make.¹⁷ For the institutional reader, the maximum-potential-future-payments disclosure under ASC 460 plays the same role as the maximum-exposure-to-loss disclosure under ASC 810: it is the mechanism by which a contingent, off-balance-sheet commitment is rendered in a number the reader can weigh. A guarantee covering a substantial portion of the residual value of a multibillion-dollar campus for the first sixteen years of its life is not a footnote curiosity. It is a material commitment, and the standard requires that its magnitude be disclosed.
The same is true of the purchase and capacity commitments that frequently accompany these arrangements. An operating company that contracts to take a fixed quantity of computing capacity, or to make minimum payments to a vehicle or a vendor over a period of years, has entered an unconditional or take-or-pay obligation that the commitments literature—principally ASC 440, Commitments—directs it to disclose, and a loss that becomes probable and estimable under such a contract is a contingency governed by ASC 450, Contingencies.¹⁸ In a cycle in which capacity is contracted years in advance and in which the line between a vendor relationship and a financing relationship has blurred, these commitment disclosures are not boilerplate. They are frequently the only place in a filing where the multi-year cash obligations created by the build-out are quantified.
None of this is to suggest that issuers are concealing these obligations. The standards require their disclosure, and reputable issuers disclose them. The point is narrower and more practical: the obligations that the off-balance-sheet structure removes from the balance sheet do not vanish from the financial statements; they reappear, if the reader knows where to look, in the variable-interest-entity footnote, in the guarantee footnote, and in the commitments-and-contingencies footnote. An institutional process that reads the balance sheet and the income statement but treats the footnotes as ornamental will systematically understate the leverage and the exposure of the issuers most central to the present cycle. That is not a defect in the standards. It is a defect in how the standards are read.
Briefings on markets, governance, and our investment philosophy—delivered directly.
There is, however, one development in the disclosure framework that deserves the institutional reader's particular attention, because it ran in the opposite direction to the trend it now governs. On November 19, 2020, the Securities and Exchange Commission adopted amendments to Regulation S-K that, among other changes, eliminated the separately captioned "Off-Balance Sheet Arrangements" disclosure that had been required by Item 303(a)(4) of Regulation S-K and replaced it with a principles-based instruction to discuss off-balance-sheet arrangements within the broader context of Management's Discussion and Analysis; the same amendments eliminated the tabular disclosure of contractual obligations that had been required by Item 303(a)(5).¹⁹ ²⁰ In place of the dedicated caption and the table, the amended rule requires a registrant to discuss its material cash requirements, including commitments and obligations arising from arrangements with unconsolidated entities, where those arrangements are reasonably likely to have a material effect—even when the arrangement results in no obligation reported on the balance sheet.²⁰
The Commission's rationale was reasonable on its own terms: it judged that a prescriptive, separately captioned disclosure and a rigid table had become rote, that they duplicated information available in the financial statements, and that a principles-based instruction would better elicit information useful to investors.²⁰ Reasonable people supported the change. But the institutional reader should appreciate the timing and the consequence. The dedicated, comparable, tabular presentation of off-balance-sheet arrangements and long-dated contractual obligations was removed from the required disclosure framework in late 2020. The explosion of off-balance-sheet data-center financing began in earnest roughly four years later. The result is that the single most ambitious off-balance-sheet financing campaign in the recent history of the public markets is unfolding under a disclosure regime that no longer requires the issuer to gather its off-balance-sheet arrangements and its long-term contractual commitments into one comparable place. The information is still required to be disclosed where it is material; it is simply more diffuse, more narrative, and less comparable across issuers than it was when the table existed. An investor who once could turn to a single table must now assemble the picture from the variable-interest-entity footnote, the guarantee footnote, the commitments footnote, the lease footnote, and the liquidity discussion, and must do so issuer by issuer, with no assurance that two issuers have organized the same facts the same way.
There is a related measurement problem that compounds the disclosure problem. Even the capital expenditure figures on which the entire debate rests are sensitive to how finance leases are treated, because a data center occupied under a finance lease contributes a right-of-use asset and a lease liability rather than a conventional cash outlay, and aggregations that capture cash spending but not new finance leases will understate the true investment in productive capacity. The research that reconstructs hyperscaler capital expenditure from the filings has acknowledged precisely this, noting that its estimates combine cash spending with new finance leases and that even so they may understate the total because operating-lease right-of-use assets are excluded.¹ The magnitude of what such timing can defer is itself considerable: one credit-rating agency's analysis of the five largest providers found total commitments on the order of $969 billion, of which roughly $662 billion represented data-center-related leases that had not yet commenced—obligations that are disclosed in the notes but that will not be recognized as right-of-use assets and lease liabilities on the balance sheet until the leases begin.²¹ If the headline capital-expenditure number is itself an artifact of which obligations are captured and which are not, the institutional reader is well advised to treat it as a starting point for inquiry rather than as a settled fact.
It would be a mistake to dismiss the present structures as a rerun of the operating-lease debates of the prior decade, for two reasons. The first is the retention of risk. In the classic operating lease, the lessee's exposure was, in substance, the stream of rental payments, and the lessor bore the residual and the financing risk. In the structures now prevalent, the operating company frequently designs, operates, and occupies the asset, and guarantees a substantial portion of its residual value—so that the economic risk the operating company has shed is considerably less than the debt it has kept off its balance sheet would suggest. The phrase that best captures the arrangement is not "off the books." It is off the balance sheet, on the hook. The institutional reader who concludes that an issuer has reduced its risk in proportion to the debt it has externalized has misread the structure.
The second reason is the financing channel. The debt that funds these vehicles is increasingly held not by banks subject to regulatory capital and consolidated supervision, but by private-credit funds and the life-insurance balance sheets that increasingly stand behind them. That channel is less transparent, less liquid, and less uniformly marked than the public bond market, and it concentrates exposure to a single sector in a part of the financial system that the post-2008 supervisory architecture was not primarily designed to monitor. This is the specific concern that the macroprudential authorities have articulated. The International Monetary Fund warned in its October 2025 Global Financial Stability Report that asset valuations had returned to stretched levels, that the concentration of those valuations in a handful of artificial-intelligence-related firms stood at historical highs, with the forward price-to-earnings ratio of the broad United States index near the ninety-sixth percentile of its range since 1990, and that the deepening interconnection between banks and nonbank intermediaries—including the private-credit funds that are financing this build-out—could transmit and amplify a shock.⁸ ²² The Bank of England's Financial Policy Committee judged in October 2025 that the risk of a sharp market correction had increased and that equity valuations appeared stretched, particularly for technology companies focused on artificial intelligence, and elaborated in December 2025 that United States equity valuations were close to their most stretched since the dot-com peak, that industry estimates put artificial-intelligence infrastructure spending over the following five years at more than five trillion dollars with roughly half expected to come from external financing, and that debt markets, including the rapidly expanding private-credit sector, would carry the largest share.⁷ ²³
Read together, the macroprudential warnings and the accounting architecture describe a single phenomenon from two vantage points. The regulators are warning that a great deal of debt is funding a concentrated bet on a single technology, that the debt sits increasingly in the least transparent part of the financial system, and that a repricing of the bet would transmit losses through that system. The accounting question is where, in the financial statements of the issuers at the center of the bet, that debt and that exposure can be found. When the answer is "in the variable-interest-entity footnote, the guarantee footnote, the commitments footnote, and the maximum-exposure-to-loss disclosure, if the issuer has presented them clearly and if the reader has thought to look," the institutional investor has been given a system that is disclosed but not transparent. Disclosure and transparency are not the same thing, and the gap between them is where the next cycle's surprises will be found.
As issuers file their quarterly reports for 2026 and prepare the annual reports that will be filed in early 2027, and as allocators evaluate the private-credit and securitized vehicles that increasingly fund this sector, Buxton Helmsley believes there are seven questions that institutional investors, audit committee members, and limited partner allocators should be prepared to ask. They are organized to move from the issuer's own books outward to the financing structures themselves.
First, does the issuer disclose, in a single and comparable place, the aggregate amount of its off-balance-sheet data-center commitments—including its interests in unconsolidated financing entities, its residual value and other guarantees, and its long-term capacity and purchase obligations—or must the reader assemble that figure from separate footnotes? The elimination of the dedicated off-balance-sheet caption and the contractual-obligations table in 2020 means the issuer is no longer required to gather these items for the reader; an issuer that nonetheless does so is providing a meaningful service, and an issuer that does not should expect the reader to assemble the figure and to treat any difficulty in doing so as itself a finding.
Second, for each material unconsolidated financing entity, does the issuer disclose its maximum exposure to loss under ASC 810, and is that figure reconciled to the issuer's equity interest, its guarantees, and its other involvements? Maximum exposure to loss is the number that translates the off-balance-sheet arrangement into balance-sheet terms. Its absence, or its presentation at a level of aggregation that prevents the reader from associating it with specific vehicles, should be treated as a material limitation on the usefulness of the disclosure.
Third, what is the substance of the consolidation conclusion? Where the issuer designs, operates, and occupies an asset owned by an entity it does not consolidate, the reader is entitled to understand the basis on which the issuer concluded that it does not hold the power to direct the activities that most significantly affect that entity's economic performance, and does not hold economics that could be significant to it. A conclusion of non-consolidation that rests on the formal allocation of decision rights to a majority equity holder, while the issuer retains the operational reality of the asset, is a conclusion that should be probed rather than accepted, and is exactly the kind of conclusion the Commission's staff has historically required registrants to defend.
Fourth, what is the maximum potential amount of future payments under the issuer's guarantees, and over what period are they outstanding? A residual value guarantee, a completion guarantee, or a debt service guarantee is a real obligation whose magnitude ASC 460 requires the guarantor to disclose. The institutional reader should locate that figure, should understand the triggering events, and should add it to the issuer's reported leverage when forming a view of the issuer's true financial commitment.
Fifth, what are the issuer's unconditional purchase and capacity commitments, on a multi-year basis, and to whom are they owed? In a cycle in which capacity is contracted years in advance and in which suppliers, customers, and financing counterparties increasingly overlap, the commitments footnote is frequently the only quantified record of the cash the issuer has promised to spend. An investor who does not read it will not understand the issuer's forward obligations, and an investor who reads it should ask whether the counterparties to those commitments are independent of the issuer or are themselves entities in which the issuer holds an interest.
Sixth, where the financing has been placed in the private-credit or securitized markets, what does the pricing and the rating of that financing reveal about where the market believes the risk resides? When the senior debt of a single-tenant vehicle prices at a yield materially above the tenant's own cost of funds, the market is expressing a view—that the credit of the vehicle is not the credit of the tenant—that the institutional reader should take seriously even where, and especially where, the issuer's own disclosures emphasize its limited legal obligation.
Seventh, and for the limited partner allocator specifically, what is the underlying collateral and the contractual cash flow that supports a private-credit or asset-backed vehicle marketed as exposure to artificial intelligence infrastructure, and how is that collateral valued in the absence of an observable market? The same questions that Buxton Helmsley has raised in other contexts about the valuation of private credit apply with particular force here, because the assets are novel, the useful lives are contested, and the contractual cash flows depend on the continued occupancy of facilities by tenants whose own demand is the very thing the macroprudential authorities have flagged as uncertain.
The question is not whether the off-balance-sheet financing of the artificial intelligence build-out is lawful. By the public record, it is the product of careful and competent work, and it complies with the standards. The question is whether the institutional investor has done the work required to understand it. Where issuers present their unconsolidated interests, their guarantees, their commitments, and their maximum exposure to loss clearly and in a way the reader can assemble into a coherent picture, the answer can be yes, and the resulting financial statements, however much leverage they reveal once read in full, are at least subject to informed evaluation. Where those disclosures are diffuse, where maximum exposure to loss is aggregated into uselessness, and where the consolidation conclusion rests on form rather than substance, the answer is no, and the reported balance sheet should be treated as a partial account of the issuer's true financial commitment rather than as a complete one.
Buxton Helmsley has long maintained that the most important disclosures in a set of financial statements are the ones that require the reader to do work. In the present cycle, the obligations that finance the largest concentrated capital program in the history of the public equity markets are precisely those disclosures. They have been moved, lawfully and by design, from the face of the balance sheet into the footnotes, the guarantee schedules, and the financing vehicles that the issuers do not consolidate. The capital is no less deployed for having been externalized, and the economic exposure is no less retained for having been kept off the balance sheet. The institutional investor who learns to read these structures with the rigor the moment requires will understand the leverage the market is presently financing. The institutional investor who reads only the balance sheet will not—and will discover the difference at the least convenient time.
May 28, 2026
Read
May 21, 2026
Read
May 18, 2026
Read
Briefings on markets, governance, and our investment philosophy—delivered directly.
Referenced Sources:
Epoch AI, "Hyperscaler capex has quadrupled since GPT-4's release," data insight reconstructing combined capital expenditure of Alphabet, Amazon, Meta, Microsoft, and Oracle from company financial filings (cash spending plus new finance leases), reporting growth averaging approximately 72 percent per year since Q2 2023 and a projection of approximately $770 billion in 2026 if the trend continued; combined 2025 spending neared half a trillion dollars. Available at https://epoch.ai/data-insights/hyperscaler-capex-trend. The 2022 and 2025 aggregate figures of approximately $162.3 billion and $448.3 billion are drawn from the same Epoch AI dataset as presented in Visual Capitalist, "Visualizing Big Tech's AI Spending (2022–2025)," Apr. 20, 2026, https://www.visualcapitalist.com/visualized-big-tech-ai-spending/.
"Tech AI spending approaches $700 billion in 2026, cash taking big hit," CNBC, Feb. 6, 2026, reporting that the four largest hyperscalers were projected to increase capital expenditure by more than 60 percent above 2025 levels, with combined 2026 spending approaching $700 billion. https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html.
Morgan Stanley, "Credit Markets' Role in AI Financing Gap," Thoughts on the Market (Vishy Tirupattur and Vishwas Patkar), Aug. 6, 2025, identifying a financing gap of approximately $1.5 trillion to be met by external capital; and Morgan Stanley, "Credit Is Key for Data Center Financing," Thoughts on the Market, Aug. 19, 2025, projecting approximately $3 trillion of data-center capital expenditure required by 2028, roughly half to be funded by hyperscaler cash flows. https://www.morganstanley.com/insights/podcasts/thoughts-on-the-market/credit-markets-ai-financing-gap-vishy-tirupattur-vishwas-patkar and https://www.morganstanley.com/insights/podcasts/thoughts-on-the-market/data-center-financing-vishy-tirupattur-vishwas-patkar-carolyn-campbell. The allocation of the external gap (approximately $800 billion to asset-based private credit and approximately $350 billion to other channels) is summarized from the Morgan Stanley report Bridging a $1.5trn Data Center Financing Gap (July 2025), as cited in Apollo Global Management, "Spotlight: Financing the Digital Infrastructure Surge," Aug. 6, 2025, https://www.apollo.com/insights-news/insights/2025/08/spotlight-financing-the-digital-infrastructure-surge.
"The $3 Trillion AI Data Center Build-Out Becomes All-Consuming for Debt Markets," Feb. 3, 2026 (as published by Insurance Journal), reporting, on data attributed to IJGlobal (a Green Street company), that of approximately $950 billion of debt issued across the market in 2025 about $170 billion went to data-center-related loans, an increase of roughly 57 percent from the prior year, and that global convertible-bond issuance reached a 24-year high of approximately $167 billion in 2025; the same article attributes to Morgan Stanley the projection of $250 billion to $300 billion of issuance in 2026 from the hyperscalers and related joint ventures. https://www.insurancejournal.com/news/international/2026/02/03/856623.htm.
Carmen Arroyo and Laura Benitez, Bloomberg, "Meta, Blue Owl seal $30 bn private capital deal for Hyperion data centre" (as published by Business Standard), Oct. 17, 2025, reporting that technology companies raised approximately $157 billion in the United States bond markets through late September 2025, up roughly 70 percent year over year. https://www.business-standard.com/technology/tech-news/meta-blue-owl-seal-30-bn-private-capital-deal-for-hyperion-data-centre-125101700310_1.html.
Morgan Stanley research, as summarized in Brandywine Global, "Brave New World of AI Capex: Giving Credit Where Credit Is Due," Nov. 12, 2025, reporting that data-center asset-backed and commercial-mortgage-backed securities issuance exceeded $20 billion in 2025, eclipsing the full-year 2024 total of $11.6 billion, with more than $50 billion raised since 2018. https://www.brandywineglobal.com/around-the-curve/2025/brave-new-world-of-ai-capex-giving-credit-where-credit-is-due.
Bank of England, Financial Stability Report, December 2025, Financial Policy Committee, noting that many risky asset valuations remained materially stretched, particularly for technology companies focused on artificial intelligence; that United States equity valuations were close to their most stretched since the dot-com bubble and United Kingdom valuations since the global financial crisis; that industry estimates put artificial-intelligence infrastructure spending over the following five years at more than $5 trillion, with around half expected from external financing and debt markets (including the private-credit sector) playing the largest role. https://www.bankofengland.co.uk/financial-stability-report/2025/december-2025. See also the Financial Policy Committee's observation, in its responses to the Treasury Committee, that artificial-intelligence infrastructure investment had been financed largely by the cash flows and equity of large, profitable technology companies but that debt financing was increasing quickly, including a material contribution from private markets, https://publications.parliament.uk/pa/cm5901/cmselect/cmtreasy/1791/report.html.
International Monetary Fund, Global Financial Stability Report, October 2025, "Shifting Ground beneath the Calm," finding that equity and corporate credit valuations had returned to stretched levels, that the concentration of valuations in a handful of artificial-intelligence-related firms stood at historical highs, and that the growing role of nonbank financial intermediaries—including as intermediaries in private credit—could transmit stress to the core banking system. https://www.imf.org/en/publications/gfsr/issues/2025/10/14/global-financial-stability-report-october-2025.
Financial Accounting Standards Board, Accounting Standards Codification Topic 842, Leases (issued February 2016; effective for public business entities for fiscal years beginning after December 15, 2018), requiring lessees to recognize a right-of-use asset and lease liability for leases with terms exceeding twelve months, with classification as a finance or operating lease determined under the criteria in ASC 842-10-25-2 (including whether the present value of the lease payments and any lessee residual value guarantee equals or exceeds substantially all of the fair value of the underlying asset, and whether the asset is of such a specialized nature that it is expected to have no alternative use to the lessor). See FASB ASC 842; effective date and classification criteria summarized in FinQuery, "Operating Lease Accounting for ASC 842 Explained," https://finquery.com/blog/operating-lease-accounting-asc-842-explained-example/.
Dan Rabb, "Meta Pushes Its Largest Data Center Project Off Its Books With $27B JV," Bisnow, Oct. 22, 2025, reporting the formation of an approximately $27 billion joint venture to develop and own a data-center campus, with funds managed by the asset manager owning 80 percent and the operating company owning 20 percent, the operating company providing a residual value guarantee for the first sixteen years of operation. https://www.bisnow.com/national/news/data-center-capital-markets/meta-pushes-its-largest-data-center-project-off-its-books-with-27b-joint-venture-131490.
Carmen Arroyo and Laura Benitez, Bloomberg (as published by Business Standard), Oct. 17, 2025, reporting that under the special-purpose-vehicle structure the financing entity—not the operating company—is the borrower, allowing the operating company to avoid placing the debt on its balance sheet while serving as developer, operator, and tenant, and noting that a separate artificial intelligence developer was pursuing a comparable structure of approximately $20 billion in which it would rent rather than own the underlying chips. https://www.business-standard.com/technology/tech-news/meta-blue-owl-seal-30-bn-private-capital-deal-for-hyperion-data-centre-125101700310_1.html.
"Meta's $27 billion bet turns AI compute into Wall Street's hottest new investment," Fortune, Oct. 31, 2025, reporting that the vehicle's debt carried a single A+ rating and priced at a yield of approximately 6.58 percent, a level closer to high-yield territory, and that the structure allowed the operating company to build the facility without placing the full amount of debt on its own balance sheet. https://fortune.com/2025/10/31/metas-27-billion-bet-turns-ai-compute-into-wall-streets-hottest-new-investment/.
Financial Accounting Standards Board, Accounting Standards Codification Topic 810, Consolidation, including the definition of a variable interest entity in ASC 810-10-15-14 and the disclosure requirements in ASC 810-10-50 applicable to holders of variable interests in entities they do not consolidate (including disclosure of the nature of the involvement, the carrying amounts of related assets and liabilities, and maximum exposure to loss). See FASB ASC 810; summarized in BDO, "Control and Consolidation Under ASC 810" (May 2024), https://www.bdo.com/getmedia/8a458199-b9e0-4445-8b0e-f8f1bd7a180f/Control-and-Consolidation-Under-ASC-810.pdf.
FASB ASC 810-10-25-38A, providing that a reporting entity with a variable interest in a variable interest entity is the primary beneficiary, and consolidates the entity, if it has both (a) the power to direct the activities that most significantly affect the entity's economic performance and (b) the obligation to absorb losses or the right to receive benefits that could potentially be significant to the entity. See FASB ASC 810-10-25-38 through 25-38J; summarized in Deloitte, "Roadmap: Consolidation—Identifying a Controlling Financial Interest," ch. 7, https://dart.deloitte.com/USDART/home/codification/broad-transactions/asc810-10/roadmap-consolidation/chapter-7-determining-primary-beneficiary/7-1-introduction.
FASB ASC 810-10-25-38G (directing increased skepticism where a variable interest holder's economic interest is disproportionately greater than its stated power to direct the entity's activities). Summarized in PwC, "Identifying the primary beneficiary of a VIE," https://viewpoint.pwc.com/dt/us/en/pwc/accounting_guides/consolidation_and_eq/consolidation_and_eq_US/chapter_5_identifying_the/5_1_identifying_the.html.
Deloitte, "Roadmap: SEC Comment Letter Considerations," consolidation section, collecting examples of Securities and Exchange Commission staff comments requiring registrants to support conclusions regarding variable-interest-entity status and primary-beneficiary determinations, including a comment regarding a registrant accounting for an 80 percent ownership interest under the equity method. https://dart.deloitte.com/USDART/home/publications/deloitte/additional-deloitte-guidance/roadmap-sec-comment-letter-considerations/chapter-2-financial-statement-accounting-disclosure/2-2-consolidation.
Financial Accounting Standards Board, Accounting Standards Codification Topic 460, Guarantees, requiring a guarantor to recognize a liability at the inception of certain guarantees, generally measured at fair value, and to provide disclosures under ASC 460-10-50, including the nature of the guarantee, its term, the events that would require performance, and the maximum potential amount of future payments under the guarantee. See FASB ASC 460-10; summarized in PwC, "Accounting for a guarantee under ASC 460," https://viewpoint.pwc.com/dt/us/en/pwc/accounting_guides/financing_transactio/financing_transactio_US/chapter_2_guarantees_US/22_accounting_for_a__US.html, and PwC, "Guarantees—disclosure requirements," https://viewpoint.pwc.com/dt/us/en/pwc/accounting_guides/financial_statement_/financial_statement___18_US/chapter_23_commitmen_US/236_guarantees_US/2361_asc_460_disclos_US.html.
Financial Accounting Standards Board, Accounting Standards Codification Topic 440, Commitments (requiring disclosure of commitments such as unconditional purchase obligations), and Accounting Standards Codification Topic 450, Contingencies (governing the recognition and disclosure of loss contingencies that are probable and reasonably estimable). See FASB ASC 440 and ASC 450.
U.S. Securities and Exchange Commission, "SEC Adopts Amendments to Modernize and Enhance Management's Discussion and Analysis and Other Financial Disclosures," Press Release 2020-290, Nov. 19, 2020, announcing the replacement of the off-balance-sheet arrangements requirement (former Item 303(a)(4)) with an instruction to discuss such arrangements in the broader context of MD&A and the elimination of the tabular disclosure of contractual obligations (former Item 303(a)(5)). https://www.sec.gov/newsroom/press-releases/2020-290.
U.S. Securities and Exchange Commission, Final Rule, "Management's Discussion and Analysis, Selected Financial Data, and Supplementary Financial Information," Release No. 33-10890 (Nov. 19, 2020), adopting the amendments to Items 301, 302, and 303 of Regulation S-K, including the principles-based off-balance-sheet instruction and the requirement under amended Item 303(b)(1) to disclose material cash requirements from known contractual and other obligations, including obligations arising from arrangements with unconsolidated entities even where no obligation is reported on the balance sheet. https://www.sec.gov/files/rules/final/2020/33-10890.pdf. See also Alston & Bird, "SEC Adopts New Rules to Modernize Items 301, 302, and 303 of Regulation S-K," Dec. 2020, https://www.alston.com/en/insights/publications/2020/12/sec-adopts-new-rules-to-modernize.
Moody's analysis (February 2026), as reported by Fortune (Mar. 7, 2026), finding that the five largest cloud and artificial intelligence infrastructure providers had made total commitments on the order of $969 billion, of which approximately $662 billion represented data-center-related leases that had not yet commenced. https://www.fortune.com/2026/03/07/big-tech-trillion-dollar-borrowing-ai-century-bonds.
International Monetary Fund, Global Financial Stability Report, October 2025, Chapter 1, reporting that the 12-month-forward price-to-earnings ratio of the broad United States equity index had climbed to approximately the 96th percentile of its range since 1990. https://www.imf.org/-/media/files/publications/gfsr/2025/october/english/ch1.pdf.
Bank of England, Record of the Financial Policy Committee meeting of October 2, 2025, stating that the risk of a sharp market correction had increased and that, on a number of measures, equity market valuations appeared stretched, particularly for technology companies focused on artificial intelligence. https://www.bankofengland.co.uk/financial-policy-committee-record/2025/october-2025.
This article is published by Buxton Helmsley USA, Inc. for general informational and educational purposes only. It is analytical in nature, describes accounting and disclosure structures in the abstract, and does not identify, and is not intended to make any allegation regarding the conduct of, any particular issuer. It does not constitute investment, legal, accounting, or tax advice, and it is not an offer or solicitation with respect to any security. Investors should conduct their own due diligence and consult with qualified professionals before making investment decisions.
Briefings on markets, governance, and our investment philosophy—delivered directly.
