Off Book

In 1986, Hyman Minsky described a structural tendency in capitalist financial systems that he called the Financial Instability Hypothesis. The argument was precise: stability is destabilizing. When an economy has been growing steadily for a long period, the very steadiness of the growth encourages participants to take on more risk. Financing structures transition from what Minsky called hedge finance, where cash flows service both interest and principal, through speculative finance, where cash flows service only interest and the borrower must refinance the principal, to Ponzi finance, where cash flows cannot cover even interest and the borrower depends on asset appreciation or new borrowing to remain solvent.

The transitions are not driven by fraud. They are driven by the rational responses of individual participants to an environment in which the recent past has been consistently profitable. Each layer of the system is solvent on its own terms. The instability is systemic. It becomes visible only when a stress event forces the system to demonstrate that the aggregate risk exceeds the aggregate capacity to absorb it.

Meta Platforms generated $164.5 billion in revenue in 2024. Net income was $62.4 billion. Free cash flow was $52.1 billion. The company held $77.8 billion in cash and marketable securities at year-end. By any conventional measure, Meta is a hedge-quality borrower: its cash flows can comfortably service both interest and principal on virtually any amount it might need to borrow.

It is also paying approximately $6.5 billion in additional interest expense over 24 years to finance its largest data center project through a special purpose vehicle rather than borrowing directly on its own balance sheet. The premium is approximately 100 basis points above what Meta would pay on equivalent corporate debt. On $27 billion in principal over nearly a quarter century, that premium accumulates to $6.5 billion in additional cost.

That is the price of invisibility.

In August 2025, Meta formed a joint venture with Blue Owl Capital for Project Hyperion, a four-million-square-foot AI data center megacampus in Louisiana. The entity, registered as "Beignet Investor," issued approximately $27.3 billion in bonds and raised $2.5 billion in equity. PIMCO led the debt placement, purchasing roughly $18 billion of the initial issuance. Blue Owl holds 80% of the equity. Meta retains 20%.

The structure is designed to satisfy a specific accounting requirement. Under ASC 810, a company must consolidate a variable interest entity (VIE) on its balance sheet if it is the "primary beneficiary," meaning it has both the power to direct the VIE’s activities and the obligation to absorb its losses. By ceding 80% of the equity and structuring the lease so that Meta is a customer of the SPV rather than its controlling entity, the $27 billion in debt does not appear on Meta’s balance sheet. Meta’s quarterly filings show lease payments. They do not show $27 billion in obligations.

The bonds were rated A+ by S&P, investment grade, but yielded 6.58% at issuance, a rate typically associated with high-yield instruments. PIMCO’s position traded above 110 cents on the dollar within weeks, generating an estimated $2 billion in mark-to-market gains. The profit-taking was immediate. In Charles Kindleberger’s taxonomy of speculative cycles, this is the fourth stage: the moment when early participants begin extracting returns before the underlying assumption, that AI will generate revenue sufficient to justify the infrastructure, has been validated.

Meta’s rationale for paying the premium is strategic, not accidental. The company anticipates years of additional AI capital expenditure and needs to preserve its corporate credit rating and borrowing capacity for future raises. By keeping Hyperion off-book, Meta can later issue corporate debt at lower rates for subsequent projects. The premium paid today is an investment in future optionality. It is individually rational. The systemic question is what happens when every major participant applies the same logic simultaneously.

Meta is not alone. Oracle has raised approximately $38 billion through JPMorgan-arranged debt for two data center campuses in Texas and Wisconsin, part of the $500 billion Stargate partnership with OpenAI and SoftBank. JPMorgan has encountered diminished investor interest as it attempts to sell down the debt, with lenders growing wary of exposure to a company whose credit rating sits below peers like Microsoft and Google.

Elon Musk’s xAI raised $20 billion through a structure that demonstrates the pattern in its purest form. The financing splits into approximately $7.5 billion of equity and $12.5 billion of debt, channelled through an SPV. Nvidia contributed $2 billion in equity. The SPV uses the capital to purchase Nvidia GPUs, which xAI then rents over five years. Nvidia is simultaneously the investor, the hardware supplier, and the beneficiary of the SPV’s purchasing decisions. Apollo Global Management and Diameter Capital Partners supplied the debt side.

The pattern compounds. Nvidia invested $100 billion in OpenAI in September 2025, an investment made in cash, the majority of which will be used to lease Nvidia chips. Nvidia receives no voting power in return. What it receives is revenue: the cash flows back to Nvidia as hardware procurement. Fortune asked the question directly: "How much of the AI boom is just Nvidia’s cash being recycled?"

The circularity extends through CoreWeave, a GPU infrastructure company that went public in March 2025 at a $26 billion valuation. Nvidia invested $2 billion in CoreWeave. CoreWeave plans to spend $20-23 billion in 2025 on AI infrastructure, predominantly Nvidia GPUs. CoreWeave then leases that compute capacity to Microsoft, which accounted for 62% of CoreWeave’s 2024 revenue, and to OpenAI, which signed an $11.9 billion five-year deal with CoreWeave and received $350 million in CoreWeave stock as part of the arrangement.

The capital flows in a circuit. Nvidia invests in CoreWeave. CoreWeave buys Nvidia chips. CoreWeave leases capacity to Microsoft and OpenAI. Their lease payments support CoreWeave’s revenue. Nvidia books the chip sales as revenue. Nvidia marks up its CoreWeave equity investment. Every balance sheet in the circuit appears independently healthy.

The Verge’s analysis of CoreWeave described a company with "no obvious path toward profitability except in the absolute best-case scenario of fast AI adoption," whose function in the ecosystem is partly to absorb risk that larger participants prefer not to carry.

CoreWeave generated $1.9 billion in revenue in 2024, a 737% increase from the prior year. It reported an $863 million net loss. Its revenue backlog stood at $25.9 billion as of March 2025. It continues to issue debt: $2 billion more in convertible notes as of late 2025. In Minsky’s taxonomy, this is speculative finance: the borrower can service interest from current cash flows but must refinance principal through new borrowing. Whether it transitions to Ponzi finance, where even interest coverage requires new capital, depends entirely on whether AI demand grows fast enough to close the gap between revenue and expenditure.

In aggregate, more than $120 billion in AI infrastructure obligations have been shifted off balance sheets across the sector. UBS estimates that approximately $450 billion in private credit is tied to big tech AI infrastructure, with roughly $125 billion flowing into long-term data center project finance. The capital is supplied by PIMCO, BlackRock, Apollo Global Management, Blue Owl Capital, and major banks. The risk is no longer on the balance sheets of the technology companies. It is in the private credit markets, the insurance portfolios, and the pension allocations of the institutions that funded the SPVs.

In September 2025, Moody’s downgraded Oracle’s credit outlook from stable to negative, assigning a Baa2 rating, the lower end of investment grade. The agency identified "significant counterparty risk" arising from Oracle’s $300 billion five-year contract with OpenAI to provide 4.5 gigawatts of compute capacity. Moody’s forecast that Oracle’s debt would increase faster than EBITDA, with leverage reaching 4x and free cash flow remaining negative "for an extended period."

The concentration is stark. Oracle’s data center buildout depends on a single anchor tenant. OpenAI, valued at $500 billion, remains unprofitable and has outlined plans to spend more than $1 trillion by 2030. Oracle’s five-year credit default swaps have reached record highs as investors price the risk that the market can read but the balance sheet does not show.

CoreWeave faces comparable concentration. Microsoft accounted for 62% of 2024 revenue. The top two customers represented 77%. The company’s $25.9 billion revenue backlog depends on Microsoft and OpenAI honouring contracts that extend years into the future. If either renegotiates, delays, or cancels, CoreWeave’s ability to service its debt becomes immediately questionable.

The aggregate exposure is concentrated at a second level as well. OpenAI has reportedly signed over $1.4 trillion in long-term compute commitments across multiple providers, including Oracle, CoreWeave, and Microsoft Azure. A single entity’s financial trajectory, an entity that has never been profitable, anchors obligations dispersed across dozens of SPVs, private credit facilities, and project finance structures. If OpenAI’s revenue growth disappoints, the impact does not arrive at one counterparty. It arrives at all of them simultaneously.

This is the specific mechanism through which individual rationality produces systemic fragility. Each participant’s exposure appears manageable in isolation. Oracle’s Moody’s downgrade reflects Oracle-specific concentration risk. CoreWeave’s customer concentration reflects CoreWeave-specific business risk. The SPVs holding the project debt are non-recourse to the parent companies, ring-fencing the obligations. But the dependencies converge on the same set of assumptions: that AI will generate enough revenue, at enough scale, on a fast enough timeline, to justify the infrastructure commitments. If that assumption fails, it fails everywhere at once, and the ring-fencing that protected the parent companies' balance sheets becomes the mechanism through which the loss is transmitted to the bondholders, private credit funds, and institutional investors who funded the SPVs.

The structural parallel most frequently invoked by analysts is subprime mortgage-backed securities. The comparison captures something real, the separation of risk from the entity that appears to control the asset, but it is imprecise. The underlying assets in the current cycle are not subprime. They are data centers built for Meta and Oracle, entities with investment-grade credit profiles and substantial cash reserves. The counterparties are not low-income borrowers with adjustable-rate mortgages. They are the most profitable technology companies in history.

The more precise parallel is the 1990s telecommunications infrastructure buildout.

Between 1996 and 2001, telecommunications companies invested over $500 billion, approximately $2 trillion in 2025 dollars, in fiber optic cable. The investment was debt-financed. The thesis was that internet traffic was doubling every three to four months and that capacity would be exhausted without massive new infrastructure. The thesis was wrong. Actual demand doubled roughly every twelve months, a fourfold overestimate of the growth rate. By 2002, only 2.7% of the fiber that had been laid was in active use. Between 85% and 95% of the capacity remained dark, unused, for years after the bust. Companies like Ciena, JDS Uniphase, and WorldCom collapsed or were restructured. The infrastructure eventually found use, two decades later. The debt did not wait two decades.

The parallel is not metaphorical. It is mechanical. In both cases, the investment was debt-financed infrastructure built on demand projections that had not been validated. In both cases, the demand projections were extrapolated from a genuine technological breakthrough (wavelength division multiplexing for telecom; transformer-based AI for the current cycle). In both cases, the companies making the projections had strong incentives to overstate demand because their revenue depended on the infrastructure being built. In both cases, the structure of the financing separated the risk from the entities best positioned to assess it: the borrowing was done by SPVs, conduits, and project finance vehicles whose investors had limited visibility into the accuracy of the demand forecasts.

There is a second structural parallel, closer to the specific accounting mechanism in use. Enron’s off-balance-sheet structures, the vehicles with names like LJM, Chewco, and Raptor, employed the same principle that Meta’s Beignet Investor and Oracle’s Vantage structures employ: control without consolidation. Enron maintained operational control of assets while structuring the SPVs so that they would not be consolidated on Enron’s financial statements. The mechanism was the same, the manipulation of the consolidation rules under GAAP, even though the specific standards have been updated since. Enron’s structures were fraudulent because the SPVs were not genuinely independent; they were capitalized with Enron stock and managed by Enron executives. The current structures are not fraudulent. The SPVs are capitalized by independent third parties (Blue Owl, PIMCO, Apollo) and the accounting treatment is technically compliant with current standards. But the effect on investor perception is structurally similar: the obligations exist, but the balance sheet of the entity that controls the asset does not show them.

What distinguishes the current cycle is that the hyperscalers are genuinely profitable. Meta’s $62.4 billion in net income is not fiction. But a critical asymmetry exists in the telecom parallel that applies here. In the 1990s telecom cycle, the efficiency of fiber optic transmission was improving exponentially (64x capacity improvement through wavelength division multiplexing alone), meaning that a smaller amount of infrastructure could serve a larger amount of demand. In the AI infrastructure cycle, the trend is reversed: GPU power consumption has increased from 300 watts (Nvidia V100) to 1,000-1,200 watts (Nvidia B200), and each generation of chips requires more power and cooling infrastructure, not less. The efficiency curve that saved telecom from permanent overcapacity, by eventually creating enough demand to fill the fiber, runs in the opposite direction for AI compute. If demand disappoints, the infrastructure does not become cheaper to operate. It becomes more expensive.

The system has not been stress-tested.

The off-balance-sheet structures are built on 20-year leases and multi-decade debt instruments. They assume that AI infrastructure demand will persist at current or higher levels for the duration. The assumption has not been validated. Several categories of stress would test it simultaneously.

Custom silicon. Google’s TPU, Amazon’s Trainium, and Microsoft’s Maia chips represent in-house alternatives to Nvidia’s hardware. If these accelerators reach competitive performance, the demand for Nvidia GPUs, which supports CoreWeave’s business model, xAI’s SPV, and the revenue circularity that sustains the broader ecosystem, would decline. CoreWeave’s entire infrastructure is Nvidia-based. It has no published plan for hardware diversification. A shift in the GPU market would impair both CoreWeave’s revenue and the collateral value of every SPV whose assets consist of Nvidia hardware.

Revenue realization. Moody’s has flagged that OpenAI will not be profitable before 2029. OpenAI has committed over $1.4 trillion in long-term compute obligations. These obligations are supported not by current revenue but by the expectation of future revenue from products and services that have not yet been brought to market at the required scale. If AI monetization disappoints, the revenue that supports the lease payments that support the SPV bonds that support the private credit allocations will be insufficient. The question is not whether the technology works. The question is whether it generates enough revenue, quickly enough, to service the obligations that have been incurred in its name.

Interest rate sensitivity. The SPV bonds carry fixed rates (6.58% in Meta’s case) over multi-decade terms. If interest rates rise further, the market value of existing bonds declines, creating unrealized losses for the institutional holders. If rates decline, refinancing becomes attractive, but the single-tenant structure of the leases limits the SPV’s flexibility. In either scenario, the duration mismatch between the long-term project debt and the liquidity needs of the funds holding it creates pressure.

Asset illiquidity. The SPV assets are data centers. They are large, purpose-built, geographically fixed, and in many cases leased to a single tenant. If the anchor tenant defaults or does not renew, the residual value of the asset depends on finding another tenant for a facility optimized for a specific workload. A four-million-square-foot AI data center in rural Louisiana is not a fungible asset. The recovery value in a distressed scenario is deeply uncertain.

The contagion path is specific. The SPVs use similar underwriting, correlated exposures (all dependent on AI demand), and shared counterparties (OpenAI, Microsoft, Nvidia). A stress event at one SPV would trigger reassessment of all similarly structured vehicles simultaneously. The private credit market, now valued at approximately $1.7 trillion with $125 billion in data center project finance, has limited transparency and limited secondary market liquidity. Insurance companies and pension funds that hold these instruments may be unable to exit positions without crystallizing losses. $120 billion in obligations dispersed across vehicles whose performance is correlated by a single assumption: that AI demand sustains current growth rates for two decades.

Minsky described the structural conditions under which financial systems become fragile. The companies building AI infrastructure are profitable. The technology is real. The accounting is compliant. The financing structures are designed to make the system’s aggregate leverage invisible to the balance sheets that investors read. The risk has not been eliminated. It has been relocated: into SPVs whose bondholders have claims on physical infrastructure but limited visibility into demand forecasts, into private credit funds whose investors bear concentration risk they may not fully understand, into pension and insurance allocations whose duration profiles are mismatched with the liquidity of the underlying assets.

The question is not whether AI will generate revenue. It almost certainly will. The question is whether it will generate enough revenue, at enough scale, on a fast enough timeline, to validate 20-year obligations undertaken on the assumption that current demand growth rates are permanent. The telecom precedent suggests that when the answer is no, the discovery is abrupt. And the risk is found where it always was: not on the balance sheets of the companies that control the infrastructure, but in the portfolios of the institutions whose capital was used to make the infrastructure invisible.