The AI Boom's Invisible Balance Sheet

What public filings reveal about who funds AI data centers and who bears the risk, and what oversight should look like

Stijn Van Nieuwerburgh's “Financing the AI Buildout” paper describes an investment cycle of nearly $9 trillion over 2025–32 that is increasingly funded through leases, joint ventures, project bonds, private credit and special-purpose vehicles. These structures expand funding capacity, but they also raise leverage at the level of the asset and keep much of it off the balance sheets of the firms that use the capacity. The template is Meta's Hyperion campus in Louisiana. A joint venture, 80 percent owned by Blue Owl funds and 20 percent by Meta, owns the campus and leases it back to Meta; Beignet Investor, the Blue Owl vehicle that holds the 80 percent stake, sold about $27 billion of bonds to pay for it.

The paper's central worry is that the resulting exposures are layered, correlated and hard to observe. This post asks how hard. I tried to map the financing web from public filings, and I use what that exercise reveals to set out what oversight could look like.

What the public record shows

What follows is my own analysis, not the paper's. To test how much of this a determined outsider can see, I built a network map of AI data-center financing using only public sources, mainly SEC filings: annual and quarterly reports, 8-Ks, bond indentures, business development company (BDC) loan schedules and mutual-fund holdings reports. It traces 23 financing structures across eight tenants. That comes to about 99 entities and 115 links, each tied to its source, with roughly 80 percent of links resting on SEC filings. Each exposure is drawn so that the arrow points from the party that would lose to the party it depends on; ties within one corporate family are drawn as plain lines. Leases, funded debt and equity, and guarantees are kept in separate buckets and never added together. The interactive map lets you follow any claim back to its filing.

Four things stand out.

The protection behind the bonds is the tenant itself

Hyperion's structure is shown below. In a stylized scenario where Meta does not renew its leases at Hyperion and El Paso, the model's shortfalls come to about $35.4 billion. Under the illustrative assumptions, Meta's residual value guarantees absorb all of it and bondholders are unimpaired. Meta's filings cap the guarantees at about $28 billion for Hyperion, a threshold that declines over time, and about $13 billion for El Paso. Both exceed what is at stake: Hyperion's estimated $27 billion cost and El Paso's $12.3 billion of bonds. Non-renewal is an option Meta holds, not a default; the guarantee is what makes exercising it costly. The result depends on Meta remaining creditworthy, on the guarantees being enforceable and paid on time, and on the assumed recoveries. What it shows is that substantial economic risk stays tied to Meta even though the project debt is legally separate from its balance sheet. If Meta itself were in distress, the bonds would lose their main protection at the moment it was needed. That scenario matters most, and a fuller model should price it.

Mapping these links does not measure the statistical correlation the paper worries about. What it does is make concrete the channels through which those correlated exposures could pass losses between firms and financial institutions.

Diagram of the Hyperion financing chain from bondholders and Blue Owl funds through Beignet Investor and the joint venture to Laidley and Meta — Hyperion: who has a claim on whom. Arrows point from the party exposed to the party it depends on. Sources are linked on each edge in the interactive map. Tap to enlarge.

GPU collateral is where private credit is exposed

Meta's projects are backed by a guarantor. GPU-backed lending is not. The most central node in the map is CoreWeave, a GPU cloud provider, with about $35.6 billion of mapped debt against $5.0 billion of book equity. That debt combines about $13.6 billion of delayed-draw facilities secured on GPUs and customer contracts, $16.6 billion of senior and convertible notes, and $5.1 billion of equipment financing. CoreWeave's filings do not say which customer contract backs which facility, so the map does not divide the debt among its customers. In a scenario where GPU collateral values fall 60 percent, the model allocates about $7.2 billion of $18.3 billion in modeled losses to credit investors once borrowers' equity is used up. The equity of CoreWeave and of Valor Compute (the vehicle leasing NVIDIA GPUs to xAI) is wiped out. These are illustrative scenarios on a partial map, not forecasts. But they show where losses would land first.

Bar chart showing who absorbs balance-sheet losses in four stress scenarios — Illustrative allocation of modeled losses in four stylized severe scenarios. Balance-sheet losses only; guarantee payments count against the guarantor; lost compute-contract revenue is shown separately. An OpenAI walk-away is not shown: no filing ties its compute contracts to mapped debt or capex, so it moves no mapped balance sheet. Tap to enlarge.

How the stress numbers are built

Losses are allocated with a simple priority-of-claims cascade on the mapped network, run at corporate-group level. (1) A shock lowers an asset's value. The shock is either non-renewal of leases with an assumed re-letting recovery (10 percent in the severe case), a fall in GPU collateral values, or a 20 percent refinancing haircut on debt due by 2030. (2) Guarantees and backstops pay first, up to their caps. (3) The borrower's equity absorbs the next loss, using filed book equity where available and stated assumptions elsewhere. (4) Creditors take any remainder pro rata, up to their exposure. (5) Losses then pass to whoever is exposed to an impaired entity. Each dollar of loss is counted once. A compute contract reaches a balance sheet only up to the capital a filing ties to it; the rest is reported as revenue at risk. These rules approximate, but are not, the legal waterfall of each vehicle.

Suppliers are financing their own demand

The map finds the circularity the paper describes, written into filings. NVIDIA holds equity in CoreWeave and has agreed to buy up to $6.3 billion of CoreWeave's unsold capacity. AMD granted OpenAI a warrant on up to 160 million shares that vests as OpenAI buys AMD chips. Google backstops about $4.6 billion of lease payments owed by Fluidstack, a GPU cloud provider with no public balance sheet, to data-center landlords in which Google itself holds equity. None of these arrangements is improper. Each, however, ties the supplier's or backstopper's balance sheet to the same demand it is helping to create.

Most of the system is hard to see from outside

Even for the best-documented deals, holders are hard to see. Registered funds report only about 34 percent of the Hyperion bonds and 3 percent of the El Paso bonds. The rest does not show up in public holdings data. Some is probably held by insurers, whose statutory filings regulators can see but which cannot be publicly linked to the transaction. Some may sit in separate accounts or vehicles that report under other frameworks, or not at the transaction level at all. The problem the map documents is fragmented observability, not total invisibility. Guarantees are the least visible claims: 8 of 13 contingent links have no public amount, and nearly three-quarters of contingent dollars rest on a single estimate (NVIDIA's financing program of up to $125 billion). Altogether, the leases I could trace come to about 4 percent of the $969 billion of hyperscaler lease commitments Moody's estimates. That is the strongest argument for the measures below. A careful search of public filings finds only a sliver of the system, and the sliver it finds is the best-disclosed part.

Charts showing the share of each bond issue reported by registered funds and the share of mapped dollars by source type — Panel (a): share of each bond issue held by registered funds that report their holdings (Form N-PORT). Panel (b): shares within my mapped sample, not the whole market. Tap to enlarge.

Caveats: public disclosures only. Bilateral private credit, syndicate allocations, insurers' private placements and rating-agency reports are mostly outside the map. Contract values and guarantee caps are upper bounds. Stress results depend on stated assumptions about recovery rates and equity buffers, all documented with the map.

What oversight could look like

The goal should not be to stop off-balance-sheet finance. Separating the owners of infrastructure from its users is how railroads, aircraft fleets and pipelines have long been financed, and it draws in capital that hyperscaler balance sheets alone could not supply. The goal is to make the commitments and risk transfers visible, and to make sure the institutions that ultimately hold the risk are capitalized for it. The mapping exercise points to three gaps, and seven measures would go a long way toward closing them.

Gap 1: Who owes whom?

Rebuilding the network showed how hard it is to piece together ownership, funding, contractual obligations and ultimate claims, even for the best-documented deals.

1. Build a look-through map of AI infrastructure credit

Regulators need a single view of who owns each campus and who holds its debt. That view should also show who bears demand, refinancing and residual-value risk, and which tenant all of it ultimately depends on. Two design choices matter most: a common project identifier, in the spirit of the legal entity identifier, and a clear rule on who reports (the sponsor, the issuer and the arranger). The building blocks exist. The FSB's non-bank data task force is working on private-credit data gaps, and its May 2026 report proposes a core set of comparable metrics. The Fed has now started measuring private credit on the national balance sheet. The map above shows how far public data can go and where it stops. The Office of Financial Research is a natural home for a project-level registry covering data-center debt, joint-venture equity, leases, guarantees and GPU-backed lending.

2. Report economic leverage, not just accounting leverage

Hyperscalers should disclose, project by project, their future lease commitments, residual value guarantees, construction guarantees and renewal options. Today most not-yet-commenced leases stay off the balance sheet, and the details are scattered across footnotes, if they appear at all. Two questions should be kept apart. Disclosure, in notes and in management's discussion, is something the SEC can require now through guidance. Recognition on the balance sheet is a question for accounting standard setters, who should revisit how commitments that take effect only in the future, and the guarantees attached to them, are treated. A guarantee that will certainly bind under one of two outcomes is an obligation, whatever its legal start date.

3. Give data-center debt deal-level transparency

Registered securitizations of mortgages and auto loans already carry asset-level disclosure requirements under the SEC's Regulation AB II. Most data-center debt is not registered. The Hyperion bonds were sold as a Rule 144A private placement, and much of the rest is syndicated or private loans. Comparable deal-level information should be available to supervisors as standardized reporting, and to investors through offering documents and post-issuance updates. That means tenant, lease term, renewal dates, residual-value guarantees, power arrangements and the collateral valuation behind the rating. Investors in Beignet at least received a public rating and a rating-agency report. Most bank and private-credit facilities disclose almost nothing publicly.

Gap 2: Who bears the tail risk?

The stress scenarios show how much the allocation of losses depends on guarantees, project equity and recovery assumptions. Supervisors should be running those scenarios with real data rather than leaving them to outsiders with partial maps.

4. Stress-test the joint shock

The scenario that matters is not one tenant defaulting. It is a simultaneous fall in utilization, collateral values and refinancing capacity, which is exactly how an AI demand disappointment would play out. The scenario set should include the failure of a major guarantor, since the analysis above suggests that is where project bondholders are most exposed. Bank supervisors can run this through the exploratory scenarios in their stress-testing programs, and insurance supervisors through their own scenario analysis of insurers' holdings. Both should be able to see which institutions are exposed to the same tenants and collateral. The FSB's 2025 recommendations on non-bank leverage already ask authorities to strengthen how banks manage counterparty risk to leveraged non-banks. AI infrastructure is a timely place to start.

5. Match collateral practice to the life of the asset

Lending against GPUs is not lending against buildings. Advance rates and haircuts on chip-backed facilities should reflect three- to six-year economic lives and the speed of technological change. That is first a matter for lenders' own underwriting. Supervisors should check that banks lending to GPU financiers, and insurers buying their debt, apply haircuts consistent with those lives before any formal prudential rule is considered. They should also flag structures where long-dated debt is serviced by short-lived assets. Vendor-provided residual value guarantees, like NVIDIA's, should be disclosed as contingent liabilities by the guarantor, with the counterparties named.

Gap 3: Who is holding the risk?

The holdings exercise showed how difficult it is to identify the ultimate holders of project debt from public data. Much of that information exists somewhere in the regulatory system, but in pieces that are not linked.

6. Scrutinize private ratings where the risk lands

The FSB flags insurers as rating-reliant investors whose use of private ratings in private credit warrants monitoring. Their capital charges depend on those ratings. Since January 2026 the National Association of Insurance Commissioners has reportedly had stronger authority to challenge private ratings for capital purposes when they diverge from its own analysis, and it is said to be reviewing insurer exposure to data centers. That is the right instinct. Ratings on data-center and GPU-backed debt should come with their obsolescence and residual-value assumptions, so that supervisors and investors can test them.

7. Link the data that already exists

The SEC has pushed the compliance date for the 2024 Form PF amendments to July 2027 while it considers a proposal from April 2026 to revise them. There are reasonable arguments about reporting burden. Whatever version emerges should keep the fund-level information on private-credit leverage, liquidity and sector concentration. Even a complete Form PF would not show who holds a given data-center bond, because it reports fund-level aggregates, not deal-level positions. That is why linking existing sources matters as much as any single form: insurers' statutory holdings, registered-fund and private-fund filings, and bank exposure data, joined through common entity and project identifiers.

None of this requires designating private-credit managers as systemically important. FSOC's March 2026 proposal would make an activities-based review the default first step and reserve entity designation for a higher bar. If that approach is adopted, AI infrastructure finance is a natural activity to review: identify the risk, coordinate the primary regulators, and act on the activity rather than the firm.

There is a serious counterargument. Heavier disclosure and capital requirements could raise the cost of capital for a buildout that may deliver large productivity gains, and could push activity into even less visible corners or offshore. Tenants are investment grade, and many private-credit funds are structured to withstand redemption pressure; the Fed's May 2026 Financial Stability Report judged recent redemption requests manageable. That is why the emphasis above is on measurement and disclosure first, with capital and collateral rules aimed at the places where leverage actually concentrates. The aim is not to price the buildout out of existence but to make sure its risks are priced.

Explore the data

Interactive map: AI Financing Network · Interactive model: AI Buildout Return Model

The map shows every entity and link described here, with each link tied to its source filing and filters for links that are funded, committed or contingent. The model rebuilds the paper's revenue requirement and lets you test how returns change with GPU prices, utilization and chip lifespans. Part 1 covers the paper itself.

None of this shows that the AI buildout is a bubble. What it shows is that the people best placed to judge that cannot yet see the whole picture. A careful outsider working from public filings can trace a few dozen structures and a small fraction of the lease commitments. Supervisors could see far more if the pieces they already hold were linked. That is a modest ask, and the time to make it is before the cycle turns, not after.