How Worried Should We Be About AI Debt?

For much of the past year, alarm about the credit taken on to fund the boom in artificial intelligence has centered on software—a concern I examined in June. The focus is understandable. Private-credit lending to software companies had exceeded $500 billion by the end of 2025, and fears that AI-related debt could crowd out other borrowers have pressured valuations and software-heavy business development companies alike.

This exposure is relatively visible because many private-credit vehicles disclose their portfolios. But other AI infrastructure financing, however, is much harder to observe.

The physical build-out of AI—chips, data centers, power connections, servers, and specialized compute capacity—has mostly been financed off the consolidated balance sheets of the largest technology companies via project and construction financing, equipment-backed lending, leases, asset-backed securities, and, yes, private credit.

There are even loans secured by GPUs, with the chips that run AI themselves pledged as collateral. As it works with a mortgage secured by a house, if a borrower stops paying, its lender takes the chips and sells them. But while most houses hold their value for decades, GPUs are being displaced every two or three years as faster ones are developed, and nobody yet knows what a used one is worth.

Investors commonly measure AI-related leverage by looking at the balance sheets of the largest cloud-computing companies, the five big hyperscalers: Alphabet, Amazon, Meta, Microsoft, and Oracle. For these companies, conventional leverage remains manageable, but that is only one part of the financing system. An institution may hold technology equities, infrastructure funds, private-credit vehicles, securities issued by insurers, and real estate debt, and regard those exposures as diversified. Legally, they are separate investments, but economically, they may share risks related to future AI demand.

In a working paper, I mapped this financing system and performed a stress-test on it. The question I had is how much financing depends on the same underlying cash flows—and where the associated risk ultimately resides.

The results of this test suggest that today’s situation is not a rerun of the 2008–09 financial crisis: First-loss positions sit mostly outside the regulated banking system. But losses could still reach $140 billion and could grow as AI-infrastructure credit expands.

Are the hyperscalers overextended?

Through 2024, the largest technology companies financed most of their aggregate capital spending from operating cash flow. That has changed. The big hyperscalers undertook approximately $380 billion of capital spending in 2025 and are expected to spend roughly double that this year, putting capital spending on course to overtake operating cash flow.

Debt issuance has risen in step. The hyperscalers issued approximately $120 billion of corporate bonds last year, compared to an average of about $28 billion annually between 2020 and 2024. Issuance in the first half of 2026 has already exceeded the 2025 amount.

Let’s be clear: The hyperscalers aren’t in financial difficulty. But the marginal financing of AI infrastructure has begun to extend well beyond their own balance sheets.

A hyperscaler may lease a data center from a developer that has financed the property, construction, and equipment separately, borrowing from a range of private-credit funds, insurers, pension funds, and retail-oriented vehicles.

Commitments also extend beyond funded debt: S&P Global Ratings, the credit rating agency, has identified about $675 billion of signed but not yet commenced lease obligations across the hyperscalers—an amount that’s excluded from the funded-debt totals in my analysis but that demonstrates why reported corporate debt is an incomplete measure.

Where the financing sits

The financing system divides into several broad channels. The most visible is investment-grade corporate debt: My research identifies about $520 billion of senior unsecured hyperscaler bonds held by bond funds, insurers, pension funds, and other institutions at the lowest-risk end of the spectrum. Project and data-center finance adds another $250 billion or so, typically tied to specific facilities and contracted tenant cash flows, with materially shorter maturities. Infrastructure and asset-backed securities account for roughly $60 billion.

Private credit is another significant source of financing, although the market is tough to measure. The approximately $200 billion figure used in my analysis is my estimate rather than an observed total. Specialist compute providers have borrowed about $35 billion secured directly against GPUs and other assets, a structure that carries greater risk.

Vendor financing and credit enhancement are growing in importance. They’re among the least transparent parts of the market and cannot be measured reliably. For example, in September 2025, Nvidia agreed on a $6.3 billion deal to buy AI cloud-computing company CoreWeave’s unsold data-center capacity through early  2032. The supplier underwrote the residual value of its own hardware, converting GPUs into bankable collateral.

Even single projects can be complex. In June 2026, asset management firms Apollo and Blackstone created a Special Purpose Vehicle to raise $35 billion of debt for Anthropic’s computing capacity. The money went not to the AI giant but to buy the chips and lease them back—keeping the debt off Anthropic’s books. Apollo’s financing arm owns the SPV, and the debt is serviced from Anthropic’s lease payments. If those stop, the chips are sold. If they sell for less than the debt outstanding, Broadcom, which manufactures them, has promised to cover the shortfall. That promise let the senior lenders price close to Broadcom’s investment-grade cost.

Investors in the same deal own quite different things. The senior notes are a claim on Broadcom’s promise. The junior notes are a claim on chips. This means Broadcom’s credit rests on building AI chips. Broadcom’s guarantee is weakest in precisely the conditions that trigger it, which is what happened to the bond insurers who built structured credit in the run-up to the 2008–2009 crisis.

If Anthropic fails to honour its commitment, the SPV runs out of cash flow and might be liquidated to repay investors. Its assets are chips, so any liquidation would depend on the resale value of used chips, just as with a car loan, except that chips depreciate faster than cars and if several large AI labs fail to honour their commitments a lot of chips may hit the secondary market at once, depressing prices and recovery values for creditors.

Does diversification still diversify?

Asset allocation classifies investments by legal form or strategy. Public tech equities are separate from infrastructure funds and private credit, while data-center debt can appear under real estate, infrastructure, or structured credit. These classifications are useful for governance and reporting, but they don’t clarify the underlying economic dependencies.

Consider an institution holding a large hyperscaler through stock, a data-center fund through its infrastructure allocation, a private-credit vehicle that finances specialist compute providers, and an insurer with infrastructure-backed securities. Those investments differ in seniority, liquidity, and contractual protection, but their performance may still be linked to a common set of variables: demand for AI, the durability of long-term capacity contracts, the creditworthiness of a few counterparties, the residual value of specialized infrastructure, and the ability of borrowers to refinance at maturity. The same contractual relationship can even appear differently depending on where it sits. A long-term compute agreement may be a backlog to the provider, evidence of future cash flow to a lender, and a purchase commitment to the customer.

Diversification across asset classes isn’t an illusion. But it should be scrutinized for common sources of repayment and common counterparties. A portfolio can hold several distinct legal claims while remaining significantly exposed to one economic assumption, namely that demand for AI computing will grow enough to support the contractual cash flows on which those claims depend.

Does a big equity loss mean a big credit loss?

Events in July illustrated how quickly the assumptions underpinning the AI build-out can change. A Chinese laboratory released a freely available AI model whose performance approached that of leading US systems, while US lawmakers proposed giving the government authority to shut down the most powerful models. Both of these—technological diffusion and regulation—could weaken projected demand for paid AI capacity.

According to my research, a severe debt re-rating could reduce the value of AI-linked equities by about $10 trillion–$14 trillion. This figure is an illustrative stress calibration, not a forecast of either overvaluation or the probability of a correction. That equity loss wouldn’t all become a credit loss. A decline in market capitalization does not cause bonds to default, and a data center with a financially sound tenant can service its debt through a tech-sector sell-off.

The fallout depends on the financing structure. Equity-backed borrowing can react rapidly through loan-to-value tests; asset-backed financing through collateral revaluation and advance rates; and project financing when tenant credit deteriorates, covenants are breached, or a refinancing is required. For strong investment-grade companies, the first effect of a selloff is likely to be wider spreads rather than credit impairment. I calculate this would lead to approximately $60 billion–$140 billion of realized credit loss.

The more important result concerns how losses get distributed. Under the assumptions used in the analysis, first-loss and junior positions sit mainly outside the regulated banking system. The permanent losses would fall first on private-credit funds and dedicated AI-infrastructure credit vehicles. They would come to rest with the limited partners in those funds—principally pension plans and endowments—with private equity–owned insurance platforms, and, to a lesser extent, with private investors of business-development companies.

Banks remain important providers of project and senior secured financing, but their positions tend to sit higher in the capital structure. Traditional insurers and pension funds holding senior asset-backed securities are generally expected to take lower realized losses, though mark-to-market losses could still be substantial. The result would be different from 2008: Aggregate credit losses would be smaller than the accompanying equity loss and far less concentrated in highly leveraged banks.

What investors should ask

In short, AI infrastructure is not a single risk category. And to understand the financing structure beneath each investment, investors should ask five questions.

1. Who ultimately supplies the cash flow? Whether a loan is labeled infrastructure or private credit matters less than if repayment depends on a single tenant or customer.
2. How durable are the commitments? A 10-year contract offers little protection if it can be canceled after 90 days. Termination rights, volume requirements, guarantees, and repricing terms determine what survives stress.
3. What happens at maturity? A borrower may service its debt today but fail to refinance if collateral values fall or tenant credit weakens.
4. Where does the investment sit in the capital structure? Senior bonds, project loans, and residual equipment interests carry different risks.
5. Where else does the portfolio depend on the same cash flow? Exposures reported as infrastructure, private credit, or structured finance may still rely on the same tenants, customers, or contracts.

These questions should prompt limited partners to press for greater disclosure of underlying counterparties and contractual concentrations. They should encourage insurers to take a broader view across affiliated private-credit, structured-credit, and infrastructure portfolios. And boards should consider stress tests that extend beyond direct AI holdings to the financing structures that depend on them.

Why losses could be quiet

Unlike banks in 2008, many AI-infrastructure lenders do not rely heavily on short-term funding. Private loans trade infrequently, project vehicles can be restructured privately, and valuations adjust periodically. Losses would therefore be absorbed largely by fund investors and recognized gradually rather than become visible within days.

A deterioration in AI-infrastructure credit could therefore appear as a sequence of financing events rather than a single crisis: refinancing difficulties, weaker fundraising, lower private valuations, restructurings, delayed projects, and losses recognized gradually across portfolios. That structure may reduce the likelihood of contagion from AI financing to the broader system. It also makes the accumulation and distribution of risk harder to assess in real time.

Don’t conclude that AI investment is excessive or that a correction is inevitable. Technological and investment success are separate propositions. After all, fiber networks built during the technological boom of the late 1990s became indispensable infrastructure, e many of the capital structures that financed their first owners did not survive.

Investors should keep an eye on how much AI-related debt has been issued, as well as identify the cash flows supporting that debt, the structures through which the claims are held, and the investors who ultimately bear the junior risk. The market has good visibility on some of this exposure and limited visibility in other areas. That may matter as much as the amount of debt outstanding.

Stefan Hepp is adjunct assistant professor of entrepreneurship at Chicago Booth and the author of Private Capital: The Complete Guide to Private Markets Investing (Wiley, 2024).