AI Infrastructure Is Becoming a Credit Story

For the last several years, most of the AI conversation has centered on models, chips, software, and capability. Increasingly, the more consequential question is becoming: who is financing the physical infrastructure underneath AI, and who ultimately carries the risk?

A recent Brookings Papers on Economic Activity analysis by Columbia University professor Stijn Van Nieuwerburgh projects that U.S. AI investment in data center buildings, power systems, networking infrastructure, specialized chips, and related equipment could total $10.3 trillion from 2025 through 2032, averaging 3.63% of U.S. GDP per year. The paper argues that the buildout would be larger relative to the economy than historic U.S. canal, railroad, electrification, highway, and telecommunications investment booms.

That is an extraordinary amount of capital. But the more interesting development may be how the financing is changing.

From balance-sheet spending to a broader capital stack

Early AI infrastructure spending could largely be absorbed by the enormous operating cash flows and balance sheets of the hyperscalers. Morgan Stanley estimates roughly $2.9 trillion of global data-center-related capex through 2028. Its broader framework assumes about $1.4 trillion can be covered by hyperscaler cash flows, leaving an estimated $1.5 trillion to come from outside capital.

That means the structure is moving beyond:

Big Tech spends its own cash.

And toward a more familiar infrastructure-finance chain: banks lend, private-credit funds lend, special-purpose entities borrow, assets are securitized, developers assume completion risk, tenants sign long-term commitments, and investors buy the paper.

The technology may be revolutionary. The financing is still finance.

The underlying technology can be unprecedented while the financial mechanics remain familiar. Leverage still matters. Counterparty risk still matters. Completion guarantees still matter. Tenant concentration still matters. Power availability still matters. Construction overruns still matter. And someone always has to absorb the loss when assumptions fail.

Morgan Stanley’s estimate of the $1.5 trillion external financing gap illustrates how quickly the capital stack is broadening. Its framework assigns the largest share to private credit and asset-based finance, with additional funding from corporate debt, securitized credit, and a mix of sovereign, private-equity, venture, and other capital.

Debt is not the thesis. Credit selection is.

Andreessen Horowitz’s September 2026 State of Markets II offers a useful counterweight to the idea that rising debt automatically makes the AI buildout uneconomic. Its analysis argues that hyperscaler capex is consuming free cash flow and bringing debt markets more deeply into the buildout, while hyperscaler returns on invested capital still remain above their cost of borrowing. At that level of the stack, borrowing can still be economically rational.

That does not eliminate credit risk. It makes the underwriting question more specific. The issue is increasingly which projects, counterparties, and capital structures can support the debt through construction, power delivery, utilization, refinancing, and changing compute economics.

AI infrastructure is becoming a credit-selection story.

Risk is migrating, not disappearing

That migration is already visible. SoftBank recently launched more than $11 billion of high-yield bonds, with expected yields around 9% to 10%, in part to fund its OpenAI investment. The financing is notable not because one bond sale proves a systemic problem, but because it shows how AI investment is reaching farther into leveraged capital markets.

At the same time, Oracle-linked infrastructure financing has become a useful case study in how execution risk and credit risk can meet. Reuters reported that roughly $18 billion of loans tied to Project Jupiter, a New Mexico data-center development, were trading below par amid concerns about regulatory delays and Oracle’s growing debt exposure. Bloomberg separately reported that Oracle sent the project developer a force-majeure notice designed to protect Oracle against payment obligations if delays prevent the project from coming online as planned.

None of these developments individually proves that AI infrastructure financing is unstable. Together, however, they demonstrate something important:

The AI boom is no longer confined to technology-company balance sheets.

It is spreading into credit markets, project finance, private capital, structured finance, real estate, utilities, and construction. That changes the risk map.

October 2026 update: the credit market is repricing the risk

New reporting puts market prices behind the broader thesis. In an October 16 Forbes investigation, Giacomo Tognini reports that U.S. data-center-related bond issuance reached approximately $80 billion in 2026 year to date, of which $37 billion, or about 46%, was below investment grade. Barclays data cited by Forbes put 2025 issuance at approximately $36 billion, including $8.6 billion of junk-rated bonds, or about 24%.

The distinction matters: more issuance below investment grade is not the same as every bond becoming distressed. But it does show that a larger share of the financing is being placed in riskier credit categories, with a greater dependence on investors willing to accept higher default risk and financing costs.

Forbes also reports that approximately $4 billion of CoreWeave corporate debt was trading below 80 cents on the dollar, while yields on bonds financing several data centers leased by CoreWeave averaged roughly 11% following a selloff. Its analysis of FactSet data identifies 16 data-center projects financed by below-investment-grade bonds and leased to, or backed by, a range of major technology firms and neocloud operators.

These are different exposures. A neocloud's corporate debt, a project-level bond backed by lease cash flows, and an investment-grade hyperscaler's direct obligations should not be treated as interchangeable. A familiar tenant or investor may improve demand visibility without providing a legal guarantee of the debt.

The underwriting pressure point is the mismatch between the economic life of compute equipment, the durability of customer demand, the cost and timing of hardware replacement, and the term of financing obligations. In a leveraged neocloud model, long-lived lease or financing commitments may be supported by shorter-duration compute contracts and revenue assumptions that can change quickly. That creates refinancing and residual-value risk even when AI adoption continues to grow.

The relevant stress test is not simply whether contracted revenue covers today's interest expense. It is whether cash flow still supports debt service after utilization declines, GPU pricing resets, replacement capex, power or construction delays, and higher refinancing rates. Lenders also need to identify exactly which entity bears each obligation and what contractual remedies exist if the project is delayed or the customer weakens.

The market is beginning to distinguish demand for AI compute from the credit quality of the structures financing its delivery.

That is a more useful signal than treating the entire AI buildout as either a guaranteed success or an inevitable collapse. Read the October 16 Forbes investigation.

The bottleneck is not just capital. It is energized capacity.

Even when financing is available, projects still depend on generation, transmission, substations, interconnection, permitting, equipment, labor, fiber, and construction execution. As AI-driven power demand rises, constraints in one layer can create investment opportunity in another.

That is particularly interesting in the mid-market, where fragmented infrastructure, utility carve-outs, behind-the-meter generation, and operationally complex assets can become strategically valuable precisely because the larger system is constrained.

a16z makes the same physical constraint visible from another direction: its market work argues that demand has spilled beyond chips into power, networking, skilled trades, manufacturing, and other capital-intensive infrastructure, while noting that even when compute hardware is available, powering it can remain an open question. That is why energized capacity matters as much as financial capacity.

Before credit spreads move, the municipal record may already be warning you

Financing is only one part of the development timeline. A data-center project may have a committed lender, an anchor tenant, and a credible construction budget, yet still miss its commercial operation date because local approvals, utility agreements, or community objections delay the work.

Those risks often surface first at the local level: planning commission hearings, city council agendas, zoning amendments, water and power discussions, development agreements, environmental reviews, and requests for moratoriums. By the time a delay becomes visible in a lender update or bond price, the underlying conflict may have been documented publicly for months.

Resources such as LocalGovIndex’s data-center issue page offer a way to look for local-government signals alongside conventional financial reporting. Such records are potential due-diligence inputs, not proof that a particular project is delayed or impaired. Coverage, timeliness, and the underlying primary documents should be verified jurisdiction by jurisdiction.

For lenders, developers, contractors, and CFOs, that suggests adding an explicit entitlement-and-energization review to the credit file: Which permits and zoning decisions remain outstanding? What public commitments have been made on power, water, roads, and infrastructure? Are there active objections or pending hearings? Which approvals sit on the project's critical path, and who bears the financial consequences if they slip?

The earliest signal of a future credit problem may be in a municipal meeting packet, not a bond quote.

Local-government intelligence does not replace engineering, legal, or credit underwriting. It can, however, help investors and operators identify execution risk while there is still time to reprice it, mitigate it, or decline the exposure.

For builders, lenders, and investors, the logo is not the credit analysis

A hyperscaler or major AI company may sit somewhere in the economic chain. That does not necessarily mean it is your direct obligor.

The underwriting questions become much more traditional:

• Who owns the project?
• Who borrowed the money?
• Who signed the lease?
• Who guarantees completion?
• Who absorbs cost overruns?
• What happens if the tenant delays occupancy?
• What happens if power delivery slips?
• What happens if compute economics change before the debt matures?
• What collateral actually supports the financing?

Those questions can get obscured when an enormous technology brand sits somewhere near the project. But a famous tenant is not the same thing as payment security.

AI data centers should increasingly be analyzed for what they are: large, capital-intensive infrastructure projects with technology-demand risk layered on top.

The opportunity is enormous

None of this means the buildout is irrational. Quite the opposite. AI may become one of the most important infrastructure investment cycles in modern history.

The opportunity spans power generation, transmission, substations, fiber, cooling, construction, semiconductors, networking, real estate, equipment finance, private credit, and project finance.

The mistake would be assuming that enormous opportunity eliminates financial risk. Usually it does the opposite. The larger the capital requirement becomes, the more important underwriting becomes.

AI may be new. Capital cycles are not.

We have seen versions of this story before. Transformational infrastructure attracts capital. Capital attracts leverage. Leverage enables faster construction. Faster construction creates excess capacity somewhere. Then markets discover which projects were economically durable and which were built on assumptions that did not survive contact with reality.

AI may ultimately justify every dollar invested and more. Or some parts of the capital stack may be significantly overbuilt before demand catches up.

Both can be true:

AI can transform the economy, and some AI infrastructure investments can still lose money.

That is the distinction I think executives, CFOs, lenders, builders, and investors should be watching.

The AI conversation is rapidly evolving from What can the technology do? to Who is financing what it takes to do it, and who owns the downside?

That may become one of the most important finance questions of the AI era.

Sources

• LocalGovIndex — Data Centers: local-government issues and records (research starting point; verify primary records)
• Forbes, Giacomo Tognini (October 16, 2026) — Why Debt-Laden Neoclouds Could Become a Trillion Dollar WeWork 2.0
• Brookings Papers on Economic Activity — Financing the AI buildout
• Andreessen Horowitz — State of Markets II — full presentation
• Morgan Stanley Research — Bridging the Data Center Financing Gap
• Morgan Stanley — Credit Markets’ Role in the AI Financing Gap
• Financial Times — SoftBank high-yield bond financing for OpenAI investment
• Reuters — Oracle-linked Project Jupiter debt under pressure
• Bloomberg — Oracle force-majeure notice on Project Jupiter