The AI buildout is running into a financing constraint, and it is surfacing in the bond market: nearly half of the bonds issued to fund United States data centers this year carry junk ratings.

The figures come from a Forbes analysis of FactSet data, conducted with input from Columbia Business School professor Stijn Van Nieuwerburgh. Of the roughly $80 billion in bonds issued for US data centers in 2026, approximately $37 billion is rated below BBB — the threshold at which investment grade ends and speculative grade, or junk, begins. The distinction carries practical weight: below-investment-grade bonds generally must pay higher coupons to compensate lenders for default risk, and many institutional investors are restricted to holding investment-grade paper, which can narrow the pool of buyers for junk-rated deals.

The numbers behind the boom

The year-over-year increase is the striking element. In all of 2025, junk-rated data center bond issuance totaled approximately $8.6 billion; this year's approximately $37 billion is more than four times that figure.

The proceeds are flowing into 16 projects tied to some of the biggest names in AI, including Amazon, Meta, and Nvidia. The bonds themselves, however, are not necessarily backed by those companies' balance sheets. The junk ratings reflect the creditworthiness of the developers and financing structures doing the borrowing — a distinct matter from the tenants those developers hope to sign. In this corner of the market, financing terms often hinge on the strength of tenant commitments: leases with highly rated renters are commonly used to support cheaper debt, while projects without them borrow against the developer's own credit profile.

Investors start asking for more

By mid-October 2026, average yields on junk-rated data center bonds had climbed to nearly 10%, with the move reportedly occurring within the span of a single week.

Stress is also visible in corporate debt tied to so-called neoclouds — newer cloud providers that rent out GPU capacity to AI customers. CoreWeave's corporate debt was reportedly trading below 80 cents on the dollar.

Heavily indebted developers Firmus and AirTrunk have postponed or canceled initial public offerings amid the turbulence.

Why tenant quality now matters

The underlying tension lies in a split over who is actually renting these facilities. At one end sit the hyperscalers: giant cloud companies with stable cash flows and strong credit. At the other are riskier developers and younger operators whose revenue depends on AI demand holding up for years.

Layered on top are soaring construction costs and energy supply constraints. The research also notes that broader AI-related debt issuance has reached hundreds of billions of dollars so far this year, with data centers accounting for a significant portion.

What this means for the AI buildout

A near-10% average yield indicates that lenders are no longer treating data center debt as a safe proxy for the AI boom. Higher borrowing costs eat into project economics, and with IPO windows narrowing, lower-rated developers may find capital harder to access at precisely the moment they need it to finish construction.

Three signals are being watched in the coming months: whether junk-rated data center yields stabilize or keep climbing; whether distressed pricing spreads beyond CoreWeave to other neloud borrowers; and whether developers such as Firmus and AirTrunk revive their listing plans or keep them on ice.