Data center bonds shift to project finance as spreads widen 30 bps

Bond investors who bought data-center debt on the strength of Microsoft's name are learning that a lease is not a guarantee.

Investors are repricing AI data center bonds as project finance rather than corporate credit, a shift that pushed spreads on a Microsoft-linked QTS facility 30 basis points wider since April while Microsoft's own bonds barely moved.

"The market is now doing what project-finance investors are meant to do, testing the cash flow and asking who is left holding the risk if the day of refinancing is unkind," Citigroup credit analysts Daniel Sorid and Mathew Jacob wrote in a June 10 note.

The bond at the center of the repricing is a $4.6 billion senior secured note issued by Blackstone-owned QTS Data Centers, due 2036 and secured against a single campus in Fayetteville, Georgia, built to house servers for Microsoft. The deal drew about $12.5 billion of peak demand at issuance and carried a Baa2 rating from Moody's, two notches into investment grade. Its bullet structure — meaning the entire principal falls due at maturity rather than amortizing over time — leaves a refinancing cliff that investors are now pricing explicitly.

The shift matters because the borrowing is still accelerating. Global debt tied to AI is on course to reach roughly $570 billion in 2026, close to double last year, with about $236 billion raised by the end of May, according to a Morgan Stanley forecast. Citi counted five investment-grade data-center bond sales worth more than $50 billion since October. A market expanding that fast on structures that vary widely in quality was always going to start charging for the difference.

The QTS deal was structured as a non-recourse single-asset project bond, the first of its kind in the data-center space under Rule 144A, according to the law firm that advised the banks. For two years, the presence of a large technology tenant pulled project debt toward the pricing of the tenant's own balance sheet. That era is ending. Investors are now reading the document rather than the logo.

Other deals show the same discipline arriving from a different direction. CoreWeave's $8.5 billion delayed-draw term loan, which closed in March, was built around GPU infrastructure tied to named customer commitments and carried an investment-grade rating. Amazon entered into a $17.5 billion senior unsecured delayed-draw term loan on June 8, with Citibank as administrative agent, drawable until September 30 and maturing three years after each draw. That is a direct claim on Amazon itself — a different instrument from a bond secured on a single leased building.

What the repricing means for developers

For operators that issue this debt, the practical message is that bullet structures will face a higher bar. Developers may have to offer more amortization, pay wider spreads, lean on private credit, or write tenant commitments that give bondholders cleaner protection. The build-out still needs enormous amounts of capital, and it will keep getting it. What is changing is the price of convenience.

TeraWulf, the bitcoin miner pivoting to AI infrastructure, is exploring leveraged loans to finance a large-scale buildout after a successful junk bond sale, working with Morgan Stanley and other banks to expand its data center footprint. The company's rapid 24/7 construction at its Lake Mariner site shows how quickly operators are repositioning themselves — and how much capital they need to do it.

The investor takeaway

The repricing carries implications beyond the bond market. If data center developers face higher borrowing costs, the pace of AI infrastructure buildout could slow. Hyperscalers that have relied on off-balance-sheet structures to finance capacity may need to commit more capital directly. Amazon's $17.5 billion term loan is one model; direct corporate borrowing may become more common as project-finance discipline tightens.

For bond investors, the lesson is that a hyperscaler lease can carry a deal a long way, but it can no longer carry it the whole distance. The name on the door is worth less than a claim on the cash — an old rule in lending that the AI build-out is relearning in real time.

This article is for informational purposes only and does not constitute investment advice.