Banks are now trying to offload the AI data center debt they just underwrote

Major banks that eagerly underwrote huge construction loans for AI data centers are now quietly trying to shed that exposure. JPMorgan, Morgan Stanley, SMBC and MUFG have been marketing significant risk transfers - structures that let them keep loans on their books while selling the riskiest slices to outside investors - because they are hitting internal concentration limits. One stalled effort involves more than $38 billion of construction debt for a Texas and Wisconsin campus leased to Oracle; lenders have spent six months trying to place pieces of that package and are even willing to sell at discounts. Oracle’s financing stress is a key driver: its five‑year CDS widened sharply, S&P cut the rating to BBB‑, and the company is pursuing tens of billions more in debt and equity after massive bond sales and a dispute in Wisconsin that could require large collateral postings.

Market conditions that once fueled the AI buildout have shifted, thinning demand and repricing risk. Order books for hyperscaler and AI-linked bond offerings have shrunk, investors are rotating toward data centers leased to investment‑grade companies like Microsoft and Google, and weaker credits trade at materially higher yields (e.g., CoreWeave‑leased paper at roughly 7-9.9% versus 5.7-7.9% for hyperscaler‑leased). Rising Treasury yields and tougher placement dynamics mean banks are no longer comfortable holding concentrated AI infrastructure exposure; the special‑purpose vehicle model moved risk off corporate balance sheets but funneled it back to the lenders, testing the assumption that cheap, ample credit would keep funding the AI buildout indefinitely.