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

A risk transfer lets a bank remain the formal lender while shifting potential losses to outside investors. The bank keeps the loan on its books, but another investor absorbs an agreed share of the riskiest exposure. This matters because banks can reduce concentration without immediately removing the entire loan from their records.

The article describes these transactions as a way to move a loan’s riskiest slice. A bank might arrange financing for an AI data center, then transfer part of the credit risk through a separate deal. The loan still exists, but the bank is less exposed if the borrower or project performs badly.

Here, JPMorgan, Morgan Stanley, SMBC and MUFG are exploring significant risk transfers because they are approaching internal limits on AI-linked exposure. The move shows that banks are hedging their own financing bet. It also signals that outside investors are being asked to price risks that banks no longer want concentrated on their balance sheets.

There is no single total for banks’ current AI-debt exposure in the article. The clearest broad figure is $120 billion of AI infrastructure spending spread across special purpose vehicles. Those vehicles were designed to keep debt away from companies’ own balance sheets, leaving financing banks exposed instead.

The most concrete example is a group of banks, including JPMorgan and MUFG, trying to distribute $38 billion of construction debt tied to Texas and Wisconsin data centers leased to Oracle. Oracle is also targeting more than $40 billion in additional debt and equity this fiscal year. It previously sold $18 billion of bonds in one day.

These figures are related, but they should not be added together as one exposure total. The article identifies several large funding pools, not a complete banking-industry balance sheet. The scale is still significant: banks are financing projects and corporate borrowing large enough to push against internal sector limits.

JPMorgan, Morgan Stanley, SMBC and MUFG helped finance the AI buildout, but their exposure has grown faster than their internal limits allow. The article says they are exploring risk transfers because too much of each bank’s balance sheet is tied to one sector. Selling the riskiest slices helps spread that risk.

The mechanism is straightforward. A bank underwrites a loan, keeps it legally on its books, and transfers part of the potential loss to outside investors. That can free room for other lending while reducing the damage from a weak project or borrower. Some lenders are even willing to sell pieces at a discount to move them.

The timing shows a change in confidence and capacity. Six months earlier, banks were competing to underwrite loans tied to Oracle, Meta and other AI companies. Now they are trying to distribute debt that they never intended to hold forever. Their actions suggest caution about concentration, pricing and the ability of investors to absorb more AI risk.

The $38 billion deal is difficult to distribute because the market’s appetite for AI-linked debt has weakened. Investors have become more selective, while banks are holding unusually large construction loans tied to a single AI-related project. The article says the distribution effort has lasted more than six months.

The project involves Texas and Wisconsin data centers leased to Oracle. Oracle’s five-year credit default swaps rose to about 151 basis points in early October, and S&P cut its rating to BBB-, one notch above junk. Oracle is also spending heavily while its AI investment outpaces cash flow. Those developments make the project harder to price.

Demand has weakened across AI corporate debt. Order books for major technology issuers shrank from about five times the amount offered in February to under two times by July. As investors demand more compensation, some lenders are willing to sell parts of the Oracle-related debt at a discount rather than keep waiting.

When investors demand discounts or higher interest rates, borrowers must pay more to raise the same amount of money. A discount reduces the cash received from selling debt. A higher yield raises future interest payments. Either way, the cost of building and financing a data center increases.

The article gives a clear market comparison. CoreWeave-leased paper prices at yields from 7.00% to 9.875%, while bonds tied to hyperscalers command 5.70% to 7.875%. The 10-year Treasury yield also reached 5.304%, making the risk-free base for borrowing more expensive before company-specific risk is added.

That pressure can narrow projects’ returns and make lenders more cautious. Capital is already rotating toward data centers leased to investment-grade Microsoft and Google, away from weaker credits. Future construction may therefore require stronger tenants, more equity, higher promised returns or discounts. The article does not state specific cancellations, but financing conditions are clearly becoming tougher.

Investors view Microsoft- and Google-linked debt as safer because the article identifies them as investment-grade hyperscalers. A stronger tenant generally gives lenders more confidence that lease payments and project financing can continue. That reduces the return investors demand compared with weaker credits.

The pricing difference makes the mechanism visible. CoreWeave-leased paper yields between 7.00% and 9.875%. Bonds leased to hyperscalers yield between 5.70% and 7.875%. The higher yield on CoreWeave-linked debt is compensation for greater perceived credit risk, according to the market data cited in the article.

CoreWeave is described as smaller and less capitalized than Oracle, while the article emphasizes the balance sheets and existing customer bases available to larger companies. Investors are therefore rotating toward stronger tenants and away from weaker credits. A smaller AI infrastructure company raising debt next quarter must compete for the same cautious lenders, but pay more to attract them.

Interest rates set the basic cost of borrowing. When the 10-year Treasury yield rises, companies pay more on top of that baseline to finance risky projects. Credit ratings then signal how much additional compensation lenders may demand. Stronger ratings usually support cheaper debt, while a downgrade can narrow investor demand.

Oracle illustrates the interaction. S&P cut its rating to BBB-, one notch above junk, while Oracle’s AI spending outpaced its cash flow. Its five-year credit default swaps climbed to around 151 basis points. At the same time, Oracle reportedly sought more than $40 billion in additional debt and equity. Heavy borrowing becomes harder to sustain when cash generation lags spending.

For infrastructure, safe borrowing depends on whether expected cash flow can cover interest and debt payments. The article does not provide a specific coverage ratio or safety threshold. It does show the warning signs: higher market rates, weaker ratings, rising insurance costs and investors demanding greater yields for AI-linked debt.