JPMorgan, MUFG and SMBC Are Trying to Sell the AI Debt They Just Wrote

JPMorgan, Morgan Stanley, SMBC and MUFG are shopping pieces of the AI debt pile they helped build, according to reporting compiled by Startup Fortune from Financial Times and Bloomberg coverage. The instrument of choice is the significant risk transfer, a deal that keeps a loan on the bank’s books while moving its riskiest slice to outside investors. The reason is mechanical: the lenders closest to the money are bumping against internal limits on how much AI-linked exposure any one institution can hold.

That is the news. Six months ago the same banks were competing to underwrite data center loans tied to Oracle and Meta. Now they want out, and the market is not making it easy.

The $38 billion that will not move

The clearest evidence is a single syndication. A group of banks including JPMorgan and MUFG has spent more than six months trying to distribute $38 billion in construction debt tied to a Texas and Wisconsin data center project leased to Oracle, per Bloomberg. Some lenders are now willing to sell pieces at a discount to non-bank buyers just to clear the position. Six months is a long time to hold debt nobody underwrote to hold forever.

Matthew Moniot, co-head of credit risk sharing at Man Group, told the Financial Times the deal sizes are “out of scale to anything we’ve thought about, ever.” Read that carefully. This is not a fund manager talking down a trade he wants no part of. It is someone whose job is pricing exactly this risk saying the numbers no longer fit any model he has used.

Then look at Oracle itself. Its five-year credit default swaps, the cost of insuring against default, climbed to around 151 basis points in early October, levels last seen in 2009, according to market data cited by Bloomberg. A separate report puts the CDS spread at 203 basis points, an 18-year high, after S&P cut Oracle’s rating to BBB- in July. That is one notch above junk. Oracle is reportedly targeting more than $40 billion in additional debt and equity this fiscal year, on top of the $18 billion in bonds it sold in a single day last September, to keep funding its Texas, Wisconsin and New Mexico sites. It is also suing Wisconsin’s Public Service Commission over a rule that could force it to post more than $7 billion in collateral for its Port Washington campus.

A company spending faster than it earns, litigating with a state utility regulator, and paying 2009 prices to insure its debt is not a company the credit market is relaxed about.

The demand side thinned out first

The banks are not reacting to nothing. Bonds from Alphabet, Amazon, Meta, Microsoft and Oracle now trade at a premium to similarly rated non-AI corporate debt. Order books for their offerings have shrunk from roughly five times the amount on sale in February to under two times by July, per Bloomberg. Meta’s latest bond drew a third less investor demand than its sale a year earlier.

Banks do not sell down debt they are confident in. They sell down debt when too much of the balance sheet has become one bet on one sector.

The macro backdrop compounds it. The 10-year Treasury yield hit 5.304% this month, its highest since 2002, after a $39 billion auction cleared at 5.30%, the highest auction yield since 2000. The Federal Reserve’s September meeting minutes named AI-related corporate borrowing as a contributor. Capital Economics argues the real drivers are inflation and deficits, not tech bond sales. Either way, the risk-free rate that every data center financing is priced against just got more expensive.

The SPV trick is coming back around

Here is the part worth sitting with. Spreading roughly $120 billion of AI infrastructure spending across special purpose vehicles was supposed to make this risk easier to swallow. Oracle, Meta, xAI and CoreWeave have all used the structure to keep debt off their own balance sheets. Instead the risk is landing at the banks that financed the SPVs in the first place, and they are now trying to pass it further downstream.

The investors still willing to play have gotten picky about where. Capital is rotating toward data centers leased to investment-grade hyperscalers like Microsoft and Google and away from weaker credits. That is why CoreWeave-leased paper prices at 7.00% to 9.875%, against 5.70% to 7.875% for hyperscaler-leased bonds, per bond market data compiled by CCIR. A smaller, less capitalized AI infrastructure company raising debt next quarter now competes for the same nervous pool of lenders as Oracle, without Oracle’s balance sheet or customer base to point to.

What this means for anyone training a model

The AI boom runs on an assumption that credit stays open indefinitely at reasonable terms. Six months of a bank failing to place $38 billion in loans is the clearest evidence yet that the assumption is being tested now, not in some theoretical downturn years out.

For AI builders the transmission channel is straightforward. Compute is financed. If the cost of that financing rises, or if lenders start discriminating between hyperscaler-adjacent capacity and everything else, then the price of GPU hours, reserved capacity, and long-term training contracts moves with it. A lab that signed a multi-year compute deal at 2025 prices may find the counterparty renegotiating, or the SPV behind the deal struggling to syndicate.

Watch three things. Whether the $38 billion Oracle-linked syndication clears, and at what discount. Whether Oracle’s CDS spread holds above 200 basis points or widens further. And whether the next tier of AI infrastructure borrowers, the neoclouds and the model labs leasing from them, can still get a deal done at all. The banks have already answered the question for themselves. They are trying to sell.

{/* TODO: comment sought from JPMorgan, MUFG, SMBC, Morgan Stanley, and Oracle */}