Banks Rush to Offload AI Data Center Debt They Just Finished Underwriting

JPMorgan, Morgan Stanley, SMBC, and MUFG are exploring significant risk transfers to offload portions of the AI debt pile they themselves arranged. The same banks that spent months competing to underwrite data center financing now want someone else to hold the risk. Meanwhile, AI-related debt issuance fell from $113 billion in June to $23 billion in September.

TL;DR: Major banks including JPMorgan, Morgan Stanley, SMBC, and MUFG are exploring significant risk transfers (SRTs) to shift some credit risk from AI-related lending. Deutsche Bank is reportedly working on an SRT tied to a project-finance book that includes data-center loans. AI debt issuance fell from $113 billion in June to $23 billion in September; an Oracle-linked loan package was quoted below face value; and Taiwanese lenders made NT$4 trillion (about $126 billion) in loans in the first seven months of 2026, with much of that money flowing to the AI supply chain.

Why Are Banks Suddenly Trying to Offload AI Data Center Debt?

Banks are offloading this debt because their own exposure limits are biting. Sources report that JPMorgan, Morgan Stanley, SMBC, and MUFG have been exploring significant risk transfers to move portions of their artificial intelligence debt pile off their own balance sheets. The debt in question is the same paper these institutions recently underwrote for data center construction.

The mechanics are straightforward. Under a significant risk transfer, a bank packages loans and shifts a portion of the credit risk to outside investors. Bloomberg reported on September 7, 2026 that Deutsche Bank was working on an SRT tied to about €2 billion of project-finance debt, including data-center loans. The report described work in progress, not a completed transaction, and did not say the entire book consisted of data-center loans.

Why does this matter now? Because concentration. A handful of lenders absorbed enormous AI-related exposure in a short window, and internal or regulatory limits cap how much more they can hold. Bloomberg reported in December 2025 that Morgan Stanley had held preliminary talks with investors about offloading some of the credit risk tied to its data-center lending activities. The pattern is consistent: the boom produced balance sheets that are fuller than the banks want them to be.

Is this a sign of distress, or simple prudence? Sources suggest both dynamics are in play, with banks acknowledging financing risks while continuing to fund deals.

Which Banks Are Exploring Significant Risk Transfers?

The names are the biggest in the business. JPMorgan, Morgan Stanley, SMBC, and MUFG have all been exploring significant risk transfers to shed portions of the AI debt they originated. Deutsche Bank is also reported to be working on an SRT tied to about €2 billion of project-finance debt that includes data-center loans; it has not been reported as a completed deal.

Morgan Stanley’s case is instructive. The bank has held preliminary talks with investors about offloading some of the credit risk associated with its data-center lending. Preliminary is the operative word — these are early-stage conversations, not completed transactions, according to Bloomberg’s reporting.

The reported Deutsche Bank plan illustrates one possible approach. Rather than selling loans outright, an SRT can transfer part of the credit risk while the bank retains the lending relationship. The reported €2 billion is the size of the project-finance book, which includes data-center loans, not a claim that the full amount is data-center debt.

There is also a separate bond-market takeout route, not an SRT. A group led by Morgan Stanley reportedly plans to refinance a $15 billion debt package for a Google-backed, Anthropic-leased Texas data-center campus through bond sales after the loans are drawn, according to Financial Times reporting republished by Yahoo Finance. That is a plan, not a completed sale. It is another way lenders may reduce exposure while retaining the client relationship.

How Has the AI Debt Market Cooled in Recent Months?

The cooldown has been sharp and fast. AI debt issuance fell from $113 billion in June to $23 billion in September — a drop of roughly 80% over three months, according to Coinpaper. The decline was sharp, but issuance did not stop.

What changed? Bond markets are getting tough for lower-rated borrowers, sources report. One example stands out: CleanSpark, which is developing a data center for Meta Platforms, had to offer investors big concessions to land financing, according to The Information. When a borrower with a Meta anchor tenant has to sweeten terms to sell bonds, pricing power has clearly shifted toward investors.

This cooling creates the context for the banks’ behavior. Issuance at $23 billion versus $113 billion means less new fee business to go around, while existing exposure remains on balance sheets. The incentive to move risk off the books grows precisely when the market for new deals shrinks. The change from the June peak to September took three months.

What Is Happening With Oracle-Linked Data Center Loans?

Oracle-linked loans were quoted below face value. Coinpaper reported that about $18 billion in loans backing Project Jupiter, a New Mexico data-center campus associated with Oracle, were quoted at 89–91 cents on the dollar in September, while banks struggled to distribute them.

That combination points to investor caution: the reported loans were quoted at a discount to face value. The $18 billion figure describes the size of the loan package; the reporting does not establish $18 billion in completed secondary-market trades.

For banks exploring significant risk transfers, this is relevant market context. SRT pricing and investor appetite depend on how the underlying credits are perceived. When a major data-center loan package is quoted below par, hedging conversations may become harder and more expensive.

Does a below-par price mean default is expected? Not necessarily — discounts can reflect illiquidity, tenor, and sector sentiment as much as default probability. But it does mean the market is no longer treating AI data center debt as an automatic pass. The buffer between underwriting assumptions and market reality has narrowed visibly.

How Much Money Is Actually Involved in This Buildout?

The numbers are large, but the categories matter. One CryptoBriefing estimate puts exposure to AI-adjacent sectors at around $800 billion. It includes traditional bank lending as well as money flowing through private-credit structures and dedicated project vehicles; it is broader than direct bank loans to AI data centers alone. The same report says executives did not signal an immediate retreat and cites fee income as a reason to continue financing. This estimate provides context for risk transfers, but it does not mean banks are unable to stop funding the sector.

Asia’s contribution is also significant. Taiwanese lenders extended a record NT$4 trillion — approximately $126 billion — in loans during the first seven months of 2026, with much of that money flowing toward the AI supply chain, according to CryptoBriefing. This is an aggregate lending figure, not an AI-only loan total. CryptoBriefing describes the region’s banks as having found their favorite customer, one that runs on GPUs.

Individual deals also illustrate the scale: lenders on a Google-backed, Anthropic-leased Texas data-center project reportedly plan to refinance a $15 billion package through bond sales after drawdown. Separately, Deutsche Bank is working on an SRT tied to a €2 billion project-finance book that includes data-center loans.

Why does the aggregate matter? Because it shows how financing is spread across several channels. The biggest players in the industry need massive capital infusions, whether the label is artificial intelligence or super intelligence. Broad estimates of AI-adjacent exposure, Taiwan’s record aggregate lending, and a separately reported Oracle-linked loan package quoted below par all help explain why the financing structure of the buildout — not just the technology — has become part of the story.

What Role Are Asian Banks Playing in AI Financing?

Asian banks have become some of the largest funders of the AI build-out, according to sources. Taiwanese lenders extended a record NT$4 trillion — roughly $126 billion — in loans during the first seven months of 2026. The reporting says much of that money flowed to the AI supply chain; it does not characterize the entire total as AI lending.

That scale matters. Japanese institutions SMBC and MUFG appear alongside JPMorgan and Morgan Stanley in reports of banks exploring significant risk transfers to shed portions of their AI debt piles.

So the pattern is global. Taiwan’s reported lending growth includes activity flowing to the AI supply chain, while Japanese institutions SMBC and MUFG appear among banks exploring risk transfers. The customer, as one source put it, “runs on GPUs.” Whether Asian regulators and investors tolerate that exposure as patiently as the banks hope remains an open question for the rest of the cycle.

Why Did Banks Underwrite This Debt in the First Place?

Competition and fee income explain most of it. Sources describe the biggest industry players as needing “massive capital infusions,” whether the goal is framed as artificial intelligence or superintelligence. Banks that refused to underwrite simply lost the mandates to rivals who said yes.

The result is a tension, not an inability to exit. One estimate puts exposure to AI-adjacent sectors around $800 billion, including traditional bank lending and money flowing through private-credit structures and project vehicles. The cited reporting says executives did not signal an immediate retreat and pointed to fee income; it does not say the exposure makes continued funding impossible.

There is also the classic underwriting incentive at work. Banks earn fees at closing; losses arrive years later, if at all. The current offloading effort suggests many lenders now judge the risk heavier than they priced it — but the commitments were already signed. The underwriting happened. The second-guessing came after.

Are Lower-Rated Borrowers Already Feeling the Squeeze?

Yes, and the evidence is concrete. AI debt issuance fell from $113 billion in June to $23 billion in September, a drop of roughly 80% over three months. At the same time, an Oracle-linked loan package of about $18 billion was quoted at 89–91 cents on the dollar in September, according to Coinpaper.

Bond markets are getting tougher on some borrowers. One example cited by The Information: CleanSpark, which is developing a data center for Meta Platforms, had to offer investors big concessions to land financing. This is evidence of more expensive capital, not by itself proof that CleanSpark stopped construction.

The mechanism is straightforward. When refinancing windows narrow, weaker credits pay more or walk away empty-handed. The borrowers least able to absorb higher coupons are the ones building the physical infrastructure. That is where stress shows up first.

How Are European Lenders Hedging Their Exposure?

Through significant risk transfers, among other tools. Bloomberg reported that Deutsche Bank was working on an SRT tied to about €2 billion of project-finance debt, including data-center loans. The report described a transaction in progress, not a completed hedge, and did not say the full book was data-center debt.

An SRT works by shifting a portion of the credit risk on a loan portfolio to investors — often through credit protections — while the bank keeps the client relationship and the assets on its books. The lender reduces regulatory capital requirements; investors earn yield for taking the tail risk.

For AI data center debt specifically, the appeal is obvious. Banks can keep servicing hot borrowers without concentrating balance-sheet risk in a single, unproven asset class. Whether SRT investors fully understand what they are buying in a downturn is the question European regulators will eventually have to answer.

What Could a Broad Retreat From AI Debt Financing Look Like?

The issuance numbers show a sharp slowdown: AI-related debt fell from $113 billion in June to $23 billion in September, a three-month change. Separately, an Oracle-linked loan package of about $18 billion was quoted below par. If those conditions persist, refinancing becomes the pressure point: maturing loans either reprice sharply higher or fail to roll.

A broad retreat is only one possible scenario, not an established sequence. CleanSpark reportedly had to offer investors concessions for Meta-linked data-center financing. Separately, several banks are exploring or working on SRTs, while a group led by Morgan Stanley plans to refinance a $15 billion loan package for a Google-backed, Anthropic-leased Texas campus through bond sales after drawdown. That planned bond-market takeout is separate from an SRT and is not a completed sale.

The banks that underwrote the boom would, in effect, be passing the risk to SRT investors and bond buyers. Whether those buyers exist in sufficient size at acceptable prices is the central uncertainty of the next phase.

Frequently Asked Questions

What is a significant risk transfer (SRT)?

An SRT is a transaction in which a bank shifts part of the credit risk on a loan portfolio to outside investors, typically reducing the capital the bank must hold against those loans. Bloomberg reported that Deutsche Bank was working on an SRT tied to about €2 billion of project-finance debt, including data-center loans; the report did not describe it as completed.

Which banks have been reported as exploring AI debt offloading?

JPMorgan, Morgan Stanley, SMBC, and MUFG have been reported exploring significant risk transfers to offload portions of their AI debt piles. Morgan Stanley has held preliminary talks with investors about offloading some credit risk from its data-center lending, and Deutsche Bank is reportedly working on an SRT tied to a €2 billion project-finance book that includes data-center loans. Separately, banks led by Morgan Stanley reportedly plan to refinance a $15 billion loan package for a Google-backed, Anthropic-leased Texas campus through bond sales after drawdown; that is a planned takeout, not a completed sale or an SRT.

How sharply has AI debt issuance fallen?

AI debt issuance fell from $113 billion in June to $23 billion in September, a decline of roughly 80% over three months. In September, an Oracle-linked loan package of about $18 billion was quoted at 89–91 cents on the dollar, and CleanSpark reportedly offered investors concessions to secure financing.

Why are banks continuing to fund AI despite acknowledged risks?

One estimate puts exposure to AI-adjacent sectors around $800 billion, including traditional lending and funding through private-credit structures and project vehicles; it is not a measure of direct bank loans to AI data centers alone. The cited reporting says executives did not signal an immediate retreat and pointed to fee income, but it does not say banks are unable to stop financing. SRTs are one way lenders can reduce concentration while retaining lending relationships.

Summary

Major banks — JPMorgan, Morgan Stanley, SMBC, MUFG, and Deutsche Bank — are exploring significant risk transfers to shed AI data center debt they recently underwrote.
New AI debt issuance fell from $113 billion in June to $23 billion in September; an Oracle-linked loan package of about $18 billion was quoted below par.
CleanSpark reportedly offered investors concessions for its Meta-linked data-center financing, an example of more expensive capital.
Taiwanese lenders made NT$4 trillion (~$126 billion) in loans in the first seven months of 2026, with much of that money flowing to the AI supply chain; the total is not AI-only lending.
An estimate of exposure to AI-adjacent sectors is about $800 billion and includes nonbank funding channels; it is not direct bank lending alone. Watch the SRT market and secondary loan pricing for the answer.