Wall Street Wrote the AI Boom's Checks. Now It Wants Out.

📖 Read the full version with charts and embedded sources on ComputeLeap →

The banks that underwrote the largest infrastructure buildout since the transcontinental railroad are trying to leave the table before the cards are turned. JPMorgan Chase, Morgan Stanley, SMBC, and MUFG — the institutions that assembled tens of billions in AI data-center construction debt — are now actively seeking to offload that exposure to anyone willing to take it. Rising rates, a market selloff, and binding concentration limits have turned what was supposed to be lucrative deal-making into a balance-sheet problem.

View original post on X →

This is not a story about whether AI is overhyped. The banks are not making a call on AI's future. They are making a call on their own exposure — and the distinction matters enormously. When the financial plumbing of a trillion-dollar buildout starts leaking, the question shifts from "will AI work?" to "who pays when the construction bill comes due?"

The Numbers Behind the Exits

The scale of AI infrastructure financing has broken records that the industry did not know existed. Data center financing reached approximately $56 billion year-to-date in 2026, compared with $26 billion for all of 2025. Hyperscaler capex guidance for 2026 sits around $730 billion — roughly $825 billion if you include Oracle — and Bain projects cumulative spending could reach $6 trillion by 2031.

The flagship example is the $38 billion Oracle-linked construction loan for data centers in Texas and Wisconsin, tied to the Stargate commitment with OpenAI. A syndicate led by JPMorgan and MUFG has spent more than six months trying to distribute that debt to a wider investor base. Some banks have reportedly offered pieces at a discount to non-bank lenders — the financial equivalent of marking down inventory to clear the warehouse.

ℹ️ The Deal Tracker: Key AI Infrastructure Debt in Play

• $38B — Oracle/Stargate construction loan (TX and WI). 6+ months to syndicate. Priced ~250bps over benchmark, 4-year maturity.
• $15B — Morgan Stanley-led bond sale for Google-backed, Anthropic-leased data center in Hubbard, TX. Expected speculative-grade rating.
• $4B — Blue Owl/CoreWeave Lancaster, PA facility. Lenders passed. CoreWeave rated B+ (below investment grade).
• €2B — Deutsche Bank synthetic risk transfer on project finance portfolio including data center loans.
• $18B — Oracle-leased New Mexico data center. Syndicate banks quoting at 89-91 cents on the dollar.

It is not just one deal. Bloomberg Law reported in early October that five-year credit default swaps on hyperscaler debt were nearing their widest levels of the year. All four high-yield data center bond deals since July priced at a discount, compared with only 3 of the prior 10. Oracle, whose credit rating S&P downgraded to one notch above junk in July, has seen the cost of insuring against its default jump sharply.

View original post on X →

Read the full analysis on Startup Fortune →

Why Banks Want Out (It Is Not What You Think)

The easy narrative is that banks are spooked by AI. The reality is more mechanical — and more consequential.

Concentration limits are binding. Banks have strict internal and regulatory caps on how much they can lend to a single borrower or sector. The deals were so large and so heavily tied to Oracle that many buyers could not take on more exposure. One analyst told eMarketer that "a lot of banks are pretty much up against their concentration limits in their data-center lending." When you are maxed out, every dollar of stuck debt is a dollar you cannot deploy elsewhere.

Syndication is the pressure valve, and it is stuck. Normally, banks underwrite large loans with the intent to sell slices to other lenders and institutional investors within weeks. That model works when buyers exist. For AI data center debt, the buyer pool has shrunk because so many investors already hold the same names. The $38 billion Oracle deal went to construction lenders, insurance companies, pension funds — and many said they were full.

Risk transfer is the new game. The tools banks are turning to tell you everything about the state of play. Significant Risk Transfers (SRTs) let banks keep loans on their books while paying outside investors to absorb the default risk. Deutsche Bank executed an SRT on roughly €2 billion of project finance debt including data center loans — its first ever tied to a project finance portfolio. Morgan Stanley has explored similar structures. When a bank's first instinct is to buy insurance on its own loans, that is a signal.

Rising rates compound everything. The Federal Reserve's recent hike to 3.75-4% has made AI infrastructure debt more expensive to carry and harder to place. Construction loans for data centers are typically "mini-perm" structures — two to five year terms that need to be refinanced. If rates stay elevated, the refinancing math changes for every project in the pipeline.

The Structural Shift: From Banks to Shadow Lenders

Here is where the story gets interesting — and, for practitioners, consequential.

The money is not disappearing. It is changing hands. Private credit funds now hold an estimated $135 billion in outstanding data center debt globally, up from roughly $40 billion in 2023. As banks shed exposure, non-bank lenders are absorbing it — but at higher rates and with less regulatory oversight.

This is the same playbook that played out in commercial real estate, leveraged buyouts, and (yes) mortgage-backed securities: when regulated lenders hit their limits, the debt migrates to the shadow banking system. The question raised by Disruption Banking is not whether individual loans default — it is whether the opacity of these structures creates blind spots that regulators cannot see until it is too late.

Columbia professor Stijn Van Nieuwerburgh estimates that banks have about $800 billion riding on AI when you add up all the indirect exposures — mortgages, syndicated loans, project debt. The Chicago Fed puts direct exposure at a more modest 0.8% of bank assets. The gap between those numbers is where the risk lives: in the structures, derivatives, and off-balance-sheet vehicles that are hard to trace.

Read on American Banker →

⚠️ Contrarian Corner: Is This Really 2008?

No. The subprime comparison is provocative but misleading. Subprime mortgages went to unqualified borrowers with no income verification. AI data center loans are backed by 10-20 year lease contracts with hyperscalers — the most cash-rich companies on Earth. Oracle's credit is weak, but Microsoft's, Google's, and Amazon's are not.

The real parallel is not subprime mortgages — it is shadow banking. The risk is not in the quality of the underlying asset. It is in the daisy chain of who holds it, who insured it, and who has to cover the gap if demand for AI compute comes in at 60% of projections instead of 100%. Bank exposure at 0.8% of assets will not cause a banking crisis. But $135 billion in private credit, growing fast, with limited regulatory visibility? That is the pressure point nobody is stress-testing.

The Oracle Problem

No discussion of AI data center financing is complete without Oracle's outsize role. The company sits at the center of the largest deals in the space, and its credit profile is the single biggest source of lender anxiety.

S&P downgraded Oracle to one notch above junk in July 2026, reflecting the company's aggressive borrowing to fund its data center ambitions. Oracle's capex rose to $55.7 billion in fiscal 2026 — up from $21.2 billion the prior year — and the company guides to $90-95 billion for fiscal 2027.

The New Mexico data center, where Oracle triggered a force majeure clause that alarmed both lenders and local officials, is emblematic. Syndicate banks including Santander and Jefferies have been quoting the $18 billion in associated loans at 89-91 cents on the dollar — meaning buyers are demanding a 9-11% discount to take on Oracle-linked risk.

Regulators Are Watching

Central banks and financial supervisors are paying attention. Bank of England Governor Andrew Bailey, writing as chair of the G20's Financial Stability Board, warned G20 finance ministers that markets remain "exposed to stretched valuations, rising debt, and growing concentration in the AI sector." The Bank's July 2026 Financial Stability Report explicitly named three AI-related risks: a potential stock market bubble, cybersecurity vulnerabilities, and "increasingly complex and opaque corporate debt."

In a Bank of America fund manager survey, 34% of respondents named hyperscaler capex as the likeliest source of a future systemic credit event — a share that doubled in a single month. The Bank of England has also launched a formal review of data center lending specifically, examining how AI companies connect to the broader financial system.

What the Community Is Saying

The story hit a nerve on Reddit, where a post titled "Banks are now trying to offload the AI data center debt they just underwrote" drew over 15,000 upvotes and 1,100 comments on r/technology. The top comments oscillate between vindication and genuine concern about what happens when the music stops.

Meanwhile, the Reuters article on rising rates testing the AI data center financing boom landed on Hacker News with a discussion focused on structural mechanics — whether syndication difficulty reflects credit risk or simply unprecedented deal size, and whether the shift to private credit is healthy diversification or opacity by another name.

View discussion on Hacker News →

On X, Jason Calacanis resurfaced a conversation with CoreWeave's leadership about the company's debt structure, drawing over 300,000 views. The timing is notable: CoreWeave's total debt load has ballooned to nearly $30 billion, with Q2 2026 interest expense ($640 million) exceeding the company's net income.

View original post on X →

What This Means for You

If you are building products or services that depend on AI infrastructure, the financial stress in data center lending translates directly to your roadmap:

1. Capacity timelines will stretch. Construction financing is getting harder and more expensive. Projects that need refinancing at higher rates will slow down. The gap between "announced" capacity and "available" capacity is about to widen.

2. Lease rates are heading up. When the cost of capital rises for data center developers, they pass it through. If your cloud or colo costs have been stable, expect upward pressure over the next 12-18 months, especially for new commitments.

3. Counterparty risk matters. Not all data center operators are equally well-financed. Oracle-linked and CoreWeave-linked facilities carry measurably more credit risk than those backed by Microsoft or Google. If you are signing long-term infrastructure commitments, the developer's balance sheet is now a first-order concern.

4. Geographic concentration is a risk. Community opposition — like the Pennsylvania residents refusing $10,000 buyouts for data center construction — adds permitting risk to projects. Maine nearly passed a moratorium on large data centers. These are not just NIMBY stories; they are financing risks.

5. The capex-to-capability math is changing. As we covered in our analysis of AI scaling law economics, the relationship between spending and capability gains is not linear. If financing costs rise 200-300 basis points, the ROI threshold for new data center projects shifts materially — meaning some planned facilities simply will not pencil out.

💡 The Bottom Line for Builders

Do not plan your infrastructure roadmap around announced capacity. Plan around financed capacity — and right now, those two numbers are diverging fast. Build optionality into your cloud commitments, maintain relationships with multiple infrastructure providers, and keep a close eye on your counterparties' balance sheets. The AI capability curve may be exponential, but the financing curve just developed a kink.

Looking Ahead

The immediate question is whether the syndication logjam clears or hardens. If rates stabilize and equity markets recover, banks may be able to distribute their stuck inventory. If the selloff deepens, expect more forced sales at steeper discounts — and potentially some smaller data center developers running into covenants.

The larger question is whether the structural shift from bank to non-bank financing is healthy. Private credit brings flexibility and appetite for risk. It also brings opacity, leverage, and limited regulatory oversight. The Bank of England is right to be asking questions. Whether anyone gets satisfactory answers before the next round of financing closes is another matter entirely.

Goldman Sachs CEO David Solomon expects "bumps and recalibrations" over the next 6-18 months. Jensen Huang, speaking to the Bg2 Pod, argues that revenue will ultimately justify every dollar of capex. They are probably both right — but the people in between, the banks and developers holding construction risk during the "recalibration," are the ones feeling the squeeze today.

The AI buildout is not stopping. But who finances it, at what cost, and with what visibility into the risk chain — those answers are changing right now, in real time, and they will shape what infrastructure actually gets built over the next 3-5 years.

---

Related reading on ComputeLeap:

• The Capex-Capability Math Is Breaking
• Oracle's Force Majeure Data Center
• The AI Safety Selloff: Trading Desks Are Buying the Dip
• The AI Token Economics Subsidy Clock
• Data Center Backlash: The Water Bill

---

Originally published at ComputeLeap