AI data center debt lenders just changed their posture. On November 9, 2026, Reuters reported that banks and credit investors have become reluctant to add exposure to AI-related borrowers, and that some lenders are now hesitant to finance deals even when they are backstopped by names such as Anthropic and Google. According to two of Reuters' banking and credit sources, the shift happened in just the past few days. For CRE developers, powered land owners and landlords courting AI tenants, that matters more than any stock chart: the guarantee was supposed to be the thing that made the debt easy. For the wider set of tools investors use to underwrite these assets, see our guide to AI tools for real estate investors.
Key Takeaways
- Reuters' sources say some lenders now hesitate to fund AI deals even with Anthropic or Google backstops, which weakens the credit support model large AI leases have relied on.
- Lenders are stress-testing cash flows for a downturn and reading service obligations, operational deadlines and tenant reduction or termination rights more closely.
- Data center loans are now a large line in bank CRE books: JPMorgan's category that includes data centers reached $29 billion, or 12% of its CRE loans, at the end of Q3.
- Riot Platforms' new Anthropic lease shows the cost of a financing gap in writing: if Riot cannot raise the money on acceptable terms, it owes rent credits that carry 12% interest.
What Reuters Actually Reported
The facts in the report are narrower than some of the headlines it will generate, so it is worth separating them.
- The pullback is about adding exposure. Banks and credit investors are "showing reluctance to increase exposure to borrowers," according to Reuters. That is a capacity and appetite story, not a report of loans being called or defaults.
- Guarantees are no longer decisive for some lenders. Some banks that were "lining up to fund AI deals" a few weeks ago now say they would be reluctant to finance additional deals even with guarantees from major technology companies. Reuters did not name the lenders, and Anthropic and Google declined to comment.
- The trigger is market-driven. Reuters ties the shift to a selloff that erased $2.5 trillion in Nasdaq 100 market value in one day and pushed the index into a correction, plus anxiety about more data center oversight after the midterm elections.
- Pricing has already moved. Bank of America credit analyst Yuri Seliger told Reuters that the spread on 10-year hyperscaler debt has widened by about 50 basis points this year, compared with roughly 20 basis points for high-quality industrial companies not tied to AI.
- The weakest credits feel it first. The pullback is most pronounced for neoclouds and standalone data center operators. Reuters reports that Nebius explored a high-yield bond in September and dropped the plan, then signed a $3.5 billion senior secured facility with Citigroup in October.
Demand for capital is not shrinking. Reuters cites Goldman Sachs Research projecting global hyperscaler debt issuance of $420 billion in 2027, more than 60% above its $250 billion estimate for 2026. More borrowers chasing more selective lenders is the setup CRE developers need to plan around.
Why a Guarantee Pullback Hits CRE Directly
Five days ago, we covered how Google's backstop of Anthropic's lease obligations at TeraWulf's Kentucky campus works as AI tenant credit support: a highly rated third party stands behind a younger tenant so the landlord can raise project debt against the lease. That structure assumes lenders treat the guarantee as close to the guarantor's own credit. Reuters' sources describe lenders who are no longer stopping there.
Reuters also explains why so many deals carry this support. Chipmakers typically backstop the senior debt used to buy advanced chips for an AI company and also guarantee payments on the data center lease. Banks and private credit were attracted to these loans because the guarantees secured debt service. If some lenders are now reluctant even with that protection, the reason is not that the guarantor became weaker overnight. It is concentration: lenders are asking how much more of their book they want tied to one sector, whoever signs the guarantee.
For a developer, that changes the order of operations. A signed lease with a strong guarantor used to be the hard part, with financing following. The financing is now its own gating item, and it can lag the lease.
The Clauses Lenders Are Now Reading
Reuters reports that lenders are scrutinizing "service obligations, operational deadlines and the conditions under which large technology customers could reduce or terminate commitments." Riot Platforms' quarterly report, filed November 4, gives a public example of exactly those clauses in a live AI lease.
According to Riot's 10-Q for the quarter ended September 30, 2026, its subsidiary signed a 15-year powered land lease with Anthropic PBC on October 28 for 910 MW of critical IT capacity at Corsicana, Texas, effective November 4. The filing discloses:
- Base rent of $130 per kilowatt of critical IT capacity per month for each delivered block, escalating 3% a year, on a triple net basis.
- Delivery credits: if Riot delivers capacity more than 60 days late, the tenant earns escalating rent credits, capped at 270 days of base rent for the undelivered blocks in that phase.
- A partial termination right: if a phase is not delivered within 270 days after its target date, the tenant may terminate the lease as to that phase on 30 days' notice, subject to Riot's cure right, with extensions for force majeure of up to 365 days.
- A financing contingency: Riot will reimburse up to $4.0 million per MW of certain data center module costs "subject to Riot Corsicana obtaining financing on market terms reasonably acceptable to it." If that financing is not available, Riot instead provides rent credits that amortize the amount "plus interest at 12% per year" over the rest of the initial term.
That last clause is the one to study in a tighter credit market. It does not say Riot will fail to raise the money, and nothing in the filing suggests it expects to. It shows how the lease prices the risk if debt is unavailable: the cost of the tenant's modules moves out of a loan and into reduced rent, at a 12% rate.
A Worked Example: What a Financing Gap Costs
The following is a hypothetical illustration using the 10-Q terms, not a forecast of Riot's results. Assume the full reimbursement is earned and that the credits amortize as level annual amounts over 15 years. The actual remaining term will differ by phase.
- Maximum reimbursement: $4.0 million per MW times 910 MW equals about $3.64 billion.
- Annual credit at 12% over 15 years: the annuity factor is about 0.1468, so the credit is roughly $534 million a year.
- Starting base rent: $130 per kW per month times 910,000 kW times 12 equals about $1.42 billion a year.
Under those assumptions, the credits would offset roughly 38% of year-one base rent. Across 15 years, the landlord would give up about $8 billion of rent to settle a $3.64 billion obligation. A lender sizing project debt to that rent would size against the net figure. A financing gap therefore reduces the income stream behind any later financing attempt.
The lesson for CRE developers is not specific to Riot. Any AI lease that ties landlord-funded tenant improvements, fit-out reimbursements or infrastructure delivery to the landlord's financing needs a clear answer to what happens if that financing arrives late or costs more. We walk through how to model these layers in our guide to AI data center financing and capital stack modeling.
Where This Shows Up in Bank CRE Books
Reuters reports that at JPMorgan Chase, loans backed by "other income-producing properties," a category that includes data centers, stood at $29 billion at the end of the third quarter. That is the bank's second-largest CRE exposure at 12%, behind apartments, up from 8% three years ago when the category ranked fourth after apartments, industrial and office. Note the category is broader than data centers alone, so the figure is an upper bound on JPMorgan's data center exposure rather than a precise count.
That concentration is the plain explanation for the pullback. When one property category climbs from fourth to second in a loan book, credit committees ask for more on each new deal regardless of the guarantor. We covered the systemic version of this risk when the Kansas City Fed flagged AI debt moving into CRE loans and CMBS. Reuters' report is the lender-side behavior that follows.
It also connects to the secondary market. ING's auction of roughly $500 million of data center loans, which we covered in what the data center loan BWIC means for CRE debt investors, is about banks selling existing exposure. Reuters describes banks hesitating to originate new exposure. Taken together, the two reports show lenders both selling existing data center loans and slowing new ones.
What CRE Developers and Landlords Should Do Now
- Treat financing as a separate milestone in the lease. If you fund infrastructure or tenant reimbursements, negotiate what happens when debt is late or unavailable. A fallback rate like Riot's 12% is a real cost, so model it before you sign.
- Underwrite the guarantee and the concentration. A strong guarantor still matters, but lenders are asking about sector exposure. Expect questions about how much of your rent roll depends on one AI tenant or one guarantor. Our guide to hyperscaler lease underwriting and tenant credit covers how to separate the two.
- Map every delivery deadline to a dollar amount. Delivery credits and phase-level termination rights are exactly what Reuters says lenders are reading. Build a schedule that shows the rent at risk for each month of delay, and share it with lenders before they ask.
- Line up more than one capital source. Reuters reports a Ropes & Gray finance partner saying debt investors' books "have become full." Approach banks, private credit and, where the asset fits, the secondary market in parallel rather than sequentially.
- Use AI to read the contracts. Lease abstraction tools can extract termination, reduction, service-level and delivery clauses across a portfolio of AI leases into one comparable table. That is the same work lenders are now doing, and doing it first gives you a better position in the negotiation.
If you are negotiating an AI tenant lease or preparing a data center financing package, The AI Consulting Network helps CRE teams set up AI-assisted lease abstraction and scenario models that answer lender questions before they are asked.
What We Do Not Know Yet
Reuters' sources did not name the lenders, and the report describes reluctance to add exposure, not a market-wide halt. Citigroup's October facility for Nebius shows secured credit is still available to some borrowers. The CEO of alternative asset manager AIP told Reuters he expects most banks and credit investors to take a "wait and see approach" after the midterms, which means the posture could ease as quickly as it tightened. Watch for whether announced AI leases close their project financing on the timelines disclosed, and whether new leases start to include financing contingencies like Riot's.
Frequently Asked Questions
Q: Are banks refusing to lend on AI data centers?
A: Not across the board. Reuters reports that some banks and credit investors are reluctant to increase exposure, and some hesitate even with Big Tech guarantees. Secured financing is still closing for some borrowers, but terms and selectivity have tightened.
Q: Why would a lender hesitate if Google or Anthropic backs the deal?
A: The concern is concentration, not the guarantor's credit alone. When a sector takes up a growing share of a loan book, lenders limit new exposure and look harder at termination, service and delivery terms that could reduce the cash flow behind the loan.
Q: What is the main lease risk for CRE developers?
A: Clauses that link the landlord's obligations to its ability to raise money. In Riot's Anthropic lease, unavailable financing turns a tenant reimbursement into rent credits at 12% interest, which reduces the rent available to support later debt.