An AI infrastructure project can have a compelling growth story. The loan still needs a repayment story, preferably one that survives a delay.
On November 12, the OCC released its Fall 2026 Semiannual Risk Perspective, highlighting credit, market, compliance, and operational risks in the federal banking system. Our read: the infrastructure discussion deserves a place in credit review, not just in the innovation committee's reading folder.
What the OCC actually says
The report's discussion of infrastructure financing on pages 2 and 3 identifies potential risks from construction delays, power availability, community sentiment, technology depreciation, opaque financing connections, and slower-than-expected revenue growth, while noting that bank lending is typically senior and often secured.
The OCC uses the term “Super Intelligence,” or SI, in this discussion. It also identifies current support for credit quality, including banks' senior positions, processing demand, commitments from SI labs, and low data-center vacancy rates.
Keep both sides of that assessment. This is not a declaration that AI infrastructure lending is inherently unsound, and potential vulnerabilities are not evidence of widespread losses.
The report expressly says it does not set policy or interpret statutes or regulations. It generally reflects data as of June 30, 2026, unless otherwise indicated; the November release date should not be mistaken for the observation date of every underlying measure.
Underwrite the dependencies, not just the borrower
For a proposed facility, ask the credit team to map what must happen before the borrower can generate the cash expected in the repayment model. Put construction, power, customer commitments, and financing dependencies on the same timeline.
Then challenge their interaction. A hypothetical delay in power availability should not be tested only as an engineering inconvenience; examine how it would affect revenue timing, liquidity needs, and the next financing decision.
This is our recommended analysis, not an OCC-prescribed scenario. The aim is to expose where several apparently manageable assumptions depend on the same event going right.
A long loan deserves a hard look at a short-lived asset
Do not let the word “secured” end the collateral discussion. Ask which assets support recovery, how their usefulness could change, and what evidence supports the assumed value at the relevant date.
Consider the building, the power connection, and the computing equipment separately. A credit memo should explain the proposed recovery path rather than treating every part of a project as having the same useful life.
The OCC describes some institutions using shorter tenors, selective client choices, and potential repurposing to mitigate risk. Those are observed approaches, not a universal prescription or proof that a particular structure is safe.
For your transaction, test whether the mitigation still works under the downside case. Shortening a loan is not a complete answer if repayment simply moves to an untested refinancing assumption.
Separate contractual support from independent repayment capacity
Map the relevant borrower, customer, guarantor, and financing relationships. Ask where cash originates and which commitments depend on another participant in the same ecosystem.
The question is not whether a recognizable name appears in the structure. It is what that party has actually committed to do, under which conditions, and with what capacity to perform.
At the portfolio level, group exposures by meaningful shared dependencies and legal borrower. Review whether several transactions could be affected by the same customer, infrastructure constraint, or refinancing environment.
Treat that exercise as a request for evidence, not an assumption that every connected structure conceals a problem. Precision is more useful than either enthusiasm or alarm.
Make the credit policy visible in the decision.
Policy is the moat when the institution can show how it reached a decision. Record the assumptions accepted, the evidence reviewed, the exceptions approved, and the conditions that would trigger reassessment.
Build on existing credit and monitoring infrastructure where it can support that record. A new sector does not automatically justify a parallel approval process with different ownership and weaker traceability.
Use institution-specific scenario models to rehearse delays and changing assumptions before approving exposure. Keep proprietary borrower information within authorized boundaries; testing the policy does not require pooling customer records across institutions.
The executive task is not to reject the AI buildout or finance it by slogan. It is to make each repayment thesis, dependency, and downside response defensible.
FAQs
Did the OCC issue a new AI-infrastructure lending rule?
No. The report expressly says it does not set policy or interpret a statute or regulation; it is a supervisory risk assessment, not a new lending rule.
Is the OCC reporting widespread losses on these loans?
The infrastructure discussion identifies potential vulnerabilities, not widespread realized losses. It also describes factors currently supporting bank credit quality.
Does senior, secured lending eliminate the risks?
No such conclusion follows from the report. Our recommendation is to assess repayment, collateral, contractual support, and correlated exposures separately rather than treating seniority as a complete answer.
Where should credit teams begin?
Start with the assumptions connecting construction and operation to repayment. Test the evidence for those assumptions, the effect of delays, and who must approve a change in the credit view.