Banks that want more AI data-center exposure can now sell part of the default risk to an insurer. The invoice is large.
Barclays has bought credit insurance for AI data-center exposure, HSBC has explored or used it and JPMorgan considered the product but did not proceed, according to the Financial Times. The reported structures typically cover 20% to 50% of a bank's exposure. Brokers cited by the FT say annual premiums can reach 50% to 70% of what banks earn on the covered portion of their loans.
CTOL's view is that this swaps lending income for loss protection and, in some cases, room to lend. The cover leaves the data centers' ability to service debt unchanged. It moves a covered default loss to the insurer.
Half the covered income goes to the insurer
Start with a loan portfolio generating 100 units of annual loan income, on the same basis as the brokers' premium comparison. Assume the insured and uninsured portions earn income in proportion to their share of exposure.
- Insure 20% of exposure and that slice earns 20 units. A premium of 50% of that income costs 10 units, leaving 90 of the original 100 before other costs.
- Insure 50% of exposure and pay a premium of 70% of that slice's income, and the cost is 35 units, leaving 65.
The general rule is a product. The whole-book premium burden is the insured exposure share times the premium share of income on that exposure. So 20% × 50% = 10%, and 50% × 70% = 35%.
Treat these as endpoints. They combine separately reported coverage and premium figures. They are neither prices from disclosed transactions nor a forecast for the market. They say nothing yet about net profit, expected credit losses or regulatory-capital relief. At those terms a bank gives up a sizeable share of the income the insured loans generate. The benefit has to earn that income back.
The premium has to buy something
The bank's economic test has two sides. The cost is the premium plus any risk the policy leaves with the lender. The benefit is the value of covered losses the bank no longer expects to bear. It adds the risk-adjusted income from additional lending the cover enables, plus any defensible value from keeping client relationships. A policy pays off only when the benefit outweighs the cost.
A bank hitting an internal concentration limit may place a high value on extra capacity. A bank with no binding limit can still value default protection, but cannot count hypothetical new lending as a benefit. That explains why banks could decide differently. The FT account leaves the reasons behind Barclays buying and JPMorgan passing unstated.
Regulation adds another dividing line. As the FT reported, U.S. banks generally get less regulatory-capital benefit from these arrangements than European and Canadian lenders. Economic loss protection, relief from an internal exposure limit and recognized regulatory-capital relief are three different outcomes. A bank should count only the benefit its own transaction and rules provide.
The insurer takes the default and leaves the bottleneck
The FT describes protection that responds to borrower default whatever caused the project to fail. A permitting setback, local power shortage, political opposition or weaker AI computing demand could damage a project's cash flow. That can erode its ability to repay. Disruption alone is not an insured default, and a claim must meet the policy's terms.
Insurance therefore leaves every project bottleneck in place. It reallocates some of the financial loss if the bottleneck ends in a covered default. The bank keeps its uncovered exposure and any losses outside the contract, and the insurer must price the share it accepts.
That matters most where projects depend on common electricity markets, sponsors or AI demand. Failures can be correlated even when the underlying loans have different borrowers. The FT counted more than 70 project-finance credit insurers, compared with a handful a decade ago. That shows a larger field of participants. Net exposure to shared risks may still be concentrated. For an insurer, the relevant unit is the combined portfolio exposed to a common shock, and the number of policies written matters less.
Insurers may diversify better than banks, and executed deals can show it
The strongest counterargument is that insurers can diversify these risks more effectively than individual lending banks. Substantial premiums can then pay for themselves when they free scarce bank capacity. That is plausible. Executed transactions would confirm it if two gains exceed the all-in insurance cost: the documented reduction in retained risk and the incremental risk-adjusted return from enabled lending. A rising count of banks buying policies would leave it unconfirmed.
CTOL's current assessment, at moderate confidence, is that the reported terms can cost banks a material share of insured-loan income. In return they redistribute AI infrastructure default risk. Valuable lending headroom remains a bank-specific possibility, and no universal result has been shown. The bank choices, quoted premium terms and default-transfer mechanism support that judgment.
At CTOL's 31 March 2027 review checkpoint, the observation that would change the capacity assessment is an executed AI data-center insurance transaction. It would disclose its actual premium, any recognized lending or regulatory-capital headroom and a positive risk-adjusted economic gain after the insurance cost. That date is CTOL's own review point, and no disclosure deadline has been announced.
Until that comparison exists, price the protection against the lending opportunity and the loss it truly transfers. Count insured notional as exposure moved, and leave it out of profit. Treat regulatory capacity as restored only where the bank's own rules recognize it.