Banks turn to credit insurance to stretch AI lending limits
Banks are buying credit insurance on data-centre and AI-infrastructure loans so they can lend more than their internal limits allow. Barclays has bought cover and HSBC has looked at it. JPMorgan Chase considered it in recent weeks but decided against buying (all according to unnamed people). Brokers say a bank lending hundreds of millions of dollars to a project typically passes 20–50% of the exposure to insurers, which makes larger loan packages possible. More than 70 providers now write project-finance credit insurance, up from a handful a decade ago; they include AIG, Chubb and Axa. Annual premiums run at 50–70% of what the bank earns on the insured slice. The policy pays out on default, and the bank does not have to show a political cause. The FT ties this to the 38 anti-data-centre ballot measures that passed in last week's US midterms. Caveats: US banks generally get no regulatory capital relief from this cover; the benefit is internal risk capacity (European and Canadian banks have long used it for capital relief). The piece names no borrower, project, loan size or insurer deal, and no regulator has commented. It does not show that hidden leverage is actually building.
Why it matters
It shows how the AI build-out is being financed beyond the hyperscalers' own balance sheets. Banks are moving part of their data-centre risk to insurers so they can keep lending. That gives more room for data-centre developers and neoclouds, but the risk ends up in a less visible place. Insurers taking 50–70% of the bank's margin also tells you how the market prices data-centre default risk.
Executive takeaways
1) Credit insurance has become a way to grow bank lending to AI data centres: 20–50% of a project's exposure moves to insurers. 2) Barclays has bought cover, HSBC has explored it and JPMorgan passed, so big lenders disagree on whether it is worth it. 3) In the US the gain is internal limit headroom, not regulatory capital relief, so this is about appetite rather than solvency rules. 4) The insurer market has grown from a handful of providers to 70+ (AIG, Chubb, Axa), and strong demand keeps premiums at 50–70% of the bank's return on the covered portion. 5) Cover pays on default whatever the cause, which matters now that 38 anti-data-centre ballot measures passed in the midterms.
What Financial Times may be missing
The piece names no borrower, project or loan size, so we cannot tell whether the cover sits on hyperscaler-backed projects (low risk) or on speculative neocloud and developer builds (high risk). It does not look at concentration risk on the insurer side: if one wave of data-centre defaults hits many banks' insured slices at once, the risk is correlated, not diversified. It also misses the counterparty question (insurer ratings, claim disputes over policy terms) and gives no regulator view (PRA, Fed, ECB) on whether this kind of limit-stretching should be disclosed. Finally, it does not size the market: how many billions of AI loans are already insured.