The lender that helped arrange the debt now wants to be insured against it. Barclays has been buying protection against defaults on more than $100 billion of loans tied to AI infrastructure, according to people familiar with the matter, moving part of the exposure off its own books and into the insurance market.
The instrument is credit risk insurance. The bank pays a premium to one or more insurers, which agree to absorb losses on a defined slice of a loan portfolio once a threshold is crossed. If the protection is written by a highly rated counterparty, regulators may allow the bank to hold less capital against the insured portion, which is the entire point of the trade.
Barclays is not retreating from the business. It is one of the principal lenders to data center construction, and the reason it needs protection is that demand for that lending has grown faster than its balance sheet can absorb. Hyperscale customers now ask for facilities that would once have been considered portfolio-sized commitments to a single borrower, and a bank that says yes to several of them runs into concentration limits long before it runs out of appetite.
The pressure is structural rather than cyclical. A data center campus is financed before it has a tenant in some cases, and before it has revenue in all of them. The credit depends on the tenant’s lease, the tenant depends on demand for computing, and demand for computing depends on a business model that is still being priced. Every lender in the market is underwriting the same assumption at the same time.
That is why the risk is being passed along rather than held. For most of the past century, a commercial bank originated a loan and kept it until maturity, collecting interest and accepting the loss if the borrower failed. The last three decades replaced that model with distribution: syndicate the loan, securitize the pool, sell the tranches to investors with different tolerances for loss.
AI infrastructure is following the same path, faster. Data center loans are being packaged into asset-backed securities, moved into collateralized loan obligations, and now insured. Each step pushes the exposure further from the institution that made the credit decision and closer to a buyer who sees a rating and a coupon rather than a construction site.
Insurers are willing buyers because the premiums are attractive and the losses, so far, have been zero. The question is capacity. A single policy on a $100 billion book is larger than most insurers can write alone, which means the risk is spread across several carriers and, in practice, into the reinsurance market. Reinsurers price catastrophe and longevity; they are less accustomed to underwriting the construction schedule of a building full of graphics processors.
There is also a duration mismatch. Insurance protection is typically written for a few years and renewed, while the underlying loans may run longer. A bank that has bought protection against a five-year exposure and holds a seven-year loan is unhedged for the tail, which is precisely the period in which a chip generation becomes obsolete and a tenant’s lease comes up for renewal.
None of this means the loans are bad. The largest borrowers are among the most profitable companies in the world, and their commitments are contractual. But the structure of the market has changed in a way that makes the risk harder to see. When a bank holds a loan, its examiners can look at it. When the loan is insured, securitized and sold in tranches, the exposure sits in several places at once, and each holder believes someone else is carrying the first loss.
The pattern is repeating across the industry. Chip suppliers are taking equity in their customers. Vendors are guaranteeing residual values on equipment they sell. Private credit funds are writing construction loans that banks would once have kept. Each arrangement solves a financing problem for one participant and adds a link to a chain that has never been tested by a downturn.
For Barclays, the arithmetic is straightforward. Buying protection costs a premium, and the premium is cheaper than the capital the bank would otherwise have to hold against the exposure. As long as spreads stay tight and defaults stay near zero, the trade improves returns on equity and lets the lending desk keep growing.
The test arrives if a tenant fails. Then the insurers pay the insured tranche, the bank absorbs whatever it retained, and the securitization investors discover what the collateral is worth when a building full of servers has no customer. The market has priced that scenario as unlikely. It has not priced what happens if two of them occur in the same year.
Barclays declined to comment on the arrangements, according to the people familiar with them. The bank’s disclosures will show the premium line and the insured exposure, but not the name of the borrower behind each policy, which is the part that matters most.
