The unpriced line item in AI data-center finance: insurers/reinsurers are underwriting grid-interruption and residual-value “tails” that lenders can’t see

In AI data-center buildouts, the financed stack often assumes that (1) construction can be insured for the full build cost and (2) operational downtime is manageable. Swiss Re’s research on insuring AI data-center risks argues the opposite: financing institutions demand limits that cover full construction cost even though probable maximum loss scenarios are much lower—so re/insurers can only take a fraction of those limits at competitive rates.

Grid reliability issues are turning “interruptions” into underwriting events. In the PJM footprint, PJM said it won’t start curtailing supply until June 2027 and that the curtailment mechanism would apply only to data centers 50MW or larger. Even though this is a grid-operator program (not a hurricane policy), it matters because business interruption coverage has to decide how it treats regulated power curtailments and the resulting downtime, loss of rent, and service degradation.

Beyond property and BI, AI data-center finance increasingly uses structures that compensate owners if leases aren’t renewed or terminate early. While the construction/build contract may be insured, the residual-value tail can sit elsewhere (e.g., guarantees, floors, or landlord-protection structures) and still behave like an insurance liability: it’s effectively a payout triggered by tenant behavior and asset resale stress.

That’s why Swiss Re’s research emphasis on BI duration and the mismatch between demanded limits and what re/insurers can take is so relevant: it points to the underwriting capacity constraints that can force deal structures to get “creative” (more guarantees, more retention by sponsors) when traditional P&C capacity doesn’t fit the demanded economic limit.

When insurers ration capacity for construction and BI limits, the brokerage value shifts toward (1) placing risk into markets that still have appetite and (2) engineering policy structures to align exclusions/definitions with the actual tail scenarios. That’s the practical reason brokers like Arthur J. Gallagher can benefit in a world where more AI builds require more negotiation around terms—because the “coverage fit” problem grows as sites scale.

Swiss Re flags several technical mechanisms that increase BI exposure in AI data centers. It highlights that lithium-ion battery backup units introduce an ignition source that “did not previously exist”, and also quantifies fire-cost concentration: fire accounts for 10.9% of events but 42.3% of loss costs in a referenced FM study. It further points to liquid-cooling “escaped liquids” exposure—nearly 24% of data-center loss costs—with sprinkler leakage and new-cooling escaped liquid damage each contributing material portions.

Near-term (next 1–2 quarters), watch for: (1) higher attachment points and tighter underwriting terms on BI for large-MW data centers, (2) more demand for surety/default coverage in complex construction, and (3) increased use of contractual structures that move residual-value economics into separate risk buckets. Over 1–3 years, the key thesis test is whether underwriting capacity normalizes as projects standardize (design-to-underwrite playbooks) or keeps ratcheting because AI-density architectures keep increasing BI technicality.