The data-center build-out is reinsurers' unseen hail risk

A hailstorm over a data-center campus in Dallas or Northern Virginia would now land on a concentration reinsurers have never had to map: PWD's tracking shows twenty US locations hold about 80 percent of storm-damaged data-center floor space, and the next wave of construction is moving deeper into the same hail and tornado belt. The loss the market can already see, Bald Range, likely lands in cedents' results rather than triggering catastrophe bonds. That division—visible wildfire inside retentions, unseen convective-storm aggregation—defines the property-cat trade in late September. A Canadian wildfire that stays below attachment points is, for reinsurers, a comfortable outcome: it says more about where January's repricing will be fought than about any cat bond trigger.

But the same week produced three capital commitments that push the unseen aggregation further up the exposure curve: on September 16, Serverfarm closed a $3.89 billion deal, Microsoft announced a $465.5 million deal, and Exa Infrastructure announced a deal with no disclosed size. Taken with the concentration data, none of that capital appears to be leaving the corridor; a $3.89 billion close in a market where the same twenty locations hold most storm-damaged floor space is not a diversification trade but an addition to the same aggregation.

That concentration is pulled together by existing power, fiber routes, cheap land and a construction workforce that knows how to build a hall, and those economics also push reinsurers into a severe-convective-storm exposure they have historically priced for warehouses and light industrial, not for buildings whose users cannot tolerate downtime. The Microsoft announcement is the more instructive of the two priced deals: at $465.5 million it is not a flagship campus by itself, but it shows the marginal dollar is still buying compute capacity in geographies where severe convective storm is a recurring event. The same concentration that makes data centers efficient also places floor space under the hail swaths and tornado tracks that reinsurers have spent decades modeling for retail and distribution properties rather than mission-critical computing.

Bald Range offers a false sense of where the risk sits: a wildfire in Canada that stays inside retentions is, for the moment, exactly what primary carriers want—a loss big enough to justify rate but small enough not to hit the reinsurance tower. The data-center exposure does not present itself as a single event; it is an accumulation of many small and mid-sized hailstorms, each of which damages roof systems, cooling equipment and backup power on buildings that likely cost more to repair per square foot than the warehouses reinsurers have spent decades pricing.

Serverfarm's $3.89 billion close is the larger of the two priced transactions, and it arrives at a moment when hail-driven replacement-cost exposure on data halls is only beginning to be priced. A close of that size, in a market where construction is moving deeper into the hail and tornado belt, suggests the new capacity is being added to the same twenty locations that already hold most storm-damaged floor space. The $465.5 million Microsoft announcement points the same direction, and Exa Infrastructure's undisclosed deal adds queue without adding diversification.

That pattern is what should worry a reinsurer: catastrophe models are built on geography, and geography is the one thing the data-center boom is not changing. Power and fiber dictate site selection, and those constraints have concentrated the build-out in a small number of locations, so a peril model built for retail and industrial risk is being asked to digest a new, high-value target in the same hail swaths. Every additional campus in Dallas or Northern Virginia likely raises the probable maximum loss for a spring hailstorm, and the models will not show it until the claims do. The 80 percent figure is a starting point: every new campus in one of the twenty locations raises the share that a single hail event can damage, and the next wave of construction is moving deeper into the same belt rather than away from it.

The replacement-cost problem is not just the IT equipment inside the hall. Hail that breaches a data center's roof does not have to destroy servers to produce a claim; it may only have to interrupt cooling or damage rooftop units, forcing a shutdown while repairs are made from the inside. For a facility designed to run continuously, business interruption may begin at the first minute of downtime, which suggests the loss could exceed the visible property damage. Reinsurers have spent two decades refining business-interruption models for hotels and manufacturing plants, but data halls are a different exposure because the tenant's tolerance for interruption is effectively zero.

The question for reinsurers is whether the models treat a data hall as a high-value, single-tenant property with near-zero tolerance for interruption, or as another warehouse. The deal flow suggests the market is still pricing the latter, which would make this underpricing unusually quiet. The January renewal will be the first place where that gap shows up: Bald Range gives cedents a retention-level loss to discuss, but it does not force a debate about cat bond triggers, and the data-center concentration does not force one either because it is not a single event. Reinsurers have no loss to point to, no attachment point to test, and no headline to justify a repricing; they have only a slow accumulation of exposure in twenty locations, backed by a capital pipeline that keeps clearing deals. Serverfarm, Microsoft and Exa Infrastructure all acted on September 16 without changing the geography, and as long as that remains true the severe-convective-storm aggregation reinsurers cannot see will keep growing. The next test will be a routine hailstorm over one of the twenty locations.