Insurers Turn to Wall Street to Cover AI's Data-Center Boom

Data centers built to train and run artificial-intelligence systems have grown so large and so expensive that the insurance industry is running out of room to cover them on its own balance sheets. Insurers, reinsurers and risk-modeling firms are now moving to bring Wall Street's catastrophe-bond market — the mechanism long used to spread hurricane and earthquake risk to bond investors — into the business of insuring AI infrastructure.

The shift became concrete this month. On September 3, Verisk, the risk-analytics firm whose AIR Worldwide unit builds catastrophe models used across the insurance industry, launched a U.S. Data Center Exposure Database covering more than 2,500 facilities, giving insurers building-level detail — location, construction type, floor area, power capacity — on a class of property that barely existed as an underwriting category a decade ago. Two days later, at the reinsurance industry's annual Rendez-Vous de Septembre gathering in Monte Carlo, Swiss Re Institute told underwriters that AI data centers and renewable-energy infrastructure could generate roughly $200 billion in cumulative insurance premiums between 2026 and 2030, an opportunity the reinsurer's economists compared to the biggest growth waves the commercial property market has seen in decades.

The Numbers Behind the Buildout

The scale driving the concern is best captured in dollars. A single hyperscale AI data-center campus can now cost up to $50 billion to replace, according to Swiss Re Institute's Sigma research, a figure that dwarfs the low single-digit billions of per-risk capacity that most large property insurers are willing to put up for any one facility. The five largest U.S. hyperscalers are expected to spend close to $800 billion combined on AI-related capital expenditure this year alone, part of a global data-center investment wave that Federal Reserve staff economists estimate will reach an annualized $370 billion by the second quarter of 2026, up sharply from pre-pandemic levels, based on project-level construction data rather than company announcements alone.

Geography compounds the exposure. Texas and Virginia together host roughly 40% of U.S. data-center capacity, and about a quarter of that footprint sits in hail-prone zones while a similar share experiences three or more tornado days a year on average, industry risk modeling cited by Insurance Business magazine shows. Much of that exposure sits inside what meteorologists have long mapped as Tornado Alley, the corridor where warm Gulf air collides with cold air off the Rockies. Insurable values per campus now commonly run $20 billion to $30 billion, well above what a handful of primary insurers can absorb without reinsurance backing.

How Data Centers Outgrew Traditional Coverage

For most of their history, data centers were underwritten as fairly conventional industrial property: warehouses full of racks, insured much like any large commercial building. That changed as generative-AI training runs pushed operators toward campuses holding tens of thousands of specialized chips, each facility concentrated in a handful of markets with abundant power and land. The result is an asset class combining enormous, spiky value with limited claims history, few comparable losses to price against, and heavy dependence on the same electrical grids, cooling systems and network links across many nearby sites.

That combination is precisely what has pushed reinsurers toward capital markets. Traditional reinsurance capacity is finite and reset annually; catastrophe bonds instead let pension funds, hedge funds and dedicated insurance-linked-securities managers take on a slice of the risk directly, in exchange for a coupon, with principal at risk if a specified disaster occurs. The catastrophe-bond market has itself been expanding rapidly, with roughly $65.6 billion in bonds outstanding by the middle of 2026, a record, as investors have piled into a corner of fixed income that has historically offered strong returns uncorrelated with stocks and bonds. Data-center risk, insurers argue, could become the next major peril category layered onto that structure, following the path wildfire and cyber risk took into the cat-bond market in prior years.

Who Carries the Risk

The most immediate parties affected are the hyperscale cloud and AI companies themselves, which need insurance in place to satisfy lenders and to keep financing costs down on campuses that can take years and tens of billions of dollars to build. Reinsurers such as Swiss Re, and specialist catastrophe modelers such as Verisk, are positioning to profit from underwriting and pricing that risk. Insurance-linked-securities managers, including firms such as Brookmont Capital Management, are exploring dedicated data-center catastrophe-bond funds, betting that artificial intelligence itself can help price a peril with almost no historical loss data.

"We are seeing the digital economy become a real economy. AI needs data centers, power grids and increasingly complex infrastructure – and all of it needs insurance," said Gianfranco Lot, chief underwriting officer for property and casualty reinsurance at Swiss Re.

Further down the chain, the costs are likely to reach corporate customers renting cloud and AI capacity, and eventually consumers, if insurance premiums are passed through in the price of cloud services. Communities near major data-center clusters in Texas, Virginia and Arizona also have a stake, since insurers' willingness to underwrite new campuses affects how quickly local utilities and governments approve additional construction.

What Happens Next

Executives in the sector describe the market as still assembling the tools it needs before the first dedicated data-center catastrophe bond can be sold to investors. Rob Newbold, president of Verisk's catastrophe and risk solutions unit, said the new exposure database reflects a broader shift in how the industry views computing infrastructure. "AI is increasingly a physical infrastructure story," Newbold said, one with implications "for insurers, reinsurers, investors and capital markets participants," while Jay Guin, Verisk's executive vice president and chief research officer, framed the underlying problem more simply: "If you can't identify the exposure, you can't effectively measure or manage it."

Analysts covering the insurance-linked securities market say a first dedicated data-center catastrophe bond could still be twelve to eighteen months away, largely because catastrophe modelers need more loss history and clearer building specifications before capital-markets investors will accept the risk at a price sponsors can afford. In the meantime, Swiss Re Institute's chief economist, Jérôme Haegeli, has argued that getting the insurance question right is central to keeping the broader AI buildout financeable at all. "A new investment era is taking shape, with unprecedented amounts of capital flowing into the infrastructure that will power future economic growth," Haegeli said, adding that insurance is what will make that capital "resilient and financeable" as the buildout continues. For an industry accustomed to insuring hurricanes and earthquakes, the next major test may be a data hall full of servers.