Since the launch of ChatGPT in 2022, artificial intelligence has captured the imagination of the business world, and insurance has been no exception.
While there’s been much talk about LLMs speeding up underwriting, powering new catastrophe models and leading to efficiencies in handling claims, AI has also offered an exciting new sub-sector as the developers of multi-billion dollar data centers look to protect their investments.
As this publication reported just last week, tech heavyweight Meta is in the final stages of placing multiple towers of coverage to protect the vast new data center, Hyperion, it is building in Louisiana. Sources said that the total insured value of the project is $20bn+, with Marsh acting as the placement broker for the builders’ risk cover.
But while some in the industry are scrambling to provide capacity for the increasingly pricey infrastructure, others are acknowledging there could be some clouds on the horizon in the form of a possible bubble.
Spending on AI data centers has ballooned in recent years, with the top hyperscalers known as the Magnificent Seven spending over $100bn on data center projects in Q2 2025 alone.
According to McKinsey, the projected spending on data centers is expected to hit $6.7tn by 2030 in order to keep pace with anticipated demand for computing.
The data centers present a complex underwriting risk, said one source, observing that there are few underwriters who have developed specialization in the field. The risks are a result of the complex environment that makes up a data center.
The chatbots that have become AI’s most public face require a ton of computing power.
To provide that computing, hyperscalers and companies, known as colocation providers, that rent out data center space to smaller entities, are funding the construction of multi-billion dollar facilities across the US.
The buildings themselves are required to be state of the art, due to the intensive power and hygiene needs required to keep the delicate circuits functioning optimally. They also require complex cooling systems, due to the heat generated by the GPUs.
Those GPUs themselves can be fantastically expensive. The latest Nvidia models can cost in the tens of thousands of dollars.
Meta, for instance, has estimated that upon the completion of its mega-data center in Louisiana, the company’s total number of GPUs will come out to over 1.3 million.
And according to CEO Mark Zuckerberg, the company will be spending up to $50bn on that facility.
Preet Gill, EVP and global technology risk practice leader at Lockton, said there could be concerns about providing capacity for a property valuation of that size.
Assets of these size extend well beyond the typical available market limit, particularly given the massive compression in market limit that has taken place since 2019.
Brian Hearst, Aon’s data center and life sciences builders risk leader, noted that data centers have become attractive for many carriers, as thus far, there have been few major losses, aside from some notable electrical fires and damage from malfunctioning water-based cooling systems.
The risk of loss is significant, given that a malfunction inside could lead to billions in both property and business interruption losses.
That’s resulted in a strong emphasis on comprehensive mitigation in the design of the centers. An insurer source observed that numerous contingency plans must be put in place, including sectioning off portions of the center to prevent the spread of fire and flood.
“The business income impact of a loss at a data center is significant,” they said. “That's the kind of challenging occupancies that we're really seeing.”
Spending a fortune in a cat-prone zone
The lack of losses thus far is, of course, not an indication that the sub-sector will run clean in the future.
Many data centers are being placed in rural areas that are prone to severe convective storms. The Meta facility for example is being built in Richland Parish, an area that’s known for both SCS and floods.
While the possibility of major damage to an expensive facility may seem like an obvious reason not to construct one in an area, Gill observed that exposure to catastrophic events is just one variable when developers are choosing locations.
The availability of enough land for the enormous projects, the availability of state subsidies, and feasibility of building the power infrastructure needed to generate the enormous amounts of electricity needed can be larger factors. That can lead to, at the very least, SCS exposure becoming an acceptable risk.
“You're building these data centers in places you probably would never think about building anything, and it's led to more catastrophic exposure, more convective storm exposures,” he said.
“The Louisiana parish is a windstorm risk, and it's also a flooding risk.”
Another possible source of risk is the electricity required to power the thousands of GPUs and other circuits within the centers. Gill said many campuses also have their own power plants, which can be renewable, natural gas or coal-powered and come with their own sets of loss risks.
“If you're looking at geothermal, there's not a really good understanding of what that risk exposure is,” he said.
“If you're looking at these renewable energy projects, there's all kinds of new risks that are being introduced from a contamination standpoint and a chemical standpoint and a gas standpoint. So, there's environmental concerns as well that are creating new risks.”
What if there’s a bubble?
While the data centers themselves are growing increasingly expensive there have been growing doubts about whether generative AI’s growth can continue at the current rate.
Microsoft, Amazon and other hyperscalers have pulled back investment in data centers and Sam Altman, CEO of OpenAI – the company behind ChatGPT – has publicly said the sector is in a bubble, comparing market conditions to those ahead of the dotcom crash in the early 2000s.
Aon’s Hearst acknowledged that if AI is a bubble and it bursts, there would be downstream effects on insurers who had deployed capital. That would not, however, likely result in any sizeable losses, he said, but would just force insurers to find new markets in which to deploy capacity.
“Is any one insurer so exposed to this right now that it's going to take them down? I don't think so,” he said, adding that, at least for builder’s risk, carriers have reduced their limits. Thus, if any one project fails, none are likely to be hit too hard.
While a bubble popping would be bad for tech firms, Jim Dunn, US construction practice leader at Marsh, said that the construction industry as a whole is “very resilient” and has weathered numerous economic downturns in the past.
He added that even if generative AI does prove to be the revolutionary tech that some predict, the level of growth of data centers is not sustainable and the amount of capital entering the space will eventually level off, and with it, the amount of builder’s risk premiums carriers are currently enjoying.
Not everyone believes so much caution is warranted. Lockton’s Gill expressed skepticism about a bubble, saying that as the amount of devices and their complexity expands, so will the need for computing.
“They're even talking about building data centers out on the moon,” he pointed out.
“Eventually you're going to run out of space. It's hard to envision why it would slow down when the opportunity set only keeps growing.”
Got a tip for us? Reach out confidentially here.