AI data centers are becoming a “stress test” for insurance companies, as rapid technological advances and the use of increasingly complex financial structures pose unique challenges and opportunities for this sector.
According to McKinsey, global data center spending could reach $7 trillion by 2030, and much of that spending can no longer come from hyperscalers alone. Instead, Big Tech companies are increasingly turning to private equity, private credit, and using debt to finance the construction of capital-intensive facilities.
Private infrastructure data center transactions consistently exceeded $10 billion last year, according to Preqin data. The largest deal was valued at $40 billion, with Nvidia, Microsoft, BlackRock, and Elon Musk’s xAI forming part of the investor consortium acquiring Aligned Data Centers.
The fact that data centers cost so much money to build, build and operate has been a “real stress test” for major insurance companies over the past four to five years, said Tom Harper, data center leader at the insurance brokerage. gallagherhe told CNBC.
“When you put more than $10 billion to $20 billion into one location, you have market capacity issues. The market is always concerned about these risks because they are very high quality buildings. They have state-of-the-art technology and are AA plus plus construction sites, but the capacity, the ability to provide insurance capacity in these locations, has been tough.”
Harper said it would have been nearly impossible to reasonably insure the $20 billion campus in 2023. But in 2026, it has become a weekly conversation.
We are talking about trillions of dollars in funding and we are almost back to the same cycle with little transparency regarding funding structures. Its scale is astronomical.
rajat rana
Partner at Quinn Emanuel Urquhart & Sullivan
Estimated spending on AI data centers is said to be the largest peacetime investment project in history. Rajat Rana, a partner at Quinn Emanuel Urquhart & Sullivan, took this a step further, stressing to CNBC that this is “the largest peacetime investment project in human history that is largely financed off-balance sheet.”
Mr. Rana, who worked on structured finance litigation in the wake of the housing crisis sparked by the 2008 financial crisis, said tracking funding trends for AI data centers feels like a “deja vu” experience.
“We’re talking about trillions of dollars in funding and we’re almost going back to the same cycle with very little transparency around the funding structure. The scale is astronomical,” he said.
The AI boom is not only surging demand for facilities, but also spurring rapid advances in power generation and chips, key technologies powering data centers. Advances and large inflows of capital into this field present both risks and rewards for insurers and lenders.
Tailor-made policy
Data centers require a specialized approach from insurance companies that include both real estate and technology assets. Gallagher’s Harper said some of the world’s largest insurance companies are creating data center-specific ways to manage projects.
Harper told CNBC that the facilities have unique challenges because of their concentrated value, required amount of power generation and “state-of-the-art technology,” which typically allows for favorable pricing, making them “highly desirable.”
Insurance companies want to spread risk and reduce costs. But problems arise when $20 billion worth of assets are concentrated in high wind and hurricane zones, he added.
Disruptions in the supply chain can increase complexity by concentrating expensive equipment that has not yet been installed. Customers import large amounts of cargo from overseas, often storing it in facilities they don’t own or operate, creating additional risks, he said.
The M&A boom has kept transactional lawyers busy, with Kirkland & Ellis noting that many companies are building specialized data center teams and recruiting experts in real estate, power, communications, finance, insurance, trading, private equity and cybersecurity.
professional services company marsh We have launched a dedicated Digital Infrastructure Advisory Group aimed at supporting our clients as contracts become increasingly complex.
Last year, Marsh also launched Nimbus, a €1 billion ($1.2 billion) insurance scheme to cover data center construction in the UK and Europe. Seven months later, we expanded our facility to offer up to $2.7 billion in limits.
“Private credit can meaningfully complement banks and support non-hyperscale contract offtake,” said Alex Wolfson, senior vice president of credit specialties at Marsh Risk.
As data center lending increases, insurance companies are starting to reach their limits in protecting lenders if borrowers don’t pay, Wolfson explained. Marsh is working on solutions that support lenders.
But Quinn Emanuel’s Rana cautioned that when it comes to data centers, it is difficult for insurers to fully understand the risks as funding moves off-balance sheet.
He noted that in January, four US senators called on the government to investigate how Big Tech companies are increasingly turning to “complex and opaque bond markets to borrow large amounts of cash.” In an open letter, the senators warned that the huge debt burden could lead to “volatile losses” for financial institutions and trigger a broader financial crisis that would harm the economy.
Rana said increased funding opacity could lead to secondary litigation risks for downstream investors, such as pension funds, insurance companies and asset managers that invested in private credit funds, who later found out they were not fully aware of concentration risks.
He told CNBC that some PE funds have contacted him with concerns about commercial leases and real estate valuations.
Tenants are trying to negotiate expansions on their properties, while landlords are fighting over the value of their AI data centers, demanding higher prices.
“I am not an apocalypse who says it will collapse. What I am saying is that whether it collapses or not, conflict is inevitable. And we are already witnessing such conflicts,” Rana said.
“GPU Debt Treadmill”
An important discussion about the potential cracks in the financial center regarding GPUs and the risk that the lifecycle of GPUs may not match the longer useful life of the facilities that house them.
coreweaveThe company, which sells AI technology in the cloud, was the first to secure GPU-backed loans, essentially using the value of high-performance chips as collateral. The company announced last week that it had secured $8.5 billion in its first investment-grade rated GPU-backed deal. The company’s stock price rose 12% on the day.
While data center lifecycles typically span decades, the average lifecycle of a GPU is approximately 7 years.
“There are different data centers that are raising debt by disclosing different lifecycles to investors,” Rana said. He called this problem the “GPU debt treadmill,” a term coined by AI critic Dave Friedman.
“This is like a treadmill with an AI data center running,” Rana told CNBC. Even if the financing structure is ring-fenced and backed by an investment-grade counterparty, the real risk may lie in whether an equity issue today will later develop into a credit issue over time.
“As these new chips come out, data centers are going to feel pressured to raise more debt and have to build new infrastructure. That essentially creates a multibillion-dollar question: How quickly can we build these facilities? How quickly can we raise credit?”
The cost of financing these projects has increased the amount of commercial mortgage-backed securities sold to investors, and the recent growth in asset-backed securitization transactions is likely to continue to accelerate, Harper said.
For some insurers like Gallagher, the changing dynamics in this space are more of an opportunity than a challenge. Harper said GPU lifecycles are getting longer. If the value of an item declined rapidly, Mr. Gallagher had to get creative and write a bespoke insurance policy that agreed in advance how the asset would be valued.
“Given the size and scope of these (facilities), determining (the value of) individual units would be a nightmare,” he said.
Harper also emphasized that GPUs are replaceable. The company has seen carriers anticipate relatively short lifecycles and build more modular facilities in response.
“There is a core tension in data center project financing. Typically, lenders want an asset life that comfortably exceeds the term of the loan, but the short lifespan of GPUs calls that assumption into question,” said Marsh Risk’s Wolfson.
Therefore, lenders are becoming more cautious in making loans to protect themselves.
