AI data centers have become a "stress test" for insurers as rapid technological change and increasingly complex financing structures create a new set of challenges and opportunities for the sector.
Global spending on data centers could reach around $7 trillion by 2030, according to McKinsey, with AI‑optimized facilities expected to account for most new capacity. Much of that spend can no longer come solely from hyperscalers. Instead, Big Tech is increasingly tapping private equity and private credit and using debt to finance the capital‑intensive build‑out of new facilities, often via special‑purpose vehicles and off‑balance sheet structures.
Private infrastructure data center deals were consistently above the $10 billion mark last year, according to Preqin. The largest transaction reached $40 billion, with Nvidia, Microsoft, BlackRock and Elon Musk’s xAI among a consortium agreeing to buy Aligned Data Centers.
The scale of capital tied up in building and operating these sites has been a “real stress test” over the last four to five years for major insurers, Tom Harper, data center leader at Gallagher, told CNBC.
“When you put $10- to $20 billion plus in a single location, it creates capacity issues in the marketplace. The marketplace has always had an appetite for these risks because they are such high-quality builds … but the capacity - the ability to provide the insurance capacity at these locations - has been tough,” he said.
Harper said it was nearly impossible to reasonably insure a $20 billion campus in 2023. By 2026, however, it has become a “weekly conversation” as more global and London market carriers commit dedicated line size to AI‑heavy campuses and consortia or facility‑style placements become more common.
Estimated spending on AI data centers has been described as the biggest peacetime investment project in history. Rajat Rana, partner at Quinn Emanuel Urquhart & Sullivan, told CNBC he would go further, calling it the “largest peacetime investment project in human history, which is financed largely off balance sheet.”
Rana, who worked on structured finance litigation after the 2008 financial crisis, said tracking AI data center financing feels like “deja vu”.
“We’re talking about trillions of dollars, and almost going back to the same cycle where there’s almost no transparency about the financing structures - the scale is astronomical,” he said.
For insurers, that scale has a macro dimension. AI data centers are expected to account for a rapidly rising share of global electricity demand, with some projections suggesting they could represent around 8% of US power demand by 2030, up from about 4% in 2023. That raises questions about power‑availability risk, grid stability and sustainability targets, all of which are increasingly factored into underwriting and pricing.
Data centers demand a specialized approach from insurers, encompassing both real estate and technology assets. Some of the largest insurers are creating data center‑specific units or products, Harper said, while brokers are building multi‑disciplinary digital infrastructure teams.
The facilities pose particular challenges because of the concentration of values, the power required and “bleeding edge tech”, which can make them attractive but also complex to insure. Accumulation control has become a key focus as large AI campuses cluster in a limited number of power‑rich locations, including areas exposed to US wind, European flood and other natural perils.
Supply chain issues add another layer of risk when high‑value equipment is imported and stored before installation, often in third‑party facilities. Those stockpiles can fall between traditional marine, cargo and property wordings, forcing insurers to revisit how they structure transit, delay in start‑up and inland storage cover.
Growing M&A activity is also driving demand for warranty and indemnity insurance on data‑center‑heavy deals, as well as coverage linked to long‑term power purchase agreements and offtake contracts. Marsh has launched a dedicated digital infrastructure advisory group and the Nimbus facility for data center construction in the UK and Europe, recently expanded to offer limits of up to $2.7 billion.
As lending into the sector grows, insurers that protect lenders if a borrower defaults are starting to hit capacity limits, Marsh’s Alex Wolfson said, prompting work on excess‑of‑loss and portfolio solutions for banks and private credit funds.
Rana cautioned that as financing moves off balance sheet, it is harder for insurers to fully assess risk. US senators have already urged federal agencies to examine how Big Tech is using “complex and opaque” debt markets, warning of potential “destabilizing losses” for financial institutions. Rana said that opacity can translate into second‑order litigation risks for downstream investors, including pension funds and insurers, particularly around concentration, disclosure and valuation.
A key concern around potential financing cracks centers on GPUs and whether their shorter life cycles align with the longer lifespan of the facilities that house them.
CoreWeave, which sells AI compute in the cloud, is the first company to secure GPU‑backed loans, using high‑performance chips as collateral. Late last month, it announced an $8.5 billion, investment‑grade‑rated GPU‑backed deal.
While data centers typically have a decades‑long life, the average life cycle of a GPU is around seven years, and some analysts expect chip generations to turn over even faster.
“There are different data centers that are raising debt by disclosing different life cycles to investors,” said Rana, calling it the “GPU debt treadmill”.
“This is almost like a treadmill that these AI data centers are running on,” he told CNBC. As new chips arrive, operators may feel pressure to raise more debt and build new infrastructure, raising questions over how fast facilities can be delivered and how quickly credit can be obtained.
For insurers and reinsurers, the “GPU debt treadmill” crystallizes a mismatch between rapidly changing technology assets and long‑dated project finance and lease structures. That tension is feeding directly into thinking on asset valuation clauses, agreed‑value or replacement‑cost provisions and business interruption periods following a major loss at a heavily leveraged campus.
The cost of funding these projects is also fueling growth in asset‑backed securitization, including commercial mortgage‑backed securities and GPU‑backed notes. Harper said that while GPU life cycles are lengthening and modular, interchangeable designs are emerging, lenders are structuring loans more cautiously - and insurers are being pushed to refine how they assess, price and aggregate these fast‑evolving AI infrastructure risks.
