Data Center Late Fees Hit $5 Million or More Per Day, Spawning New Insurance

This spring, Mike Mathews, who leads the global digital infrastructure practice at insurance broker Marsh, raced to design a policy to protect a data center developer from a $1 million-a-day late fee. The developer’s tenant was demanding damages if electricity and facilities weren’t available on time during the first phase of a new gigawatt-scale facility.

Mathews’ team arranged the novel coverage, a bespoke form of “delay-in-delivery” coverage just for data centers, only to see penalties escalate sharply in a matter of months: Tenants in new deals began asking for upwards of $5 million for each day of delay, citing even higher estimates of lost revenue.

Welcome to the wild, and no longer boring, world of insurance whisperers trying to keep data center developers alive. Marsh, one of the two leading insurance brokers alongside Aon, is designing policies for AI infrastructure risks where data can be scant and financial exposures are growing to a scale never contemplated.

Bankers and bondholders are asking for more of this insurance to get comfortable underwriting record amounts of spending by first-time developers and money-losing AI startups. The recent quintupling of this type of late-delivery fee, and escalation of others, “sent a shockwave,” says Mathews. “I’m not saying investors were hands off on insurance, but they’re all hands on now.”

Swiss Re estimates that data center insurance premiums could approach $24 billion by 2030. S&P Global estimates premiums are already running around $10 billion 2026, which is twice the size of the global aviation insurance market. Data center construction growth has been so rapid —now roughly $2 trillion in insurable value and growing $300 billion a year—that S&P warned in a recent report, “We expect capacity constraints to limit the industry’s ability to fully insure these hyperscale data center projects, as total insurable values reach $20 billion-$30 billion per location.”

When every party in a buildout relies on borrowed money, a delay can create cascading problems. “My bet is late penalties will bankrupt more developers than interest rates ever will,” KKR’s global head of digital infrastructure, Waldemar Szlezak, wrote this month in a recap of a J.P. Morgan infrastructure investment forum.

Mathews compares insurance to a clock: “Over the past 100 years, insurance has maybe changed five minutes. But if you look at digital and data centers, it spins like a stopwatch.”

GPU Insurance

Mathews has hired mostly data center development and operations veterans like himself, rather than career insurance brokers, to staff his team in recent years, and just brought on a former risk manager from Meta. They are helping insurers design coverage for emerging risks, including service interruptions caused by new types of equipment, potential declines in the value of Nvidia graphics processing units, and financial guarantees for multibillion-dollar utility upgrades, even when developers may not use the full power capacity. Efforts to insure the residual value of GPUs are still nascent, Mathews said. Older chips are holding their value or even rising in a scarce market, so pricing such protection is tricky.

Meta said Marsh is its insurance broker on a data center project in El Paso, Texas. Marsh has its own role in arranging insurance solutions for Meta’s large AI campus in Monroe, Louisiana. Marsh recently said it found insurance providers for more than $5 billion in multiple forms of credit protection for a data center developer making long-term commitments to pay for utility infrastructure. Oracle has had to satisfy such commitments in Wisconsin, but Marsh did not identify its client.

The AI risks that need insurance today do not fit neatly into the categories that builders and cloud providers have traditionally been able to insure. Conventional policies cover construction delays or business interruptions caused by physical events, such as fires or storm-related water damage. They did not cover the risk that a utility cannot provide power on a promised timeline or that an on-site power plant is delayed.

Mathews offers an analogy: Imagine building or renovating a house that the developer promises will be ready by Christmas, only to find you must wait until February for appliances or a power hookup. That aggravating delay is usually tough luck—not an insurable event. If, after you move in, a leak or air-conditioning failure forces you out during repairs, it might be on you to spring for a few nights in a hotel. Annoying, but it won’t bankrupt you.

Now imagine a data center firm that promises to hand over as much as a 1 GW powered data center shell to an AI developer or cloud provider, only to postpone delivery from Christmas to the end of February. Late fees of up to $5 million a day could consume the entire daily rental income that the facility’s landlord was expecting, says Christian Okoye of Occam Edge, which quantifies delay costs for investors, lenders and developers. Over those two months, the data center firm could owe $300 million in penalties while also forgoing rent—a damaging combination for loan repayments.

And an AI lab could forgo multiples of that $300 million in monthly revenue if it can’t get its chips energized, as we recently reported, or if the data center goes down.

Even for risks that are insurable, like storms, owners may make a business decision to only insure part of a gigawatt scale campus given that the facilities plus the pricey chips and servers cost around $50 to $60 billion per gigawatt. Mathews says he hasn’t found any project to be uninsurable and points out that reinsurers have excess capital they want to put to work in this market.

Or, to satisfy bondholders, some developers may go above and beyond. The relatively young bitcoin-miner-turned-AI-infrastructure developer Hut 8 is building a data center in Corpus Christi, on the Gulf Coast of Texas where hurricanes are a regular risk. Sean Glennan, CFO of Hut 8, said “we went well above maximum and probable foreseeable loss on these policies…We said, ‘Let’s get very robust insurance policies to give bondholders peace of mind.’”

Hut 8 also bought 19 months of “delay in startup” insurance to provide some revenue protection if insured damage causes a delay. And Glennan says he negotiated a lease that can’t be terminated if the Texas campus experiences certain delays. Hut 8 used the insurer FM and the broker Aon.

Pooling Risk

Marsh and Aon have both been arranging multibillion-dollar insurance facilities, which are arrangements where insurers pool risks across multiple projects. Pooled risks are necessary because no single insurer is set up to absorb tens of billions of dollars of risk on a single campus. Because portions of sites must be “self-insured or only partially insured,” S&P said in an April report that it expects growth in alternative arrangements like insurance-linked securities.

Tim Walker, Marsh’s digital infrastructure construction leader, points out that some challenges just aren’t insurable. “What happens when you don’t have enough people to build all these things? To me that’s one of the biggest risks.”

The demand just keeps getting bigger as penalties rise: RBC Capital Markets recently reported that a data center where temperatures deviated from safe targets for one month risked a $150 million penalty. A report by Gallagher Re found that a single 45-minute outage could wipe out 40% of a data center’s annual cash flow in penalties, if it is held to the traditional industry promise of 99.999% uptime, or about five minutes of downtime a year.

An old risk, that rats could eat through data center-related cables, has re-surfaced as an even bigger worry. A certain fiber maker used to make a cable in a shade of orange that rats loved, leading to new insurance add-ons for “vermin.” Vermin could cost providers a whole lot more as they rush back into cities to serve up AI inferencing. “You can build a $10 billion, state-of-the-art facility, and Chuck E. Cheese comes and eats your fiber and you’re out.”

Ann Davis Vaughan is the author of the AI Infrastructure newsletter for The Information. She is a former senior Wall Street Journal investigative reporter turned investment strategist who has tracked the energy, industrial and financial sectors for three decades.