Building the infrastructure for the AI era, it turns out, is easier than paying for it. Major global lenders are finding that data centers — the physical backbone of every chatbot query and cloud computing workload — present a tangle of risks that existing banking playbooks were never designed to handle.

The scale of the exposure is striking. US data center-related debt has ballooned to at least $1.3 trillion as of mid-2026. The top 15 lenders alone hold roughly $196 billion of that exposure — about 15% of the total — with TD Bank at the front of the pack at $26.8 billion. Such numbers would be impressive for a mature asset class. For one that bankers openly describe as "entirely novel," they are borderline alarming. The borrowing mirrors the economics of the AI buildout itself: with single campuses carrying price tags once reserved for national infrastructure projects, developers have leaned heavily on credit markets that are still learning how to value the assets behind the loans.

The insurance gap nobody planned for

The core problem is deceptively simple: you cannot insure what you cannot fully understand. Construction costs for a single mega-project now routinely land between $10 billion and $30 billion. Meta's Hyperion campus, budgeted at $30 billion, managed to secure only about $4 billion in insurance coverage — leaving an enormous gap between what these facilities cost and what traditional insurers are willing to backstop.

That gap has real consequences at the deal table. Both KKR and Blackstone have walked away from multiple data center debt deals because insurance capacity simply was not there to match the risk profiles they required.

The risks themselves are layered. Asset longevity is a genuine question mark: a data center built around today's GPU architecture could become functionally obsolete within a decade if chip design takes an unexpected turn. Supply chain disruptions — particularly for specialized cooling systems and high-voltage electrical components — add another dimension of uncertainty. Concentration risk looms as well, since a single campus failure could wipe out billions in value overnight.

Marsh, the insurance brokerage giant, is trying to address part of the problem. On August 26, it launched an initiative called the Stratus exchange, designed to unlock an additional $10 billion in capital earmarked specifically for digital infrastructure risks. The global data center insurance market is projected to more than double, surpassing $24 billion by 2030. Whether that capacity arrives quickly enough to keep pace with current construction schedules is one of the open questions now hanging over the sector.

Community opposition is derailing projects

Insurance is not the only headache. In the first quarter of 2026 alone, at least 75 major US data center projects — valued at more than $130 billion collectively — were either delayed or scrapped entirely. The primary culprit was not financing costs or construction bottlenecks. It was community opposition.

Local resistance to data centers has shifted from a nuisance to a material credit risk. Residents near proposed sites are pushing back against the massive water and electricity consumption these facilities demand, along with concerns about noise, environmental impact, and strain on local infrastructure. Lenders, in turn, are starting to bake these social factors directly into their underwriting models. For an industry whose growth model depends on a steady pipeline of new sites, that introduces a bottleneck that sits outside the financial system altogether.

Banks improvise under pressure

Faced with insurance shortfalls and rising project risk, some lenders are improvising. JPMorgan Chase has reportedly been exploring synthetic risk transfers as one mechanism to manage its data center exposure — packaging and redistributing slices of risk to other investors rather than holding it all on its own balance sheet.

Those improvisations point to a deeper gap: data center lending still lacks the standardized tools — mature insurance products, proven risk-transfer markets, settled valuation methods — that other asset classes built up over decades. How quickly those tools emerge is now a central question for the sector, and for the $1.3 trillion of US debt already tied to it.