AI Data Centers Are Maxing Out P&C Limits, AIG's CEO Says

Eric Andersen, ten weeks into running AIG, told Bloomberg TV this week that AI data centers "need everything from project finance all the way through to operations... That's construction, that's cyber, that's property, that's liability," before landing the line that should reorganize a few capacity discussions: the buildout "is absolutely maxing out the P/C insurance industry in terms of the limits that are required" (Insurance Journal, August 2026). S&P Global Ratings puts numbers behind the claim: roughly $10 billion of new insurance premium in 2026 alone, about twice the entire global aviation insurance market (S&P via Beinsure, 2026).

The Exposure, Sized

S&P counts roughly 11,000 data centers in operation globally, an insurable asset base above $2 trillion, with annual investment in new construction on track to pass $300 billion by 2027. The number that matters for capacity, though, is per site: a single hyperscale campus carries estimated total insurable values of $20 billion to $30 billion, with the construction phase alone running $10 billion to $30 billion (Artemis, 2026). Those are values the property market last confronted in offshore energy and mega-infrastructure, and it confronted them with syndication structures built over decades. The AI buildout is asking for comparable limits on dozens of concurrent projects, on construction timelines compressed by hyperscaler urgency, for occupancies whose loss behavior has a few years of history at current scale.

Andersen's inventory of the coverage stack is worth taking literally. On AIG's second-quarter earnings call he described roughly 30 distinct insurance products a data center project needs across its life, from permitting and financing through construction into operations: builder's risk, marine cargo for equipment in transit, general and excess liability, cyber, business interruption, and power-generation covers among them (AIG Q2 2026 earnings call). AIG writes across that stack and insures the power providers feeding the sites, which is the basis of his claim that few carriers can serve the whole program. The claim is self-interested and also, on the current market's evidence, true.

The Only Hard Market Left

The strain Andersen describes lands in a market that is softening nearly everywhere else. Property catastrophe rates fell 16% at the mid-year renewals, casualty programs are fighting for rate against social inflation, and this site has spent the summer documenting reserve releases and pricing declines across commercial lines. Data center capacity is the conspicuous exception: demand for limits is growing faster than carriers can prudently supply them, which is the textbook precondition for rate adequacy, the thing the rest of the P&C market currently lacks.

That divergence creates a portfolio-steering question actuaries will recognize from every prior localized hard market. Capital flows toward the one segment paying for it, and the discipline question is whether the rate being charged reflects genuine uncertainty loading or merely scarcity. Both produce high premiums; only one produces adequate ones. The honest actuarial position is that nobody has credible frequency and severity distributions for a $25 billion campus of GPU halls: the fire experience of legacy server farms transfers imperfectly to liquid-cooled, high-density AI clusters, the business-interruption profile depends on chip supply chains with single-vendor concentration, and the liability exposure of facilities training frontier models has no case law. Pricing without loss history is what the profession did for cyber in 2015 and gene therapy in 2023. Some of those prices turned out to be scarcity masquerading as rigor.

What "Maxing Out the Limits" Means Mechanically

Translate Andersen's phrase into placement arithmetic and the strain becomes concrete. Large individual carriers typically deploy line sizes in the hundreds of millions on a single complex property risk; even the most aggressive rarely put more than a billion of net limit on one location. Building a $25 billion tower at those line sizes requires stacking dozens of participations across quota shares and excess layers, and every carrier in the tower is watching its aggregate across every other data center program it has already joined. The market ran this playbook for decades on refinery and offshore-energy schedules, but those exposures arrived a handful of projects at a time. S&P's $300 billion annual construction pace means the syndication machinery is being asked to assemble multiple such towers per quarter, and each new tower draws from the same finite pool of carrier capacity, the same reinsurers behind them, and increasingly the same facultative markets. Scarcity at the top layers is where it shows first: the last billion of a mega-tower is the hardest to place, carries the least premium per unit of limit, and is exactly where a capacity-constrained market starts declining. When a CEO whose company can write across the whole stack says limits are maxed out, he is describing the top of those towers.

The uninsured remainder is the confirming evidence. ENR reports billions of dollars of data center construction risk currently proceeding without full coverage, which in a rational market means the price of the missing layers exceeded what sponsors would pay, or the capacity did not exist at any price. Hyperscalers with trillion-dollar balance sheets can self-insure gaps that would stop a conventional developer cold, and that option disciplines how much rate carriers can extract from scarcity. The bargaining structure is unusual: the buyers are larger and better capitalized than the sellers, which caps the hard market's pricing power even while demand outruns supply.

The Business-Interruption Problem Nobody Has Priced at Scale

Inside the coverage stack, business interruption is the layer where actuarial methods have the least to stand on. BI severity for an AI campus depends on variables no historical fire book contains: replacement lead times for GPUs sourced from effectively one supplier's production schedule, the tenant's revenue model (training compute lost for months versus inference capacity rerouted in days), and contingent exposure through power infrastructure that Andersen notes AIG also insures, meaning one grid event can hit the carrier from two directions in the same portfolio. Standard BI valuation starts from historical revenue and recovery periods; a facility whose tenant's revenue did not exist three years ago and whose critical equipment has a supply queue rather than a spot market defeats both inputs. S&P expects exactly this slice, business interruption and technology losses, to stay substantially self-insured or migrate to captives (Artemis, 2026), which is the market conceding that it cannot yet price the tail. Carriers writing the BI layer anyway should be reserving against the possibility that their recovery-period assumption, not their frequency assumption, is the one that fails.

Aggregation Is the Actuarial Problem Inside the Growth Story

A $10 billion premium opportunity distributed across $20-30 billion single-site values is an accumulation problem wearing a growth costume. A hyperscale campus stacks multiple correlated perils on one footprint: construction all-risk during a multi-year build, property and equipment breakdown in operation, business interruption whose severity depends on GPU replacement lead times, cyber with physical consequences, and liability that follows whatever the tenant trains inside. A single convective storm, fire, or grid failure can trigger several coverages in the same tower of limits. Carriers modeling this as uncorrelated line-of-business growth will discover the correlation at claim time, the way marine and property markets discovered it in Tianjin and Baltimore.

S&P's own read is that the capacity gap will not be closed by traditional balance sheets alone: significant operational-phase exposure, especially business interruption and technology losses, will stay self-insured or move through captives, and the rating agency expects insurance-linked securities to take a growing share (Artemis, 2026). The broker side is already building the plumbing. Aon expanded its Data Center Lifecycle Insurance Program to $3.5 billion in April, adding a billion of capacity in one step (ENR, 2026), and ENR reports billions of dollars of data center construction risk currently running uninsured. Andersen spent his career until June at Aon, most recently as its president; the CEO now describing carrier capacity limits helped design the broker structures that aggregate demand against them. He knows exactly where the market's seams are.

What This Means for Pricing and Capital Work

For actuaries at carriers entering or expanding in this segment, the work divides into three problems. First, exposure-based pricing with disciplined uncertainty loads, because experience rating is unavailable: rate on values, construction characteristics, redundancy engineering, and named-peril accumulation, and resist the temptation to let scarcity pricing substitute for a loading you can defend in three years. Second, PML and accumulation control at the campus and the counterparty level: the same handful of hyperscalers stand behind a large share of global projects, so counterparty concentration compounds geographic concentration, and a carrier's aggregate across builder's risk, property, cyber, and liability to one tech company's ecosystem is a number someone should be computing quarterly. Third, reinsurance and alternative capital strategy: if S&P is right that ILS takes a growing share, ceding actuaries will be structuring covers for an exposure class the cat-bond market has never modeled, and the modeling firms' first data center vulnerability curves will deserve the same skepticism their first cyber curves earned.

AIG's own quarter shows why the growth is worth the trouble: the company beat expectations in Andersen's first full quarter, and its General Insurance book has the underwriting result to fund expansion. The buildout Andersen describes is the rare story where the insurance industry's constraint is not demand or rate but engineering: how much correlated, historyless, multi-billion-dollar exposure a balance sheet can carry per site. That is an actuarial question, and for once the market is paying whoever answers it carefully.

Four markers will show how the segment matures from here. The first 144A catastrophe bond with named data center exposure would confirm S&P's ILS thesis and put a market price on the tail; watch the sponsor, because a hyperscaler sponsoring its own bond bypasses the insurance tower entirely. The catastrophe modeling firms' first data center vulnerability curves will set the industry's loss picks for a decade, the way early cyber curves did, and deserve adversarial review before they harden into pricing. The size of Aon's lifecycle program at its next expansion will measure how fast brokered capacity is scaling against the $10 billion premium pool. And AIG's own disclosures are worth reading closely: if the company starts breaking out data center or infrastructure premium in its supplementary schedules, the segment has become material enough to reserve-watch, and the first accident year of that book will be the profession's first real loss datum in an exposure class currently priced entirely on judgment.