AI campuses hit the insurance ceiling, and outage risk sits outside policies

Noise and signal

The noise runs like this: data centres are insurance's new growth area, and the premium pool will soon double. In March the Swiss Re Institute wrote that global data centre premiums would rise from $10.6 billion to $24.2 billion by 2030. The picture implies that where demand exists, capacity will follow.

The signal is different: the problem is not the size of the premium but the concentration of value in one place. Swiss Re's sigma 3/2026 report, published on 7 September, puts the replacement cost of a single AI campus at up to $50 billion. S&P Global Ratings estimates the insurable value of such a campus at $20–30 billion. Large insurers' per-risk limits, meanwhile, stay in the low single-digit billions.

Why the ceiling sits at $5 billion

Aon's data centre lifecycle programme opened at $1.5 billion in June 2025, rose to $3.5 billion in April 2026 and to $5 billion in July. Capacity is drawn from Lloyd's, the company market and other facilities. The programme provides a further $1 billion for terrorism and $400 million for cyber and technology professional liability.

Even this limit does not reach a quarter of S&P's $20–30 billion campus value. In July Kelly Kinzer, Zurich's global head of construction and surety, said the market lacks the capacity to insure the largest projects at full value. Lenders, meanwhile, are increasingly making full replacement-value cover a condition of financing.

Swiss Re identifies four sources of accumulation: the size of single assets, geographic clustering, supply chain dependence and shared physical and digital networks. More than 40% of existing and planned US data centre capacity is concentrated in Texas and Virginia. More than 25% of capacity sits in hail zones, and about 40% in areas with at least three tornado days a year.

Outages fall outside the policy

According to a Howden study announced on 11 November, half of all facilities reported a significant outage in the past three years. Each outage costs between $500,000 and $1 million, implying global losses of about $2 billion a year. Of major outages, 45% stem from power and uninterruptible power supply failures, 14% from cooling and 11% from networks.

Under a conventional property policy, business interruption cover only responds if there is physical damage. Howden writes that only a small share of outages produces insured physical damage. Most of the $2 billion annual loss therefore falls outside policies today; the broker proposes parametric cover tied to service level agreements to fill the gap.

The root cause is the grid. In the US, data centre power demand has grown 19% a year since 2000, while grid capacity has grown 2% a year. Data centres' share of US grid capacity rose from 0.1% in 2000 to 5% in 2025. Howden expects about 10 GW of off-grid generation to be built by 2030, adding a new power plant risk for insurers.

The war exclusion has reached the data centre

Howden writes that up to five data centres in the UAE and Bahrain were struck by Iranian drones in 2026, the first known deliberate wartime targeting of commercial data centres. In 2025, capacity built within 10–15 km of conflict zones reached about 60% of the 2020–24 total.

This shows the war exclusion in property policies is no longer an abstract clause. On 10 November Rachel Turk, Lloyd's chief of markets strategy, urged insurers to state clearly what they cover on war, AI and sanctions. She said ambiguity works against insurers in court.

Who closes the gap

At a London panel published by Artemis on 10 November, speakers said capital markets can only tackle the insurability problem, the ability to transfer risk at a reasonable price, from the reinsurance and retrocession end. Retrocession is the protection a reinsurer buys to pass on its own risk. The panel listed four gaps: risk transparency, correlation with investor portfolios, standard structures and the rapid depreciation of chip value.

Swiss Re expects about $200 billion in total premiums from AI data centres and renewable infrastructure over 2026–2030. Meeting that demand depends on how quickly accumulation modelling matures before the 2027 renewal. Accumulation that cannot be modelled may be the one item that keeps prices high even in a softening property catastrophe market.

What would falsify this

If a broker or insurer announces in the first half of 2027 that it has placed more than $10 billion of cover on a single campus, the ceiling thesis weakens. If two major insurers add non-damage business interruption to standard property policies for an extra premium, the outage gap thesis fails. In the other direction, a hail, tornado or war loss above $1 billion at a single campus would quickly pull capacity back.