AI can feel weightless and abstract: A cloud of algorithms and models drifting somewhere on the internet. Its insurance risk, however, is anything but. Generative AI depends on vast physical infrastructure, including hyperscale data-centre campuses, dense server halls, specialised electrical and cooling systems, and extensive power and fibre connections. The land and climate surrounding each facility also shape its exposure to hazards, such as flooding, wildfire, and extreme heat.
What makes this exposure particularly challenging is not simply the value of individual facilities, but how much risk can accumulate around them. Billions of dollars of assets can sit on a single site, while multiple campuses may depend on the same power infrastructure, equipment suppliers, or network connections.
Those facilities are being built at a speed and scale the insurance market has rarely seen. In the US, data-centre construction starts reached $77.7 billion in 2025, a 190% increase on the previous year. By June 2026, year-to-date starts had already reached $81.5 billion, exceeding the total recorded for 2025.
The premium implied by that build-out is a new pool rather than a marginal one. S&P Global Ratings estimated in April 2026 that data centres would generate around $10 billion of new premium in 2026 alone, roughly double what the entire global aviation industry produces annually. Aon's January 2026 renewal report puts cumulative global data-centre premium at $134 billion between 2026 and 2030; Swiss Re Institute projects annual premium rising from $10.6 billion to $24.2 billion by 2030. The estimates differ by a wide margin, and that spread highlights that nobody yet knows how much of this exposure the market will end up carrying, or on what terms.
Beneath the premium sits capital expenditure at a different order of magnitude. McKinsey estimates $6.7 trillion of cumulative data-centre capex by 2030, more than $4 trillion of it in computing hardware, with over 40% of the total spent in the United States. Aon's range is wider at $5 trillion to $10 trillion, across more than 2,100 facilities under construction or planned. Individual campuses can require limits of $10 billion or more.
This raises a fundamental question for insurers, how do we underwrite a risk that is both highly concentrated and deeply interconnected — and who has the capacity and expertise to lead?
For risks of this scale and complexity, the London Market is well positioned to shape the answer.
AI’s physical footprint concentrates an unusual amount of risk
From an insurance standpoint, one of the defining challenges of AI data centres is aggregation. One fire, flood, grid failure or construction loss can quickly run into 10-figure numbers. The same event may also affect several facilities across a portfolio. This creates a question that goes beyond traditional property underwriting. When a loss occurs, is it fundamentally a property loss, a cyber loss, or something that cuts across both?
Delay in Start-Up (DSU) cover highlights the challenge. During construction, a damaged transformer, generator, cooling system, switchgear component, or shipment of specialist hardware may be difficult to replace. Competition for critical equipment can extend lead times, turning a physical loss into a prolonged interruption and a potentially very large DSU claim.
Once operational, the risk profile changes but does not become simpler. Data centres depend on continuous power, advanced cooling, connectivity and tightly integrated building-management systems. For many tenants and downstream businesses, their own business interruption exposure also depends on a small number of cloud and data-centre providers.
A regional outage or service-provider failure could therefore create correlated losses across multiple policies and portfolios. The risk is no longer confined to what happens inside an individual building; it extends to the network of infrastructure and businesses that depend on it.
Cyber adds another dimension. Operational technology, including power-management, cooling and building-management systems, is increasingly connected to corporate networks and managed remotely. A compromise could disrupt operations or cause physical damage by manipulating temperature, power, or safety controls, creating a potential gap between property and cyber cover.
The challenge for insurers is therefore not simply understanding each individual exposure. It is understanding how those exposures interact or compound.
What should AI infrastructure insurance look like?
The market could approach AI data-centre risk in several ways.
A new specialty class. Data-centre and AI infrastructure could be treated as a defined segment, with its own appetite, limits and accumulation controls. Existing property, engineering, DSU, business interruption, cyber, and cargo coverages could be brought together within a broader programme. Aon’s Data Centre Lifecycle Programme and Marsh’s Nimbus illustrate this type of integrated structure.
A cyber specification. A dedicated cyber extension could address service-provider outages, dependent business interruption and attacks on operational technology that cause physical damage or disruption. Property and engineering cover would still be needed for non-cyber events such as fire, flood, equipment breakdown, and construction delays.
A standalone AI infrastructure cover. A single product could combine relevant property, engineering, DSU, business interruption and cyber protections under one proposition, while retaining separate limits, triggers, and underwriting considerations for each exposure.
Whatever structure emerges, the London Market is well positioned to shape it.
It has centuries of experience developing cover for complex, international and emerging risks, alongside specialist expertise across energy, engineering, marine and cyber. With individual risks likely to cost billions, London Market’s layered placements present a clear advantage.
Data-centre geography makes that relevance concrete. The US is the clear centre of gravity for global data-centre capacity, as the graph reported by Insurance Insider shows. That aligns with the London Market’s international footprint and its substantial exposure to US business. London therefore offers more than specialist expertise. It also provides an established route to international capacity, assembled across multiple insurers through its subscription model.
And brokers and insurers are already moving. Aon launched its Data Centre Lifecycle Programme in June 2025 with $1.5 billion of capacity. It reached $2.5 billion in January 2026, $3.5 billion in April, and $5 billion in July. Others are building too, with Marsh's Nimbus facility, backed by a group of Lloyd's and company insurers, growing to $2.7 billion. Fidelis has set up a consortium at Lloyd's for construction risk. Additionally, AIG UK covers the whole life of an asset, from building it to running it.
How can insurers manage the complexity?
Underwriting AI data-centre risk will require more than capacity. Insurers will need a consistent way to bring together information that traditionally sits across different parts of the underwriting process. This includes site information, engineering reports, policy wordings, supply-chain dependencies, hazard data, and cyber intelligence.
The right technology can help create that connected view. It can support a more streamlined underwriting process, from submission and risk assessment through policy administration and claims.
AI data centres are bringing a new combination of property, engineering, climate, supply-chain and cyber-physical exposure into the insurance market. Whether the market responds through a new specialty class, a cyber specification, or an integrated standalone product, the challenge will be connecting these exposures and managing accumulation at scale.
London Market’s history and specialist expertise provide the foundations for that work. The next step is giving underwriters the data, workflows, and decision support to turn those foundations into sustainable capacity.