The global data centre construction pipeline has passed the point where it can be treated as a property risk with some technology overlay. AI infrastructure investment — running at hundreds of billions of dollars annually across hyperscaler programmes alone — is generating a category of exposure that sits uncomfortably across the existing boundaries of construction, property, casualty, and specialist lines. The London Market, with its structural capacity for complex, layered risk, is being asked to respond. Whether it responds well depends largely on how clearly underwriters and their clients understand what they are actually insuring.
A Construction Risk Unlike Prior Infrastructure Cycles
Data centre construction has always carried distinctive risk characteristics — high-value M&E content, tight commissioning schedules, condensed supply chains. But the AI-driven build cycle introduces several features that break materially from historical precedent, and which demand that the architects of insurance programmes treat prior loss data with significant caution.
The power density problem is the most immediate. A standard enterprise data centre of five years ago might have been designed around 5–10kW per rack. GPU clusters supporting large language model training are now operating at 50–100kW per rack, with liquid cooling increasingly required rather than optional. This is not an incremental change in thermal load — it is an order-of-magnitude shift in the concentration of electrical and mechanical risk within a given floor plate. The implications for CAR policy construction, for sublimits on M&E, and for the assumptions embedded in fire suppression engineering are substantial. Policies that were adequate for the prior generation of facilities are structurally mismatched to the current one.
The supply chain dimension compounds this. Nvidia H100 and H200 GPU clusters — the primary compute infrastructure for AI workloads — carry lead times that, at peak demand, have exceeded twelve months. A single hardware loss event on a commissioned facility does not carry the same recovery profile as a conventional property loss. Business interruption valuations must account for the cost of replacement hardware at current market pricing, the queue position for delivery, and the revenue impact of compute unavailability during a period of acute commercial competition among hyperscalers and enterprise AI programmes. Standard BI indemnity periods and reinstatement assumptions, derived from conventional property loss scenarios, will routinely understate the actual exposure.
There is also a geographic concentration dynamic that is only beginning to register in portfolio-level underwriting. The AI infrastructure build is not distributed evenly. It is clustering around power availability, land, water access, and political incentive — which means Northern Virginia, parts of the UK and Ireland, Singapore, and a small number of other locations are absorbing disproportionate concentrations of insured value. Accumulation management tools built for the prior generation of data centre risk need recalibration.
Operational Risk and the Failure Mode No One Has Priced
The construction phase, for all its complexity, is at least familiar territory for the engineering lines market. The operational risk profile of AI infrastructure is less well understood, and this is where the most significant unpriced exposure sits.
AI compute infrastructure operates under sustained, continuous high load in a way that conventional data centre hardware does not. Training runs for frontier models can sustain near-100% GPU utilisation for weeks at a time. The failure distributions for hardware operating under these conditions are not well-characterised in actuarial terms — the installed base is too new, the operating profiles too different from prior enterprise IT environments. Underwriters are being asked to price risks where the historical loss data is, at best, weakly predictive.
The installed base is too new and the operating profiles too different from prior enterprise IT environments. Underwriters are being asked to price risks where the historical loss data is, at best, weakly predictive.
Cooling system failure is the operational scenario that deserves most attention. Liquid cooling — direct-to-chip and immersion — is being deployed at scale because air cooling cannot manage the thermal density of modern GPU clusters. But liquid cooling infrastructure introduces failure modes that are largely absent from air-cooled facilities: coolant leaks in proximity to high-value electronics, biological contamination in open-loop systems, and maintenance complexity that most facility operators are only beginning to develop competence around. The loss potential from a significant cooling failure in a high-density AI compute hall is materially larger than an equivalent event in a conventional facility, and the triggering events are less well-understood.
Climate risk adds a further layer. A significant proportion of the global data centre pipeline is being developed in locations chosen for power availability rather than climate resilience. Facilities in regions with constrained water supply face growing tension between cooling requirements and regulatory or physical water access constraints. Facilities in coastal or low-lying zones face long-duration flood and storm exposure that is shifting as climate patterns change. Property catastrophe modelling for this asset class needs to account for the interaction between physical climate risk and the operational dependencies of the facility — power, water, connectivity — in ways that standard CAT models do not currently reflect.
Programme Architecture for a Risk That Doesn't Fit Clean Lines
The structural challenge for any insurance architect working on this class of risk is that the exposure profile spans lines that are traditionally managed separately and, in the London Market, placed through different channels. CAR, property, machinery breakdown, business interruption, cyber, and environmental liability all have legitimate claims on material portions of the risk. Designing a programme that addresses the exposure coherently — rather than reproducing the fragmentation of the insurance market in the structure of the coverage — requires deliberate integration at the placement design stage.
This is not an abstract structural point. There are specific coverage gaps that arise predictably when these lines are placed independently. Cyber exclusions in property policies, and property exclusions in cyber policies, create a gap precisely around the physical-digital interface where AI infrastructure risk is most concentrated. A cooling system failure caused by a software fault — or a firmware vulnerability exploited to cause physical damage — sits in the gap between property and cyber coverage unless the programme has been explicitly engineered to close it. The same logic applies to the boundary between machinery breakdown and property when GPU hardware fails under sustained operational load.
Contingent business interruption coverage for AI workloads requires particular care. The value at risk is not just the revenue of the facility operator — it is the downstream commercial impact on the enterprises and hyperscalers whose AI programmes depend on uninterrupted compute availability. CBI sublimits and trigger definitions that were adequate for conventional technology infrastructure will not hold in scenarios involving large-scale AI training programme disruption.
For London Market firms — carriers, brokers, and the managing agents responsible for syndicate appetite — the immediate implication is clear. The data centre boom is not a passing cycle that existing products can accommodate with minor adjustment. It represents a structural shift in the nature of the risk being presented to the market. Firms that invest now in genuine technical understanding of AI infrastructure — the engineering, the failure modes, the supply chain economics — will be better positioned to write this business profitably and to hold their positions when losses occur. Those that treat it as conventional property with a higher sum insured will find, in time, that they have been writing something rather different from what they believed.