The next insurance policy protecting an AI data centre may be issued by a company owned by the business that needs protection. That can be a sensible way to manage a difficult risk. It also means the word insured tells an investor less than it first appears to about who ultimately pays after a loss.
Bloomberg reported on September 12 that captive insurance was emerging as a coverage model for AI infrastructure, citing Marsh's captive-solutions leadership. That is an industry assessment, not a published census of adoption. The financially important question is how the arrangement divides losses between the operator, its captive and independent insurers.
Follow the claim back to the owner
The NAIC defines a basic captive as an insurance subsidiary owned by the non-insurance business it covers. This is a form of self-insurance. The subsidiary may issue policies and settle claims, but ownership means that a payment within the group does not by itself make the group's economic loss disappear.
Imagine an operator whose equipment is damaged and whose captive pays the repair claim. The operating subsidiary receives cash, while the captive loses cash. At the combined-owner level, money has moved between pockets and resources still have to be spent restoring the equipment. This is a hypothetical accounting explanation, not a description of a reported accident.
The structure can nevertheless be useful. A dedicated insurance vehicle can make retained exposure explicit, accumulate resources for claims and tailor the coverage to the business. It can also participate in a wider programme with outside insurers. The mistake would be to infer that establishing the vehicle automatically transfers every adverse outcome away from shareholders.
Independent risk transfer requires another party to assume a covered obligation. The NAIC's reinsurance explanation describes the transfer of some or all policy risk from an insurer to a reinsurer. Applied to a captive programme, the relevant questions are which losses remain retained, where external cover begins, and what contractual limits apply. An insurance label cannot answer them.
The same outage can reach several policies
Data-centre exposure is not just the replacement cost of a building. An interruption can also affect the service the facility supplies. Depending on the contracts, different parties may bear different consequences. A property policy, a business-interruption policy and an agreement with a customer need not recognise the same event in the same way.
That creates a potential aggregation problem. If several assets depend on a common power source or other shared infrastructure, their losses may occur together. Adding more sites does not necessarily diversify that particular dependency. This is a scenario for assessing risk concentration, not a claim that a specific operator has failed to diversify.
Swiss Re's September research explicitly places accumulation risk alongside the opportunity created by infrastructure investment. The analytical implication is that the amount insured cannot be considered independently of how losses cluster. A programme spread across many individual policies can still concentrate economic exposure in one event.
Capital must be available when claims arrive, not only when premiums are collected. The NAIC notes that captives face requirements such as capital, reserves and reporting. For an owner, the associated commitment is part of the cost of retaining risk. Money supporting insurance obligations is not economically interchangeable with unrestricted cash available for the next construction project.
This does not establish that captives are weaker than commercial insurance. A well-designed programme can deliberately retain manageable losses and transfer more severe ones. The quality of that design depends on coverage, capital and counterparties. Comparing premiums alone would miss precisely the features that determine whether the arrangement is robust.
Premium growth leaves the loss bill open
The potential market is substantial, but its measurement needs care. Swiss Re Institute estimates that AI data-centre construction and operation could generate about $91 billion in cumulative insurance premiums over 2026–2030. That is a forecast across several years, not annual revenue already earned and not profit attributable to a particular insurer.
The same research separately estimates renewable-energy premiums. Combining the categories into one AI insurance figure would blur what is being counted. More fundamentally, premium income precedes the expense of claims, administration and the capital supporting the business. A larger market can create opportunity while also increasing the amount at risk.
For a data-centre owner, the meaningful comparison is the total cost of coverage and retention, including claims it ultimately funds. For an outside insurer, the corresponding test is whether pricing and contractual terms compensate for the assumed exposure. Both sides can report growth while reaching very different economic outcomes.
Evidence of clearly defined loss layers, credible external reinsurance and adequate resources for retained claims would support the case for a captive. Evidence of concentrated dependencies or unexpectedly large retained obligations would challenge it. The insurance boom therefore deserves a balance-sheet reading: the decisive fact is where a covered loss ends up, not simply whose name appears on the policy.
