Marsh AI Data Center Captive Insurance Is Surging as Traditional Capacity Falls Short

Marsh says AI data center captive insurance is heading for explosive growth as billion-dollar computing campuses strain traditional insurance capacity. The shift puts more financial risk inside the companies building and operating AI infrastructure.

A captive is an insurance company owned by the business it covers. Instead of transferring every exposure to commercial insurers, the parent company funds selected losses through its captive.

The structure is not new. Mining, energy, manufacturing, and transportation groups have used captives for decades. What changed is the scale and concentration of assets entering data centers.

AI campuses combine buildings, high-voltage equipment, advanced cooling, networking systems, and expensive computing hardware. They also depend on power grids, water systems, telecommunications networks, contractors, and specialized component suppliers.

This concentration creates an uncomfortable choice. Operators need substantial insurance limits, but commercial markets cannot always provide enough capacity at acceptable terms.

Captives offer another source of coverage. They also leave the owner responsible for losses that would otherwise sit on an insurer’s balance sheet.

That tension is the real story behind the captive insurance forecast. The AI infrastructure boom is not eliminating risk. It is changing who finances that risk, how losses are modeled, and where the consequences land.

Marsh Sees Captives Moving Into AI Infrastructure

AI data centers are turning captive insurance from a specialist financial tool into a central part of infrastructure planning.

Michael Serricchio, Marsh’s US and Canada captive solutions leader, expects rapid growth in captives covering data center portfolios. His argument reflects the mismatch between project values and available commercial insurance limits.

A captive can cover a deductible, fill an excluded layer, or retain an entire category of exposure. It can also buy reinsurance, which transfers part of its accumulated risk to another insurer.

This flexibility matters because a data center does not pass through one simple risk period. Its exposure changes from land acquisition through construction, commissioning, operation, and eventual expansion.

Construction introduces contractor failures, design changes, equipment delays, workplace injuries, and natural catastrophe exposure. Delayed completion can also postpone tenant revenue or trigger contractual penalties.

Operational facilities face different problems. Fire, cooling failures, transformer damage, power interruptions, cyber incidents, and equipment breakdowns can stop services even when the main building survives.

Marsh describes further dependencies in its overview of data center risks. A disruption involving power, fiber, contractors, or suppliers can produce significant losses without directly damaging the data hall.

That feature separates modern AI campuses from many conventional commercial properties. Their value comes from a functioning system rather than the building alone.

A warehouse can suffer a partial interruption while inventory moves elsewhere. A large AI cluster depends on synchronized computing hardware, networking, cooling, power quality, and software services.

Even a short disruption can affect multiple tenants and downstream customers. Business interruption losses can therefore exceed the cost of repairing damaged equipment.

Traditional insurers can still cover parts of this exposure. The difficulty comes when several risks accumulate at one location or across interconnected facilities.

An insurer underwriting multiple campuses might face correlated losses from the same storm, grid failure, equipment defect, or cloud service interruption. Reinsurers must also account for that accumulation.

A captive does not remove those connections. It gives the owner a vehicle for retaining exposures that the external market will not absorb completely.

The parent company funds the captive through premiums. The captive builds reserves, pays covered claims, and can purchase reinsurance above an agreed threshold.

This arrangement can make insurance spending more predictable across several years. It can also give the owner greater control over claims data and coverage language.

However, the captive needs real capital. It must satisfy regulatory, governance, actuarial, and reporting requirements in its chosen domicile.

That makes captive formation different from simply raising a deductible. The owner is creating a regulated risk-financing entity with obligations that continue after a project opens.

Marsh-managed captives wrote $79.1 billion in gross premiums during 2025, according to the broker’s 2026 benchmarking findings. Captive premium volume among participating Fortune 500 companies grew by 9 percent.

Those figures cover many industries, not only technology. They still show that captives were expanding before the latest wave of AI campuses entered operation.

The data center buildout gives that broader movement a new source of large, concentrated exposures. It also gives brokers and reinsurers a reason to design facilities around captive participation.

The Insurance Gap Starts With Project Scale

The defining problem is not whether insurers want data center business. It is whether they can support the required limits without concentrating too much risk.

AI data centers are becoming larger, more expensive, and more dependent on infrastructure built specifically for each campus. These characteristics create demand for coverage across property, construction, cyber, liability, and business interruption.

Swiss Re Institute estimates that global insurance premiums tied to data centers will reach $24.2 billion by 2030. Its estimate begins from $10.6 billion, showing how quickly insurance demand is expected to expand.

The reinsurer says a large data center construction project can exceed $20 billion. Installed GPUs and other technology can double the total value exposed at the site.

That combination creates a capacity problem. Lenders can demand insurance limits reflecting the full construction value, even when modeled maximum losses are considerably lower.

Commercial insurers and reinsurers must decide how much capital they can place behind one campus. They must also consider other facilities exposed to the same weather systems, suppliers, grid regions, or equipment designs.

Aon estimates that hyperscale project values commonly range from $10 billion to $50 billion. It expects global data center investment to reach between $5 trillion and $10 trillion by 2030.

The firm argues that conventional markets often provide only a fraction of the limits lenders require. Its analysis of alternative risk capital identifies captives alongside parametric coverage, insurance-linked securities, and segmented program structures.

Parametric insurance pays when a measurable trigger occurs, such as wind speed or rainfall crossing an agreed threshold. It does not require the same damage assessment as traditional indemnity coverage.

Insurance-linked securities transfer defined insurance risks to capital-market investors. These tools can add capacity beyond conventional insurer balance sheets.

Captives can connect both approaches. A corporate captive might retain a primary layer, purchase parametric protection, and use commercial reinsurance for severe losses.

That structure allows external insurers to participate above a level the owner can finance. It also reduces the amount of commercial capacity required for routine or predictable claims.

The mechanism does not guarantee lower total costs. Captive owners still pay claims, operating expenses, reinsurance premiums, and the cost of committed capital.

The attraction is greater control over where that money goes. A well-performing captive can preserve underwriting profit that would otherwise remain with a commercial carrier.

The owner can also adjust coverage around its operational data. That matters when insurers lack long claims histories for new cooling systems, high-density computing layouts, or behind-the-meter power generation.

Behind-the-meter generation produces electricity at or near the campus instead of relying completely on the public grid. It can speed development in regions with long interconnection queues.

It also transfers more responsibility to the owner. Fuel supply, generation equipment, emissions compliance, maintenance, and power reliability become part of the facility’s risk profile.

A captive can finance selected exposures from these systems. Yet it cannot make the underlying engineering safer or produce outside capacity when a severe event arrives.

This distinction is critical. Insurance solves a financing problem after the owner has addressed design, construction, operations, and emergency planning.

Large limits without credible engineering controls will remain difficult to place. Captives can support an insurance program, but they cannot substitute for risk reduction.

AI Data Center Captive Insurance Changes Who Holds the Loss

Captive growth reverses the usual insurance promise because owners gain control by accepting more direct financial exposure.

A commercial policy transfers defined risks to an outside insurer. The buyer pays a premium and expects the carrier to absorb covered losses above the deductible.

A captive keeps part of that process within the corporate group. The parent pays premiums to its own regulated insurance subsidiary, which becomes responsible for covered claims.

The captive can retain manageable losses and transfer severe layers through reinsurance. This arrangement creates a portfolio rather than treating each facility as an isolated placement.

That portfolio approach is especially relevant for hyperscalers, developers, and operators managing several campuses. Risk can be distributed across locations, construction phases, and operating units.

The owner gains access to detailed claims information. It can use those records to identify recurring equipment failures, contractor problems, or interruption patterns.

Better data can support negotiations with commercial insurers. It can also help the captive set retention levels according to actual experience rather than broad market assumptions.

However, portfolio diversification has limits. Several supposedly separate risks can fail together.

A transformer design flaw might affect equipment installed across multiple campuses. A regional grid disturbance could interrupt several facilities at once.

A common cloud dependency can spread business interruption beyond the damaged location. The same applies to shared network providers, cooling technologies, and hardware suppliers.

Geographic clustering adds another problem. Data centers often concentrate where land, power, fiber connections, and tax incentives are available.

Those advantages can place several valuable facilities within the same storm, flood, wildfire, hail, or earthquake zone. A captive holding risks from every site can face accumulated claims.

Swiss Re warns that modern infrastructure contains greater value concentration and stronger dependencies. Its premium forecast estimates $91 billion in cumulative AI data center premiums between 2026 and 2030.

Renewable energy infrastructure could generate another $111 billion over the same period. The two categories together exceed $200 billion in expected commercial insurance premiums.

The connection matters because AI campuses increasingly depend on dedicated energy projects, storage systems, transmission upgrades, and new generation capacity.

An interruption at the power asset can affect the data center without damaging its computing equipment. Coverage must therefore address dependencies outside the insured building.

Business interruption becomes harder to model when several entities own the connected infrastructure. The campus operator, utility, energy developer, tenant, and cloud provider can each carry different policies.

Contracts determine who bears penalties after an outage. Insurance language determines which losses qualify for payment.

Captives can cover gaps created by those contracts. They can also incubate risks until commercial insurers develop enough experience to offer broader terms.

That capability explains why captives appeal to sophisticated operators. It does not mean every exposure belongs inside one.

The owner must decide which losses it can absorb without harming investment plans, debt obligations, or daily operations. Retaining too much risk can turn one incident into a balance-sheet problem.

The main opponent in this market is therefore not one insurer or broker. It is captive control versus genuine risk transfer.

Captives promise flexibility, customized coverage, and access to reinsurance. Traditional insurance promises an external balance sheet that absorbs covered losses.

Most major data center programs will likely combine both. The unresolved question is how much exposure remains inside the corporate structure.

Commercial Insurers Are Building Their Own Answer

Traditional capacity is not disappearing, but brokers are reorganizing it around larger and more structured placements.

Marsh launched Stratus in August 2026 as a property insurance exchange for operational digital infrastructure. The structure offers up to $10 billion of capacity on one placement.

Thirty traditional and alternative capital providers can assess each risk independently. Marsh says that design helps participants measure, diversify, and manage accumulated exposures.

The Stratus exchange initially focuses on property coverage for global exposures belonging to US-domiciled companies.

Marsh plans to extend it into inland marine, cyber, and casualty lines. Stratus also connects with Nimbus, the broker’s construction-phase insurance facility.

The two offerings address a transition that can expose gaps. Construction insurance protects the project while workers and contractors build it.

Operational property insurance begins after commissioning. The asset value, equipment mix, contractual obligations, and interruption exposure can change sharply during that handoff.

A project might start with conventional buildings and electrical systems. Its value rises when operators install computing hardware and begin serving tenants.

The transition also changes the loss scenario. A construction delay affects completion, while an operational outage affects revenue and customer commitments.

Stratus shows that commercial markets still want this business. The exchange is designed to assemble multiple sources of capital around individual placements.

Captives can complement that capacity. An owner might retain a lower layer while Stratus participants support larger property losses.

This blended structure spreads exposure across the captive, commercial carriers, reinsurers, and alternative capital. It also places greater importance on consistent engineering and exposure data.

Insurers need to know what sits inside each campus. They also need information about power systems, cooling designs, fire protection, suppliers, regional hazards, and operating procedures.

The data challenge grows when hardware values change quickly. A facility’s insured value can increase after equipment upgrades or tenant installations.

Replacement costs can also move differently from conventional construction costs. Specialized chips, transformers, generators, and cooling equipment may face long delivery times.

A delayed replacement can amplify business interruption. The physical damage might remain limited while revenue losses continue for months.

Commercial carriers price that uncertainty into limits, deductibles, exclusions, and policy wording. Captives can assume the portions that external markets price conservatively.

Aon describes another route through parametric backstops. In one published case, the firm moved a client’s primary property layer into a captive.

It then added $25 million in parametric wind protection from third-party capital. The structure reduced dependence on traditional coverage without increasing the client’s retained exposure.

That example did not involve an identified AI data center. It still illustrates how captives and external capital can work together.

The market response is therefore broader than self-insurance. Brokers are building exchanges, captive programs, parametric covers, and reinsurance arrangements around the same capacity constraint.

Competition will focus on data quality, policy certainty, available limits, and claims performance. A facility with clear engineering controls should attract more options than one with poorly documented dependencies.

Captives Cannot Solve Correlated AI Infrastructure Risk

The strongest argument against unchecked captive growth is that ownership changes the financing structure, not the probability of failure.

A captive can reserve money for expected claims. It can negotiate tailored wording and purchase reinsurance for severe events.

It cannot prevent several campuses from depending on the same grid, hardware vendor, cloud platform, or cooling design. Those correlations define the hardest AI infrastructure risks.

Cyber exposure makes the problem clearer. A vulnerability in a shared service can affect many tenants without causing visible physical damage.

Coverage disputes can then involve cyber policies, technology errors and omissions insurance, property coverage, and business interruption provisions.

Aon found that 24 percent of surveyed captive owners underwrote cyber risk through their captives in 2025. That share was only 1 percent in 2014.

The firm’s cyber captive adoption data shows how companies already use captives for emerging exposures.

Cyber captive premiums managed by Aon grew 58 percent between 2022 and 2023. The growth reflects demand for deductible coverage, policy-gap protection, and access to reinsurance.

AI data centers combine cyber risks with physical infrastructure. An incident can start in software but produce operational consequences through cooling, access controls, or power management.

The reverse can also happen. Physical damage can interrupt digital services and create downstream contractual claims.

Captives face another limitation when loss histories are thin. Actuaries need credible assumptions about frequency, severity, and dependencies.

Historical records from conventional data centers might not reflect new AI campuses. Higher rack density, liquid cooling, custom accelerators, and dedicated power systems change the exposure.

Insurers and captive owners can compensate with engineering models and scenario analysis. Models remain estimates, especially when technologies and operating patterns change quickly.

A captive that underestimates loss severity can require additional capital from its parent. That demand might arrive when the underlying business already faces an operational disruption.

The parent must therefore evaluate captive risk alongside project debt, capital spending, and customer commitments. Insurance cannot sit in a separate planning silo.

Regulation also matters. Captives operate under the rules of their domicile and must meet capital, governance, reporting, and solvency requirements.

A structure created mainly to obtain favorable accounting treatment can fail its broader purpose. The captive needs a documented insurance rationale and credible claims-paying resources.

Tax authorities can also examine whether premiums and risk transfer reflect genuine insurance arrangements. Operators need qualified legal, actuarial, accounting, and regulatory advice.

These requirements favor large companies with diversified portfolios and mature risk teams. Smaller developers might use protected cells or group structures instead of forming standalone captives.

A protected cell segregates one participant’s assets and liabilities within a larger insurance entity. It can reduce administrative barriers, but the underlying exposure still requires funding.

Lenders will scrutinize these arrangements. They need confidence that the captive can pay claims and that reinsurance responds under severe scenarios.

Customers may also care. An AI provider promising high service availability depends on the financial and operational resilience of its infrastructure partners.

Insurance details rarely appear in product announcements. Yet coverage gaps can influence recovery speed after a fire, grid event, equipment failure, or cyber incident.

The skeptical view is not that captives are ineffective. It is that rapid adoption can create false confidence when risk models remain immature.

A properly capitalized captive can strengthen resilience. An underfunded captive can concentrate losses inside the same corporate group already facing the disruption.

That difference will only become visible through transparent capitalization, credible stress tests, disciplined reinsurance, and actual claims performance.

Three Signals Will Show Whether the Forecast Holds

The next phase will be measured by real captive commitments, commercial capacity, and losses rather than optimistic premium forecasts.

The first signal is captive premium growth among technology and communications companies. Marsh reported $79.1 billion across all managed captives in 2025, but industry-specific movement will matter more.

A clear rise in property, construction, cyber, and interruption premiums tied to data centers would support Marsh’s forecast. New formations alone would provide weaker evidence.

A company can establish a captive without transferring substantial exposure into it. Premium volume and retained limits reveal whether captives are becoming central to financing.

The second signal is how commercial facilities such as Stratus perform. The exchange offers substantial stated capacity, but individual providers still evaluate each risk.

Placement activity will show whether outside capital can support operational campuses at scale. Consistent participation would reduce pressure on owners to retain more risk.

Restricted limits, broad exclusions, or expensive reinsurance would strengthen the case for captive expansion. They would also suggest the capacity gap remains structural.

The third signal is the industry’s early loss experience. Fires, equipment failures, construction delays, grid interruptions, and cyber events will test policy wording and risk models.

Claims will reveal whether owners classified exposures correctly. They will also show whether interruption losses exceed physical damage, as reinsurers have warned.

A severe correlated event would put captive capitalization under immediate scrutiny. It could also cause commercial insurers to reduce capacity across similar projects.

Conversely, manageable claims and improving engineering data could attract more insurers. Greater outside capacity would let captives retain targeted layers instead of absorbing broad portfolios.

These signals will unfold as more AI campuses move from construction into operation. That transition shifts attention from completion risk to uptime, asset concentration, and downstream service obligations.

For developers, the immediate task is mapping dependencies before choosing a financing structure. Power, cooling, fiber, equipment, contractors, and customer contracts belong in one exposure model.

For enterprise AI buyers, infrastructure insurance is not a remote back-office issue. It can affect service continuity, recovery resources, and the financial stability of critical suppliers.

Procurement teams should ask providers how they manage concentration and interruption risk. They do not need confidential policy documents, but they need credible resilience answers.

Investors should watch whether operators retain increasing losses through captives while continuing aggressive capital spending. Both activities draw on the parent’s financial resources.

Regulators will need to evaluate whether captive capital and reinsurance remain adequate as facility values rise. They must also consider risks that cross state and national boundaries.

Marsh AI data center captive insurance is likely to grow because the underlying capacity problem is already visible. The decisive question is whether that growth produces stronger protection or simply relocates exposure.

Owners now have more ways to finance risk, including captives, commercial exchanges, reinsurance, and alternative capital. The durable programs will combine those tools with engineering controls and honest stress tests.

As the next generation of campuses opens, watch where operators place their largest loss layers. That decision will show whether captives are supporting outside insurance or quietly replacing it.