Marsh AI Captive Insurance Forecast Shifts Data Center Risk Back to Tech Owners

Marsh says AI infrastructure owners are turning to captive insurance as data center campuses become too concentrated for conventional coverage alone. The Marsh AI captive insurance forecast signals a reversal for technology companies accustomed to transferring physical risks outside their balance sheets.

A captive is an insurance company owned by the business whose risks it covers. It lets that parent retain selected losses while purchasing commercial insurance or reinsurance for more severe events.

The shift does not eliminate traditional insurers. It changes which risks they accept, how much capacity they provide, and what data they demand from technology companies.

Marsh’s Michael Serricchio expects “explosive growth” in captives supporting AI infrastructure, according to a September 12 captive growth report. That forecast arrives as Swiss Re and Aon warn that enormous campuses can turn one physical disruption into losses across several facilities and policies.

This is not simply an insurance industry story. Captive adoption exposes how technology companies plan to finance outages, construction delays, equipment failures, and natural disasters within their AI expansion programs.

It also places more responsibility on operators. A captive can retain a risk that commercial markets will not absorb efficiently, but the parent company ultimately funds that retained exposure.

The central contest is therefore clear. AI infrastructure owners want enough insurance capacity to keep construction and financing moving. Traditional insurers want protection from clusters of assets whose dependencies remain difficult to measure.

What the Marsh AI Captive Insurance Forecast Actually Changes

Captives are moving from a specialized corporate tool toward a central component of AI infrastructure financing.

Captive insurance has long served companies with unusual, expensive, or poorly understood risks. Mining, energy, transportation, and industrial businesses often use captives when standard policies offer limited capacity or restrictive terms.

Marsh now sees the structure becoming a preferred option for AI data center portfolios. The change reflects the scale of modern campuses, not a new variety of software liability.

These campuses combine buildings, processors, electrical equipment, cooling systems, telecommunications links, and backup generation. They also depend on contractors, utilities, chip suppliers, lenders, cloud operators, and tenants.

A traditional policy can cover parts of that system. However, insurers must consider what happens when one event affects several insured parties or coverage lines simultaneously.

Captives let an owner place selected risks inside a controlled insurance vehicle. That vehicle can cover deductibles, absorb predictable losses, or bridge exclusions in commercial policies.

It can also purchase reinsurance, which transfers defined portions of the captive’s exposure to other carriers. That approach separates risks the owner can retain from extreme losses requiring outside capital.

The immediate change is not that companies have stopped purchasing commercial coverage. Instead, the captive can become the organizing layer around a larger insurance program.

That arrangement gives owners more control over policy wording, claims information, and loss histories. It can also help different subsidiaries apply consistent coverage across a data center portfolio.

The tradeoff is equally important. Premiums paid to an independent insurer generally transfer contractual exposure away from the buyer. Money placed in a captive remains economically connected to the parent company.

A severe loss can therefore consume capital that the owner expected to use elsewhere. Captive growth represents risk retention as much as insurance innovation.

The Marsh AI captive insurance forecast matters because it shows owners preparing for limits within the conventional market. Their data centers still need coverage acceptable to lenders, investors, contractors, and customers.

Yet insurers cannot offer unlimited capacity simply because a project requires it. They must price the chance that one storm, grid failure, or equipment bottleneck affects several valuable assets together.

Marsh’s forecast puts that tension in public view. AI infrastructure is creating demand for insurance while simultaneously making portions of the exposure harder to transfer.

AI Data Center Insurance Has Outgrown the Single-Building Model

The most dangerous exposure is no longer one damaged facility, but several connected losses arising from the same event.

Swiss Re expects annual global premiums tied to data centers to increase from USD 10.6 billion to USD 24.2 billion by 2030. Its broader insurance demand forecast links that growth to construction and operational complexity.

The reinsurer says a large AI data center project can exceed USD 20 billion in construction costs. Installed processors and other technology can then double the project’s total value.

Those figures change the insurance problem. A conventional commercial building contains valuable equipment, but a hyperscale campus concentrates far more capital and operational dependency.

Construction adds one set of exposures. Developers face physical damage, contractor defaults, supply delays, commissioning problems, and postponed operating dates.

Operations create another set. A functioning campus must continuously coordinate power, cooling, networking, physical security, software systems, and tenant commitments.

A minor physical event can produce a much larger business interruption claim. Equipment might remain mostly intact while a damaged power or cooling component prevents the campus from serving customers.

The same incident can trigger several policies. Property damage, equipment breakdown, construction delay, business interruption, cyber, cargo, and liability coverage can overlap.

Aon describes this as accumulation risk, meaning multiple exposures combine into one portfolio-level loss. Its accumulation analysis identifies vertical and horizontal forms of concentration.

Vertical accumulation occurs when several policies cover different interests at one campus. The owner, contractor, operator, tenant, utility provider, and lender can each have claims connected to the same event.

Horizontal accumulation occurs when one disruption affects multiple campuses, regions, or insurance portfolios. A severe storm or shared supplier failure can create losses across otherwise separate facilities.

Time adds another dimension. A campus can enter phased operation while construction continues elsewhere on the site, causing insured values and policies to overlap.

This complexity pressures insurers because exposure data often arrives through separate business lines. An underwriter might understand one building without seeing every related tenant, network, contractor, or regional dependency.

Captives offer a place to consolidate that information. An owner can collect claims, engineering reports, asset schedules, and operating data across its portfolio.

Better information can support commercial placement because reinsurers receive a more coherent view of the retained and transferred layers. However, a captive cannot make hidden dependencies disappear.

The owner must identify shared substations, cooling designs, equipment suppliers, network routes, and construction partners. Otherwise, the captive could underestimate losses just as easily as an external carrier.

This is why AI data center insurance is becoming a data problem. Owners need a reliable map of physical assets, contractual obligations, and technical dependencies before they can allocate risk intelligently.

The challenge resembles knowledge management at an industrial scale. Teams must connect engineering records, contracts, incident reports, vendor data, and financial assumptions without losing their original context.

An inaccurate map can produce false confidence. The captive might retain exposure based on site-level estimates while missing a regional event that affects several campuses.

AI infrastructure owners are therefore being pushed beyond purchasing annual policies. They need continuous exposure management that follows each campus from planning through construction and operation.

Captives Give Owners Control, but They Also Return the Loss

A captive can improve coverage design, yet it cannot create economic capacity without capital behind it.

The attraction begins with control. A parent company can define which subsidiaries participate, which risks enter the captive, and how losses are reported.

This flexibility helps when standard policies exclude new exposures or impose deductibles that leave a meaningful gap. The captive can cover that first layer under terms aligned with the owner’s operations.

It can also retain frequent losses whose severity remains manageable. Commercial insurers and reinsurers can then focus on rarer events requiring much larger payouts.

That layered design can improve negotiations. An owner retaining a meaningful portion of risk has stronger incentives to invest in prevention, monitoring, and recovery planning.

The captive also creates a formal record of claims. Over time, that record can produce evidence about equipment failure, outage duration, contractor performance, and the cost of business interruption.

Good evidence matters because the AI data center market lacks decades of directly comparable loss experience. Current campuses use denser computing equipment and more complicated electrical and cooling systems than earlier facilities.

A captive can help bridge that evidence gap. It can collect operating information that external markets might not receive consistently from individual projects.

However, control does not equal transfer. When the captive pays a claim, the economic loss still belongs to the broader corporate group.

That distinction becomes critical during a correlated event. Several campuses might submit claims at once, drawing on the same captive capital.

The parent may then need to contribute additional funds, reduce coverage, or purchase more reinsurance. Each response can compete with spending on construction, processors, energy, or networking.

Captives also introduce governance demands. The company needs actuarial assumptions, reserves, underwriting rules, claims procedures, regulatory compliance, and independent oversight.

A captive built mainly to fill a difficult placement can become vulnerable if its pricing assumptions lag actual exposure. Rapid construction makes that danger more acute.

New buildings, installed processors, revised power arrangements, and additional tenants can change insured values within a policy period. Static asset schedules will not capture that growth reliably.

Aon recommends consistent exposure tagging across coverage lines and both deterministic and probabilistic loss modeling. Deterministic analysis tests defined events, while probabilistic modeling estimates many possible events and their likelihoods.

Neither method provides certainty. Their value comes from exposing how different assumptions change the loss distribution and required capital.

The Marsh AI captive insurance forecast should therefore be read as a shift in financial responsibility. Owners gain more influence over coverage, but they also accept a clearer share of adverse outcomes.

That arrangement can work when the retained layer matches the parent’s balance sheet and risk tolerance. It becomes dangerous when captive capacity is treated as an accounting substitute for outside capital.

The decisive question is not whether a company owns an insurer. It is whether that insurer has enough capital, reliable exposure data, and credible access to reinsurance.

Geographic Clusters Turn Efficiency Into Insurance Concentration

The same clustering that improves data center economics can magnify losses across the insurance market.

Data center developers favor places with available land, electricity, water, fiber connections, skilled contractors, and supportive permitting. These requirements naturally concentrate campuses in a limited number of regions.

Swiss Re reports that Texas and Virginia represent more than 40 percent of current and planned United States data center capacity. That concentration creates efficiency, but it also links facilities to common hazards.

Several campuses can depend on the same grid, transmission corridor, equipment suppliers, and construction workforce. A regional disruption can therefore spread without directly damaging every building.

Natural catastrophe exposure makes that concentration more visible. Swiss Re estimates that more than one-quarter of United States capacity sits in locations experiencing at least three large-hail days annually.

About 40 percent lies in areas potentially exposed to at least three tornado days each year. These estimates describe exposure patterns, not predictions that every facility will suffer a loss.

Data centers have design features that require careful modeling. Large roof areas, surface penetrations, outdoor equipment, and moisture-sensitive electronics can increase vulnerability to hail and water.

A single storm can also interrupt construction across multiple projects. Contractors may compete for replacement materials, specialized labor, and limited electrical equipment after the event.

High-voltage transformers present a particularly difficult dependency. Swiss Re says lead times for critical equipment can extend across multiple years.

The resulting loss can move beyond repair costs. A delayed transformer can postpone revenue, trigger contractual penalties, and extend financing obligations.

Traditional insurers respond by scrutinizing limits, deductibles, policy wording, and geographic aggregation. They must consider their exposure across all insured clients, not just one operator.

A captive views the problem from the opposite direction. It can see one corporate portfolio in detail but may not understand the full regional insurance market.

Reinsurers connect those perspectives because they accept risks from multiple insurers and captives. Their willingness to provide capacity depends on credible information about overlapping exposure.

This is where the primary conflict sharpens. Owners want insurance that recognizes their engineering controls and separates them from less prepared facilities.

Insurers want evidence that those controls remain effective during a regional event. They also need to know whether supposedly independent sites share hidden failure points.

A campus with redundant generators may still depend on one fuel network. Two facilities with separate electrical feeds may share an upstream substation.

Multiple sites might use identical cooling equipment containing the same design defect. Geographic distance alone does not guarantee independence.

Captives can support this analysis because the owner controls detailed operational information. The structure can encourage engineering teams and financial teams to use the same exposure model.

Still, internal ownership can create pressure to accept optimistic assumptions. A project team focused on opening capacity quickly may discount low-frequency events that complicate financing.

Independent modeling and reinsurer review provide a necessary counterweight. They test whether the captive’s retained layer reflects actual loss potential rather than budget preferences.

The insurance shift therefore reaches into site selection and architecture. Risk costs can influence where companies build, which suppliers they diversify, and how much redundancy they install.

That is a meaningful change for AI strategy. The cheapest megawatt or fastest construction site may not remain attractive after correlated exposure enters the calculation.

Alternative Coverage Expands, but the Verification Gap Remains

New insurance structures increase available options, yet every option still depends on measurable triggers and credible exposure data.

Captives are only one response to constrained conventional capacity. Commercial insurers, reinsurers, brokers, and specialist firms are developing several ways to distribute data center risk.

Parametric insurance is one example. It pays when an objective measurement crosses an agreed threshold, rather than waiting for a full assessment of physical damage.

Descartes Underwriting introduced parametric coverage for construction and operational data centers in January 2026. Its products address natural hazards including floods, hurricanes, earthquakes, tornadoes, extreme heat, and deep freezes.

The company says its United States capacity can reach USD 140 million per policy for hurricane and earthquake risks. That limit remains modest beside a campus valued in the billions.

Its usefulness comes from speed and flexibility. A measured flood depth or a tornado’s intensity and proximity can trigger payment without a traditional adjustment of every damaged asset.

That liquidity can support repairs, debt obligations, or delayed construction. It can also cover losses arising from interruption when direct physical damage remains limited.

However, parametric coverage introduces basis risk. That occurs when the measured trigger and the policyholder’s actual loss do not align.

A facility might suffer a serious interruption without meeting the defined threshold. Conversely, a trigger could produce payment even when the loss remains relatively small.

Traditional insurance still has an essential role. AIG offers coverage spanning construction, property, business interruption, equipment breakdown, cargo, cyber, environmental liability, and political violence.

Its data center program illustrates how commercial carriers are building integrated offerings rather than abandoning the sector. AIG also says more than USD 4 billion in annual premium flows through captives using its network.

Those options can sit beside a captive. The owner might retain predictable losses, purchase conventional coverage above that layer, and use parametric protection for specific hazards.

Reinsurance can then provide additional capacity. Facultative reinsurance covers an individual risk, while treaty reinsurance covers a defined portfolio accepted by an insurer.

Alternative capital may eventually take a larger role through insurance-linked securities or other structures. Such investors generally require clear models and carefully defined loss conditions.

Every layer depends on verification. Insurers need accurate asset values, engineering standards, geographic coordinates, dependency maps, and realistic interruption scenarios.

The available public evidence does not show which hyperscalers have moved specific data center exposures into captives. Marsh’s forecast describes a market direction, not a disclosed list of completed transactions.

It also does not establish how much risk each owner will retain. Captive structures vary widely, and a parent can use one for a narrow deductible or a much broader portfolio.

Readers should therefore avoid interpreting adoption as proof that commercial insurance has failed. The more defensible conclusion is that no single coverage mechanism can absorb the entire exposure efficiently.

The skeptical question concerns capitalization. Captive growth helps only when retained losses are backed by funds that remain available during stress.

It also requires governance capable of challenging internal expansion targets. A captive should not assume that additional campuses automatically diversify risk.

Common technology, suppliers, grids, and locations can create correlation across a portfolio. More sites can mean more concentration when the dependencies repeat.

The verification gap will narrow only when owners disclose structures or loss experience produces stronger data. Until then, forecasts should remain forecasts.

Who Faces Pressure as Risk Returns to the Balance Sheet

Captive growth forces technology owners, insurers, lenders, and customers to confront costs that conventional coverage once obscured.

Hyperscalers face the clearest pressure. Their construction schedules require insurance acceptable to financing partners, contractors, landlords, and local authorities.

If commercial markets limit coverage, the owner must retain more exposure or find another source of capital. Either choice can increase the resources tied to each project.

Risk managers also gain a larger role. Their work now influences capital allocation, site selection, supplier diversity, engineering standards, and operating resilience.

Finance leaders must decide how much captive capital the group can commit. They also need to test whether several simultaneous losses would strain liquidity.

Engineering teams face new reporting requirements. Insurers cannot model accumulation without consistent information about equipment, construction phases, power sources, and interconnections.

Cloud customers have indirect exposure. A severe campus outage can interrupt hosted applications even when the customer’s own systems remain intact.

Service-level agreements can transfer part of that cost back to an operator. Yet contractual compensation rarely captures every operational consequence experienced by customers.

Developers and enterprise buyers should therefore examine redundancy claims carefully. Two advertised regions may still share suppliers, network routes, or weather exposure.

The insurance program can offer useful clues. Underwriting questions often reveal dependencies that product documentation treats as implementation detail.

Traditional insurers face pressure from both sides. They want access to a growing premium pool, but they must avoid accepting several versions of the same underlying event.

Swiss Re estimates that AI data centers and renewable energy infrastructure together can generate about USD 200 billion in cumulative commercial insurance premiums from 2026 through 2030.

That opportunity comes with four structural concerns: larger assets, geographic clustering, supplier dependencies, and shared physical or digital networks.

Insurers that price too cautiously risk losing business to captives and alternative capital. Insurers that price too aggressively may discover hidden concentration after a major loss.

Brokers such as Marsh occupy the middle. They help clients combine captive retention, primary insurance, reinsurance, and specialized products.

Their influence grows when programs become more complicated. However, complexity can also make accountability harder to follow during a claim.

Lenders represent another pressure point. They often require limits connected to a project’s full construction value, even when modeled probable losses are much lower.

Swiss Re warns that the conventional market can support only a fraction of some required limits at competitive terms. Captives can fill portions of that gap, but lenders must accept their credit quality.

Regulators will also watch the shift. Captives must meet the rules of their domicile, maintain appropriate reserves, and document transactions with affiliated companies.

A weakly capitalized captive can transfer risk only on paper. Supervisors and rating agencies will focus on whether the parent can support the vehicle during adverse conditions.

The pressure ultimately reaches AI users. Infrastructure costs influence cloud contracts, computing availability, and the economics of training or serving models.

Insurance will not determine token costs by itself. Still, risk capital joins electricity, chips, construction, networking, and financing as a real input.

For knowledge workers and enterprise buyers, the practical lesson concerns continuity. AI services depend on physical systems whose resilience cannot be evaluated through model benchmarks alone.

Vendor reviews should include recovery arrangements, geographic dependencies, and contractual protections. The rise of captives shows that infrastructure owners themselves consider those exposures material.

Three Signals Will Test the Captive Insurance Shift

The next evidence must show whether captives improve risk allocation or merely relocate exposure within corporate groups.

The first signal is disclosed captive participation in major data center programs. Named transactions would reveal which risks owners retain and which layers remain with commercial insurers.

Such disclosures would strengthen the Marsh AI captive insurance forecast if captives cover meaningful property, interruption, or construction exposure. Narrow deductible arrangements would suggest a more limited shift.

The second signal is improved portfolio-level modeling. Aon says insurers need consistent exposure identification across campuses, policies, and project stages.

Tools that map common grids, suppliers, contractors, and network dependencies can improve underwriting confidence. Their adoption would support larger and more stable insurance programs.

Incomplete reporting would weaken that outlook. Capacity providers will remain cautious if they cannot distinguish genuine diversification from repeated exposure to one failure point.

The third signal is the behavior of insurers and reinsurers during renewals. Limits, deductibles, exclusions, and demanded retentions will show whether conventional markets view AI campuses as manageable.

Stable capacity would indicate that captives are becoming one coordinated layer within a broader market. Sharply reduced limits would suggest owners are retaining risk because outside appetite remains constrained.

Claims will eventually provide the hardest test. A significant storm, outage, equipment failure, or construction delay would reveal how different policies interact.

The crucial questions concern payment speed, overlapping coverage, capital adequacy, and recovery. A complicated program is valuable only when it functions under pressure.

Technology leaders should also watch site selection. Projects moving toward less concentrated regions would show that risk pricing has started influencing infrastructure strategy.

The same applies to supplier decisions. Longer transformer lead times and repeated equipment designs can turn one disruption into a portfolio problem.

Captives can encourage better decisions because retained losses give owners a direct financial reason to reduce dependency. That benefit disappears if the captive relies on optimistic models or weak governance.

The AI data center risk shift is therefore neither a retreat from insurance nor a simple expansion of it. It is a renegotiation of who funds uncertainty.

Marsh expects captives to absorb more portfolio exposure. Swiss Re expects insurance demand to grow while concentrations become harder to manage.

Both views can be correct. The market can expand rapidly while transferring a smaller share of each project’s total risk to independent carriers.

The final outcome will depend on evidence, not structure alone. Captives need sufficient capital, outside review, current asset data, and reinsurance that survives a regional event.

Enterprise buyers should ask how their AI providers separate facilities that appear independent on a status page. Investors should ask how much risk remains inside affiliated insurers.

Infrastructure teams should connect engineering decisions with insurance assumptions before construction begins. Fixing an unidentified dependency after deployment costs more and may not restore available capacity.

The Marsh AI captive insurance forecast deserves attention because it exposes a physical constraint beneath the AI boom. Computing expansion requires somebody to finance the consequences when infrastructure fails.

Over the next several months, watch for named captive placements, stronger accumulation tools, and material changes in renewal terms. Those signals will show whether the market is expanding its capacity or shifting the burden back to technology owners.