How Should Operators Set AI Infrastructure Insurance Limits in 2026?

Direct Answer to the AI Infrastructure Insurance Limits Question

There is no defensible universal dollar limit for AI infrastructure insurance in 2026. The correct limit is the loss the operator can financially survive, translated through a detailed maximum-per-occurrence analysis for the facility, its equipment, power dependencies, business-interruption exposure, and any contractual indemnities. A stand-alone equipment limit should cover replacement cost, but property damage is only one layer: a realistic program also needs business interruption or contingent business interruption cover, cargo protection during construction and transport, machinery breakdown, cyber risk, natural catastrophe, and liability coverage.

Also worth reading: How Is Scaling Insurance AI Agent Infrastructure Transforming Modern Broker Operations? · How Do Equipment Insurance Claim Limits Work, and How Much Coverage Do AI Companies Need in 2026? · What Are California Rideshare Insurance Limits for Uber and Lyft Drivers in 2026?

The market issue is that conventional data-center limits were often selected when facilities contained fewer, less capital-intensive servers and faced different supply-chain and customer dependencies. AI clusters can concentrate extraordinary value in power, advanced processors, networking equipment, cooling systems, and long-term customer contracts. Existing property limits can therefore become misaligned with both physical replacement costs and the income exposed to a prolonged outage. Aon’s reported expansion of a data-center insurance program to $5 billion illustrates how capacity is being scaled for the sector, but a program’s total capacity does not tell an individual operator how much insurance it should buy.

The safest answer is to set limits site by site rather than by adopting a global “AI data-center” figure. Insurers may impose sublimits for high-value equipment, generators, transformers, cooling, or business interruption, while exclusions and deductibles can materially change the economics. Operators should obtain at least two independently constructed valuations and several broker-led loss scenarios, then document how each scenario affects liquidity, debt covenants, customer service-level agreements, and restart schedules. A limit that protects the balance sheet after a catastrophic event is not necessarily the limit that produces the most efficient premium; those are separate decisions.

Why Traditional Data-Center Limits May No Longer Fit

AI infrastructure is not simply an ordinary data center with more racks. High-density accelerators increase electrical load, heat rejection requirements, component replacement costs, and the operational consequences of a single failed upstream system. A facility may also depend on contracted grid capacity, on-site generation, fuel, liquid cooling, specialist technicians, replacement processors, and software reconfiguration. These dependencies must be treated consistently because the absence of a covered building does not make the financial loss disappear.

Demand growth amplifies the mismatch. Research cited in the supplied context reports that AI data-center infrastructure could require approximately $2.8 trillion of capital expenditure by 2030 under one estimate, while McKinsey has estimated a figure approaching $7 trillion for broader AI infrastructure investment. These are forecasts, not insurance values, and they differ in scope and methodology. They nevertheless demonstrate why the insurance market cannot safely assume that historical asset values, depreciation rates, and interruption patterns will remain stable. Equipment may depreciate rapidly on an accounting schedule while remaining physically essential to a multi-year customer contract.

A second problem is the distinction between values that can be replaced quickly and values that cannot. Servers may be available only by allocation, logistics may take months, and a damaged site may need a newly built replacement elsewhere. Business interruption can consequently outlast the property restoration period. Willis’s reported call for data-center operators to rethink insurance buying, and Risk & Insurance’s discussion of data centers overbuying insurance, both point to different forms of inefficiency: some buyers purchase layers that are highly unlikely to contribute after a severe event, while others leave underinsured gaps.

Limits should therefore reflect the actual recovery model. If a damaged cluster can be migrated to another cloud region in 14 days, income interruption may be limited even when equipment value is high. If workloads are tied to a specialized site, a scarce grid connection, or customer processes that cannot tolerate substitution, the insured period may be much longer. The underwriting question is not merely “What did the equipment cost?” It is “How long, where, and at what cost can the operation resume?”

How to Calculate Property and Equipment Limits

Begin with a replacement-cost valuation as of the proposed policy inception date, including installation, engineering, racks, cabling, power distribution, cooling, generators, uninterruptible power systems, and decommissioning costs where relevant. Do not limit the account to server invoice prices. Replacement cost can rise during a claims period, foreign equipment can lose availability, and specialized components may carry freight, duties, rigging, testing, and expediting charges. Where components are imported, values should be tested under both ordinary replacement and severe supply-chain disruption scenarios.

Next separate categories that may attract different treatment. General IT equipment, motors and generators, transformers, buildings, tenant improvements, and cargo in transit may not be treated identically by every insurer. The reported launch of dedicated cargo cover for AI infrastructure by Overhaul and Navium also shows a growing distinction between coverage for equipment moving toward a facility and coverage after arrival. Project cargo, installation testing, and testing-in-service provisions deserve attention because damage during assembly or initial operation can fall into gaps between builders’ risk, property, machinery breakdown, and marine cargo policies.

A useful rule is to insure declared values on a replacement-cost basis without settling automatically for the first sum insured. The first sum insured is a claim-control tool, but repeatedly declaring “whatever the claim is” can weaken the relationship with the insurer. Subject to policy terms, the declared amount should represent the insurer’s maximum obligation under the selected valuation basis. A common approach is to include a separately stated inflation or replacement-cost margin, such as 10% to 20%, only where the valuation and current cost trend justify it; this is not a universal recommendation. Organizations increasingly acquire capacity to replace in the same location, but cheaper alternative-site recovery can change the appropriate basis and may require separate business-interruption analysis.

## Designing Business Interruption and Dependency Limits

The business-interruption limit should equal the financial loss reasonably expected during a defined period if the insured facility or dependency suffers specified physical or operational damage. That amount can include lost gross profit, continuing operating expenses, emergency power or cloud costs, restart expenses, and contractual service credits. It should not be set equal to annual revenue automatically, because operating margins, variable costs, salvage income, and the ability to redeploy workloads all affect the loss.

A rigorous model starts with time to damage, time to inspection, time to notify and adjust, time to order, time to deliver, time to install, time to test, and time to return workloads to production. Each interval should be stress-tested rather than placed around a single vendor’s optimistic schedule. For an accelerated deployment, the model should examine 30-, 60-, 90-, 180-, and longer restoration bands. Publicly available premium and loss data are limited, so three severity scenarios—probable, severe, and extreme—can provide a clearer basis for limit selection than one forecast that appears precise without being reliable.

The indemnity period must be linked to actual recovery. A 12-month indemnity period is not automatically sufficient for a site with grid constraints, and a five-year period may overprice risk if redundant capacity exists. Dependence on a single utility, cloud platform, construction contractor, or component supplier should be modeled separately. Coverage should align with outage scenarios that are legally and contractually insurable, because speculative lost profits based on hypothetical future contracts are generally more difficult to substantiate than documented losses during the indemnity period.

Sublimits deserve particular attention. Property policies and cyber policies may use different definitions of the insured peril, and equipment breakdown, utility interruption, ingress, and cyber recovery can fall into separate sections. If power equipment is scheduled for only 20% of a total account value, operators should verify whether that figure follows exposure or simply reflects insurer capacity. The program should also distinguish between the maximum foreseeable loss and the amount available after deductibles, waiting periods, contractual recoveries, and limitations elsewhere in the structure.

Cargo, Supply Chain, and Project Risks

AI infrastructure is especially exposed before it becomes a fixed part of the insured site. Accelerators and other high-value components may cross oceans, enter ports, move by road, and remain in temporary warehouses before installation. Damage at each stage creates a different claim location, and marine cargo policies commonly provide cover on an international journey basis while domestic movement requires separate attention. Domestic cargo limits, project cargo sums, storage extensions, and delay-in-start-up cover should therefore be checked against the actual itinerary.

Logistics limits should reflect the loss of the component and the contractual consequences of missing the commissioning date. A replacement delay does not have to involve accidental physical damage to qualify under every policy, so delay-in-start-up and business-interruption terms must be examined side by side. Where the project is financed through construction or project-wrap policies, the relationship between cargo cover and builders’ risk also matters. Double insurance is not necessarily ideal when competing definitions of “property” produce disputes after a claim; a coordinated schedule is usually more useful than stacking broad names indiscriminately.

Performance guarantees, loss of production, and late-delivery risks may be difficult to place or may require specialist underwriting. Buyers should not assume that a high-accelerator value automatically produces a corresponding delay cover. Insurance should fit the economics, but excessive hypothetical limits can increase premiums and create false confidence. A better structure often starts with robust physical and transit cover, then uses contractual risk allocation, contingency reserves, alternate suppliers, and carefully selected project-delay protection.

Comparing the Main Alternatives

Operators can buy a broad all-risks property program, a layered specialist program, a wrap around construction and operational risks, or a financially sized program supplemented by self-insurance. There is no universally best arrangement. The broad program is administratively simple and may suit a diversified real-estate owner, while a specialist placement can accommodate unusual equipment values and international cargo. A wrap can protect a mortgage lender and contractor during development, but it may introduce different definitions, deductibles, and consent requirements after completion.

Self-insurance can be appropriate above a substantial retention when the operator has stable cash flow, multiple sites, and enough remaining exposure to justify a captive layer. A $1 million retention may be routine for one company and disruptive to another. Similarly, a $250 million insurance limit is not “enough” for a lower-value cluster, while it could still be inadequate for a replacement and interruption model that supports a multibillion-dollar project. Pricing is normally negotiated rather than published because rates depend on geography, construction status, equipment availability, protection systems, fire performance, flood exposure, management controls, and claims history.

The correct comparison is total risk cost, not premium alone. A cheaper quotation with low physical-damage limits may be expensive if the company retains power, equipment, and contingent interruption exposure. Conversely, buying several billion dollars of highly remote catastrophe cover can crowd out useful capacity. Brokers should compare limits, sublimits, deductibles, exclusions, reinstatement terms, valuation, insured perils, and recovery assumptions on one page for each option.

Practical Steps and Timing for Buyers

Start at least 120 to 180 days before the first material shipment for a large build-out, and 90 to 120 days before renewal for an operating site. A complex international program may need six to twelve months, particularly if the buyer is seeking new carrier capacity. Secure specialist broker support before delivery dates become non-negotiable. Insurers need time to inspect sites, model utility dependencies, engage engineers, and attract reinsurance, whereas a last-minute declaration can reduce available terms or produce broad exclusions.

The first internal step is to reconcile the asset register with finance, procurement, facilities, engineering, and legal records. The second is to obtain a replacement-cost report that names equipment categories, serial or batch-level values where relevant, and replacement lead times. The third is to model at least three interruption scenarios, including loss of a single critical zone, loss of the whole facility, and loss of an external dependency. The fourth is to compare that output with liquidity, debt service, contracted minimum-capacity payments, and management’s maximum tolerable outage.

Limits should then be challenged in workshops rather than concluded from dashboards. Property values should be compared with contractor replacement quotes, utility equipment plans, and logistics schedules. Business-interruption assumptions should be reviewed by engineers as well as accountants. Legal review should test service-level agreements, indemnities, liquidated damages, mortgage requirements, governmental restrictions, and the obligation to maintain alternative capacity. Finally, create a claim-playbook section defining evidence needed to prove lost profits and reconstruction costs.

During deployment, update the program whenever a new site, processor generation, grid dependency, customer contract, or major equipment shipment is added. Quarterly reviews are sensible during a rapid build-out, while annual reviews remain necessary after the portfolio stabilizes. Immediately before policy inception, freeze the valuation and dependency assumptions long enough for all parties to verify them. Insurance should move with the engineering design; a model that was appropriate for a 40-megawatt phase may fail when two later phases share a constrained substation and common cooling system.

Common Mistakes and When to Act Immediately

The most damaging mistake is using a generic data-center limit because it is easier than performing a site-specific valuation. Another is counting the entire capital project as insurable property value without distinguishing equipment, buildings, land, and costs that will not be paid again. A third is equating annual revenue with business-interruption exposure. Buyers also make errors by assuming “all risks” includes every cause of shutdown, overlooking sublimits, and treating a scheduled component such as a generator or transformer as fully covered by the headline equipment limit.

Cargo is another frequent gap. A facility can have excellent operational coverage while suffering an uninsured loss before equipment arrives. In addition, the program may not respond to the same event under property and contingent business-interruption sections if the definitions are inconsistent. Buyers should also avoid selecting a long indemnity period merely to feel secure, because an overlong period can increase cost and reduce carrier appetite. Insufficient records are equally dangerous: replacement-cost protection is difficult to enforce without current invoices, engineering plans, specifications, and proof of installation.

Immediate action is warranted when existing limits fall below 70% to 80% of the independently calculated replacement value, when a single insured event could consume more than the company’s approved cash reserve, or when contractual penalties exceed recoverable insurance. A lower ratio may be acceptable where the equipment is newly delivered, readily replaceable, and protected by a complete project program; context matters. Operators should also act if a new accelerator generation doubles concentration in one location, if grid or cooling constraints add more than a stated recovery threshold, or if uninsured cargo alone exceeds the organization’s retention.

The date for action is before the risk changes, not after. Renewing the old limit while a $2 billion project is already under construction is a weak position. Waiting for a claim, shipment, or public announcement to justify urgency can restrict available capacity and weaken disclosure. If a deadline cannot be met, the operator should seek a written binder, interim cargo extension, increased sublimits, or a staged declaration from a capable broker rather than simply proceeding without evidence that coverage will respond.

How to Evaluate Cost Without Chasing the Headline Limit

A public premium range for AI infrastructure is not dependable because comparable disclosures are scarce. Pricing is driven by the insured values themselves, so a nominal rate can conceal either serious underinsurance or substantial overpayment. The cited report that data centers are overbuying insurance is especially relevant: remote, high-limit layers may be quoted at attractive rates while weaker layers, deductibles, or exclusions dominate the eventual claim outcome.

Brokers should express the proposal in total cost of risk, including premium, deductibles, uninsured sublimits, expected retained loss, loss-prevention expenditure, and financing effects. Alternative-site deductibles or coinsurance should be converted into worst-case balance-sheet outcomes. A placement that is 5% cheaper in premium but leaves a $500 million uninsurable interruption gap is not a lower-risk solution. Conversely, adding limits to remote perils when the company has only a $25 million deductible may be less useful than strengthening the operational recovery plan.

Loss-prevention work should be evaluated alongside coverage. Arc detection, fire compartmentation, water and smoke protection, generator testing, fuel storage, cooling redundancy, access control, and documented maintenance can improve both resilience and terms. No insurer should promise that these systems eliminate a loss, and savings should not be presumed without a proper engineering assessment. The commercial value comes from lower uncertainty, better carrier appetite, and fewer scenarios in which protection fails.

As of 27 September 2026, the defensible approach is an annually verified, scenario-driven limit rather than a single AI benchmark. Buyers should use current replacement cost, realistic restoration intervals, explicit dependency sublimits, and a financial survival test. The strongest program may be the one whose terms remain relevant when the accelerator market, insurance market, and operating design are all changing faster than the policy wording.