Gigawatt Campuses on a Cyclone Coast: Reading Cat Accumulation in Gujarat and Visakhapatnam

India's data centre build has stopped being an incremental story about colocation halls near Mumbai and Chennai. Energetica India reported on 21 August 2026 that live IT capacity has quadrupled to 1.7 GW, with a further 1.3 GW under construction and 3.2 GW of projects that have already secured land, power and permits. The same report, citing BloombergNEF, puts about 95 percent of announced Indian data centre investment into large hyperscale AI campuses, and projects data-centre electricity demand growing more than eightfold to 91 TWh in 2035 from 11 TWh in 2025.

Two announcements make the geography concrete. Google is developing an AI and data centre hub at Visakhapatnam in Andhra Pradesh with AdaniConneX and Bharti Airtel, part of a five-year 15 billion dollar investment plan running 2026 to 2030 and delivering gigawatt-scale computing power. Data Center Knowledge described the campus in 2026 as Google's largest AI hub outside the United States, integrating compute capacity, renewable energy infrastructure and expanded fibre connectivity. Separately, Gujarat has unveiled a Data Centre Policy 2026-29 targeting Rs 6 lakh crore in investment and 7.5 GW of capacity.

Read together, the risk shape is clear. Capacity is not spreading thinly across twenty inland cities. It is landing in a few very large campuses, and two of the biggest destinations are cyclone-exposed states: Andhra Pradesh, where the Visakhapatnam hub sits on the Bay of Bengal, and Gujarat, whose coastline takes Arabian Sea landfalls. That concentration was decided by power, land, cable landings and state incentives, and it hands the property market a nat-cat accumulation problem it has not priced at this size before.

Most Indian industrial property placements are still built as fire-led programmes with nat-cat riding along as an add-on. On a coastal hyperscale campus that ordering is wrong.

A modern AI hall is a poor fire risk in the ordinary sense. Combustible load is low, the building is steel, concrete and mineral-fibre panel, detection is early-warning aspirating, suppression is clean agent or pre-action, and compartmentation between halls stops a single event travelling. A fire loss on a well-built hyperscale site is usually a sub-hall event, which is why the fire probable maximum loss underwriters model rarely approaches the full sum insured.

Named windstorm and flood do not respect compartmentation. A cyclone at landfall attacks the envelope: roof membrane, rooftop chiller and dry-cooler arrays, external switchyard, transformer bays, cable trays and the fuel-oil day tanks feeding standby generation. Once the membrane lifts on one bay, the loss is water and debris across racked accelerator hardware whose replacement value per square metre is an order of magnitude above conventional industrial contents. Surge and rainfall flooding hit the lowest and most expensive systems: the HV intake, the chilled-water plant room, the fuel farm and below-grade cabling.

The accumulation question here is not how many buildings a corporate owns, it is how much value one wind footprint contains. A gigawatt-scale campus concentrates equipment value that a conventional manufacturing portfolio would spread across several states, and almost all of it sits inside one insurable location code. That has three consequences for programme structure.

1. The named-windstorm sub-limit becomes the operative limit. If the fire limit is written at full reinstatement value but windstorm is capped at a fraction of it, the cap is what the client actually owns for the peril most likely to produce a total-hall loss. Placing that sub-limit by reference to a modelled 250-year or 500-year event on the specific site is defensible; placing it as a round percentage copied from an inland portfolio is not.
2. Flood and storm surge need to be separated from windstorm. Cyclones deliver both, and a single combined nat-cat sub-limit lets one event exhaust the whole allocation. Separate sub-limits with a clear hours clause definition control how a landfall, its rainband flooding and the surge are aggregated into one occurrence.
3. Contents and equipment need their own treatment. Building reinstatement is a known quantity. The IT load is not, because replacement depends on a constrained global supply chain and pricing that moves. A policy wording that indemnifies equipment at book value or at a stale schedule leaves the client short at exactly the moment replacement is hardest to obtain.

The point is to stop a programme from looking adequate because the headline limit is large, while the peril that will cause the loss sits behind a sub-limit nobody stress-tested. That is the discipline set out in nat cat accumulation management for multi-location corporates, applied where the accumulation lives inside one fence line instead of across a map.

A gigawatt campus is not a single construction event. It is phased halls delivered over years against contracted capacity dates. The Visakhapatnam hub sits inside a five-year plan running 2026 to 2030 and Gujarat's policy window runs 2026-29, so both cross several cyclone seasons with works still exposed.

During construction the material damage cover sits under erection all risks and contract works, and wind and flood exposure is worse than in operation. A partly clad building has no envelope. Roof decking, standing tower cranes, laydown areas holding switchgear and transformers and open excavations for the HV yard are all vulnerable in a way the finished asset is not. A landfall in the wrong quarter can destroy temporary works, damage stored equipment never protected to operating standards, and leave the site inaccessible for weeks.

The material damage repair bill is usually the smaller number. Delay in start-up is the larger one. DSU indemnifies the gross profit or fixed costs lost because commissioning slipped beyond the scheduled date, and on a hyperscale campus the daily value of a delayed hall is set by contracted capacity revenue, not by the cost of a replacement roof sheet. A six-week slip caused by cyclone damage plus post-event supply chain congestion is a six-week DSU claim on a very large daily figure.

Three structural points decide whether that cover responds usefully:

• The time excess must be set against realistic post-cyclone recovery, not a generic 30 days. Coastal recovery involves port congestion, road access, power restoration and surveyor availability. A time excess longer than the typical delay simply removes the cover.
• The indemnity period has to survive re-procurement. Replacing a damaged transformer or switchgear set carries a lead time measured in months. An indemnity period sized on construction rework alone will expire before the delayed item arrives.
• DSU triggers only on insured material damage. Delay caused by an access restriction, a supplier's own loss elsewhere or a regulatory hold after a cyclone will not respond unless extensions for prevention of access and supplier's premises are bought and sub-limited deliberately.

The construction-side mechanics for these builds, including how EAR, DSU and GPU cargo interlock, are worked through in insuring India's AI factories.

Between the factory and the rack, the most valuable contents of the campus travel as cargo. A single shipment of accelerator systems can carry a value that would be a full-limit claim in most conventional marine cargo programmes, and it moves through freight, customs, inland haulage and a coastal last mile into what may still be a construction zone. Two failure points recur.

The first is storage in transit. Equipment often arrives ahead of hall readiness and waits in a warehouse or site container. Standard marine cover ends on delivery or after a fixed number of days, and the goods fall between the cargo policy and a construction policy that may not accept them until installation begins. That gap is where a storm surge finds unmounted switchgear at grade.

The second is valuation and duty. Insured value should be CIF plus duty plus a landed-cost margin. Insuring at invoice value leaves the client bearing duty and freight on a total loss, which on a high-value accelerator shipment is not a rounding error.

Both are cheap to fix at placement. Compare the delivery schedule against hall handover dates, buy an explicit storage-in-transit extension wherever they diverge, and confirm the storage location's flood elevation.

On nat-cat-exposed data centre risks the reinsurance market prices with deductibles as much as with rate. A named-windstorm or flood deductible on this class is usually a percentage of the affected location's value rather than a flat rupee figure, so on a campus with a very large declared value even a modest percentage is a large absolute retention. A client comparing quotes on rate alone can accept a structure where the first several tens of crore of any cyclone loss sits with them, which is a balance-sheet decision that belongs with the CFO and not inside a placement comparison spreadsheet.

Three details worth negotiating explicitly:

• What the percentage applies to. Percentage of the total declared value at the location, percentage of the damaged item value, and percentage of loss are three very different retentions on the same event. The first is the most expensive to the insured and the most common ask from reinsurers.
• Minimum and maximum caps. A percentage deductible with a stated maximum converts an unbounded retention into a known one, and it is often obtainable when the overall structure is otherwise acceptable to the panel.
• How the DSU time excess interacts. A material damage percentage deductible plus a long DSU time excess can mean a genuine cyclone event produces no recovery at all on a mid-sized loss. Model a realistic event through both deductibles together, not separately.

None of this is unusual for cat-exposed property internationally. It is new for Indian buyers whose previous programmes carried flat deductibles because their exposures were inland and their values were spread.

The most useful thing a broker can tell a data centre client early is that the Indian insurer whose name goes on the policy is not the party setting the windstorm sub-limit, the percentage deductible or the DSU time excess. A single campus of this value cannot be retained domestically. It is placed with treaty support and, above a point, facultative capacity, and the terms are effectively written by the reinsurance panel. The broader capacity picture for the class is set out in the data centre insurance capacity crunch.

That has consequences for how the submission is built.

The cat model output has to be credible. Panels underwrite from vendor cat model results run on specific coordinates, elevation, construction class and occupancy. A submission without geocoded locations, elevation relative to known surge levels, roof construction and attachment detail, and a clean split of building against equipment values gets rated conservatively, because the reinsurer prices the uncertainty.

Engineering evidence moves terms more than negotiation does. Roof attachment and uplift rating, envelope and louvre wind ratings, elevation of the HV yard and fuel farm above design flood level, chiller tie-down detail, and a documented cyclone preparedness plan covering shutdown and post-event restart are what convert a defensive quote into a workable one.

Programme structure should be settled before the market is approached. Whether nat-cat sits inside the main property tower or in a separate layer, how DSU attaches during phased handover, and how the construction and operational policies dovetail as halls go live one by one all need answering first. A panel approached with an unresolved structure will propose several, and the resulting programme has gaps at the seams.

For the operating-phase view of these sites, including machinery breakdown, cooling failure and downstream liability, see hyperscale data centre operator insurance, and for the underlying cover mechanics, engineering insurance.

With 1.3 GW under construction and 3.2 GW permitted and funded, most of these placements have not happened yet. The work is front-loaded.

1. Geocode and elevate every site before quoting. Latitude, longitude, ground elevation, distance to coast and design flood level are the first four fields a reinsurer looks at. Producing them late costs terms.
2. Split values properly. Building, plant, IT equipment and stock in transit should be declared separately, with the IT equipment valued at current replacement including duty and freight rather than at depreciated book value.
3. Model the cyclone scenario end to end. Material damage, business interruption or DSU depending on phase, both deductibles, and the realistic recovery timeline including port and supply chain congestion. Present that scenario in the submission rather than waiting for the panel to construct its own.
4. Close the transit and storage gap explicitly. Match delivery dates to hall readiness and buy storage-in-transit where they diverge.
5. Set client expectations on retention early. Percentage nat-cat deductibles on a gigawatt campus are large in absolute terms. That conversation goes better before terms arrive than after.

Coastal siting is settled. Power availability, cable landings and state incentives decided it, and no underwriting comment will relocate a 7.5 GW policy target or a gigawatt-scale hub. What is still open is whether the programmes attached to them are structured around the peril that will actually cause the loss.