Insuring India's AI Factories: GPU Cargo Values, Erection All Risks and DSU on a Rs 15,000 Crore Build

On 13 August 2026, Larsen & Toubro announced a mega order, reported at up to Rs 15,000 crore, to build India's largest NVIDIA B300 AI factory as part of a strategic partnership with Together AI. It landed in a month that also saw Adani propose a Rs 1 trillion AI data centre in Odisha on 250 acres, and Gujarat announce a policy targeting 7.5 GW of green AI data centre capacity backed by an expected Rs 6 lakh crore of investment. Wood Mackenzie projects Indian data centre capacity reaching 12 GW by 2030, while Brookfield sees 6.5 GW of AI data centre capacity coming online in India. A large fleet of GPU-dense construction projects is entering procurement at the same time.

The operational side of this market is reasonably well understood by Indian insurers; the capacity strain on that book is covered in our note on the data centre insurance capacity crunch. The construction phase is not, because an AI factory is not a conventional building project with some IT equipment at the end. On a hyperscale AI build the accelerators, networking and power train can account for the large majority of total project cost, and that inverts almost every assumption inside a standard contractors all risks or erection all risks programme: where the value sits, when it arrives, what the dominant peril is, and what a lender actually loses.

This piece takes the construction phase apart into its three insurable stages: the cargo leg, the erection and energisation on site, and the delay exposure that sits over both.

A conventional CAR or EAR programme assumes value accretes roughly in line with construction progress: civil works first, structural steel and services next, plant arriving and gaining value steadily across the schedule, with premiums, deductibles and escalation provisions all rated on that curve.

An AI factory does the opposite. The building shell, the chillers, the switchgear and the diesel or battery backup are a minority of project cost. The majority arrives late and fast: populated GPU racks, networking fabric and storage, delivered, installed and energised in the final months. Three consequences follow for the underwriter and the buyer:

• The estimated contract value at inception is not the exposure. A policy rated on the civil and MEP contract value, with IT fit-out added by endorsement later, spends most of its life underpriced and then faces a step change in values the insurer never contemplated. The full project sum insured, including the imported IT equipment at landed cost with duty and freight, has to be declared from day one, or the average clause will cut any partial loss.
• The peak accumulation window is short and late. For most of the schedule the site holds concrete and copper. For the last stretch it holds imported accelerators at a value density no conventional project site approaches, often before the fire suppression and security systems that will protect them in operation are fully commissioned. That window, not the two years of civil works, is what the material damage rating should be built around.
• The perils change. Civil-phase losses are storm, flood, collapse and third-party damage. Late-phase losses are theft of high-value portable equipment, water and dust damage to installed electronics, mishandling during rack installation, and electrical events during first energisation. The deductible structure and the warranties (storage, security, dust and moisture control) need to be written for the second list, not the first.

The closest analogue in the Indian market is the semiconductor fab, with the same late concentration of imported, moisture-sensitive, long-lead equipment; we covered that profile in fab construction-phase insurance. The AI factory adds one thing the fab does not: a global supply constraint on the core equipment itself.

Before anything reaches the EAR policy, it has to survive the journey, and this is where the numbers stop resembling conventional project cargo. A power plant project might ship a transformer worth a few tens of crores as its single high-value item. A single container of populated B300-class servers can carry comparable value in a box that handles like ordinary electronics freight, and a build of this scale means a sustained flow of such containers over months.

That does three things to the marine cargo placement:

1. Per-sending and per-location limits become the binding constraint. Open covers written for general project cargo typically carry per-sending limits sized for machinery, not for accelerator shipments. If a consolidated air or sea sending exceeds the limit, the excess is simply uninsured. The shipment plan (GPUs per container, containers per vessel or aircraft) has to be reconciled against the limit before logistics are fixed, not after.
2. Accumulation at port and inland container depot is the silent exposure. Customs clearance, duty payment and site readiness rarely line up, so containers dwell at the discharge port, the inland container depot or bonded warehousing. Several sendings accumulating at one location can quietly exceed the location limit on the open cover, and a single warehouse fire or theft event then hits a limit set for one sending. The cover needs an accumulation clause sized to realistic worst-case dwell, with declared-value reporting to match.
3. Theft and mishandling displace maritime perils as the working loss drivers. GPUs are compact, anonymous in transit and immediately saleable. Warranties on GPS-tracked transit, sealed containers, approved warehouses and defined routes stop being boilerplate; they decide whether a claim is paid or disputed.

How the marine delay-in-start-up cover attaches to the transit leg is set out in project cargo DSU for capital equipment.

On plant-heavy projects the concentrated losses occur in testing and commissioning, and an AI factory sharpens that pattern. The dangerous period is the sequence in which the medium-voltage system is energised, the UPS and battery systems are charged, the cooling loop is commissioned, and tens of megawatts of IT load are stepped up rack by rack. A protection relay set wrong, a phase fault during switchover, a cooling failure during a load ramp with liquid-cooled racks running hot: each can damage installed equipment whose value dwarfs everything else on site.

The testing and commissioning extension therefore deserves more negotiation than any other clause in the programme:

• Scope the extension to the actual commissioning method. AI factories commission in tranches: one data hall is energised and loaded while the next is still in fit-out. The wording must be clear that testing cover applies per system or per hall, and that a hall which has passed testing does not silently move the whole project into a maintenance-only basis while adjacent halls are still being energised.
• Time-box it against the real programme, not the insurer's default. A default four to eight week testing period, written for a turbine hall, does not fit a phased multi-hall energisation that can run for several months. The extension should be sized to the integrated commissioning schedule with a pre-priced mechanism to extend it if the programme slips; renegotiating testing cover mid-slip is done from weakness.
• Read the defects and electrical exclusions together. The LEG or DE defects clause chosen decides whether damage to surrounding sound equipment from a defective component is paid, and the electrical breakdown carve-backs decide how a first-energisation event is treated. On a build where the surrounding sound equipment is a hall of accelerators, the gap between a narrow and a wide defects clause is measured in hundreds of crores.

The mechanics of the testing extension, the DE ladder and the maintenance period are set out in detail in our piece on EAR and ALOP for infrastructure; everything there applies here, at higher value density and with a compressed, phased commissioning sequence.

Delay in start-up (DSU, also written in the Indian market as advance loss of profits) is where AI factories differ most sharply from conventional projects, and where a transplanted template does the most damage.

On a road or a power plant, the DSU sum insured is conventionally derived from the project's financial model: debt service plus fixed costs, or projected gross profit, across an indemnity period. The instinct on a data centre build is to anchor those numbers to construction cost, and it is wrong, because an AI factory exists to sell accelerated compute. Where a build is anchored to a named compute partner, as the L&T project is to Together AI, capacity is typically contracted forward: a defined volume of GPU capacity at agreed pricing, often committed before the first rack lands. The revenue the project loses per month of delay is the contracted GPU-hour revenue of the delayed halls; it is large relative to the construction spend and starts at full rate almost immediately, because the demand is already sold.

Practical consequences for the placement:

1. Derive the DSU sum insured from the offtake contract, not the EPC value: contracted capacity, contracted rate, ramp profile, less costs genuinely saved during delay. The business interruption logic is familiar; the inputs are not, and the underwriter will need the commercial model to price it.
2. Reflect the tranche structure. A loss that delays one hall delays one tranche of revenue, not the whole project. A DSU written on a single project-level trigger will either overcharge the buyer or underpay the claim; it should respond per hall with an aggregate.
3. Size the indemnity period to equipment replacement, not reconstruction. Civil works can be rebuilt in months. The binding constraint on recovery is the lead time to source replacement accelerators, a supply-chain question taken up in the next section.
4. Check depreciation of the revenue stream. Accelerator pricing is not static. A long delay does not just defer revenue; it can push delivery into a period of lower market rates. Lenders modelling the downside should test whether the basis of settlement captures that.

The traditional hierarchy of construction perils puts fire and natural catastrophe at the top. On an AI factory the dominant peril, measured by expected financial damage, is delay, and the reason is the supply chain for the equipment itself.

A report covered by ANI on 3 August 2026 warned that supply-chain constraints may cap India's data centre capacity at around 3 GW by 2030, against the 12 GW that Wood Mackenzie projects demand and investment could support. The binding constraints named in that debate (high-end accelerator allocation, power equipment lead times, grid connection) do not just slow the market; they change the loss dynamics of every project on it:

• A damaged GPU shipment is not a priced loss, it is a queue position lost. Conventional plant can be re-ordered against a known manufacturing slot. Current-generation accelerators are allocated, and a project that loses a tranche in transit or on site goes to the back of a queue whose length it does not control. The material damage claim pays the invoice; the DSU claim carries the real weight, and the indemnity period must be set against realistic re-allocation lead times, not shipping time.
• Long-lead power equipment doubles the exposure. Transformers, switchgear and generators carry multi-year lead times of their own. A fire in a completed substation weeks before energisation is a modest property loss and a severe delay loss.
• The suppliers' premises and contingent extensions matter. Delay can originate at a contract manufacturer or integration facility abroad, before the insured transit begins; how far up the chain DSU reaches depends on suppliers' extensions that must be deliberately bought, not assumed.

A construction programme at the L&T and Together AI scale needs to be one continuous chain of cover with no seams where value moves:

1. One coordinated marine and erection placement, with per-sending and location limits reconciled to the shipment plan, an accumulation clause sized to port and ICD dwell, and a clean handover of risk at site delivery so no tranche falls between the cargo and EAR policies.
2. An EAR sum insured built on full landed value from inception, with deductibles and warranties written for high-value electronics rather than civil works, and a testing extension scoped and time-boxed to the phased hall-by-hall commissioning programme.
3. Marine DSU and erection DSU placed together, on a sum insured derived from contracted GPU-hour revenue with a per-hall trigger, a time excess the project can genuinely absorb, and an indemnity period built on documented accelerator and power-equipment re-procurement lead times.
4. Lender requirements reconciled before financial close. The insurance schedule in the facility agreement, the offtake contract and the policies as placed must tell the same story. A lender relying on a DSU sized to construction cost has mispriced its own downside; a developer whose testing extension expires mid-commissioning has an uninsured window at peak exposure.

Each point turns on wording detail: the accumulation clause, testing period mechanics, defects clause, suppliers' extension, DSU basis of settlement. Comparing how competing insurers draft these is what separates a programme that responds from a filed certificate. Sarvada gives brokers, corporate risk teams and project-finance lenders structured, searchable access to insurer wordings, so these clauses can be compared across markets and reconciled against the EPC contract and the lender's schedule before cover is bound. Teams pricing AI factory construction programmes can Request Access to evaluate the platform.