Natural Catastrophe Risk in Data Centers and the AI Supply Chain: A Guide

The AI buildout has concentrated an extraordinary amount of economic value into a small number of physically exposed places. Chip fabrication clusters in seismic and typhoon zones. Advanced packaging clusters in monsoon-prone Southeast Asia. Data centers cluster along a handful of power corridors, several of which sit in hail, tornado, or hurricane country. A single weather event in one of those places can now interrupt revenue at companies fifteen layers away that suffered no damage at all.

That last point is the one that matters for insurance. Most of the loss in this value chain does not arrive as a damaged building. It arrives as a facility that cannot operate, a supplier that cannot ship, or a grid that cannot deliver. This guide maps where those exposures sit, why they often fall outside the response of a standard property program, and how parametric structures can be built to address them.

What is the AI infrastructure value chain?

The AI infrastructure value chain is the full sequence of industries required to design, build, power, and operate the data centers that train and run AI models. It runs from semiconductor capital equipment at the base, through chip fabrication, packaging, server assembly, and networking, up through cooling and power systems, to the owners and operators of the facilities themselves.

Recent value chain mapping by Morgan Stanley Research identifies close to thirty distinct tiers across seven columns, plus five separate categories of owner and operator at the top: hyperscalers, data center REITs, private equity and infrastructure asset managers, enterprise and tier-2 clouds, and neo-clouds.

For risk purposes, the useful way to read that structure is bottom to top:

• Lithography equipment goes to foundries. 
• Foundry output goes to chip designers. 
• Chips go to server manufacturers. 
• Servers go into buildings that consume very large amounts of electricity and water. 

Each of those handoffs is a place where a physical event in one location becomes a financial loss in another.

Why is climate risk concentrated in this value chain?

Four structural features make this value chain unusually exposed relative to its size.

Geographic concentration. Almost all of the world's most advanced chips are manufactured on a single island, in a region exposed to earthquakes and typhoons. Packaging and testing, a mandatory step for every chip produced, is concentrated across Taiwan, Malaysia, and the Philippines. If those facilities stop, no alternative site can absorb the volume.  

Single points of failure. Several tiers have one dominant supplier or a small handful. When a tier has no substitute, a disruption there cannot be routed around, and the loss propagates upward in full.

Power and water dependence. Fabs require ultrapure water in very large volumes and uninterrupted, high-quality power. Data centers require both continuous electricity and functioning heat rejection. Both are utility-dependent in a way that most manufacturing is not, which means an event that never touches the site can still stop production.

Speed of capital deployment. According to the International Energy Agency's Energy and AI report, global data center electricity consumption was around 415 terawatt-hours in 2024 and is projected to reach roughly 945 terawatt-hours by 2030 in its base case, growing about 15 percent per year. Assets are being sited quickly, in locations selected primarily for power availability and interconnection speed. Hazard exposure is rarely the deciding factor in that selection.

Contact us

Which natural catastrophe perils affect each layer?

The exposure profile changes materially by layer. The table below maps the principal perils and the form the loss typically takes.

Two patterns stand out: 

• The loss form in most rows is not physical damage. It is interruption without damage, or interruption caused by damage to someone else's asset. 
• The perils involved are measurable ones. Rainfall, wind speed, ground acceleration, reservoir levels, hail size, and temperature are all observable through independent third-party data, which is precisely the condition parametric structures are built on.

Why are many of these losses invisible to traditional property insurance?

Standard property and business interruption wordings are built around damage to insured property. That works well when a building is destroyed. It works less well in this value chain, where the most common failure modes involve no damage to the policyholder's own assets.

What is non-damage business interruption?

Non-damage business interruption, or NDBI, is revenue loss that occurs without physical damage to the insured's property. A fab that halts because a municipality rations water has an NDBI loss. A data center that curtails workload because the grid cannot deliver has an NDBI loss. A construction site that stands idle through a month of heavy rainfall has an NDBI loss.

In each case, the asset is intact. Standard business interruption coverage generally requires damage as the trigger, so the loss does not attach.

What is contingent business interruption in a chip supply chain?

Contingent business interruption is revenue loss caused by damage or disruption at a supplier or customer rather than at the insured's own site. In a value chain with thirty tiers and several single-source nodes, contingent exposure extends much further than most programs contemplate.

The practical difficulty is knowledge. Contingent BI extensions typically require the supplier to be named or the tier to be defined. Few operators can name their tier-three and tier-four suppliers with confidence, and the substrate or passive component maker whose flooding stops a server line is often exactly the supplier nobody had listed.

Why does the timing of payment matter here?

Even where coverage responds, the settlement timeline may not match the financial need. Indemnity claims of this complexity, involving contested causation, supply chain forensics, and disputed quantum, commonly take many months to resolve. Where a facility is financed against contracted revenue, debt service does not pause while the loss is adjusted. This is particularly acute for the neo-cloud and private capital owners identified in the value chain mapping, whose capital structures depend on predictable cash flow rather than balance sheet depth.

What has already happened?

These are not hypothetical exposures. Four public events illustrate the pattern.

Taiwan drought, 2021. After a year in which no typhoon made landfall on Taiwan for the first time since 1964, the island experienced its worst drought in 56 years. Reservoirs serving the major science parks fell to single-digit percentages of capacity. Authorities ordered companies in the Hsinchu and Taichung science parks to cut water use by up to 17 percent, and manufacturers trucked in water by tanker to maintain production. Semiconductor fabrication depends on ultrapure water in very large volumes; one leading manufacturer's own figures put its requirement at roughly 156,000 tonnes per day. The event caused no damage to any fab. It was purely a utility availability event.

Winter Storm Uri, Texas, 2021. In February 2021, the local utility in Austin cut power to its largest industrial users as the state grid came close to collapse. Three major semiconductor plants were idled. One manufacturer subsequently estimated a revenue impact of approximately $100 million and the loss of about a month of wafer production, with facilities offline for close to four weeks. 

Thailand floods, 2011. Flooding in the Ayutthaya region submerged industrial estates for more than a month. IDC estimated at the time that the affected factories accounted for roughly a quarter of worldwide hard disk drive production. Drive prices rose sharply, and a major processor manufacturer reduced quarterly revenue guidance by approximately $1 billion as a direct consequence. Almost all of that downstream loss was contingent: the companies taking the revenue hit owned nothing in Thailand.

Repeat monsoon flooding in Southeast Asian packaging hubs. Assembly, test, and packaging operations concentrated in flood-exposed parts of Malaysia and the Philippines have been disrupted on multiple occasions. Packaging is a mandatory step for every chip produced, and capacity is not easily relocated, so disruption at this tier reaches the entire chain above it.

The common thread is that in three of the four cases, the largest losses landed on companies whose own property was undamaged.

How does parametric insurance respond to these exposures?

Parametric insurance pays a predefined amount when an objectively measured parameter crosses an agreed threshold, rather than paying against an assessment of physical damage. The trigger is set on independent third-party data: wind speed at a location, ground acceleration, accumulated rainfall, reservoir level, hail size, or temperature over a defined period.

Three characteristics make it well matched to this value chain.

It does not require damage. Because the trigger is the event rather than the damage, parametric responds to interruption caused by utility failure, supplier disruption, or access restriction. This is the single most important point for AI infrastructure exposures.

It pays quickly. Once the parameter is verified, payment follows within days or weeks rather than through an extended adjustment process. For capital-intensive assets with fixed financing obligations, timing is often as valuable as quantum.

Funds are unrestricted. A payout is triggered by the severity of the event, not by proven repair costs, so proceeds can be applied to lost margin, expediting costs, alternative capacity, contractual penalties, or debt service.

A parametric structure is only as good as the fit between the trigger and the client's actual exposure. That fit is a modeling and structuring question, and it is where the design work sits. Any parametric solution should be tested against the client's real loss profile before it is bound, and that evaluation is a normal part of due diligence for any insurance product.

Contact us

How can parametric complement an existing property program?

Parametric works best alongside a traditional program rather than in place of one. Five structures come up repeatedly with infrastructure and technology clients.

Sublimit top-up. Property programs frequently sub-limit named windstorm, flood, or earthquake well below the replacement value of a modern facility. A parametric layer sitting above the sub-limit restores economic protection for the peril without disturbing the underlying placement.

Deductible buy-down. Large campus programs often carry retentions calibrated to a corporate balance sheet rather than to a single project entity. Parametric converts part of that retained exposure into predictable, trigger-based liquidity.

NDBI cover. The clearest application. A temperature, wind, rainfall, or drought index structured around the specific operational threshold that forces curtailment, with no damage requirement.

Supply chain and contingent exposure. A trigger placed on the geography where a critical supplier operates, rather than on the insured's own site. This works even where the supplier cannot be individually named, because the trigger is a hazard measurement in a defined area.

Exclusion wrap. Where a wording excludes a defined peril or a specific zone, a parametric structure can be built to respond to precisely that gap.

In a market where clients are realizing savings on traditional premium, several are choosing to reinvest part of that saving into structural improvements of this kind rather than simply banking the reduction. The conversation is about closing gaps that were always there, not about replacing what already works.

What should a risk manager ask before the next renewal?

• Which of our facilities depend on a utility we do not control, and what happens financially if that utility is unavailable for two weeks?
• Does our business interruption cover require physical damage to attach? What happens if it does and there is none?
• How far down our supply chain does our contingent BI extension actually reach, and which tiers are named?
• What are the sub-limits for named windstorm, flood, and earthquake at our most exposed sites, and how do they compare with the economic value at those sites?
• If we suffered a covered loss today, how long would settlement take, and can we service our obligations in the meantime?
• Where is our concentration risk in geographic terms, including the concentration our suppliers have created on our behalf?

About Descartes Underwriting

Founded in 2019 by insurance veterans and climate scientists, Descartes Underwriting is the leading specialist in parametric (re)insurance, with more than $250 million in gross written premium in 2025 and over 20 offices across 10 countries. Our team of 250+ includes 150+ in-house specialists in data science, risk modeling, and software engineering, the largest scientific team in the insurance industry. More than 600 corporates and public entities rely on our solutions, backed by A+ rated capacity.

Contact us

Frequently asked questions

Does property insurance cover a data center that loses power but suffers no damage? Generally not under a standard wording. Business interruption cover typically requires physical damage to insured property as the trigger. A grid event that causes curtailment without damaging the facility usually falls outside that trigger, which is why non-damage business interruption is addressed separately.

Can insurance cover a chip plant that stops because of a water shortage? Not through standard property cover, since there is no physical damage. A parametric structure can respond, using a measurable drought parameter such as reservoir level, accumulated rainfall over a defined period, or a recognized drought index, with a payout triggered when the parameter crosses an agreed threshold.

How quickly does a parametric policy pay? Payment follows verification of the parameter against the agreed data source, typically within days or weeks of the event, rather than following a loss adjustment process.

Is parametric insurance more expensive than traditional cover? Pricing reflects the probability of the triggering event. Where a parametric price is higher, it usually indicates broader response for that specific peril. Structures can also be designed to a defined budget, and clients avoid loss adjustment expense. The relevant comparison is not premium alone but the cost of coverage that does not respond, or responds eighteen months late.

Can parametric cover supply chain disruption at a supplier we do not own? Yes. Because the trigger is a hazard measurement in a defined geographic area rather than damage to a specific asset, a structure can be built around the region where critical suppliers operate.

What is basis risk in a parametric policy? It is the possibility that the payout differs from the actual loss because the measured parameter does not perfectly track the operational impact. It is managed through trigger design, data selection, and modeling against the client's own loss history, and it should be quantified as part of any structuring exercise.

References

1. International Energy Agency, Energy and AI: Energy demand from AI, April 2025.
2. International Energy Agency, AI is set to drive surging electricity demand from data centres, April 2025.
3. The Diplomat, How Water Scarcity Threatens Taiwan's Semiconductor Industry, September 2024.
4. CNBC, Taiwan steps up curbs on water use by chip hub Hsinchu during drought, May 2021.
5. Inquirer, Drought-hit Taiwan delays further water curbs after heavy rainfall, May 2021.
6. Fortune, Taiwan's drought is exposing just how much water chipmakers use, June 2021.
7. The Water Diplomat, Taiwan Drought: Microchip Lead Times Soar As Curbs On Water Use Tighten, May 2021.
8. Datacenter Dynamics, Samsung, NXP, and Infineon chip fabs shut down in Texas amid record storm, February 2021.
9. eeNews Europe, NXP, Infineon plants hit by power outage in Texas storm, February 2021.
10. EE Times, Infineon and NXP Resume Austin Texas Fabs After Winter Storms, March 2021.
11. IEEE Spectrum, The Lessons of Thailand's Flood.
12. Phys.org, Thai flooding disrupting hard drive supplies, October 2011.
13. Forbes, After Thai Floods, Companies Reconsider Risk, December 2011.
14. Morgan Stanley Research, AI Infrastructure Value Chain Heatmap, data as of July 2026.