Authors: Theresa Lederer - Director, Risk Management - Moody's
Derek Blum - Senior Director, Product Management - Moody's
Jack Williams - Assistant Director, Risk Management - Moody's
Derek Stedman - Director, Analytics and Modeling - Moody's
Hyperscale data centers are among the most value-dense assets in commercial property. Practices for modeling their vulnerability to physical risk from natural hazards are still maturing, with no settled industry guidelines. Reducing that methodological uncertainty is one part of the broader challenge of financing and insuring the projected US$3 trillion of capital investment needed to meet global capacity demand by 2030[1].
In part one of our blog series, we explored the insurance market behind that expansion. This second post turns to the role that catastrophe models can play in building a credible, decision-grade view of risk across parties involved in financing and insuring data centers.
It explores how probabilistic modeling can help structure coverage to better align developers’ and lenders’ requirements with the available (re)insurance capacity.
A small, worked example is used throughout to anchor the discussion. While results for this example are highly illustrative and may vary notably for different perils and locations, the example underscores two key points that apply more broadly:
- Both average annual loss and tail risk metrics, including the 1-in-1,000-year probable maximum loss, are likely to be only a fraction of full replacement cost, highlighting opportunities to reduce coverage from full replacement cost to a lower probable maximum loss threshold defined by lenders’ risk appetite.
- The example shows that modeling choices matter materially, not only for expected loss but also for volatility and per-event severity—with implications for per-risk, accumulation, and capacity management by (re)insurers.
Licensing clients can also look out for our forthcoming Data Centers: Best Practices for Modeling user guide.
As discussed in part one of our blog series, the risk to data centers is multifaceted. In the following, we focus on a specific subset of risks: property damage from acute physical risks during the operational stage of a data center.
Other types of risks, including builders’ risk, cyber, casualty, and liability, can also be analyzed using Moody’s suite of analytics, but are not discussed here.
We are currently reviewing our offerings to understand how they will reflect the fast-changing risk data center landscape, including enhancements to existing products and research into risk categories outside of our current suite of models. Contact your Moody’s relationship manager to connect with us.
The unique anatomy of hyperscale data centers
Hyperscale data centers (HDCs) are the physical infrastructure behind the AI workloads of the largest cloud and technology operators.
As outlined in part one of our blog series, the replacement cost value of these mega-assets is large. A single modern hall can exceed several hundred thousand square feet, and a multi-building campus can reach tens of billions of dollars in value.
When modeling the risk profile of these multi-billion-dollar HDCs, it is important to understand their complex and unique asset structure and how different perils interact with it:
- Severe convective storms can breach roofs, potentially exposing electronics/racks to water intrusion or damage to exposed cooling plant.
- Flooding tends to expose ground-level equipment and concentrate on low-lying or sub-grade infrastructure.
- Earthquakes can leave the shell standing while disabling internal mechanical and electrical systems.
- Wildfires can expose rooftop equipment to embers and impair operations through smoke contamination.
Other perils, including man-made risks such as terrorism and cyber, can produce additional variable impacts. As a result, the balance between property damage and business interruption can also vary by peril, since relatively localized physical damage can still trigger wider operational disruption across the facility.
Although transparency around site attributes can be limited at the point of underwriting[2], models still allow underwriters to test sensitivities around those uncertainties. When well captured, detail creates value across the capital chain.
It can help underwriters assess whether and at what price to offer capacity; can sharpen portfolio and accumulation management; give insurers and capital providers a shared, defensible view of risk and help developers and lenders calibrate their own retention appetite. Our worked example below is intended as a simplified illustration of those broader points, for a specific location and peril.
A worked example sheds light on the importance of modeling choices
We use a notional hyperscale campus located in the Dallas–Fort Worth, Texas, corridor and focus on risk from severe convective storms (a key physical risk in the area) to the data center campus. We conducted this analysis using the Moody’s RMS™ United States Severe Convective Storm HD Model, in conjunction with the Moody’s RMS Industrial Facilities Model (IFM).
Released in December 2025, the Moody’s RMS United States Severe Convective Storm HD Model assesses risk from three sub-perils—tornado, hail, and straight-line wind—underpinned by thousands of years of simulated storm activity and calibrated with over US$50 billion in historical insurance claims data.
The associated IFM allows users to capture unique aspects of specialized industrial facilities, such as data centers, in the model, enabling a more tailored view of the vulnerabilities for such assets.
The IFM is not tailored specifically to HDCs, but it does provide a far more specialized view than generic industrial and commercial exposure assumptions available outside of the IFM.
Throughout, the analysis focuses on ground-up property damage before the application of insurance policy terms. This example does not propose a specific set of recommendations for data coding, but rather highlights the sensitivity of results around several key assumptions. Recommendations will be forthcoming in our Data Centers: Best Practices for Modeling user guide.
Figure 2 below summarizes two exposure assumptions explored in the worked example and their effect on modeled loss cost relative to the default view of the site, which are discussed in further detail in the following sections. In addition, we compare results for modeling an HDC as a single location with a multi-location campus coding.
While all three effects have broad applicability in modeling data centers, their magnitude and, in some cases, directionality are specific to this example, and it can be expected that these effects will vary notably for other locations or risks. Nevertheless, key takeaways from the analysis have much broader applicability, as discussed in the final section of the blog.
Figure 2: Relative ground-up loss cost driven by contents fragility and the building-to-contents value ratio.
Contents: It matters what’s inside the box
For HDCs, insured contents risk is shaped by both vulnerability and valuation. We explore this by varying two model parameters: the assumed fragility of the contents at the modeled asset and the share of the overall asset value attributed to its contents (rather than its structure).
In the worked example, the assumed fragility of the contents has a material effect on loss cost (expected loss as a percentage of value), particularly for tornadoes and straight-line winds, while hail is less sensitive.
Holding everything else fixed, moving the fragility assumption of the content in the model from its default to the most fragile setting increases loss cost to roughly 1.6x the default, while using the most resilient setting reduces it to about 0.8x. Detailed information on fragility can hence notably change model results.
The split of building-to-contents value also matters. In the example, shifting from a more building-heavy to a more contents-heavy allocation of the overall asset value reduces modeled loss cost from roughly 1.35x the default to about 0.55x. The differences in this specific example are driven mainly by hail risk to exposed roof and yard equipment, which are part of the asset building value (Fig. 3). As its value decreases by moving a larger proportion of value to contents, the overall loss cost decreases.
More broadly, the analysis shows that HDCs cannot be reduced to a single replacement-cost figure: where value sits, and how different perils reach that value, can materially change the view of risk.
Figure 3: Ground-up loss cost by sub-peril across the building-to-contents sweep, contents proportion increasing from left to right.
Modeled loss cost is only half of the story when it comes to contents risk. The other half is the absolute value of contents, and that can come with a lot of uncertainty itself. Contents value is overwhelmingly driven by graphics processing units (GPUs) and adjacent silicon, and the unit replacement cost of that silicon has swung sharply over the past 24 months[3],[4].
As a single hall holds thousands of identical SKUs, that price uncertainty directly hits nearly the entire contents value of an HDC. Unlike other industrial risks with more diversified contents located on-site, contents in HDCs will lack the spread of values and vulnerabilities.
Furthermore, most state-of-the-art HDCs rely on GPUs from the same or at best a small number of providers, leading to valuation changes being highly correlated across data centers within a re/insurer’s portfolio.[5]
The practical implication: (Re)insurers should revisit these valuations often and conduct sensitivity testing at the point of underwriting to be prepared for future changes. Depreciation along with replacement value should be considered, as technology rapidly changes; GPUs lose value more quickly than other types of contents. With an assumed useful life of 4-6 years, insurers should consider the impacts on replacement value and obsolescence terms in their policies.
Spatial layout, volatility, and why mean loss is not the whole story
For large-footprint facilities such as HDCs, their spatial layout can be an important part of the risk story. That is especially true for localized perils such as tornadoes, where a single event may strike one part of a campus while leaving the rest largely unaffected. By modeling the entire HDC as a single point asset, we can capture its ‘average’ vulnerability and split between building vs contents value.
However, we do not capture where buildings and values are actually located within the often vast spatial extent of an HDC, and how vulnerabilities may vary between different components of the center. The strongest approach is to capture individual building footprints, their associated values and vulnerabilities in the model, and to analyze a selected ‘campus’ of assets rather than a single point.
In the worked example, we introduce this spatial disaggregation, specifically when modeling the tornado sub-peril, as this has the highest spatial variability. Expected tornado loss remains broadly unchanged but reduces volatility materially, with the tornado coefficient of variation[6] falling by roughly 38%.
Events that would otherwise have been modeled as total losses at a single point become partial losses across only part of the campus, while near-miss events can become smaller losses where the footprint is clipped.
That distinction matters because pricing is not driven by mean loss alone. A more spatially realistic view of the campus can reduce both the volatility loading and per-event severity, with implications for both aggregate pricing and excess-of-loss structures. That benefit is generally greatest for localized perils such as tornadoes and less pronounced for broader-footprint hazards such as hail or straight-line wind.
Implications for property catastrophe pricing
The preceding analysis is a single, simplified example; however, it illustrates broader points about how catastrophe models can help establish a shared view of risk, tailor coverage, and ultimately improve execution speed and ease the capacity issues that have repeatedly stalled progress in the recent data center boom. Catastrophe models are by no means a ‘silver bullet’ that can single-handedly address these issues; however, they are an important tool that has a part in unlocking growth solutions for the sector.
Two key takeaways stand out from our analysis:
1. Modeled loss can sit well below full replacement cost, even for a highly exposed campus: In our specific example, average annual loss remains <0.1% of total value, while even the 1-in-1,000-year loss remains below 10% of full replacement cost. Across our model simulations, individual years with higher damage levels do occur, and it is important to recognize that this risk does exist.
However, its likelihood is diminishingly small and may well fall within the risk tolerance of certain developers and lenders. A shared understanding of the full spectrum of potential losses and their associated likelihoods for developers, lenders and insurers is critical to move negotiations beyond blanket demands for ‘full coverage’ and truly optimize coverage to individual risk appetite and pricing of the involved parties.
2. Exposure assumptions are materially important: Assumptions can change not only expected loss but also volatility, and given the high-value nature of HDCs, resulting differences can quickly equate to large dollar amounts, even if modeled losses remain a small percentage of total value. That in turn has direct implications for insurance pricing and capital use.
Using a model with sufficient fidelity to support relevant exposure choices, establishing clear best practices, and insisting on high-quality asset data to inform those choices can yield a competitive advantage. It can also increase confidence in the risk analytics used across underwriting, portfolio management, reinsurance purchase and capital planning, ultimately leading to less ‘uncertainty loading’ on premia and more leeway for sustainable capacity increases.
In the next post in this series, we’ll look at exposure accumulations and the risk transfer market for data centers.
Did you miss part one of this data center insurance blog series? Find it here.
References:
[1] Moody's, Data Centers—Global 2026 Outlook: Capacity growth remains robust. https://www.moodys.com/research/Data-Centers-Global-2026-Outlook-Capacity-growth-remains-robust-Outlook--PBC_1460236
[2] The Insurer, Data center secrecy emerges as a property underwriting challenge, insurers and brokers say https://www.theinsurer.com/ti/news/data-center-secrecy-emerges-as-a-property-underwriting-challenge-insurers-and-2026-05-12/
[3] Silicon Data, The Illusion of Stability: Unpacking H100 GPU Market Value Trends (December 2025): https://www.silicondata.com/use-cases/h100-gpu-market-value-trends/; CloudZero, H100 GPU Cost in 2026: Buy, Rent, and Cloud Pricing Compared (May 2026): https://www.cloudzero.com/blog/h100-gpu-cost/
[4] Mercatus AI, H100 Depreciation: How Fast NVIDIA H100s Lose Value (and What It Means for TCO) (2026). https://www.mercatus-ai.com/blog/h100-depreciation
[5] https://siliconanalysts.com/analysis/nvidia-ai-accelerator-market-share-2024-2026#competitive-landscape
[6] A measure of volatility relative to average loss. The coefficient of variation is defined as the standard deviation divided by the expected loss, and is a metric commonly used by (re)insurers.