Surging demand for space to house AI and cloud computing hardware is an opportunity for U.S. data center owners and developers. Risks relating to power and water requirements and financing will have to be managed.

Why it matters:   The surge in data center demand will create significant growth opportunities for data center entities. As projects multiply and average project sizes increase, power and water requirements, financing, tenant concentration, and cost inflation emerge as constraints. Navigating these risks are key considerations when assessing data center owners' and developers' credit quality.

What we think and why:   Not all data center owners are the same. Credit risks differ between hyperscalers and retail/colocation leasing models. Access to energy is an increasingly important consideration. Obsolescence risk is less of a credit issue for the near and intermediate term. Increasing risk appetite to accompany growth could affect credit performance.

This report explores our views on the risks and opportunities in data centers and their potential credit implications, regardless of their financing structure. We maintain ratings on data centers across the corporate, project finance, and structured finance practices (Table 1). See "The Four Main Approaches For Rating Data Center Financings", published June 13, 2024.

Table 1

AI Growth Underpins Hyperscale Data Center Demand

Generative AI and cloud services are strong demand factors; growth in both model training and inference (in which a trained machine-learning model draws conclusions based on data) have proven to be significant contributors to data center revenues. Computation needs for a ChatGPT query are more than 100 times higher than for traditional internet searches. Estimates indicate that daily queries on ChatGPT are in the millions while Google responds to about 8.5 billion searches every day.

S&P Global Ratings expects demand for hyperscale data centers will accelerate in line with the increasing requirements for large-scale and high-powered platforms to power AI computing. A hyperscale data center is massive, providing extreme scalability and engineered for large-scale workloads with an optimized network infrastructure, streamlined network connectivity, and minimized latency. Due to ever-increasing demand for data storage, numerous providers use hyperscale data centers globally for a wide variety of purposes that include AI, automation, data analytics, data storage, data processing and other big data computing pursuits.

S&P Global Market Intelligence estimated that the U.S. added 10 gigawatts (GW) of data center capacity from 2017-2022 (according to the 2024 U.S. Datacenter and Energy Report), and we expect nearly 50 GW of utility power capacity to be added from 2023-2028. A record number of new leases were signed in the first half of 2024, with the bulk of the demand in the U.S.

We expect strong growth in capacity across all U.S. markets, with Washington, D.C., and Virginia remaining the largest. They account for one-third of the total U.S. projects, expanding capacity to 8 GW. We also expect quite a bit of growth in smaller markets such as Des Moines, Iowa, which could provide easier access to land and low cost power.

Chart 1

Key Opportunities

We expect the leasing environment for data centers to remain strong.  This should result in positive re-leasing spreads and low vacancy. Microsoft Corp.'s 2024 annual report indicated an additional $8.6 billion of operating leases and $108.4 billion of finance leases, primarily for data centers, in the next five years, with lease terms of 1-20 years. We believe a strong demand environment, limited near-term supply, and improved pricing dynamics should bolster data center earnings and valuations in the next few years, if not longer. We expect demand for hyperscale data centers will accelerate given rapidly increasing requirements for large-scale and high-powered platforms for AI computing. A surge in demand and limited supply has strengthened rent growth in the past three years with rapid expansion of hyperscale data centers powered by AI growth in both model training and inference.

We view interconnection as a key competitive advantage for retail data center providers, with control over network dense assets. Hyperscale data center tenants also tend to be stickier due to the significant investment they make to outfit the property, and relocating the critical infrastructure can be challenging. These long-term leases average about 10 years and often with extension options. Rent escalators are generally built into lease agreements, providing a hedge against inflation.

Excess demand has spilled over to emerging markets such as Phoenix and Hillsboro, Ore., and driven remarkable rent growth after roughly a decade of flat or even falling rents for data centers in the U.S. Trends in other major data center markets globally are similar.

A supply/demand imbalance and high construction costs will support the market value of data centers in the near to medium term.  Location of data centers, like other real estate, is critical to valuation. The tier 1 data center market in Northern Virginia continues to expand due to relatively low power rates, affordable land, asset-specific tax incentives, general safety from natural disasters, and proximity to a primary internet exchange connectivity point on the East Coast. Despite abundant projects under construction, mainly driven by developers fulfilling previously signed forward-lease commitments, the vacancy rate is very low for the entities that we rate.

We expect a continued growth trend for cloud service providers investing heavily in AI ahead of the anticipated application expansion.  S&P Global Ratings expects market annual spending for AI, including traditional AI (machine learning) and generative AI, will expand to nearly $650 billion by 2028 from less than $200 billion in 2023--a compound average growth rate in the high-20% area. We project that the market will account for nearly 15% of total global IT spending by 2028. AI spending would include semiconductors, hardware, software, and IT services (see "Midyear 2024 IT Forecast Update: Robust Cloud Spending Offsets Still-Cautious Enterprise Budgets", July 17, 2024).

Chart 2

Key Challenges

Chart 3

Power and infrastructure constraints hinder development risk.  In our view, the key challenges during data center construction are managing the supply chain of equipment and interconnection with power and cooling sources to meet the construction schedule. To accommodate massive data center growth, associated infrastructure development, regulations, and policies must keep pace. Power generation, transmission build-up, grid interconnections, labor and supply chain, and water and land availability present serious challenges. Additionally, high and rising construction costs pose further threats. Indeed, these issues are relevant in assessing construction risk--completion on schedule and budget--under our project finance methodology.

Development risk is significant for rated REITs, including Digital Realty Trust Inc. (BBB/Stable/--) and Equinix Inc. (BBB/Stable/--). Still, they can mitigate this if a good portion of the development is built-to-suit or the pipeline is highly pre-leased, resulting in less exposure to speculative development. Digital Realty has a sizable development pipeline, and we anticipate it will spend about $4.5 billion (at the company's share) over the next two years. Significant infusion of joint venture capital and proceeds from equity issuance has reduced the risk. For Equinix, we expect relatively high capital expenditure (capex) requirements of 30%-35% of revenues, or about $3 billion per year, to fund expansion and purchases of real estate. Most of this is unrelated to AI demand across 54 projects.

Equinix also recently announced a large joint venture to invest in hyperscale data centers of more than $15 billion (25% minority stake). This marks a significant commitment to capture AI-related growth whereby the venture will purchase land to build new facilities on multiple campuses of greater than 100 megawatts (MW), eventually adding more than 1.5 GW of new capacity for hyperscale customers.

We see access to power is the key constraint.   Access to power is a key factor for data center operations. However, we believe the grid infrastructure will be the biggest hurdle for data center growth, notably because of long planning and permitting process (see "Data Centers: Surging Demand Will Benefit And Test The U.S. Power Sector", published Oct. 22, 2024). We estimate incremental U.S. power demand from data centers could be 150-250 terawatt hours between 2024 and 2030, about the annual consumption of New York and California. However, the lack of available power could temper and constrain actual estimates.

Even when the power generation and transmission capacity are available, further constraints on power equipment include transformers, on-site backup generators, and power distribution units, with historically high lead times of nearly two years in some cases.

Inflationary and financing needs may become a greater hurdle.  Data center construction costs have soared substantially with construction, inflation, and financing expenses including land availability and costs. When we add financing costs, more so with high interest rates in recent years, the cost of building a mega data center is substantial. While rental rates have kept pace with rising construction costs the last two years, the inability to pass through high construction costs through rental rates could derail growth prospects.

Similarly, data centers also face exposure to higher capex reserve requirements to support future maintenance capex. Developers may face funding constraints due to higher leverage or constrained profit margin/cash flow.

The strained skilled labor force is an additional challenge. McKinsey & Co. estimates a potential shortage of up to 400,000 trade workers in the U.S., based on projected data center build-out and comparable assets requiring similar skills, such as semiconductor fabrication and in battery gigafactories. These supply chain and labor availability issues drive cost escalations and completion delays.

Regulatory/sustainability issues poses additional challenges.  Data centers' potential sustainability impacts could also pose regulatory risks. These could add requirements to planning and approvals such as restrictions on water use or targets on power use and efficiency. They can also include penalties that apply when groups of data centers create an imbalance in power given the capacity of regional grid and overall demand. All of these imply potential higher costs if and when regulations tighten.

Data centers consume significant amounts of water for cooling requirements. A hyperscale facility typically uses 200 million gallons per year. As they expand in populous centers or new locations, availability of adequate water could be a challenge. However, newer data center designs utilizing air and liquid cooling may require little to no water to operate.

Key Credit Implications For Data Center Owners And Operators

Chart 4

Operating risk is higher for retail/colocation versus wholesale/hyperscale data centers.  We view wholesale/hyperscale data centers more favorably and less exposed to rising operating cost than retail colocation (without interconnection) given the more predictable rental income. Data center properties differ in size and customer base, including expansive wholesale and hyperscale facilities leased to a few large high-quality tenants as well as retail colocation properties leased to several small tenants. Hyperscale and wholesale facilities generally benefit from long-term net leases that require tenants to reimburse taxes, insurance, and electricity costs, while retail colocation data centers frequently offer shorter-duration leases with only partial pass-through clauses, exposing them to tenant rollover risk and cost variations. Typically, labor and power costs account for the largest components of a data center cost structure, but rent expense could be significant if they lease the assets. For example, 36% of Equinix's operating expenditure in the second quarter of 2024 was labor and 28% was power. Overhead expenses such as personnel, office expenses, advertising, taxes, licenses, and insurance accounted for 18%.

Data center leases are structured in various ways, including triple-net, modified gross, or gross leases. Triple-net leases require tenants to reimburse the site manager for costs including taxes, insurance, operating expenses, and electricity. Modified gross leases, the most common, only require tenants to cover their electricity expense for wholesale data centers.

Wholesale/hyperscale data centers place the entire responsibility for managing network and equipment on the tenant, whereas retail colocation data centers may offer varying levels of hands-on support and other services. This is because they tend to support tenants with shorter-term and smaller capacity needs. In either model, data center leases typically require the manager to provide uninterruptable power and cooling that are critical to avoid business disruption, especially for tenants whose services necessitate consistent connection to their networks. The manager maintains backup batteries and generators for temporary electric utility outages. In addition, data center managers are responsible for site security, maintenance, and replacement of electrical switches, chiller plants, cooling towers, motors, and compressors, and other infrastructure components.

Power is a substantial expense, but exposure to rising energy prices could vary.   For example, power cost accounts for about 28% of operating expense for Equinix. Unhedged power cost could pressure margins amid rising energy prices. Wholesale triple-net lease contracts typically include "metered power", whereby power costs are a direct pass-through to the customer, insulating these providers from rising energy costs. Wholesale data centers would get the bulk of its utility costs reimbursed.

Retail colocation providers are more directly exposed to rising energy prices given that they consume vast amounts of power to run and cool servers. Most retail contracts don't allow for metered power the way wholesale triple-net leases do. However, many have pass-through options if power costs rise above a certain baseline.

In addition, tenants increasingly prefer renewable energy sources. Green energy can be a competitive advantage, helping customers meet their sustainability goals. Many data center providers have transitioned to solar, wind, biodiesel, and fuel cells amid elevated traditional energy prices. Long-term power purchase agreements (PPA) are usually fixed and negotiated with some price escalation, for as much as 30-year terms. We view facilities that purchase directly from a green source or from PPAs as being better positioned to deal with inflation than facilities that run on more traditional energy sources (such as natural gas or coal) because of the contracts' fixed and long-term nature. However, outsize reliance on renewable power could expose the facilities to disruption given that it is intermittent without long-duration battery storage.

Another source of green energy is through the purchase of renewable energy credits (REC), which don't provide the same protection against rising costs. A data center provider still purchases electricity from a utility, which may use a variety of energy sources. By purchasing a REC, the right to claim the use of renewable power is transferred to the data center provider, signaling that renewable energy was generated on its behalf, but the price of electricity can fluctuate with the prices of traditional sources.

Equinix (96%) and Digital Realty (66%) power data centers with renewables, although the renewable energy type and method varies:

• Equinix utilizes RECs for 46%, supplier green power for 39%, PPAs for 10%, and brown power (polluting sources) for 4%.
• Digital Realty's renewable energy accounts for 66% of its global portfolio, retail renewable supply (28%), customer-sourced renewables (24%), RECs (20%), PPAs (16%), and utility-grid mix (10%).

Many smaller data centers do not have high renewable energy usage. It can be challenging for subscale operators to purchase green energy given the relatively costly nature.

Leased data centers versus owned could be more exposed to rising rent expense.  While data centers issuers structured as REITs generally own the bulk of its assets, many data center operators leases the assets and are therefore exposed to rising rent expense. For example, rising rent coupled with low utilization could constraint profitability and cash flow. Among players who lease most of their facilities, we would favor the U.S. data centers with longer-dated average terms. In contrast, those that own their facilities are not exposed to this risk and could benefit from higher rental income. Asset ownership is a distinct credit enhancement because it provides operational benefits and greater financial flexibility.

Strong demand mitigates leasing and tenant pricing risk, supporting cash flow resilience, but overbuilding risk remains long term.  We expect data center demand to outpace supply, supporting good pricing power for data center owners, at least in the near to intermediate term. Supply constraints such as lack of power and land scarcity will likely keep fundamentals in the sector strong in the coming years, pushing up market rent and keeping vacancy minimal. Still, overbuilding could be a risk longer term given the large development pipeline. Capex requirements are high for data centers given the need to upgrade facilities to meet higher power requirements. If demand falters due to slower AI adoption, we could see a delayed ramping up of development projects and a surge in vacancy. While the risk of overbuilding is limited in the next few years, uncertainty on the pace of AI development causing AI demand to wane or regulatory limits for the data center buildout are potential risks. Data center assets tend to be less differentiated than other real estate, historically situated in regions where the combined costs of land, power, and taxes are low. While we believe data center cash flow is supported by the underlying lease agreements, particularly for wholesale data centers with long-term net leases, we believe there are rollover risks at lease maturity 10-15 years out, since the demand picture is less visible in the long term. This risk is more significant for data centers in secondary or tertiary markets.

Chart 5

Insourcing is also a risk, especially for large tech companies such as Google, Microsoft, or Amazon with sufficient economies of scale to build their own data centers, which can limit the price an independent operator can charge. While many cloud service providers (CSP) own a good portion of their data centers, they have chosen to outsource data center development to complement their own build-outs because data center operators have more efficient construction capabilities and can move quicker.

Tenant concentration is another risk for data centers given that the main customers are large CSPs. A disproportionate tenant concentration could limit pricing power, particularly at lease renewal. CSPs can also exert negotiating leverage on operators with pricing concessions or entry into less-attractive markets that could increase re-leasing risk at lease maturity (including hyperscale pushing into tier 2 and 3 locations). Data center operators' customer concentration compares unfavorably to other real estate companies. For example, Digital Realty's top three customers accounted for approximately 22% of total annualized recurring revenue as of March 2024.

Limited technology and obsolescence risks.  Data centers provide power and space, which in and of itself is not a significant obsolescence risk. Long term, the obsolescence risks stem from technology innovations that may reduce the need for space or advancement that may render older data center less competitive.

Given the healthy demand, we have not seen many examples of repurposing of data centers into alternative real estate uses. However, we have seen examples of industrial and even office real estate being repurposed into data centers. We view the repurposing of data centers to be limited given the location of assets that are in many cases in more remote locations to access low cost power, particularly those in secondary locations.

Computing equipment to run hardware for AI applications is exponentially more power-hungry than others. We believe this is a manageable risk. For starters, most AI-related demand will reside in purpose-built new facilities. Sufficient demand from cloud, content, enterprise, and network providers will keep facilities highly utilized without AI demand. Still, we believe retrofitting a building to enable AI-related liquid cooling can be accomplished affordably. Most older buildings have chillers and water-based plumbing already. Therefore, sections of the building can be converted to liquid cooling without an overhaul.

Other technology disruption could include smaller server sizes, which reduce the need for space. However, high-density smaller servers tend to generate more heat, so demand for power will likely increase as equipment shrinks. Contracts are increasingly priced based on power requirements, as opposed to space. Smaller equipment means more can fit into the same building.

Separately, technology upgrade cycles can raise traffic flowing through a given port for network cross-connect, which can pressure interconnection revenues. For example, new network equipment that allows for 100 GW of data traffic has replaced older gear carrying 10 GW. This can reduce the number of cross-connects required. However, it also enhances the importance of that interconnect point and creates customer stickiness, which allows interconnected providers to monetize demand.

Finally, power usage efficiency (PUE) is becoming increasingly important. An unforeseen innovation in building materials or equipment that would drastically improve PUE ratios could render older data centers with higher ratios less competitive, particularly for wholesale facilities. Certain carrier hotels that serve as interconnection points for diverse network carriers reside in older buildings may have weaker efficiency ratios than more modern facilities. However, this is less of a consideration for retail customers that place more value on the cross-connection capabilities with fellow tenants. These customers are typically not as price-sensitive, particularly given smaller retail colocation deployment compared with a wholesale customer.

Demand Likely Supports Credit Quality

We believe healthy demand will be a key for credit metrics, but a higher risk appetite to seize opportunities could change our view. The surge in demand for data center capacity will create significant growth opportunities for data center developers and providers over the next several years. This growth will require high investment and exposure to significant constraints, with power the key constraint. As these assets stabilize, operating risks varies across the operating model, and we believe hyperscale assets face less risk than retail colocation. While we expect positive demand would support the credit quality of data center ratings, we will continue to monitor the risk appetite for data center investments and risks around supply growth and leasing prospects as the AI deployment unfolds and growth trajectory, including tail risks around lease renewal risk at lease expiration, which for hyperscale data centers will be longer term.

Related Research

• Data Centers: Surging Demand Will Benefit And Test The U.S. Power Sector, Oct. 22, 2024
• Data Centers: Welcome Electricity Growth Will Fall Short Of U.S. Data Center Demand, Oct. 22, 2024
• Data Centers: More Gas Will Be Needed To Feed U.S. Growth, Oct. 22, 2024
• Midyear 2024 IT Forecast Update: Robust Cloud Spending Offsets Still-Cautious Enterprise Budgets, July 17, 2024
• The Four Main Approaches For Rating Data Center Financings, June 13, 2024