Optimism grows as capital flows, but what if AI promises fade?

Overall, the combination of strong earnings, record-high equity prices and narrow corporate spreads is providing ample opportunity to fund large-scale spending, which we expect to add meaningfully to economic growth in the near term. 

While we have observed a relatively disciplined market since the data center boom started two years ago, an increasing imbalance between high investment demand and limited supply of quality projects may translate into more aggressive valuations down the road. Recent record-low data center cap rates of 4.4% (per Colliers/CBRE data) correspond to EBITDA multiples well above 25x, in the upper range of the best-quality infrastructure asset classes. This may result in more financial leverage, weaker security packages or liquidity buffers, and reduced spread differentiation to reflect asset quality. It may also lead to even more financial innovation to satisfy demand, including financing of a broader scope of assets, such as power generation and electrical components, with diminishing distinction of perceived risk of the underlying assets. A rule of thumb is that for each dollar spent on a data center building, three dollars are invested in equipment that goes into it, including GPUs and cooling systems. This clearly opens more doors for debt investors. Effective data center risk assessment is essential to navigate these complexities.

Despite the sector’s strong position, we observed several, mostly long-term risks that could impact data center debtholders in transactions that S&P Global Ratings has analyzed.

Overbuilding and overinvestment: This is a longer-term concern given the robust development pipeline. There are numerous sources of data center demand, including data storage, cloud computing, content, enterprise, network providers and, of course, AI-related uses. However, if demand falters due to slower-than-anticipated AI adoption, we could see a surge in vacancy. Some of the tenants may fail, reposition, or consolidate into bigger tech companies and therefore no longer need space. One potential mitigant is that sufficient demand from other providers may keep existing facilities utilized even without AI demand, although that argument may falter as the proportion leans toward more complex AI applications. Overbuilding risk could magnify because capital expenditure needs for new builds, especially among hyperscalers, are so high and may make up an increasing tranche of overall spending. This may concentrate the financing risk associated with this sector since only a small number of very large firms can handle the massive capital needs of hyperscalers, leaving firms more exposed to any disruption. Effective data center risk management can help mitigate these risks. 

Rollover risk and low residual value: Historically situated in regions where the combined costs of land, power and taxes are low, data center assets tend to be less differentiated than other real estate. While we believe the underlying lease agreements support data center cash flow, particularly for wholesale data centers with long-term net leases, we believe there are rollover risks at maturity. This rollover risk may also be heavily concentrated in the period roughly 10 years from now, given recent rapid leasing growth. With high current demand, we have not seen many examples of repurposing data centers for alternative uses. However, we have seen industrial and even office real estate repurposed for data centers. We view the repurposing of data centers to be limited, given the location of assets in more remote locations, which limits their residual value.

Technology and obsolescence risks: Long-term, we believe the main obsolescence risks stem from technology innovations that may reduce the need for space, or advancements that may render older data centers less competitive. Most AI-related demand will reside in purpose-built new facilities. AI inferencing may require data centers closer to metro centers where real estate is at a premium, leading to re-leasing challenges for large bespoke properties in secondary or tertiary locations with fewer prospective tenants.  

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

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

Regulatory/sustainability issues: These factors could pose additional challenges, likely more so for larger projects. Data centers' potential sustainability impacts could also pose regulatory risks. These could add requirements to planning and approvals, such as targets for power use and efficiency. They can also include penalties that apply when groups of data centers create an imbalance in power for regional grid capacity. All of these imply potentially higher costs if regulations tighten. Anecdotally, hyperscalers are seeking off-grid power sources to mitigate this risk.

Many technology companies’ ambitions are linked to both the pace of AI adoption and the capabilities of their data centers. We believe the firms most vulnerable to a slowdown are those that have invested heavily but have yet to turn a profit (or lack a clear path to profitability), rely on external funding, and operate in areas susceptible to rapid technological obsolescence. Similarly, companies that have engaged in debt-financed R&D, capital expenditures, and mergers and acquisitions may face heightened risks. Among software companies, those most at risk are businesses with limited product diversity. Additionally, financial sponsor-owned companies may be particularly vulnerable, as they often carry substantial debt and lack the capital necessary to invest in innovation and compete effectively. These entities also lease data centers, and their fate may hence affect data center earnings. In contrast, the largest firms (e.g., Microsoft, Alphabet, Amazon and Meta), with diversified revenue streams, are better positioned from both business and credit perspectives. Even in a slower-than-anticipated AI adoption scenario, we do not foresee a significant impact on these types of firms.