The possibilities presented by generative AI dominate headlines, business strategy and economic projections. But, however much these possibilities may depend on technological innovation and successfully developing markets around use cases, on an operational level, generative AI (and data storage and processing more generally) is inextricably tied to physical real estate: namely, data centers and their supporting infrastructure. As real estate developers and private investors seize these opportunities to respond to insatiable sector demand, they must do so thoughtfully in consideration of palpable resource constraints, public pressures and long lead times.
Healthy Demand, Lagging Supply
The United States is facing an impressive shortage of rentable data center facilities. As of the first half of 2024, the national rental vacancy for space at “colocation” centers (that is, data centers from which tenants rent space to operate their servers and processing equipment) reached a record low of 3%, with vacancy rates in certain core markets, such as Northern Virginia’s “Data Center Alley,” being as low as 1%. Moreover, the high preleasing rate for facilities under development—which currently exceeds 80%—likewise contributes to the low availability of space. Lease renewals are also ubiquitous, due not only to tight supply but also to the logistical challenges of relocating technical equipment, and, in concentrated markets, tenant rights of first offer for adjacent rental premises are common.
Unsurprisingly, high demand and low supply have allowed data center lessors to enjoy consistently increasing rental rates, and yields for data center developments remain high and outperform those for other industrial or residential asset classes. Yet the ability of real estate developers and investors to increase inventory and extract additional surplus faces several headwinds.
Power Availability and Geographic Risks
Data centers depend on the availability of affordable, consistent power. Data processing has always been power-intensive, but the meteoric rise of generative AI has amplified data centers’ power needs: AI model training runs 24/7 with no predictable usage cycles, and such processes may be impaired by even momentary curtailments in power. In addition, the processors required for AI programs require more power than their predecessors. Indeed, total power consumption in the United States is projected to increase over the coming years—reversing years of relatively flat demand—partially as a result of generative AI and data processing growth. Data center developers and investors must therefore consider the availability and reliability of the surrounding power infrastructure, which has become strained in major markets. Grid interconnection for new data center developments may require significant time and resources; lead times for data centers to secure power are increasing and extensive, currently up to three years in certain markets.
Closely related to—and oftentimes reliant on—power availability is the need for sufficient cooling capacity, as better cooling capability allows servers to be stored more closely and thus increases data center utilization per square foot. Nearby water resources greatly reduce the cost of cooling (proximity to the Potomac River is a factor in the dominance of Data Center Alley), though reliance on air-conditioning and electricity-powered cooling mechanisms are becoming more common as data center campuses move toward drier regions in search of more available and affordable grid capacity.
Beyond the availability of power and cooling, data center developers and investors should pay special consideration to other factors relating to the geographic locations of potential centers, especially natural disaster risks and proximity to populous areas. While proximate population centers offer a large pool of highly educated employees required for the operation of data facilities, nearby data centers may face nuisance complaints (as discussed below). Moreover, the unavailability of larger plots of land in or near population centers could be problematic, as building vertical, multi-story data centers—which require specialized engineering to ensure that floors are adequately weight-bearing and that ventilation systems sufficiently insulate higher levels from heat—may not be a feasible option. On the other hand, establishing data centers far from large population clusters will result in latency for end users, though latency may be less of a concern for large model training processes.
Concern Among Utilities, Governing Bodies and the Public
Some power utility companies have begun to express concerns about projected grid overload due to data center growth. Earlier this year, for example, a large public utility company in Ohio requested that the state’s utility commission approve discriminatory rate tariffs on data center customers—including the imposition of 10-year minimum billing commitments and exit fees for early termination, secured by collateral—to provide sufficient reserve capacity and lessen the risk that costs are passed onto ordinary consumers should data centers reduce their currently intended utilization or abandon their premises. Utility companies in Pennsylvania have likewise challenged a proposed interconnection services agreement between a data center and a co-located nuclear energy facility, reasoning that the agreement would divert critical grid capacity from and shift transmission costs onto other ratepaying consumers. And while the Federal Energy Regulatory Commission recently rejected a separate request for discriminatory rate tariffs, it left the door open for future requests going forward.
Areas that have become data center hubs are also seeing community opposition, and zoning commissions have started to react to perceived nuisances. Northern Virginia’s Fairfax County, for example, enacted stricter zoning codes in September requiring equipment enclosures and imposing size limitations, larger setbacks, minimum distances from metro stations, noise studies and exterior design improvements.
Extended lead times to project completion compound the risks associated with public resistance to development. While the outcomes and frequency of public challenges to data centers and associated infrastructure development remain uncertain, developers and investors face the possibility that utility markets will saddle data center customers with long-term financial commitments untethered from the actual, realized energy utilization and tenant demand at the time the facility eventually comes online.
Regulatory Support
However, public sentiment towards data centers is not exclusively antagonistic. On the contrary, several local and federal initiatives appear to be warming to this unique asset class and its supporting infrastructure. While Fairfax County in Virgina may have imposed zoning codes unfavorable to further data center development, the nearby town of Culpeper reserved several hundred acres for use by incoming data center developers, planned for compatibility with nearby populations and energy infrastructure. Farther south and in consideration of data centers, Georgia’s public service commission approved expansion of a utility’s generation capacity—including fossil fuel-based generation.
Governmental bodies are also supporting nuclear power, attractive to data center customers due to its reliability and carbon-free emissions. For example, Pennsylvania is hosting an initiative to power data services with a decommissioned reactor at Three Mile Island. At the federal level, the Accelerating Development of Versatile, Advanced Nuclear for Clean Energy (ADVANCE) Act was enacted by Congress earlier this year and seeks to promote efficient licensing procedures for nuclear power and to encourage the reclamation of brownfields and retired fossil fuel sites for nuclear power. The U.S. Nuclear Regulatory Commission has also formally pre-approved a small modular reactor design.
Developer and Investor Strategy Adaptation
Real estate developers and investors seeking to capitalize on the voracious demand for data processing facilities while balancing a range of uncertainties have several strategies available that can be tailored to their investment capital and risk appetite. At the most basic level, real estate investors can pursue a pure land play, acquiring parcels of land to later sell or lease to developers who will construct the facilities. Identifying desirable locations is key, and consideration must be given to projected access to reliable power infrastructure and cooling systems, natural disaster risk, proximity to population centers and demand once the facilities come online. Secondary markets with greater access to affordable power and capacity for renewables development may prove attractive, particularly to customers conducting large AI model training, which are less concerned with end-user latency.
At the other end of the spectrum, real estate investors may partner with specialized developers or tenants to build new data centers to suit and oversee the entire project, from land acquisition to contracting long-term power sources and outfitting all systems required for a tenant’s operation on the premises. Small-scale facilities may elect to pool power procurement contracts and arrangements with co-located power facilities. Developers and investors of large-scale facilities may further seek to integrate modular power generators into data center sites.
Looking Forward
New demands and higher expectations for generative AI and other novel data processing applications show no signs of slowing. Although there are material challenges to the development of the data centers needed to support these applications, this unique asset class will continue to present ample opportunity for investors and developers going forward.
The Private Equity Report Fall 2024, Vol 24, No 3