Key Points:

• Technological advancements like DeepSeek may paradoxically increase overall data center demand despite improved efficiency.
  
  
• Constraints on data center supply include significant requirements for electricity and water, as well as specialized expertise for construction and operation.
  
  
• The strong, long-term demand drivers for data centers underscore their growing significance as a niche asset class.

Our latest Industrial Potential Cap Rate (PCR) model shows a modest decline from the third quarter of 2024, the second consecutive quarter of decline after seven quarters of increases. This dynamic is in part due to falling construction starts and decelerating growth in vacancy rates. But within industrial real estate and related asset classes, one sector in particular – data centers - is at the “center” of a technological transformation that could reshape demand in ways that are still unfolding. 

Data centers are crucial for housing the servers and processors that enable large-scale computing. The rapid adoption of Generative AI (Gen AI) technologies and the introduction of DeepSeek, a new Gen AI model supposedly trained at a fraction of the cost and using less electricity than existing models, has sparked debate about whether this more efficient AI model could deflate data center leasing demand.

 
While it’s possible that new technologies can be more cost effective and energy efficient than old ones, a single, new Large Language Model (LLM) is unlikely to materially impact the long-term economic prospects for data centers. Strong demand for the high-tech space and the technical complexities involved in building new data centers suggests that supply limitations will be the prevailing force in the data center market in the years to come.

"The advent of more efficient Gen AI models does not guarantee a long-term decrease in data center demand… innovation could lead to broader adoption that raises data center leasing demand - a Jevon’s paradox."

Data Centers Not Only for AI

Not all data centers are used for Gen AI computing. A substantial proportion of power is dedicated to cloud computing services, like distributed storage. As shown in the top part of the following chart, only about 30 percent of data center power consumed was for AI-related purposes at the end of 2024. However, AI power consumption is expected to grow more quickly between now and 2028. 

Even in AI-related data centers, less than half of the power consumed is for AI-related computing. As shown in the lower portion of the chart, about 40 percent goes to computing, including AI applications and related services, another 40 percent is for cooling systems, and the remaining 20 percent supports operations like lighting and network equipment. While DeepSeek and other similar agents may reduce agent-level energy demands, other factors will still continue to drive considerable energy usage in data centers.

Efficiency Paradox

DeepSeek, and other Gen AI technology, may also have the opposite effect. If it successfully lowers costs and power requirements, it is more likely to be broadly adopted. Broad adoption could then create a net increase in energy consumed by Gen AI models in data centers. 

This phenomenon, where a more efficient technology leads to an overall increase in demand, is so common that it has an economic name: “Jevon’s paradox”. Named after a 19th century economist, Jevon’s paradox was originally created to describe how the steam engine drove increased demand for coal, but it could just as easily apply to Gen AI and electricity demand today. It remains to be seen whether lower per-agent energy requirements will actually result in a decline in aggregate energy used in AI computing.

Supply Constraints Abound

There are, however, several clear and meaningful constraints to adding new data centers. Data centers require vast amounts of electricity and water for cooling. Electricity must be supplied either by a utility company or generated on-site. While investing in self-generating power is no small task, some utility companies struggle to provide the needed power, which has led to discussions about building small modular nuclear reactors next to energy-hungry data centers to directly supply them with power.

Electrical service is such a primary consideration in data center design that, unlike other types of industrial and commercial properties, data center space is typically measured in terms of available power in megawatts (MW) or kilowatts (kW), rather than square feet. Over time, denser data centers have been built, providing more power capacity in the same physical space. However, this requirement for large-scale electrical and water connections means data centers are more often built at confluences of utility service lines, which limits the quantity of available space for new construction.

Additionally, deep expertise is required to build and run a data center. A data center, though technically a type of industrial commercial property, is not like a distribution warehouse. Though some warehouses need to be climate controlled, others don’t, and if you only have lights and a few large pieces of equipment, then your electricity needs could be quite small. Compare this to a data center running tens of thousands of processors that also needs spare computing capacity to ensure minimal downtime for their customers. 

Taken together, there remain several meaningful constraints on new data center supply. Even as Gen AI continues to be incorporated into more facets of our lives, creating more supply will take time.

DeepSeek, and other Gen AI technology, may also have the opposite effect.

The Long View

The advent of more efficient Gen AI models does not guarantee a long-term decrease in energy consumed for AI purposes. Data centers have numerous long-term demand drivers, which include Gen AI and other digital services, and innovation could lead to broader adoption that raises data center leasing demand – a Jevon’s paradox. 

On the supply side, constraints are easier to identify and difficult to overcome. The collision of supply limitations with multiple long-term demand drivers will make data centers an increasingly crucial niche asset class in the years to come.

About the Industrial Potential Cap Rate Model

The industrial Potential Cap Rate (PCR) Model estimates cap rates^[1] based on the historical relationship between industrial vacancy rates, e-commerce sales as a percent of total retail sales, industrial construction starts, and CRE debt flows. The industrial PCR Model uses these metrics to establish a potential cap rate level that is supported by these market fundamentals. When actual industrial cap rates are significantly above the industrial PCR, it indicates that market fundamentals supported lower cap rates than were observed. Conversely, when actual cap rates are significantly below the potential cap rate level, market fundamentals supported higher cap rates than were observed. Industrial cap rates are aggregated nationally, and the industrial PCR Model is updated quarterly.

A cap rate is a measure of estimated yield, or the return, on an investment property assuming no debt is used to purchase it. Cap rates are calculated by dividing an asset’s net operating income (NOI) by its value. NOI is an asset owner’s remaining income after covering operating expenses, but before servicing debt. Since cap rates do not take debt service into consideration, cap rates are a measure of what is called unlevered yield.

^{[1] A cap rate is one measure of return on investment provided by a building and is equal to the net operating income (“NOI”) generated by the building divided by the price of the building. For example, an industrial property purchased for $100,000 that generates income of $10,000 a year has a cap rate of 10 percent. Higher cap rates represent higher rates of return, and vice versa.}

Fourth Quarter 2024 Industrial Potential Cap Rate Model

First American’s Industrial PCR Model estimates a potential national cap rate for industrial properties based on several industrial property market fundamentals, including industrial vacancy rates, eCommerce sales, construction starts and commercial mortgage flows. 

• The industrial PCR was 6.4 percent, a decrease of 0.2 percentage points as compared with the third quarter of 2024.
  
  
• The industrial PCR increased by 0.1 percentage points as compared with one year ago.

Industrial Cap Rate Outlook Gap

The gap between the actual industrial cap rate and the industrial PCR provides insight into the likelihood of shifts in the actual cap rate. If the industrial PCR is below the actual industrial cap rate, it indicates that fundamentals supported lower cap rates than were observed. If the industrial PCR is above the actual industrial cap rate, it indicates that fundamentals supported higher cap rates than were observed.

• In the fourth quarter of 2024, the actual national industrial cap rate of 6.4 percent was the same as the industrial PCR, indicating that market fundamentals supported observed cap rates.
  
  
• The gap between the industrial PCR and the actual industrial cap rate decreased in the fourth quarter as compared with the third quarter of 2024.