The Risks of Relying on AI Growth to Drive Data Center Demand

A classic problem in CRE is to see some area come under demand, rush in to invest, and eventually see it overpopulated. The mechanism is easy to understand because it can take extended time to build projects, only to see that many others did, potentially creating an amount of supply that can’t be quickly digested.

The point is worth considering in the face of a recent Yardi Matrix report on how AI is driving demand for data center.

“Artificial intelligence has grown rapidly over the past year, requiring vast amounts of data and processing power, in turn generating increased demand for data centers,” the report says. But it also looks at factors that can make facility construction difficult and how that might affect where projects will be built.

First, there is the question of what AI means. Rather than one thing, like a ChatGPT, it’s a broad collection of technologies going back to the 1950s. Some are hosted in data centers, some embedded in software run locally.

But a great deal is found on data centers because a lot of software capability is delivered via hosted services. According to Yardi, in 120 markets they cover, there is currently 27.4 million square feet under construction and another 33.5 million square feet in planning. Even as industrial construction has slowed, data centers have picked up.

“Between 2020 and 2022, data centers comprised 2.0% of all industrial starts, but in 2023 the share grew to 4.3%. Meta has fully embraced AI and is the most active owner of data centers under construction in Yardi Matrix markets, with 7.4 million square feet currently being built,” they wrote. “Meta’s largest project underway is the 2.0 million-square-foot Eagle Mountain Expansion in Utah.”

Northern Virginia is a hotbed, with twice as much completed space as second-place Dallas. Part of its popularity is favorable taxation, but more important is the strength and speed of the telecommunications infrastructure. It can be faster to deliver data than from other parts of the U.S. The logic is similar to why stock trading firms want to be as close to Manhattan as possible to make trades faster. Small differences can mean speed advantages that translate into monetary ones.

But geographic concentration is starting to have its limitations. The AI data centers use enormous amounts of power, and the region can’t satisfy virtually unlimited demand, given the energy required to train generative AI models, plus the area is densely built. Developers are looking further west out of necessity. However, water availability is an issue. While many more traditional data centers have made advances in air-cooled servers, the amount of computing in AI means many processors, all kicking off heat. So, they turn back to water cooling and “usage has become a major concern of residents and policy makers alike in places including Phoenix—which has 4.3 million square feet under construction and another 6.3 million in planning—Salt Lake City and Eastern Oregon.”

Right now, it seems that the demand for data centers is unquenchable. And yet, will generative AI continue as it has? There are already major lawsuits in process, and it is far from clear how courts will rule on use of unlicensed data for training. Additionally, many new technologies go through a hype phase and then slow. Will generative AI continue to be used? Yes. As widely and with as much of a “cool” factor for the curious as it is now? Yet to be seen. Those investing in the product type have to closely attend to development on the usage, building, and legal fronts.