JLL analysis shows AI infrastructure power demands now drive real estate location and site selection decisions.

According to Sean Farney, vice president of data center strategy at JLL, artificial intelligence is fundamentally altering design assumptions, operational standards, and geography for hyperscalers and colocation providers. Farney compared the uncertainty of AI planning to "planning a trip to Disney World a year in advance," explaining that "while there's an understanding of the eventual goal, many of the specifics remain uncertain." The industry itself is experiencing profound transformation, as "the industry is currently in a state of flux, with AI largely still in the research phase. This rapid change in trajectory and speed has caused a significant shift in how data centers operate, leading to a complete rewrite of operational run books and design bases within just the last two years."

Currently, AI implementation remains primarily within the realm of hyperscalers constructing massive learning facilities, both hybrid and dedicated, with some GPU-as-a-service options. However, the market is shifting toward AI inferencing—real-time applications where the value of AI is monetized at the edge. Farney noted that "every company and enterprise recognizes the need for this capability, and due to technical requirements such as low latency, the geography of where this computing takes place may shift."

The most visible impact of AI on data center infrastructure is the dramatic increase in rack-level power density. "AI cabinets can require 50 to 100 kilowatts of power, with some potentially reaching the mythical 1 megawatt per cabinet," Farney explained. This increased power density results in "a shrinking overall facility footprint, with the same computing power now possible in a fraction of the space." However, these demands fundamentally reshape infrastructure requirements. The heat produced at these densities "overwhelms traditional air cooling methods, necessitating liquid cooling technologies." The industry is currently "experimenting with various cooling methods, including immersion cooling, direct-to-chip cooling, and rear door heat exchanger technology," with different providers adopting various approaches in what Farney characterized as a period of "creative destruction" where innovation is rapidly transforming traditional operational methods and equipment.

As inferencing takes center stage, location strategy has become critical. Some hyperscalers are constructing AI-dedicated facilities with future AI inferencing monetization in mind, which requires close proximity to end-users to minimize latency. To meet the critical "sub-10 millisecond" latency target, AI-native facilities are being built within a 100-mile radius of major population centers, focusing on "top U.S. data center cities such as Atlanta, Chicago, Dallas, Northern Virginia and Phoenix." Farney stated that "this approach ensures that millions of potential users are within reach, making these established markets ideal for AI-native facilities." While large language model training facilities could potentially be located in more remote areas, "the future of AI inferencing monetization is steering providers towards building in existing, well-connected markets."

Even within these traditional data center hubs, future AI infrastructure will differ significantly from current facilities. These centers will "typically have a smaller physical footprint but much higher power density, reflecting the evolving needs of AI computing infrastructure."