Data Centers: A Peek Behind the Curtain

The shift from AI training to inference could create sustained demand for data-center infrastructure, but power, permitting and deployment constraints may prevent most announced projects from being built.

• The authors toured QTS’s data center at White Oak Technology Park near Richmond International Airport, where QTS had one building when Blackstone acquired it and four subsequent buildings were designed with closed-loop cooling.
• AI training requires large server arrays running continuously and consumes substantial electricity, whereas inference uses less computing power but creates recurring demand as trained models are distributed across data centers closer to users.
• The tour guide said operators may rent less space at the site where a model is trained but ultimately rent more space across multiple data centers; the authors also note that training and inference may require different semiconductors.
• Closed cooling systems used little water during the visit, while open systems rely heavily on evaporation, leading the authors to question whether media concerns about data-center water use are overstated.
• Data-center landlords can reportedly build substations in one-third the time of power companies, but behind-the-meter generation usually relies on noisy natural-gas generators; the authors know of no small modular reactors operating in the United States and expect only a fraction of announced projects to be completed over the next seven years.

Takeaway: The authors see the strongest long-term opportunities in companies supplying infrastructure for future data centers rather than simply in the announced facilities, while flagging uncertainty around power availability, permitting, cooling technology and the economics of frontier AI companies.