Data Centers: A Peek Behind the Curtain

Data centers (DCs) have frequently been in the headlines of financial and non-financial media all year, so we were glad to take a tour of QTS’s DC at the White Oak Technology Park near Richmond International Airport. We knew of QTS because we owned the stock long ago, and it was later acquired by Blackstone (ticker BX). DCs are the landlords that own and maintain the buildings while other companies rent space in DCs to install and run their arrays of servers and network equipment. Below are some of our observations.

From an investment perspective, perhaps the most interesting observation involved the transition from “training” AI/large language models (LLMs) to using AI/LLMs for “inference.” AI training uses huge amounts of electricity because it requires massive iterations of computational processes. During training, a model learns patterns by adjusting its parameters using large datasets. It might help to think of training like studying all the content of Wikipedia as if it were written in a long-lost language until you knew the alphabet and all the words and had mapped all the ways in which the words might be combined. AI inference uses far less computing power (and far less electricity) and is the process where a trained AI model uses what it has learned to make predictions or generate output, such as answering questions about where to go for spring break to see a Beatles cover band and go fishing for trophy fish, classifying emails, or writing a song. Inference is the execution phase of AI, distinct from training, and powers real‑time applications like chatbots, fraud detection, and recommendation systems. Training is a huge one‑time cost with additional costs for periodic updates, whereas inference is a much smaller ongoing cost every time the model is used.

Our tour guide said that the compute-heavy nature of training requires large arrays of servers (computers without screens or monitors) running around the clock to crank through the data on which it is being trained. However, once a specific model is trained, the owner distributes the AI model across multiple global data centers closer to end-users to run inference efficiently at scale. The tour guide said AI model owners will rent less space in the DC where the model was trained but will ultimately rent a greater amount of space spread across many DCs. We understand that the semiconductors required for training are not necessarily the optimal chips for inference, and a shift from training to inference could create years of sustained growth, so we are looking in this area for investment opportunities.

It was fascinating to peek behind the curtain to see how the cooling systems work and learn about the different systems made by various companies, most of which are publicly traded. We were impressed by how little water the “closed cooling systems” use (some work like a large refrigerator), and how much water open cooling systems use due to evaporation. Open cooling systems are frequently cited in the media that oppose DCs.

When QTS was acquired by Blackstone there was only one building in the White Oak Technology Park, and it used an open cooling system. The next four buildings were designed with closed loop cooling systems, and we are starting to think the water usage concerns around DCs might be overblown by our sources of news.

The White Oak Technology Park was as quiet as a regular office park, but that is because there was a substation on the property. Our tour guide said the DC landlord will often build the substation according to the same processes and specifications as the power company would because that is what regulations require, but the landlord can build it in a third of the time because of a lack of red tape that causes a power company to move relatively slowly. If a DC does not have a substation from a power company, the DC may generate “behind the meter (BTM)” electricity, i.e., electricity that does not come from the power grid. Our tour guide warned that BTM power generation is usually from burning natural gas, and natural gas reciprocating generator sets and turbines can be especially loud, and BTM is really only possible in very rural areas. There has been much talk of “small nuclear reactors (SMRs),” but we are unaware of any SMRs in use in the United States. It is also possible that as DC designs get bigger and bigger, SMRs may not be able to supply enough electricity.

The vast majority of announced DC projects will require years to fulfill required permitting and source the power. It would not surprise us if only a fraction of the announced DCs were built in the next seven years. That said, we are seeing the best long-term investments coming from companies that build the infrastructure for future DCs.

Future essays may include a deep dive on computing and networking systems in a DC to assess the addressable market, the accounting standards that apply to the trillions of dollars that have been “spent” but not yet “expensed,” and the revenues and profits required from the frontier AI model companies, such as Open AI and Anthropic, to justify these levels of investment. If you would like to get our current thoughts on DCs or other investment topics, please reach out.

Redmond Asset Management, LLC October  2026

The opinions contained in the preceding commentary reflect those of Redmond Asset Management, LLC (RAM). The stated opinions are for general information only and are not meant to be predictions or an offer of individual or personalized investment advice. They also are not intended as an offer or solicitation with respect to the purchase or sale of any security. This information and these opinions are subject to change without notice. Any type of investing involves risk and there are no guarantees. Redmond Asset Management, LLC does not assume liability for any loss which may result from the reliance by any person upon any such information or opinions.

Redmond Asset Management, LLC (RAM) is an independent, SEC registered investment management firm located in Richmond, VA and is not affiliated with any parent organization. RAM was founded in 2005 and registered with the SEC on December 22, 2005. The company offers investment management services for equity, balanced and fixed income portfolios to corporate, institutional, and individual investors.