By John Ritter, Managing Director of Real Assets at Texas Municipal Retirement System
Unless you have been living under a rock, you know the current narrative around power and data centers: We don’t have enough power, we don’t have enough data centers, we don’t have enough chips.
This narrative has driven the markets, both public and private, to extreme excesses in terms of valuation and expectations over the past few years. The narrative stems primarily from the high growth in compute, mostly driven by a new computationally expensive technology, AI.
As a real asset investor[1] for over 20 years, I am always skeptical when real asset markets get this hot, and the market for power and data centers is hot, breathtakingly hot. Many of our opportunistic real estate relationships now have a power or data center strategy, even though none of them have ever invested in the sector before. This is a big red flag, and one I will address in a later article. Naturally, I start getting skeptical and I start asking the questions: what are we not seeing? What important factors are being drowned out? There are many, too many, for this one article.
What do I not hear investors talking about? Technological change and evolution. We seize on it when it arrives like artificial intelligence, then we act as though it is in a steady state. It is not. For data centers, the most important changes are in chips and processing becoming ever more efficient over time.
This efficiency is reflected in two “laws”: The first, which everyone knows, is Moore’s law, which states that the number of transistors on a chip doubles roughly every two years. The second, lesser-known law, is Koomey’s Law, which states that the number of computations per joule of energy doubles roughly every year and a half.
These two laws allow for an exponential compounding of computational density over time. For example: the amount of compute a data center can accomplish on a power budget is roughly doubling every two years. Nvidia Corp.’s most recent Blackwell B200 chip released in February advertises a 2.2x power efficiency over the previous generation. This allows for more servers and chips per data center and greater server densities. A second factor, which we have only recently seen, is algorithmic efficiencies around AI. These allow AI to train and inference with much more energy efficiency.
We have seen the effects of these efficiencies before. The rise of internet services caused explosive growth in data center development, with data center electricity use rising 90% from 2000 and 2005, 24% from 2005 and 2010, and 4% from 2010 to 2014. However, between 2014 to 2020 data center electricity use was forecast to be basically flat[2]. Over this same period, internet traffic increased 56x.
Data centers have three primary limiting dimensions: physical space, power supply, and connectivity. Data centers are first and foremost a physical box to house servers. You can only put so many servers in a building of a particular size and specification. Key factors to space include weight-bearing capacity and cooling infrastructure. Second, each data center is rated to a particular power supply delivered to the premises.
Power is so important for data centers that data center capacity is quoted in terms of power not square feet as is the case for other real estate. However, power to a data center is not fixed over the medium to long term, and represents a moveable upper limit.
Finally, the location is important for connectivity and latency vis-a-vis end users. Though connectivity is universal, it varies in importance based on the specific use cases of a data center. For example, AI training needs less connectivity than AI inference.
So, what impact does this densification have on data centers as an investment over time?
First, be wary of older data center locations. These tend to be smaller, but well-located facilities. The problem with these facilities is one of specification. They are not built to handle either the weight of current server densities or modern cooling technologies like liquid cooling. It is easy to get lulled into buying de-risked, stabilized assets for a nice yield only to be left standing when the music stops. These data centers will need substantial capex to bring them up to modern specification. Therefore, it is important to consider how future proof a data center is.
The second impact is one of location. Many large hyperscale data centers are being built in less-than-optimal locations, such as the Texas panhandle, due to power availability, cheap land, and the sense of urgency around developing AI. These areas have far less attractive connectivity given their distance from population centers. The natural tendency will be for these workloads to migrate back toward better-connected locations as power availability in more attractive markets inevitably corrects itself and processing continues to densify.
Due to the relatively short lives of the chips (three to five years) and leases (five to 15 years), vis-a-vis the data center life (30 years+), the data center tenants have a valuable option at renewals allowing them to improve connectivity and location if they choose at little to no cost. Owning less than ideally located data centers is, in the end, a bet against technological innovation.
Even with these efficiencies, demand is estimated to grow far faster than supply. Keeping in mind that most of the demand appears to be anticipated as opposed to responding to immediate supply problems, technology estimations are frequently wrong. I don’t know the future. I don’t know if these technology “laws” will continue or if compute will really grow as fast as they say. I have learned over my career that it is generally not a good idea to bet against human ingenuity.
Ingenuity got us here; it will likely solve the problems as well in wildly unpredictable ways. Understanding these potential risks leads us to require higher required returns for older and less well-located assets.
The views and opinions expressed are solely those of the individual contributors and do not necessarily represent the official views of Markets Group.
Are you an allocator who disagrees with this column, we invite you to send your column or comments for publication to news@marketsgroup.org with the subject line: “Data Center column rebuttal.” Also feel free to suggest your own column idea.
[1] I manage a large institutional portfolio of real estate, infrastructure, and natural resources with over 25 years experience.
[2] United States Data Center Energy Usage Report, Lawrence Berkely National Lab, June 2016
