Data center real estate is such a hot topic that some have wondered whether it’s overhyped. But the digital and cloud revolutions—and the scale-up of AI—are fueling truly enormous demand for data centers. In turn, there are a variety of ways that real estate companies can participate in the growth of the industry. McKinsey senior partner Pankaj Sachdeva talks with McKinsey Global Publishing’s Katy McLaughlin about what these companies should know about the places, power, and strategic alliances that can lead to success. An edited version of their conversation follows.
Katy McLaughlin: Data centers are a much-discussed topic in the real estate industry. Do they deserve the hype?
Pankaj Sachdeva: Our view is that over the next decade, the industry will go through an S-curve of demand growth to support the infrastructure that will power the digital revolution and continue to power the cloud revolution. It can also support the rapid scale-up of AI that’s likely to happen. McKinsey values AI and generative AI as a roughly $4 trillion global opportunity.1 Data center value is a representation of the physical infrastructure needed to achieve that.
In other words, this is a long-term trend, as opposed to hype or a quick, volatile frenzy.
Katy McLaughlin: Can you quantify the expected demand?
Pankaj Sachdeva: Our proprietary data-center demand model projections say that, today, the world sits at about 60 gigawatts of total data-center-critical IT demand. By the end of the decade, there will likely be demand for around 170 to 220 gigawatts. In other words, capacity needs to more than triple.
In fact, our estimate of demand unconstrained by supply is well north of 250 gigawatts. The power constraints—both globally and in the United States—and other supply chain issues aren’t trivial, making north of 250 gigawatts a challenging goal. But the demand is there.
Katy McLaughlin: Does this mean that existing data centers have to get bigger or just more powerful? Or does this mean that more or more powerful new data centers are needed?
Pankaj Sachdeva: It means all of the above. The question is power: as we put it in the data center industry, time to power is a real limiting factor for time to market. A data-center-development project can be completed soup to nuts in 18 to 30 months if it can draw from existing power sources. But new power generation projects can take a lot longer than that.
There’s likely to be a massive amount of opportunity to build from the ground up. That’s more likely to happen in locations that already have the necessary excess power or in places where there are ways to either get power or generate more power in a reasonable time frame. The subset of data centers at locations that can easily get more power are good candidates for retrofits and upgrades.
Katy McLaughlin: How can real estate companies get involved?
Pankaj Sachdeva: You could break down the opportunities into three buckets.
At the most basic level, there are real estate companies that specialize in buying the land for data centers. They buy desirable parcels of land that they think will get the power and network connections needed for data centers, as well as customer demand. Then they sell or lease this land to data center developers.
In the next bucket, there are real estate companies that do partial [data center] development and then sell the facility to either developers or the final customers.
At the furthest end of the spectrum, there are companies that buy the land, construct the building, provide access to power and connectivity, and build out the interior as well. This is the co-location model. These facilities are then often leased by enterprises or hyperscalers, which are their tenants.
Katy McLaughlin: What’s a hyperscaler?
Pankaj Sachdeva: Hyperscalers are those that build or operate data centers at a significant scale. The most obvious examples are the big cloud services players and the biggest tech companies. They commission multiple data centers every year to keep up with cloud, and now AI, demand.
Hyperscalers use well north of half of all computing power. By end of the decade, they will use 65 to 70 percent of it.
Katy McLaughlin: Let’s start by examining that first opportunity: the pure land play. What’s changing in that market?
Pankaj Sachdeva: Until recently, data centers have mainly been built in network- and population-dense areas, such as in and around Northern Virginia, Silicon Valley, and Dallas. But due to power constraints, there’s a growing shift to smaller cities and metro areas.
There’s an even more interesting shift where data centers are built in remote locations, including in Iowa, Wyoming, Wisconsin, Alabama, and other places that don’t actually have lots of local demand. Data center developers in these markets are making investments to get power and fiber networks to these sites.
Both hyperscalers and real-estate-data-center developers do something called “land banking.” They bank land in desirable locations five, seven, or ten years before they even start to build. Many developers have hundreds, if not thousands, of acres of land in their banks. They gradually bring capacity online, so they might commission a project of, say, 50 megawatts or 100 megawatts every two years. Much of it depends on how much capital they have to deploy and when they think tenancy will begin.
Katy McLaughlin: Moving on to the partial-development opportunity, what are the recent trends?
Pankaj Sachdeva: We have observed more real estate companies moving from pure land deals to building the “dark shell,” which is basically a big warehouse. Then they sell this to a developer or directly to a hyperscaler to do the rest of the build-out.
But we’re also seeing more real estate companies proceeding to a “powered shell.” This means supplying the mechanical, electrical, and cooling technology. It also includes running fiber through the building and could involve building the aisles and racks.
Hyperscalers build about 50 to 70 percent of the capacity they want themselves. But they don’t always have the time, land, or capital to build everything on their own. So they buy or lease the rest from third parties, creating opportunity for real estate companies.
Katy McLaughlin: Now for the third opportunity: fully building out new data centers. What’s changing there?
Pankaj Sachdeva: We call this the “co-location business.” In this case, real estate companies buy or lease land, build a powered shell, and then complete it to their tenants’ specifications. These specs are often demanding and customized and typically require collaboration with specialized supply-chain-procurement, engineering, and architectural companies. It’s not easy for players that don’t have experience to be able to meet these demands in capital-efficient ways.
We have recently seen more real estate companies partner with data center developers and operators in order to access the technical know-how that allows them to enter the co-location business. Some of these partnerships are more site-level or project-based, while others are at a company/enterprise level.
Five years ago, the market wasn’t as supply constrained as it is today. Hyperscalers and other customers had more leverage, and the average lease term was about five to seven years. Today’s supply-constrained market is starting to see leases of ten-plus years, though the average is around seven years.
Katy McLaughlin: What’s the average useful life of a data center?
Pankaj Sachdeva: The average useful life of data center land tends to be somewhere between 25 and 30 years. The cooling equipment is in the ten- to 15-year range. And some of the electrical work, such as with a transformer, an uninterruptible power supply, and a power distribution unit, is around eight to 20 years.
Katy McLaughlin: That makes for a complex calculation, especially with leases ranging between five and ten years. Maybe a better way to put it is, how big a risk is obsolescence, and what can investors do about it?
Pankaj Sachdeva: Obsolescence is a real risk, but it’s important to understand where it lies. Real estate companies primarily take the risk on the land parcel and the architecture of the data center itself. But they typically don’t take the risk of the IT hardware and software, such as servers and storage. That’s typically within the province of the tenant.
The risk for real estate companies is usually whether there’s good, fiber-based network infrastructure coming to a facility and whether cooling systems will stand the test of time, especially in light of global warming. Another risk is whether a facility can be upgraded, because it may need more power as computers become more powerful.
Katy McLaughlin: What about risk from competition—the shiny new data center down the street?
Pankaj Sachdeva: Moving is incredibly hard for tenants. If the data center is 80 or 90 percent utilized, tenants usually don’t take the business risk of moving all the servers and IT equipment down the road for more modern equipment or slightly lower lease rates. Renewal rates for highly utilized data centers tend to be 90 percent or higher.
Katy McLaughlin: Finally, how would you characterize the long-term opportunity for investors?
Pankaj Sachdeva: Data centers are going through a growth S-curve that’s both structural and secular. That means that the market is quite supply constrained today and may also be for the long term, indicating growth in the asset class.
Compared with other real estate asset classes, data centers historically have had higher yield. Due to supply-side constraints, we don’t anticipate that real estate players or data center developers will face major yield compressions over the long term. Real estate investors are seeking exposure to data centers to get higher growth and sustained levels of higher yield.
