What investors should know about AI & data centers

Ever since ChatGPT burst onto the scene nearly two years ago, the whole world has taken notice of the explosive growth and revolutionary power of AI. As of March 2024, ChatGPT has over 180.5 million monthly users. Meanwhile, based on U.S. Census Data, spending on data center construction in the United States is on pace to surpass $27 billion annually. That’s more than a $10 billion jump compared to last year and equals one-third of the amount spent on all U.S. office construction.

However, what’s equally staggering is the massive amount of energy that AI consumes and the strains it poses to power grids around the world. In today’s Playbook, we are going to do a deep dive into AI energy consumption and its effect on data centers. We’ll also explain why the demand for new data centers is growing rapidly around the world and how real estate investors can get into this sector using REITs.

Data centers consume massive amounts of energy

The digital world that is so integral to our daily lives depends on data centers. They are vast warehouse-like structures packed with racks of servers, routers, storage devices, and other equipment. Nearly everything we do with technology—whether it's internet searches, streaming, online shopping, or gaming—depends on these buildings.

Even before the rise of AI, these data centers required enormous amounts of energy. The International Energy Agency (IEA) estimates that data centers account for 1%-1.5% of total global electricity use. To put that in perspective, data centers consume about two to three times more energy than the six states of New England combined (they average 120-125 terawatt hours (TWh) of energy each year). For reference, a terawatt hour equals 1 trillion watt hours (Wh).

AI will cause data centers to consume even more energy

AI is going to increase both the amount and the rate at which data centers consume energy. AI models use power at a higher rate than traditional data center activities. For example, a ChatGPT query uses 10 times more energy than a normal Google search. The IEA has already forecast that global data center electricity demand will more than double from 2022 to 2026.

Two AI-related activities put enormous energy demands on data centers: AI training and systems cooling. First, let’s address AI training. The amount of energy needed for this process depends on the AI model being trained. A 2022 study revealed that training energy requirements for models varied significantly, ranging from 2-3 kilowatt hours (kWh) for smaller natural language processing and computer vision models to a staggering 103,500 kWh for a 6 billion-parameter transformer. To put that in perspective, GPT-3, the older version of ChatGPT which came out in 2022, used almost 1,300 MWh (one MWh equals 1,000 kWh) of electricity to power its 175 billion-parameter AI model. As Joseph Polidoro notes, that’s "roughly equal to the annual power consumption of 130 homes in the U.S."

Now let’s address AI cooling. AI computing relies on high-performance microprocessing chips that generate more heat and consume more energy than standard chips. In traditional data centers, HVAC systems (Heating, Ventilation, and Air Conditioning) account for roughly 25% to 40% of the total energy consumption, while other systems and equipment consume around 50%. According to researchers at U.C. Riverside, the rise in AI use will cause data centers to consume more than 1 trillion gallons of fresh water by 2027.

The number of AI data centers is growing worldwide

Currently, there are 9,000 to 11,000 data centers worldwide across every continent except Antarctica, and their numbers are growing. CBRE, the global commercial real estate services and investment firm, reports that all major global regions saw annual gains in data center inventory in Q1 2024. Year-over-year, inventory grew by 24.4% in North America, 22% in Asia, 20% in Europe, and 15% in Latin America.

In the U.S., top data center hubs include Northern Virginia, Silicon Valley, Dallas/Fort Worth, and Chicago.

Worldwide, here are the top 5 data center markets, ranked in order of megawatt usage:

1. Northern Virginia
2. London
3. Tokyo
4. Frankfurt
5. Sydney

Requirements for optimal AI performance

Data centers require stable power grids, dense fiber networks, and proximity to internet “backbone points.” Beyond these technical needs, location also hinges on government and public support. Countries like the United States, with efficient permitting processes and data-friendly regulations, are prime candidates for AI facilities. Even typically cautious countries like France are actively pursuing AI data centers.

However, the London School of Economics notes the rise of "data center activism," with global protests aiming to halt new construction. Concerns driving this activism include the environmental impact on climate change, land allocation, the allegedly minimal community benefits of data centers, and the high water usage needed for cooling.

Hyperscale data centers are the future

Now that tech firms are entirely convinced that AI is the future of computing, they are in the business of constructing hyperscale data centers. All the major tech firms—Google, Meta, Microsoft, Amazon, Oracle, etc.—are spending massive amounts on hyperscaler construction.

Hyperscalers are much bigger than standard data centers. They take up 1 million to 2 million square feet compared to 100,000 square feet for the average cloud data center. In addition to housing more equipment, hyperscalers are more energy-efficient by design. 40% of all hyperscalers in the pipeline will be built in the United States alone. Tech industry market research firm 650 Group, "calculated that the top 5 US hyperscaler developers spent $105 billion in 2023 and forecast the figure to rise to $187 billion in 2028k," according to Bloomberg.

In their most recent earnings reports, Microsoft, Meta, and Amazon indicated that they’re increasing spending on data centers to augment AI development. Amazon said it spent $30.5 billion in the first half of 2024 and "pledged to exceed that figure over the next six months". Almost all of it is going to data centers.

Interestingly, Microsoft is considering using nuclear power to run its AI data centers, possibly using an array of small modular reactors. The company has also agreed to purchase power from Helion, which is trying to build a fusion power plant.

Use REITs to get into data center investing

Given the huge costs and technical expertise involved, there is a high barrier to entry to get into data center investing. The easiest way for a retail investor to access the benefits of this growing asset class is to use REITs like Digital Realty and Equinix. But what exactly are REITs and why should any real estate investor care about them? Here’s a brief primer on REITs and why they’re attractive investment vehicles.

What are REITs? The word “REIT” is an acronym for Real Estate Investment Trust (REIT). The firms that legally organize themselves as REITs own, operate, or finance income-producing real estate assets. These assets can be any kind of real estate property type—industrial buildings, office parks, apartments, hotels, data centers, and even cell towers. One of the best-known REITs in the world is Blackstone.

What makes REITs attractive to investors? They allow anyone to invest in real estate without incurring the costs that are typically associated with buying, maintaining, leasing, and selling a real estate property. Just as you can make a profit by buying Apple stock without actually creating, say, an iPhone, you can profit from buying stock in a data center REIT such as Digital Realty or Iron Mountain without actually buying and running a data center.

How do REITs work and make money? Conceptually, it’s quite uncomplicated and easy to understand. REITs either buy or finance various real estate property types, lease them out, and then collect rent on them. Just as small-time investors try to minimize their operating costs when running a quadplex, REITs use various simple and complicated techniques (e.g., in their leasing or property management strategies) to minimize their operating costs and maximize their net operating income. By law, REITs must pay out at least 90% of their taxable income to shareholders. In turn, shareholders pay the income taxes on those dividends.

Are all REITS listed on national stock exchanges like the NYSE? No, they are not. While most REITs are traded on major stock exchanges, there are also public non-listed REITs and private REITs. Public non-listed REITs (PNLRs) are registered with the SEC but do not trade on national stock exchanges. By contrast, private REITs are real estate funds that are exempt from SEC registration and whose shares do not trade on exchanges like the NYSE. Private REITs generally can be sold only to institutional investors. To dive deeper into either of these types of REITs, see here and here.

Should you invest in REITs and how have they performed historically? At the Playbook, we don’t give advice on whether you should invest in this or that asset or investment vehicle. However, we can report that many investors find the dividend income generated by REITs and the long-term capital appreciation gained through them to be quite attractive. According to NAREIT, 170 million Americans are invested in REITs through their 401(k), IRAs, pension plans, and other investment funds. To see how REITs have performed compared to other investment vehicles, see here and here.