Investing in AI Infrastructure (Part 1: The Big Picture)
There is a lot to look at, so I'll narrow it down.
The “big five” hyperscalers — Amazon, Alphabet, Microsoft, Meta, and Oracle — are expected to spend more than $600 billion on capital expenditures this year, 36% more than last year (and about 3 times more than what was spent in 2022!).
On the surface, the demand is real: the backlog for these 5 companies has reached more than $1.5 trillion.1 They’re spending for demand that already exists!
But these five companies, while they expect a return on their investments, are spending while they build. It’s where this money goes that makes the story more interesting. Most of this capex will be supporting AI infrastructure, which means everything from microchips (like NVIDIA’s processors), to cables and wires, to computers and data centers, and anything that supports or supplies any of those inputs. There is an overwhelming range of places to look, but what caught my interest was data center construction and maintenance.2 It has grown tremendously since 2023.3
All of these new data centers require specialized electrical and HVAC systems, which means that hyperscaler investment and backlog flows into construction revenue and backlog. I would call it a “forgotten” part of the AI investment boom, but everyone already knows about it by now. That leaves some big questions that need to be answered:
How long will the AI investment boom last and how big will it be?
How likely are the backlog numbers to be realized into revenue?
What economic limitations might slow down data center growth?
Where is the next AI investment bottleneck?
What happens to these companies when the AI boom fades?
The first thing is that I see a direct connection to the infrastructure buildout that happened in the 1990s. At the time, everyone knew that the Internet would change everything, and part of that meant building new lines for Internet connections. By the end of the decade, it became clear that network capacity was built far beyond what the world needed at the time, and many of the companies providing that service went out of business. Internet connections became more of a commodity and more affordable.
I think something similar will happen with AI, but not quite the same. In the 1990s, it was all about how many potential eyeballs a business could get on their website. But the AI investments are following AI demand that already exists, and the big tech companies clearly cannot keep up with it right now. The hyperscalers’ $1.5 trillion backlog might be a fantasy, but spending estimates for the past few years have generally been too low — and they weren’t even cautious estimates!
Going further down the line, the hard part for the HVAC/electrical/construction companies involved with data centers is how much of their backlog will be converted into revenue. A significant part of their data center backlog comes from hyperscaler AI backlog and AI spending, so the hyperscalers would have to see AI demand drop off (which seems unlikely right now) or be unable to pay for their data centers (which also seems unlikely right now). The real challenge is the potential for a shortage of workers or materials (slowing down construction) or limitations on how many data centers the electrical grid can handle (slowing down construction while everyone waits for it to be upgraded).
That last one — expansion of the electrical grid — is where I see the next big AI bottleneck. It might be worth looking at — the cost of energy is already rising in many places with new data centers — but that’s a topic for another time, and one that other people have already covered.4 At the moment, the U.S. demand for data centers is projected to grow by about 17% annually over the next 4 years. Over that same time period, worldwide spending on new data centers is projected to reach almost $7 trillion.
Even if there are never enough data centers to do everything that AI-focused companies want to do, the growth will eventually slow down. This is the point where an AI themed “trade” would transform into a long-term investment. The data center infrastructure industry is growing fast enough that every company can catch a bid, so competition doesn’t matter as much right now, and even the worst companies can look pretty good during a stretch of good years. But the competition will become stronger — and margins lower — when the boom begins to fade. This is when other parts of the company become more important, such as long-term maintenance contracts, financial strength, and involvement in other sectors.
But there is one more question: who is actually building these data centers?
Since I have readers who might not be familiar with the term, Investopedia has a good explanation of backlog.
Generative Value has an excellent introduction to data centers from 2024 (and many others about the AI investment boom). I am clearly “late to the party” on this one, so I will not get as technical.
Apricitas Economics provides a good look how AI investments have affected different parts of the economy so far.
This is not a new idea. Market Sentiment wrote a great blog post describing what is happening with the energy requirements behind AI.