Solving AI’s Power Bottleneck

Our analysts Andrew Percoco and David Arcaro break down how AI’s surging energy demand could reshape the way we produce, store and use power. Read more insights from Morgan Stanley.----- Transcript ----- Andrew Percoco: Welcome to Thoughts on the Market. I'm Andrew Percoco, Morgan Stanley's U.S. Autos and Shared Mobility analyst. David Arcaro: And I'm Dave Arcaro, Morgan Stanley's U.S. Power, Utilities, and Clean Tech analyst. Andrew Percoco: Today, why the AI power crunch could transform how we generate, store, and consume electricity. It's Tuesday, November 3rd, at 10am in New York. Andrew Percoco: So Dave, the whole AI power shortfall and, you know, how data centers [are] creating a power crunch is not necessarily a new topic. I think we've had many conversations, many people listening have probably heard this is, kind of, being a bottleneck for AI deployment. But I'm curious, how have your views around the role of distributed power changed in terms of its role in solving some of the shortfalls in the market? I think there's a lot of debate in the industry around whether or not these solutions are "bridge." Or whether they're structural and they can be truly long-term sources of supply to the market. So maybe just help us think about, you know, how our views have really changed, as it relates to distributed power specifically. David Arcaro: One of the things that we have seen is that the shortfall has grown in our estimates over the last several iterations. We're now looking at upwards of an 80 gigawatt shortfall by 2029. And how do we solve that? Well, I'd say we've leaned more into on-site power, bringing your own power. What is that? You know, if you are still waiting for years for a grid connection from the utility, what can you do as an alternative? You can bring your own power. You could bring small scale power equipment to get the high reliability in a much faster timeframe than you would waiting for the grid in comparison. And so we're seeing that increasing extremely rapidly in popularity among hyperscalers and data center developers. And we're seeing them much more, being willing to consider it on a long-term basis. What used to be, "Hey, I need to cover a few years just to bridge until I have line of sight to getting a connection to my utility…" Now we're seeing actually 15, 18, 20-year contracts for these, you know, what used to be considered just smaller scale, less attractive solutions. They're now long-term, with what looks like much more permanent considerations for highly reliable data center AI power. Andrew Percoco: I think there's a preconceived notion in the market that the grid's going to catch up; and that we won't need these solutions longer term. Or that we're building too much of it and there's too much manufacturing capacity coming to the market. So maybe we can segment those into two separate questions. Like one, what is actually the bottleneck? I think everyone thinks of it as a power generation bottleneck. But I think you've done some interesting work around how there's also transmission and distribution bottlenecks that don't get solved as quickly as the power generation. So maybe you can talk to that. But then secondly, when we think about the long-term, you know, case for distributed power. Are there certain types of technologies that maybe screen better or more favorably? Or certain characteristics of certain technologies that screen more favorably when you think about this being a structural source of power rather than maybe a short-term fix? David Arcaro: Yeah. The bottleneck is very often in the utility space. Utilities are seeing such an unprecedented demand for electricity from the data center industry. They're scrambling to keep up with it. They're not used to growing this fast, and it's often their electric transmission network that is falling behind. It's the long distance power lines and substation equipment that the utilities are really trying to react [to] – and to grow alongside the data center industry. But we expect them to fall short; and pretty meaningfully short. You know, this is large scale infrastructure that takes years to develop and build. So, then when it comes to onsite power, as you're growing alongside that constraint or trying to avoid that constraint. We're seeing that data centers, they're very much prioritizing speed. The quicker they can get the power the better. But you know, more and more we're seeing other factors come