The AI Boom Is Testing America’s Power Grid

Dan Loney: There are so many questions around artificial intelligence these days. Obviously, the build-out is one of them, but there is also an environmental component to this as well. Recently, the Wharton School hosted a webinar to discuss many of these topics, and one of the experts involved in that webinar joins us right now — Wharton professor Serguei Netessine. Serguei, great to talk to you again.

Serguei Netessine: Thank you for having me again, Dan.

Loney: Let's start with the importance of talking about all of these different topics right now.

Netessine: Well, I think what's kind of unprecedented is the scale of build-out of AI data centers and hyperscalers that we are experiencing. Usually when there is a revolution in software, and usually we perceive AI as kind of a revolution in software, it's kind of a digital world. It does not necessarily bring a revolution of the physical world. But what AI does is essentially converting electrical power into intelligence, and the demand for electrical power is really unprecedented, which is kind of ironic because our brain is very efficient in converting power into intelligence.

You and I can have a cup of coffee and then solve a lot of problems, right? Unfortunately, AI and large language models are not as efficient. They demand a lot of power. And so in the United States, what we are seeing [is] there are about 1,800 active U.S. data centers, and they are spread across about 175 utilities in 43 states. All of these data centers need power. They all connect to power. And a lot of these AI data centers are really ginormous.

An average data center is actually not that big, but the ones that we see that are going to be built over the next five or six or 10 years are going to be much, much larger in size. We're [soon going to be] in a situation like in North Virginia right now, where demand from AI data centers accounts for something like 25% of [the] electrical grid.

Loney: This is already at a time over the last few years where the topic of the electrical grid here in the United States has already been brought up and what needs to be done to modernize it just for normal everyday use, whether it be an individual's home or a company's office. Now you're adding this on top of it and that just multiplies the issue, doesn't it?

Netessine: Exactly. The reason is in the time period from 2005 to, say, 2019, we've been growing in electricity demand by 0.1% annually. So essentially no growth. Essentially, 15 years it's been flat. And so you optimize [the] electrical grid for this basically no growth in electricity.

And then suddenly over the last five years, we are growing at almost 2% per year. AI data centers are responsible for about a half of that. But actually it's growing and growing and growing. It's going to grow even faster. And the AI data centers will be responsible for most of this growth. We basically increased the rate of growth in electricity consumption by about a factor of 20.

Loney: What do we know, at least so far, in terms of the cost of electricity and what's that doing in general to the United States? How are consumers being impacted by this?

Netessine: This [has become] a politically charged question. In certain states, we see prohibitions of AI data centers. We see a lot of politicians coming out and saying, “Hey, we cannot have data centers because they are responsible for growth in electricity prices.” This is not a simple question to answer because there are lots of data centers that connect to all kinds of different utilities in a variety of states. Actually, most of them right now are pretty concentrated. There are like six states where most of the deployment is happening. And figuring out what happens with prices is not easy.

But I actually do have a current research paper, which is about to be finished, where we estimate that in the last five years — for example, from 2020 to 2025 — electricity prices increased by about 30 [to] 35%. So it's real. Electricity prices are increasing. That part is absolutely real. But this increase reflects many things. For example, fuel and capacity costs, aging infrastructure, transmission, distribution investment, extreme weather (which is becoming more and more terrible), and the rising cost of capital. To build all of this, you have to borrow money and capital costs more and more.

The data centers explain a very small proportion of it. They're definitely not responsible for all the increases that we are seeing. In some of the worst scenarios, AI data centers are maybe responsible for 10% of this increase. But in most cases, they're not even responsible for any increase at all.

Loney: But one of the things you kind of alluded to a little bit in that last response is the fact that you also have to look at the environmental component of this as well, and what this is going to do to so many different regions of the country where there may already be an impact from electrical infrastructure already.

Netessine: Yes, absolutely. One thing to understand is that AI data centers do two things to the grid. One, they, of course, bring more demand. This demand is going to require some infrastructure investment and all kinds of upgrades and new transmission lines and so on. But at the same time, you are now spreading this expenditure over a bigger base. So clearly, in some cases, tariffs actually might go down because you're just dividing the cost of electricity over a much, much bigger paying base.

Moreover, there are different kinds of tariffs most of the time. There are tariffs for industrial customers and then there are residential tariffs. For example, what we saw very specifically in Oregon, regulators increased Portland General Electric's data center customers’ rates by about 30%. But at the same time, they decreased rates for residential customers. So you can actually regulate it in such a way that industrial customers like AI data centers can absorb much of the costs and that does not necessarily have to affect your residential customers.

Loney: Are there any emerging technologies in the mix here that will potentially help this process out? Not only make the electrical use more efficient, but also not have as much a negative impact as maybe a lot of people believe could occur?

Netessine: For sure. One number to kind of know is that if you take all the capacities that we currently have on [the] U.S. grid and you take all the demand that we have, our grid is only about 55 percent utilized. So we still have almost half of the supply of electricity, which is not used. A lot of technologies are needed to make use of all of these existing capacities that we already have. And the reason why we have so much capacity sitting unused is because it's all about peak demand. In certain times of the day, the sun is not shining, for example, and solar power is not available. The wind is not blowing, so your wind power is not available. But at the same time, consumption is very, very high because, for example, you have a heat wave and you end up using a lot of air conditioners.

That's why you have to build capacity for the peak. If you are able to deal with the peak, then you might actually be OK for many years to come. And how do you deal with the peak? Well, this is where you need electricity storage. Batteries of various sorts — and not just electrical batteries, but thermal storage and maybe pumped hydro and all kinds of other technologies that exist out there. But also, you have to start thinking about some more innovative models, like, for example, moving storage around — charging [a] battery in one place and maybe discharging this battery in another place — and thinking through how AI data centers can become more responsive to the grid. For example, do you have to do this [computation] on your iPhone right now or can you postpone it until tomorrow when you're going to have more electricity? If AI companies learn how to do that, then they can actually help the grid to be optimized.

Loney: Where are we right now from a policy perspective? I realize we're probably not as far along as we're going to be five years, 10 years down the line. But how much are things being looked at and being addressed in terms of what type of policy we need to have as AI continues to develop?

Netessine: We are definitely very far from where we need to be. And regulatory constraints are really, really bad right now. For example, you can build [an] AI data center relatively quickly, but building an extra power plant and getting it on the grid and getting it to work will take you years and years. Part of it is just constraints and supply. To order a simple gas turbine that can convert gas into electricity will now take you two or three years just because the supply is very constrained and everyone who could order it placed orders for the next two or three years.

The same goes for all kinds of equipment like transformers and transmission lines and so on. But one of the biggest issues is honestly the interconnection queue. In order to get into the grid, you need to submit a request to the regulator to get you connected. In some areas of this country, that alone will take you 10 years. That's why we are kind of in a critical point where something has got to happen because many AI [and high tech] companies, are beginning to say, “Hey, I don't even want to be connected to the grid. It's too complicated [and takes] too long. I'm just going to build my own supply, my own electrical power.” Maybe it's going to be a nuclear station or maybe it's a gas power plant. And if it's a gas power plant, then we have a problem because that's very, very dirty. This is what some of the players in the market have already started doing.

Loney: And then to get those approvals, you're talking about both at the state and the federal level […] and that's all kinds of issues in itself.

Netessine: Yes, absolutely. We are talking about nuclear, but nuclear power — you need so many approvals from various agencies. That becomes even more complicated, in addition to all these interconnection queues and all kinds of other interconnection issues that you are facing. So you definitely need to need to make our utilities more agile and much more innovative in how they manage these new loads, manage new demand, manage new supply, how they think about infrastructure investment and how they start thinking about more innovative technologies that are already existing in some parts of the world, but unfortunately not in the United States.

Loney: We talked a lot about how this is going to impact major businesses. But is there any discussion going on about how this is going to impact businesses in general in so many different communities with the use of the grid, with how companies in towns where these AI centers are going to be located are going to be impacted? I think we focus so much on the build-out and the grid part and the electricity part that there are still many other components that we're going to see play out that probably still have to be looked at.

Netessine: Absolutely. One of the things that we discovered [is] what happens when you add an AI data center to the grid. In some cases, prices increase. In some cases, they don't. On average, the effect is pretty minimal. But there is one big defining feature between increasing prices and not increasing prices. When you have a utility that has experience dealing with large industrial customers in the past, they tend to manage AI data centers well and prices don't increase. When you have utilities that [do] not have this kind of experience, we see an increase in residential tariffs.

That tells you that a little bit of experience in dealing with these kinds of big industrial loads is pretty much what you need to be able to move forward. What we are seeing, of course, is a big concentration of these AI data centers in a few states. And you can kind of see why. If utilities in those states already have experience onboarding AI data centers, it just becomes easier for them.

Loney: What do you hope comes from this webinar? What do you hope is the next step in that process for you and your colleagues?

Netessine: I think we need a very concentrated effort in this country to match supply and demand for electric power. We see already in China, in a single year, they built more capacity for producing electrical power than the entire United States has. This is really mind-blowing. And clearly, this is one area where you do need some kind of government regulation. It's a pretty critical infrastructure.

I see several scenarios that can emerge. What we need to push towards is this kind of coordinated growth, coordinated increase in demand, coordinated increase in supply. And at the same time, we don't want to overregulate it because we already see what happens in this case. It takes you a decade to bring a new power plant or a new AI data center to the grid. So if we do nothing, I think we're going to end up in a very, very bad equilibrium where it's going to be much harder to get out of. And we might eventually kind of lose this AI race, not because we cannot develop good models, but because we don't have good infrastructure in the country to support those models.

Loney: I'll finish up on this question then. Because of that, is there the potential concern of an AI bubble if we don't get this right and do it properly?

Netessine: I am kind of a technological optimist. I think you can always find a way to solve those problems. For example, there are countries with cheap electrical energy. Bitcoin miners discovered it some time ago, and that's where they would go. Maybe we will create a new kind of import, which will be an import of compute or training of the model. For example, you go to a country, you train your models there on cheap electricity, and then you use those models in other countries. All kinds of new models of operation for AI companies can evolve. But of course, this means less money invested locally in the United States, so I do hope that we will pull together as a country and figure out how to solve this big energy dilemma for AI.