How Will the GenAI Revolution Be Powered?

Our Global Head of Sustainability Research and U.S. Utilities Analyst discuss the rapidly growing power needs of the GenAI enablers and how to meet them. ----- Transcript ----- Stephen Byrd: Welcome to Thoughts on the Market. I'm Stephen Byrd, Morgan Stanley's Global Head of Sustainability Research. David Arcaro: And I'm Dave Arcaro, Head of the US Power and Utilities team. Stephen Byrd: And on this episode of the podcast, we'll discuss just what it would take to power the Gen AI revolution. It's Tuesday, April 16th at 10am in New York. Last summer, scientists used GenAI to find a new antibiotic for a nasty superbug. It took the AI system all of an hour and a half to analyze about 7,000 chemical compounds; something that human scientists would have toiled over for months, if not years. It's clear that GenAI can open up breathtaking possibilities, but you have to stop and think. What kind of compute power is needed for all of this? A few weeks ago, our colleague Emmet Kelly, who covers European Telecom, discussed the exponential growth of European data centers on this podcast. And today, Dave and I want to continue the conversation about this critical moment of powering the GenAI revolution. So, Dave, what is your current assessment of the global power demand from data centers? David Arcaro: Yeah, Stephen, we're expecting rapid growth in the power demand coming from data centers across the world. We're currently estimating data centers consume about one and a half per cent of global electricity today. We're expecting that to grow to almost four percent in 2027. And in the US, data centers represent roughly three percent of total electricity consumption now, and we expect that to escalate to eight per cent of the total US by 2027. And there will be even more dramatic impacts at the local and regional level. The data center landscape tends to be highly concentrated, and the next wave of GenAI data centers is likely to be much larger than the previous generation. So, the impact on specific regions will be magnified. To give an example, in Georgia, the utility there has previously forecasted just half a percent of annual growth in electricity use but is now calling for nine per cent of annual growth in electricity consumption, and that's largely driven by data centers. It's a dynamic that we haven't seen in decades in the utility space. Stephen Byrd: You know, what I find interesting about what you just said, Dave, is -- it is impressive to see growth go from one and a half to four per cent, but it's really these local dynamics where what we're seeing is just much more concentrated, and that's where we start to see the real issues with the infrastructure growing quickly enough. So, it's becoming obvious that the existing power grid infrastructure is not meeting the growth and capacity needs of data centers. And that's something that you refer to as the tortoise and the hare. How big of a mismatch are we exactly talking about here, Dave? David Arcaro: It's definitely a big mismatch. To your point before, the US electricity growth across the country has been flattish over the last 10 years. So, this is a step change in expectations now, from the impact from Gen AI going forward. And we're looking at over 100 per cent annual growth in the power consumed by data centers now in the US over the next four years. And for comparison, the US utility industry is growing at about 8 per cent a year. These data centers that are coming are huge. They can be 10 to 50 times as big as the last generation of data centers in terms of their power consumption. And this means it takes time to connect to the electric grid and get power. 12 to 18 months in the best case, three to five years plus in other locations, often because they might need to wait for the electric utility grid to catch up, waiting for grid upgrades and assessments and new power plants to get built. Stephen Byrd: Well, I think those delays are going to be fairly problematic for the fast-moving GenAI sector. So essentially there's a lot of pressure on data center developers to secure a power source as quickly as possible. And in our note, we described the mathematics around that. The time value to get these data centers online is absolutely enormous. But you've just described the power grid infrastructure as a tortoise. So, are there any other alternatives? How about nuclear power plants in this context? David Arcaro: There's a lot of urgency, as you can tell from the data center companies, to get online as fast as possible. It's a fast-moving market, very competitive, they need the power, they need to run these GenAI models as soon as possible. And the utility industry is not used to responding to demand that's coming this quickly. It's a slower moving industry. There's policies and processes and regulation that all utilities have to get through. They're not prepared strategically to move as quickly as the data center industry is moving. So, data center developers are getting creative and they're looking at all options to get power. And one that has an appealing value proposition is nuclear plants. By placing a data center at an existing nuclear plant, this can avoid the need to go through that lengthy electric grid connection process, providing a much faster timeline to get the data center powered up. And that has big benefits for the data center companies, as you can imagine. Nuclear plants also have other advantages. They have land available on site. They have water for cooling, security. It's 24x7 clean power with no emissions, and it's already up and running, so you don't have to go and build much. Over time, we do expect renewables to play a major role as well in powering data centers along with traditional power from the electric grid and even new gas plants, but the benefits of coming online quickly in this market we think, give nuclear an edge. So, Stephen, as much as I can talk about the massive power needs of Gen AI, we can't ignore the issue of sustainability. So, what have you been thinking about when it comes to assessing the potential carbon footprint of powering data centers? What concerns are you seeing? Stephen Byrd: You know, Dave, this field is evolving so quickly that we've had to evolve our assessment of the carbon footprint of GenAI quite quickly as well. You know, traditionally what we would have seen is a data center gets connected to the grid. And then that data center developer would often sign a power contract with a renewable developer. And that results in a very low carbon footprint, if zero in many cases. But going forward, we do see the potential for increased natural gas usage in power plants, higher than we had originally forecast. And that's driven really by two dynamics. The first is the increased potential to site data centers directly at nuclear power plants, which you described, and there are a lot of benefits to doing so. In effect, what's going to happen then is, those data centers will siphon away that nuclear power, so less nuclear power goes to the grid. Something has to make up that deficit. That something is often going to be natural gas fired power plants. The second dynamic that we could see happening is an increased potential for just onsite natural gas fire power generation at the data centers that could provide shorter time to power, and also provide quite good power reliability. Now, when we sum these up and we look at the projected carbon footprint of data centers going forward, we could see an additional 70 million tons a year. We're about half a per cent of 2022 global CO2 emissions for data centers. That is quite a bit higher than we had previously forecast. Now that said, a wild card would be the hyperscalers and others who may decide to consciously offset this by signing additional power contracts with new renewables that could reduce this quite a bit. So, it's very much in flux right now. We frankly don't know what the carbon footprint is really going to look like. David Arcaro: You know, there's so much urgency to bring data centers online quickly that in the past many of these big hyperscalers especially have had quite ambitious sustainability goals and decarbonization goals. I'd say it's an open question on our end as to how flexible they might get in the near term or how strictly they do apply those decarbonization … Stephen Byrd: Exactly… David Arcaro: … targets going forward as they, y’know, also try to compete in an urgent grab for power in the near term. Stephen Byrd: That's exactly right. That's… You laid that tension out quite well. David Arcaro: And finally, from your global perspective, what regions are best positioned to keep pace with the power needs of Gen AI? Stephen Byrd: You know, Dave, I am thinking a lot about what you said a minute ago, about the size of these datacentres moving from, you know, quite small – often we would see datacentres at just 10 or 15 megawatts. Now the new designs are often above 100 megawatts. And now we're starting to hear and see some signs of truly mega data centers, essentially massive supercomputers that could be a thousand megawatts, a couple of thousand megawatts, and could cost tens of billions of dollars to build. So, when we think about that dynamic, that's a lot of power for any one location. So, to go back to your question, we think about the locations. It's very local specific. The dynamics all have to line up correctly, for this to work. So, we see pockets of opportunity around the world. Examples would be Pennsylvania, Texas, Illinois, Malaysia, Portugal -- these are locations for a variety of reasons where policy support is there, the infrastructure growth potential is there, and for a number of reasons, just it's the right confluence of dynamics. Most of the world doesn't have that confluence, so it's going to be very specific. And I think we're also setting up for a lot of concentration in those locations where all these dynamics line up. David Arcaro: You know, historically, the data center industry in the US has been highly concentrated, like you say, in Northern Virginia, in Northern California, they've been data center hubs, but we're running into infrastructure constraints there, we've got to look elsewhere. And some of these factors, geographically, are going to be extremely important. Where is their local support? And one of the dynamics we think could happen is that as you build more data centers that are very power hungry, that could push up the price of power. And what kind of local pushback might you get in that situation? What's the local desire to have a data center from an employment perspective and property tax and local benefit perspective? And how does the cost benefit weigh against the potential for higher power prices in those regions? Stephen Byrd: That's a great point. I mean, in places like Northern Virginia, to your point about property taxes, the value of all this data center equipment is in the tens of billions, which does help local tax revenue quite a bit. That said, there are offsetting impacts such as higher power prices. And this is why I think your original point about the local dynamics mattering so much is so critical because you really do need to see political support, policy support. You need to see the infrastructure that's available. So that's a fairly precious lineup, a fairly rare lineup of all the attributes you need to see to support new giant data center development. David Arcaro: Definitely a delicate balance that the industry needs to tread here as these huge data centers start to come online. Stephen Byrd: Well, I think a delicate balance is a good place to end this discussion. Dave, thanks so much for taking the time to talk. David Arcaro: Great to speak with you, Stephen. And thanks for listening. If you enjoy Thoughts on the Market, please leave us a review wherever you listen to the show, and share the podcast with a friend or colleague today.