Loading summary
Transcript40 lines
- [00:06]
A
Welcome to Goldman Sachs Exchanges. I'm Alison Nathan and I'm here with George Lee, co head of the Goldman Sachs Global Institute. Together we're co hosting a series of episodes exploring the rise of AI and everything it could mean for companies, investors and economies. George, great to see you again.
- [00:21]
B
Good to be back. We had a little bit of a hiatus here while I was recovering from knee surgery, but fortunately not much happened in the world of AI over the past few months. I'm kidding, of course. It has been a cavalcade of extraordinary events. So. And be really fun to talk to our guest about what's shaping the movement right now.
- [00:38]
A
Yes, well, great to have you back, George. And today we are going to take a closer look at an area that we've touched on before in our conversations, you and I, but really deserve a closer look and that is the resources that the AI buildout could require. We are joined here by Brian Singer, head of GSUSTAIN and Goldman Sachs Research.
- [00:55]
C
Thank you for having me, Alison. Thank you, George.
- [00:57]
B
Glad to have you here.
- [00:58]
A
Brian.
- [00:58]
C
Great to be with you all.
- [00:59]
A
Brian, just to kick off the conversation, we've been talking about the tremendous amount of power that data centers will require for a while now. But the amount of spending on these data centers have just continued to grow and grow and grow. I think that was a key feature of the last few months when we think about the AI space. So first put some numbers on those developments and what it can mean for power demand.
- [01:23]
C
Sure. I mean, we have seen the hyperscalers capital budgets and R and D budgets in 2026 and 2027 combined increase by more than $300 billion. And I'm sure we can get into the debate whether there's going to be a second derivative inflection coming for slower growth going forward. But you're growing off of a upwardly Revised budget of $300 billion plus, and that is going to trickle down into power and a whole host of other uses. And so we've been focused on where will that money go and what are the investment opportunities for the enablers.
- [01:54]
B
And some of your work, Brian, I noticed notwithstanding all of that spend and there's obviously a lag effect, but you calculate that we remain in a deeply supply, demand, imbalanced state.
- [02:05]
C
Right.
- [02:05]
B
Maybe you could talk a little bit about where we stand there.
- [02:07]
C
Yeah, that's right. I mean, Alison, you and I go back to the shale revolution and arguably before as well, unfortunately or fortunately. And that was a very similar time period where we saw a massive increase in capital spending followed by a cyclical downturn. And what drove the shift then from the appraisal hopes and dreams phase into the execution phase were three factors. One, the product became oversupplied. Two, your innovators saw their return on capital erode. And third, the innovators ran out of financial flexibility. You're asking about really the first of those here, which is are we actually in oversupply? And, and it doesn't look like we're at that point where the Goldman Sachs, the big companies of the world know what they need and are going to get from a token and compute perspective. But frankly you're on the front lines of that. So I actually think you're probably a better arbiter of whether we are actually moving into a period where we're defining the demand for AI or whether as we see continued productivity gains, whether that's just going to be met with more demand for compute, more demand for tokens for the same capital budgets.
- [03:16]
B
Yep. Well, my perspective is that demand growth will continue to surprise us in contributing to that, and you observed it in your research report, is that there's more demand from enterprises for fundamental sources of token flow. Second, if you just observe the way that these agentic systems work, they are furnaces for tokens. And so you have the straightforward demand of consumers performing inference on chatbots. But what I see really opening up and what I think we'll see will ultimately swamp the amount of human AI traffic is agentic machine to machine traffic. So I think that's kind of the runaway demand driver that is keeping it really tight in supply and demand today.
- [03:57]
C
And that's really one of the reasons why we increased our power demand forecasts. We're now assuming about 220% global growth in AI. In global AI and broader non AI data center power demand 2030 versus 2023. We were at 175% before. To put what that means into perspective, that would be like adding another top 10 consuming country to the mix. It would basically be the number six or so on the list of countries consuming power. It's that significant. The reasons for the upward revisions, you highlighted one of them, which is we're starting to assume a bit more energy intensity and inference and I think there is still a long way that could potentially go. But we generally thought of inference as a much lower energy intensive period per server. But we're starting to see that start to change for some of the reasons that you highlighted. Some of the other factors that drove the increase were number one, just the sheer number of servers that are being shipped and our broader research teams forecasting more of that. And then also we're starting to see a bit more deployment of your latest gen servers as well. In addition to you're building more data centers in the world and so you have the ability to actually have servers turn into power demand.
- [05:08]
A
If you zoom out a little bit. Brian, you have a really interesting framework that George and I have both taken note for the factors that could drive or constrain power demand growth amid all of these developments. Talk us through that framework and why you think it's useful.
- [05:22]
B
Sure.
- [05:23]
C
We think of the six Ps that are driving both growth and also potentially constraining growth. Pervasiveness of AI productivity of models and servers, price of power, policy parts and people. And some of these are going to be drivers of growth. Some of these could be constraints to growth, Some of them could actually be both depending on the time period of all of these. And I'm sure we'll get into it here. We're most worried from a constraint perspective on the people side and the ones where we get a lot of questions on that, at least for now, we are less worried is the rising price or supply cost of power. And then we're less worried about the pervasiveness and the response to productivity gains for some of the reasons we just talked about.
- [06:05]
B
This is a fascinating part of your work, which is identifying something that I think is not well understood by the market, which is the human intensity of deploying these solutions. The number of folks will need to upgrade grids, the number of people to build data centers, the number of electricians. Maybe double click on that and offer some of the statistics that you often cite there.
- [06:23]
C
Yeah, we think we're going to need to see 500,000 new jobs in the US alone to be able to source the combination of building the generation that's needed to supply power to data centers, that's about 300,000 and then also the grid, the infrastructure that's required for transmission and distribution, that's the other 200,000. Now that's a big number. We are actually less worried about the bigger of those two, the 300,000 and the ability to supply that. The skill requirements are less. Where we are more concerned about is on the transmission and distribution side because there's electricians need four years of skilling and if we look today at the number of energy apprentices that are out there in the US today and it needs to rise significantly with about 45,000 energy apprentices, we think that needs to rise by another 20 to 25,000 to be able to be in the mix here to achieve some of the numbers that we're forecasting. We're not saying it's not going to happen. We're just saying that this is the major constraint that we are focused on. And I think it is driving some changes in how we're seeing hyperscalers and others invest. Because if the grid isn't going to show up for three to five years, then even if it's a bit less efficient, you're starting to see greater deployment of behind the meter solutions. And that may last for a period of time, probably not forever, because grid still seems to be the priority. But it really all kind of comes down to time parts. And to your question, people, and do we have enough skilled labor?
- [07:55]
A
Well, talk to us a little bit more about behind the meter solutions.
- [07:57]
C
What do those look like? So behind the meter solutions are largely but not entirely supplied with natural gas. And there are different types of natural gas generators. There are simple cycle generators and combined cycle generators. The simple cycle generators tend to consume more natural gas per unit of power relative to the combined cycle generators. That makes them a bit less efficient. But if that's all you can get, until we see more supply of natural gas combined cycle generators come on, which we would expect around 2029 and into the2030s, time to market is important and so is decarbonization for the hyperscalers as well. But in the near term, if you can get that data center online with a less efficient solution, we're starting to see that. And tell me if you're seeing that as well, 100%.
- [08:40]
B
And it comes from the fact, I think one of the lesser known or understood dimensions of this whole phenomenon is just how deeply capacity constrained the labs and hyperscalers are today. And so the practicality of meeting that leads you to behind the meter solutions. I think it's a persistent demand vector, as you say, it's less efficient and it is also subject to its own supply chain, inertial and gravitational forces. But I would expect to continue to see it at scale. So on one of your P's, I would add three more P's, which is policy. One of your P's, which is, I would add pricing, politics and populism, talk a little bit about what we're seeing at the state level with data center moratoria and state level legislation, concerns about pricing, going up for consumer electricity costs, et cetera.
- [09:27]
C
So affordability is a major concern and we've already seen it in the P that you added, which was politics. And we will likely continue to see that be a concern because no one wants to have to pay more, especially if their demand isn't necessarily the precise cause of it. Luckily, there are some ways, and we think we're starting to see some alignment between regulators, policymakers and industry and the hyperscalers themselves to try to ring fence the data centers where the hyperscalers are committing on a take or pay basis to whatever it is they say they need built. Which means that that should reduce the leakage risk to not necessarily have as much of an impact on affordability for broader customers. There's still more steps that likely need to be taken to ensure that, but at least we see some alignment in trying to get that done. Now, what that could mean though is the hyperscalers may be paying more for power. Our view is that is not as big of a concern as some of these other P's like people. And that also goes to the type of how power could be sourced. We did an analysis looking at sourcing power from green reliable solutions that could be nuclear, that could be combinations of natural gas with solar and wind and battery storage together. Those do cost more. They have a higher supply cost relative to natural gas combined cycle as an example, but we think that they can afford it. Not to say that that's where they're only going to go. But we did an analysis saying if Those solutions cost $40 per megawatt hour more relative to natural gas combined cycle and the hyperscalers paid all of that, not just for us power demand from data centers and that growth, but all the global growth coming from data centers, that would only have an impact of about 2.5% on their 2030 EBITDA. It would only impact their already high corporate return on capital, cash, return on cash invested by less than 1 percentage point. The point we're trying to make is not that they're not going to care about the price, but that that is less of the constraint for the hyperscalers. But it is a major issue for the broader populace.
- [11:32]
A
Are you seeing any signs that the industry is moving in that direction? What are the incentives for them to really do that?
- [11:38]
C
Well, I think the political side and the kind of NIMBY side is really starting to become louder. And so you've seen companies and you've seen some of the big hyperscalers put out community statements saying or other statements saying we are going to pay for the power that we consume. We do not want to let our demand for data and our construction of data centers impact everyday consumer prices. And I think that goes to the point on it does look like there is alignment. Maybe it's just in words, but we do think it'll translate and it is already is translating into action.
- [12:10]
B
When you think out as you have in a lot of your research to 2030 and beyond, what does the mix of energy resources for AI data centers look like? You've talked about the kind of proximate demand and convenience of natural gas solutions, but you've got a sort of Moore's law like dynamic for solar and batteries. What will the mix look like down the road?
- [12:30]
C
Yeah, we broadly think about power being sourced for data centers as roughly 60% coming from thermal sources, which is largely natural gas, 40% coming from renewable sources. And there's a little bit of extra for nuclear as well, which is going and that's through 2030. The 2000-30s could look a lot different and nuclear is going to play a major role in that. Or at least to what degree will we see nuclear be embraced? Not just from demoth balling some of the plants in the US but also building new small and larger scale reactors. But the parts availability that p that's going to really help to drive how power is going to be sourced chronologically. Because if you can't get that natural gas combined cycle turbine until 2029, 2030, 2031 and nuclear isn't going to play a meaningful role beyond the demo fallings until the 2000s and potentially mid-2030s, you have fewer options in the near term. And that could be that behind the meter simple cycle natural gas solution. It could also be solar and wind paired with battery storage paired with natural gas simple cycle to try to create that green reliable solution. And again, those come with higher costs, but that doesn't mean that they're not going to be deployed.
- [13:42]
B
They come with higher costs, but they also come with very natural leverage on volume and price and innovation curves. And so I think thinking into the 2030s, I agree with you entirely. Nuclear is hopefully going to play a meaningful role. Also, I'm really bullish on solar and battery. I think they're both on a trajectory to be surprising in terms of their efficacy and cost in that period of time, but only time will tell.
- [14:06]
C
Yeah, I think sometimes there's the perception that this is all going to be in one basket and we reject that. To use an improv comedy phrase, we are in a yes and environment here, not a no or environment.
- [14:20]
A
But let me just ask a follow up to both of you. Because as you say, the costs are higher, but these companies are for the most part extremely profitable. But there is some concern about the fact that they are eating through a lot of cash for their operations at this point. And I think we're all kind of sort of taking it for granted that they're just going to be able to fund and fund and fund. Is there some limit to that when we have these views of the future?
- [14:43]
C
I mean, there definitely is a limit. Not sure we're at that point. And not to say it has to equal the analogy that we started talking about with shale back in the 2003-2020 period. But at peak, those companies were spending more than 120% of their operating cash flow back into capex. The equivalent right now after this $300 billion upward revision is in 2026, about 87% for the hyperscalers, 87% of their operating cash flow plus R and D is being redeployed back into CapEx and R& D. And yes, that's been on a steady move up. And also, yes, there are a couple that are meaningfully higher than that that are outliers and investors have responded to that and have noticed that. But that's the free cash flow picture that doesn't speak to the balance sheet side of the equation. There is minimal net debt to EBITDA currently on books. And so yes, we're starting to see debt issuance, which is a relatively new thing because these have been free cash machines. But that doesn't mean that balance sheets today are being stretched. So we would not call the hyperscalers financially inflexible. But they're not the only ones that are part of the AI investing machine here. And so they may be in the best place because of their legacy businesses. But it's fair to not only reflect on the hyperscalers getting closer to 100%, but also that there are some that are undercapitalized relative to them.
- [16:09]
B
Also to that point, I would underscore something you said earlier, which is the actually the amount of investment in energy and electricity cost relative to the other costs of constructing data centers is relatively low. And so I don't think a price sensitivity there will be hugely determinative. But and you rightly characterize how much financial capacity the hyperscalers and other industry participants have. What about the utilities they're being called upon to make major investments. The amount of capex their deploying has gone up dramatically, almost as much on a percentage basis as the hyperscalers talk through that dynamic and the capacity that exists there.
- [16:46]
C
Sure, the utilities do not have the extent of the balance sheet strength that the hyperscalers do. But for regulated utilities that receive the permission from their regulator to invest in projects, they will likely need to finance that and that typically comes with equity and debt. So the utilities are in a much different place where they will likely be needing to tap the capital markets. And so the investors will look and say is this a viable project with a viable contract, assuming this is unregulated, or for the regulated utilities that have that approval, it will be their ability to execute. The area that I know you are very close to that is I think a bit more in flux in particular from the utilities is large scale nuclear or even nuclear broadly. And, and there seems to be a bit of a super abundant chivalry going on right now which is everyone wants to hold the door open for someone else to let those projects go through the door ahead of them. No one wants to be first, second, third or fourth. Some of that is technology driven of can we actually benefit from others improving and innovating technology before we have to do it ourselves? Some of it is nuclear hasn't necessarily always gone well and hasn't been done a lot in the US over the last 15, 20, 30 years. And so there's a hesitation to be putting capital to work where you wouldn't necessarily get those revenues for a period of time.
- [18:11]
A
Fascinating stuff. Thank you so much for joining us Brian.
- [18:13]
C
Thanks Alison. Thanks George.
- [18:15]
B
Great to have you. And you know Alison, I'm a huge admirer of Brian's work because these are very fast moving phenomena as we've been talking about. Brian's very grounded analytical frameworks I just think have been extremely valuable. So I commend the work. I thank you for it and thank you for being here today.
- [18:33]
C
Thank you for having me.
- [18:34]
A
This episode of Goldman Sachs Exchanges was recorded on Tuesday, March 3rd. I'm Allison Nathan. Thanks for listening.
- [18:43]
D
The opinions and views expressed herein are as of the date of publication, subject to change without notice and may not necessarily reflect the institutional views of Goldman Sachs or its affiliates. The material provided is intended for informational purposes only and does not constitute investment advice, a recommendation from any Goldman Sachs entity to take any particular action or an offer or solicitation to purchase or sell any securities or financial products. This material may contain forward looking statements. Past performance is not indicative of future results. Neither Goldman Sachs nor any of its affiliates make any representations or warranties expressed or implied as to the accuracy or completeness of the statements or information contained herein and disclaim any liability whatsoever for reliance on such information for any purpose. Each name of a third party organization mentioned is the property of the company to which it relates, is used here strictly for informational and identification purposes only, and is not used to imply any ownership or license rights between any such company and Goldman Sachs. A transcript is provided for convenience and may differ from the original video or audio content. Goldman Sachs is not responsible for any errors in the transcript. This material should not be copied, distributed, published or reproduced in whole or in part, or disclosed by any recipient to any other person without the express written consent of Goldman Sachs. Disclosures applicable to research with respect to issuers, if any mentioned herein are available through your Goldman Sachs representative or@www.GS.com research hedge.HTML. goldman Sachs does not endorse any candidate or any political party. Copyright 2026 Goldman Sachs. All rights reserved.