How can the grid help AI, and how can AI help the grid? Live from NYU at New York Climate Week, featuring leaders from Nvidia and Amazon | Energy Gang Live from Climate Week

21:35That's never been done before. I don't think anyone, any customer has been able to, with any level of certainty, project that far out. But we're trying to get as good as we can on that question.

21:46Amy Myers Jaffe:Yeah, I was going to ask, are you able to do that? I mean, so would you like to say with confidence where the AI industry is going to be in 12 years time from now? I mean, I think we're going to have to continue to work with our utility partners on getting those forecasts right. But these are the types of mechanisms that we're starting to see. And frankly, we are, you know, party to those conversations. I'm not sure I'd like to say with confidence where the AI industry will be in 12 days time, never mind 12 years. But yeah, yeah. So clearly it's difficult.

22:14Ed Crooks:So I got a question the other day from a student, in fact. And I thought, OK, layperson, it stumped me slightly because you just made a statement, you know, very interesting about AWS sort of committing to the, you know, large bulk of these extra costs for the new infrastructure. So is that a commitment to just the build of the infrastructure cost or is that the ongoing energy consumption that you would have where you're going to pay for that energy cost as well over the number of years? It depends on the state or the regulator. But I mean, I think by and large, what we're trying to do through the tariff design is essentially help the utility identify how much infrastructure will be needed to serve the load.

23:01And then we're still a large customer and paying for energy through rates.

23:06Ed Crooks:Both the infrastructure costs and some commitment to a tariff that you pay over time. That's right. One thing I'll say is in this country, I don't know if anybody's actually read their power bill. Like you might need a degree to read it. It's very convoluted for somebody who understands power. You go into the different lines in your power bill. It's fascinating. I invite everybody to read it because we're all going to be participating in this debate about affordability and tariffs. And it's just fascinating to read which parts of your bill are these fixed charges that's attributable to all the infrastructure built, and which part of the bill is attributable to more generation cost, which goes up and down.

23:48And that has influenced over time on overall system cost and how much people are going to pay. And also, what large loads like Amazon or any other tech companies, how much they're contributing. And when you're saying this rate-making process being convoluted and needs a revisit, it's true. My question also for you is when you go to different states, it's a very fragmented system. And you have to negotiate very differently, and everybody has a different approach. Just by nature, right? Utilities essentially report to their state regulators. And to the extent we're interested in growing in, you know, pick the number of states today, you have different venues, you know, sort of weighing different issues.

24:37Amy Myers Jaffe:Okay, so you've been talking quite a bit about contractual and regulatory fixes to these issues around affordability. I wonder also about technological solutions and what can be done in terms of improving kind of everything in the whole kind of AI ecosystem to ease these problems, for instance, by reducing power demand. I mean, maybe Josh, bring you on to this. I know this is something you're very interested in at NVIDIA. It feels like kind of every month or so we get a new announcement about how much more efficient kind of each generation of chips is compared to the one that went before. Is that, do you think, an important part of, as I say, kind of easing the concerns that people have about affordability and how much burden AI is going to place on the grid?

25:23If you look, again, at the kind of long-term time horizon, then it is critical, if you're looking at where are we going to be in 10 years, to factor in really profound variables like energy efficiency in the data centers. So just taking the data center load, as you mentioned, energy efficiency trends have just been mind-blowing, to be honest, over the past decade. We've, at NVIDIA, reduced the energy consumed for a particular inference task by 100 ,000 times in the past decade. And just in the past two years, in the past one and a half generations of our product, we've reduced the energy consumption for the same inference workload by 30 times.

26:04So that's a 97 % reduction in energy for the same exact inference workload. in just two years. So that's a very, very strong trend, and it's important to take that into account, but you still have the net, you know, the absolute energy requirements of AI growing because there's just so much demand for it. So putting that in context is important, but there is, you know, still net growth.

26:28Amy Myers Jaffe:Right, and the famous Jevons paradox, of course, which is that if your energy use becomes more efficient, you actually end up using more energy because you can do more with it. And I'm very clear, it seems to me, that that's exactly what's going on with AI, right? Which is that as the chips become more powerful and more efficient, people think of new things to do with them, the technology develops very rapidly. And as you say, net, we end up actually using more energy rather than less.

26:51Ed Crooks:But you also have the flip side, which, you know, CSAS just had a report out. I had a piece in the Wall Street Journal. IEA has done a big study on this, looking at sort of the computational trade-off between AI's energy use, which we've been talking about, the burden on electricity grids, versus what it can do to create energy efficiency. And that's in sort of a wider range. So CSIS kind of went through everything it can do to solve the electricity problems, including helping speed up interconnection cues and permitting processes by just speeding up what human beings can process in a day or model over time.

27:38Ed Crooks:But even in the fuel system, so having to do with, you know, freight and trucks, now apparently there's like a new program that if you have a large container ship company can use, you know, it's kind of like the ways for shipping so that you're not waiting at port for birth, which can be very carbon emitting, you know, sitting and idling at port. So across the whole spectrum of the economy, buildings, transportation, et cetera, you're going to have a lot of energy savings. And the tradeoff is that those energy savings might be something like two, three, or four times the energy use in training the models or doing inference activity.

28:20Ed Crooks:So it's really kind of like a mixed bag. You don't have people talk about that part because there's been so much focus on the electricity usage itself. Yeah, and I'd love to talk about that part all day long because I think that is the critical missing piece. People are starting to talk about it more, but again, if you're looking at the net, that's what matters, right? You can look at AI's footprint. It's important, but the more important thing is to look at what are the broader impacts? Where do we end up given AI's kind of second-order impacts in other industries? And I'll give you two data points to take home with you.

28:54One of them is in manufacturing, and one of them is in buildings. the two largest energy-consuming sectors in the world. In manufacturing, we have a manufacturing partner in Mexico. They make NVIDIA systems, and they used NVIDIA Omniverse Digital Twin to optimize the facility for energy before building it, and they're projecting a 30 % reduction in energy over the lifetime of the manufacturing facility because they used a little bit of energy and used some AI on the front end to optimize it. Similarly in buildings, a company I was on stage with earlier this week, Train, they do HVAC, heat management buildings.

29:36They said they're seeing 30 to 40 percent reduction in energy just by using AI to optimize how the building is heated and cooled. And so when you put those in context, those two sectors account for about two-thirds of overall energy consumption in the world. And again, AI, we're down at one and a half percent. So if we use a little bit of AI to reduce the huge footprint that we have in manufacturing and in transportation as well and buildings, then we end up very, very quickly getting to a net benefit to energy from AI. And that's where we see everything going very rapidly. Yeah. And if I may, I mean, I think we're also seeing lots of opportunity to tangibly impact the way the grid itself operates, and particularly with AI.

30:22So like one example, AWS customer Grid Unity recently entered into a partnership with the Southwest Power Pool. It's a regional transmission organization headquartered in Little Rock, runs all the way through Oklahoma, Kansas, up to the Dakotas, very large transmission planning region. They are using AWS Cloud and AI now to help expedite the interconnection queue at SPP. So this is the way that new generation projects are studied on the grid and upgrades to the grid are identified in order to integrate them. They were previously happening basically on manual PDFs in an exchange between engineers at a generator company and Southwest PowerPool.

31:02Now we're automating a lot of that and drawing in those timelines in a way that is going to allow those projects to interconnect far more quickly. and using AWS and AI as a solution to make that happen.

31:15Ed Crooks:So I have a question. I have to be like the AI skeptic in the sense that I tried teaching with large language models as the textbook this summer for an energy class. And the idea was we'd learn about what kind of things AI could do to improve optimization in energy. And then we would try to do some of those things in class. and I had guest lecturers come and someone working in a company that does permitting research and permitting for large utility scale solar came and taught us how you would use the tool and then use the tool to do your own information about the process and how to make a petition more efficient and do it faster, but then also to compare the regulations when you go to meet with the regulatory body.

32:01Ed Crooks:And it was a super interesting exercise. One of the things that we found in class is that, because all of us, or not everybody, but probably most of us are using these tools alone in a room. And you never ask the person sitting in another room, did you get the same answer? But when you're doing it 10 people together, what we found is we're not getting the same answers. And then, you know, I would try putting some elaborate prompt in students' computers, still didn't give it the same answer that I got, which, in my opinion, was the correct answer. So a question I have is, are we far enough along in the technology?

32:38Ed Crooks:I mean, in some functions, I think it'd be incredibly speedy. But in other functions, do we have to worry that we don't have enough guardrails?

32:47Amy Myers Jaffe:Yeah, no, that's a great point. And actually, I'm going to put that in a slightly different way as well, which is with quite a few of these things, it often feels like there is a lot of potential there and a lot of exciting stuff that could be done, but the results in a material sense are often not so evident. I thought that was a very good example you had, Josh, about your plant in Mexico. That's something that really happened. Those are real savings you've achieved there. How common is that kind of example, I wonder? I mean, is it really making a material difference now? Or, as I say, are we mostly talking about things that could maybe perhaps be done if we can get the technology to work right.

33:26You have to remember that our kind of retail experience, our consumer experience with AI is a tiny, tiny piece of what AI actually is. The industrial commercial applications of AI typically are really effective, especially at things like optimization, prediction, maintenance. So in driving efficiency, it's excellent. It does that very reliably. And also, something else, I mean, I use AI every day. It is very useful, but it's not perfect. They say in Colorado, if you don't like the weather, wait 15 minutes. With the chatbots, if it's not giving you a great response, if it's not consistent enough, reliable enough, wait six months.

34:05The pace of innovation is so rapid. I continue to be just shocked by how capable it's becoming, how quickly. But again, we've got to remember that that is one small piece of AI, the industrial applications that are really driving very significant improvements in emissions and energy are reliable and very real already. If I can also just tie it back to energy a little bit, related to some of these discussions about AI and manufacturing, we're starting to hear about more facilities, warehouses, manufacturing, becoming much more automated, so basically more robots operating in these facilities, which actually increases your energy consumption, as well for those.

34:48So we've seen examples of facilities asking for more power provision because of that automation. So that's also one angle that we have to manage. The other angle is also just the data centers coming into the power markets in such big load does change the way that we operate the grid and the markets. I want to use the privilege of being a host to show one slide. Excuse me for the folks that are just listening to the podcast. One of the things we do in Woodmac, we install thousands of EMF sensors and other sensors in the nation's grid. We have real-time monitoring of how power grid is operating.

35:32So this is an example of a sensor that's pointing to a transmission tower. We get every 15 minutes a data feed on whether or not there's power being going through that if there's congestion. This has traditionally been how we monitor generation facilities. Starting last year, we started installing these sensors at large load locations to monitor how much their energy they're using. And the traditional sort of wisdom, conventional wisdom, is that generally a lot of data centers don't flex and run a flat load. So this year, does everybody remember the heat wave we had at the end of June in the East Coast?

36:15So I was in D.C. It was really hot, end of June. So I asked our team, can you find me a data center in the D.C. area? I want to see if they're changing their load. So this is the power price in PJM, Dominion specifically. So you'll see in the early evening hours when the sun has set, power prices went up significantly. So on the 23rd of June, it went from$30 per megawatt hour to$1 ,700 per megawatt hour because of that evening demand. And then we overlaid the load change of a data center in Virginia. This data center belongs to Meta. And if you look at the load change during that peak power price hour, on the 23rd of June, if you just look at that data, it went from 200 megawatt load to 20 megawatt load.

37:10So it dropped 90 % during that one hour. And the reason is simple. Power prices are really high. So during that one hour alone, if they're paying wholesale power prices from that data center, they probably saved between$300 ,000 to$500 ,000 during that one hour. And so what this has led us to believe is that probably data centers are more flexible in some cases than we traditionally believe. We've already seen cryptocurrency facilities do the same thing in Texas, but as more and more data centers come into the market, we'll probably see more of this kind of behavior. And this actually helps make the power markets more efficient.

37:55If this data center does not flex during that hour, the power prices would be even higher. So there's something related.

38:03Amy Myers Jaffe:It's fascinating. It's a shame for podcast listeners that might not be able to see this job. We should make sure we put it on the web so people can go and look it up, because I do think it's a really fascinating graphic and more worth a look. So following up on that, as a general point about the importance of flexibility, this is a subject we've covered a few times on the show already this year, talking about the work of Tyler Norris at Duke University, obviously putting a lot of emphasis on this in terms of the importance of data centers being flexible load. And, Josh, this is something which you're working on, right?

38:36Amy Myers Jaffe:So we had Varun Sivaram of Emerald AI on just very recently, a couple of weeks ago. And so NVIDIA has a relationship. You are part owners. You have an investment in Emerald AI? Yeah, we invested in them, and they're part of our startup program. Right. And what are they doing then? Because they're working in this area, right? What's their goal and what is the benefit potentially of what they're trying to do? Yeah, it's a fascinating company, really good business model. I think if you care about AI and if you care about sustainability both, they're basically finding ways to do kind of what you're talking about, Shijo, in terms of making data centers grid assets because they're flexible.

39:17And the thesis of Emerald AI is that there are very few AI workloads that are truly non-preemptible. And so they can flex. And based on the Norris study at Duke, if just for dozens of hours per year, so 1 % flexibility per year, if you can ramp down by 25%, potentially unlock 100 gigawatts of energy that's just there and not being used. Because as you mentioned, Amy, we're planning for peak load. And when we're not at peak load, there's a ton of energy available potentially to be used. And what's a great thing to do with that energy if it's available? You just convert it into intelligence. You match it with some data, convert it into intelligence when it's there, and then you're making great use of that.

40:00And that, of course, does lead to more efficiency in the grid because you're able to make much better use and get paid for the energy that's actually available rather than having it go to waste. And it really does turn these data centers into grid assets. And it also helps normalize the pricing between dispatchable power and variable power. because no longer do you have to prioritize dispatchable power necessarily to the same extent that you did before because the variable power is really valuable. And so it actually can help drive greening of the grid because those variable assets like solar and wind end up becoming relatively more valuable compared to dispatchable power.

40:38Amy Myers Jaffe:Every business has priorities to protect, goals to reach, and decisions that need to hold up. Energy should support all of that, not become one more thing to manage. That's why Engie takes the time to understand your business before building energy solutions around it. Your operations, your goals, your pressures, your plans for what's next. Because while Engie knows energy, no one knows your business like you. And when that expertise comes together, energy becomes more than something that powers your business. It becomes part of what helps move it forward. Engie helps turn energy plans into outcomes with solutions built around real business needs.

41:13Amy Myers Jaffe:Learn more at engiresources.com. That's E-N-G-I-E-N-G-Resources.com. Exactly. So during those evening hours, usually in PJM in Virginia, it's the gas plants ramping up. And so if we don't use as much power during those hours, you basically don't burn nearly as much natural gas, at least given the grid generation mix today.

41:39Ed Crooks:And let me just point this out because I have a group of students here at NYU who have been working with the outage data. you know, the U.S. official outage data. And you would imagine in your minds, everybody's thinking in their minds, watching the headlines, that California CAISO and ERCOT in Texas have the greatest record of outages, constant outages, right? But actually the data does not show that. The data shows that the highest rate of outages by a wide, wide margin is actually in PGM where they do have the concentration of natural gas peaking plants. So it's really interesting because what's the difference?

42:19Ed Crooks:The difference is in the last two years, Texas has put in probably more batteries than almost any place in the country. And California has had a steady ongoing deployment of batteries. And so it really does bring up this question of new technology, some of which could be driven and enhanced by AI, like virtual power plants where a utility is using some combination of battery storage and people who are willing to have the utility reach into their thermostat and change it slightly or tell them to turn off their washing machine or what have you or bi-directionally charge their car at a different time to actually, again, take those moments where there's too much demand for an hour or two hours or a couple hours and remove that demand in other ways or meet that demand in other ways.

43:10It's so effective that, you know,

43:14Ed Crooks:in my recent conversations with people in Texas, the actual natural gas industry is thinking that they need to actually add batteries to their peaking plants to get that efficiency. That is amazing.

43:28Amy Myers Jaffe:I hadn't heard that yet. So I was actually just going to, I think, build on something that you mentioned, Amy, and that was sort of, I think you were mentioning VPPs, but also just the opportunity for new technologies to sort of help with some of the challenges that I think we just laid out. And one area that we've been really excited about at AWS is grid-enhancing technologies. And this kind of gets back to similar sensors, I think, like to the way that WoodMac is getting their data. But, you know, types of automated technologies on transmission systems that use real-time sensing, weather data, and the like to essentially, you know, determine how can we best optimize the lines that we have today in order to get more out of them, right?

44:10We actually did a study with RMI, I think it was last year, looking specifically at PJM, because we, I think, all understand the challenges in the mid-Atlantic. And we found through that study that we could unlock six and a half gigawatts of new energy projects with the deployment of GETs alone. That is without building another transmission line. Now, we know we need to build more transmission. We need more energy infrastructure in order to reduce congestion, serve economic development, and help with the continued integration of new generation assets. But fascinating that we could have technological solutions that are available today to get that level of new power onto the system without actually having to build another power line to connect it.

44:56Amy Myers Jaffe:Yeah, that is indeed fascinating. Where are you on this question of data centers as flexible load? Is that something you're working on at AWS? So certainly a very hot topic, right? And we're hearing it from, I think, regulators and policymakers alike across the U.S. I think for our part, I do want to say that there is no one-size-fits-all solution here, particularly across the data center industry. Like at AWS, we are hosting lots of customer workloads that demand high uptime. Uptime, they are built on the back of the AWS cloud. And so we're talking about hospital systems and government agencies and others that, you know, I think might have concerns about, you know, what we mean when we say flexibility.

45:40And so we are in conversations with, you know, all of our utility partners and trying to figure out what might make the most sense. But I would just caution it's just not a one-size-fits-all type of solution.

45:52Amy Myers Jaffe:And something that often I think there's confusion about, and I think probably I'm confused about it as well, is the question of when we talk about flexibility at a data center, to what extent are we talking about kind of compute operational flexibility and the kind of things that you've been talking about, Josh, and to what extent are we talking about actually just having some kind of backup generation on site, and whether it's diesel generators or some gas engines or gas turbines or batteries or whatever it might be? I mean, both are options, right? Is there one that's kind of more viable than the other?

46:22There's also spatial flexibility too. And this is something that's a little bit unique to AI workloads is that typically they're a lot less sensitive to latency than traditional data centers. So if you have a hospital that's running critical things off a data center, you probably really care about the latency. If you're asking AI to write you a research paper, it's already going to take you a few minutes probably. Having a few extra milliseconds of latency in there isn't going to be noticeable. And that's especially true of training. Our CEO, NVIDIA Jensen, likes to say AI doesn't care where it goes to school, meaning that you can train your AI model wherever you want to.

46:59You can train it in Alaska, nobody cares. And then you bring it to wherever you need to use it. And again, the latency isn't that sensitive. So flexibility in time, flexibility in space is also useful. And then, yeah, having the batteries there to absorb some of the shocks is also useful. So you combine all those together and it's pretty powerful. I like to be optimistic and I think we have more reason to be optimistic than ever before. So for an energy focused audience like this, I want to just put a hypothetical. What would we like to see ideally in terms of modernization of the grid? What do we want to get out of the grid in the future?

47:34And what conditions might create that grid for us? And I think we're actually in a perfect storm of incentives that could drive us to that place where we want to be. And the things that I would say are, number one, being able to take full advantage of variable energy sources, especially clean ones like solar and wind, being able to, of course, provide reliable service to consumers, being able to invest more in our grid because there's value there, and being able to serve the workloads like AI that are going to drive economic opportunity. If we look at where we're going, we see more batteries coming online through electrification.

48:11We see more variable energy. And AI is perfectly positioned to help match those together because that is a huge complex data problem. Trying to figure out where should electrons flow back and forth, not just unidirectional. And trying to, yeah, pull electrons from batteries when you need to and then pull it from solar when you need to. And managing that complex problem is something AI is perfectly poised to do. I think you're spot on. And, you know, this is something that we're also working on at AWS. I mentioned earlier our renewable energy program and projects that we've deployed around the world.

48:46We now have a solar and battery storage project in California that is using AI built on AWS to optimize charging and dispatch of that battery in real time. And it's actually driving down the cost of operation and improving the cost of power for that project that is grid connecting and helping to serve customers in the region. And I think that is just going to be one of many examples where we're leveraging this technology to optimize particularly our battery storage projects.

49:16Ed Crooks:Josh, when you and Craig talk about sort of like this moment of opportunity, instead of having it be like a social problem, we have this moment of opportunity. that makes me think about first of a kind. So Craig, I have you here on the stage. I mean, first of a kind, Amazon, AWS has been one of the companies that have said, okay, we're gonna give a serious look to first of a kind with SMRs, right? But the flip side is, it seems like you found a utility that's willing to work with you on that. Because when I talk to people, there are a lot of utilities that are like thinking about it, but getting a utility to actually say, yes, I'll build first-of-a-kind.

49:55Ed Crooks:They're very conservative organizations. Where are we on that? You think that someone will, we're really going to get there and have some first-of-a-kind facilities? I mean, the short answer is yes. And so what we're doing is we are, you mentioned utility partners. For our first deployments of small modular reactors, we're working with Energy Northwest, which is a public power consortium in eastern Washington state, with the goal of deploying X Energy SMRs to support our Eastern Oregon data center campus, which is essentially grid connected to the Washington site. But I think to your question, though, around first of a kind, Amy, it really brings up what most would acknowledge in the nuclear industry right now, and that is just the risk profile of first of a kind.

50:44And so I think that's why as a corporate buyer of clean energy, we decided not only to work with the utility partner and work with a SMR company in X Energy, but we actually took an equity investment in X with the goal of standing up a manufacturing pipeline of five gigawatts. That is one way that we can help with the risk profile and then work with the local utility to actually develop and deploy on that site.

51:11Ed Crooks:So the interesting thing from my recollection, reading some of the press on it, because one of the things that people have had in other locations, no names will be mentioned, is that it requires specialty fuel. But you're actually making an investment in producing the specialty crew. You're also making the investment to manufacture the specialty fuel as well, correct? Through X Energy, that's right. Right. So that's unusual because the federal government has had to try to stand up getting that fuel built in the United States. So, again, high risk, no? We're excited about the opportunity. I mean, we can just acknowledge, yeah, I mean, it is first of a kind.

51:55Like, we are trying to do hard things. But we also know if we don't step up now, we're not going to realize, you know, the ultimate deployment of these new and exciting technologies. And so that's what we've decided to do. We absolutely need SMR and nuclear to be deployed in the future as carbon-free energy. But I also just want to say the timing issue of all of this. If you recall the number I just cited, 116 gigawatts of data centers that are ready to be connected. We need that energy, like, right now. In the Woodmac forecast, our first SMR unit in the U.S. comes online in 2031. I don't know if that's reasonable to you.

52:40Amy Myers Jaffe:Did you say 2035? 2031 is the first unit, right? And you're talking like maybe 100 megawatt. So 116 gigawatt, you know, 100 or fewer megawatt in 2031. And maybe scaling up, you can do that in the 2030s if we believe in the technology. But we still have the problem of the next five years. what are we going to do about all this new demand that's coming online? And where's all of the new power supply going to come from? That's the biggest question facing the power industry in this country right now. And I think it's not too scary to call that we might have a power supply crisis if we don't do something differently.

53:27But that's sort of the five-year to 10-year time frame is really what we're worried about. and it has the potential to send energy prices much higher when these peaks become more and more frequent and disruptive and consumers are not going to like it when these prices go up. So somewhere in the policy-making circle and in the industry organization, we have to wake up to this and really think about what's going to happen.

53:58Ed Crooks:So are we in a hype cycle? Is electricity, as Ron DeSantis kind of implied, you know, is that the... It can't happen because we're just not going to have electricity. Like, where do we go?

54:11Amy Myers Jaffe:Well, those are great questions. I think we should let the audience have a chance to ask a couple of questions. Before we do, just to go back to that factual point about timetable for maybe these X-Energy SMRs coming online. Did I hear you say that 2035 was your target date for having the first one up and running? What's the... So for the entire 5 gigawatt pipeline, we would like to see that deployed by the end of the 2030s. But certainly for the first units, we're moving as fast as we can. Right, right. Yeah. And just, I mean, I hear what you say. Shijun absolutely agree with everything, as you say, about this kind of short-term time horizon.

54:50Amy Myers Jaffe:Five to 10 years is important. We're still going to need more power by the end of the 2030s, right? I mean, it's still worth starting those kind of operations, you know, getting that research and development done in order to get the technologies that we need to be commercially deployable in the future, even if it's 10 years plus off because of, yeah. We absolutely need the technology as one of the technologies of the future. Yeah, so as I said, let's take some questions from the audience then. I think we've got some mics which we can send around the audience if anyone has a question they would like to ask.

55:24Amy Myers Jaffe:Be shy. your hand up and introduce yourself. Yes, one down the front here. Thanks very much. Sure. Hi, Mary Beth Mendana, CEO of Onyx Renewables. So the question I have is, it feels like we're still in an eight-track tape environment. If we think about eight-track tapes, I don't know, maybe I'm dating myself, going to digital assets. And how do we turn around the structure of the industry? We're still, and when I say eight track tapes, we're talking about centralized utility systems. We're not talking about localized utility systems or localized solutions. Everything you guys have been talking about is localized.

56:09So why don't we turn it on its head and the main source of power become all the localized. They become the data centers. It's a storage asset. We use renewables. We don't have to build big transmission lines. We can use, we can automate everything you talked about in terms of AI usage. Like, we can optimize things. It gets a prime time to go from 8-track to digital.

56:38Amy Myers Jaffe:Yeah, so that is a great question. I'm actually going to get one better and say in some respects you could say it's like the wind-up gramophone power system because some of the industry structures and the regulatory structures go back to the 1920s, right? So I don't want to take that on. I mean, in terms of, I guess, question, do we need some kind of radically different industry structure to meet the challenges of the future? I don't know. I would just say that that's the beauty of the moment that we're living in now, is that we have all of these incentives coming together at the same time. We see real desire on the part of the government right now, and there's actually bipartisan support for expansion and modernization of the grid.

57:18We've never, at least I don't remember seeing this much motivation to cut through the red tape, the regulatory processing, as we've ever seen before, because there's such value potentially to be delivered. So we seem uniquely aligned to try to solve this problem right now, especially because we have huge economic incentives behind it. So I think if we're ever going to do that, now is the time, and AI, I think, is going to help us.

57:40Ed Crooks:And I would just say that not only is the technology there, but part of what's happening out there, maybe somewhat even in the hyperscaler world, but definitely at the retail level, is people are stepping up and just doing it themselves. One of the problems that the utilities have is they have no visibility on the fact that someone somewhere said, you know what, this is not working for me. I'm getting my own backup system for my house, or I'm getting my own backup system for my apartment building here in New York City. We had a whole program on sustainable real estate, and we had really phenomenal speakers, private equity firms and other real estate investors that have built what they call circular buildings where the building itself generates all the energy that it uses and they really don't need to take from the grid or very little from the grid.

58:31Ed Crooks:So I do think there's some momentum to having these solutions come and you're seeing that even, we're seeing it in the United States, but you're seeing it even more so in countries where the central government's not really delivering solutions, right? So you're seeing that in Africa. You're seeing that in Southeast Asia where individuals or manufacturing plants or other kinds of entities are literally doing a behind-the-meter solution. And then, you know, that's it solves the problem. I feel like behind-the-meter solution is something that a company would go and threaten a utility with right now, but you think people really go in that direction?

59:16I'll just add, in WoodMAC, we have a microgrid tracking service looking at all the microgrid development. We also look at what's called co-location of data centers with generation, which is essentially sort of behind the meter arrangement. We look at all the data centers we're tracking, only 6 % of them are going for co-location.

59:40Ed Crooks:Oh, interesting. And we can talk about why. It's a much bigger conversation. I think out there, there's regulatory issues. There's power supply issues. But the fact is that currently, most data centers are still opting to be connected to the grid for many different reasons. So there is something there. Obviously, a few data centers opted for co-location. But it's not sort of widespread phenomenon that we're seeing yet.

1:00:09Amy Myers Jaffe:Yeah, as you say, grids exist for a reason. They do have lots of advantages, and yeah, it's no coincidence that most data centers still want to be connected. But I think it's also interesting to think about whether that could change and how far that will go. Let's take another question, though. Who else? You've got one there. First of all, thank you so much for the impressive panel. And I'm an MPA student at Columbia University, and I also had the experiences to do the energy optimization for data centers at Tencent. So actually I have two questions. The first one is about the geosermal. So when we're talking about the nuclear, it's more like the sources for 2030, as you mentioned before.

1:00:51So how about the reason five years, do you think the geosermal would be a good alternative to power data center, or at least to do the cooling and the heating process for data centers? And my second question is about like the Chinese data center over sea business. So from my observation, Most of the Chinese tech companies like Tencent, Alibaba, and Bydance, they are planning to have new available zones all over the world, like in the Middle East, in the Southeast Asia, and also in Latin America. So basically, they can view the data center campus of half of the price and also offer cheaper deals to the customer.

1:01:26So how's your opinion about that? Thank you so much.

1:01:29Amy Myers Jaffe:Two interesting questions. Well, should we take them in order? Geothermal, who wants to? In principle, I'm personally very excited about geothermal. I think it has enormous potential. To me, it's not really a next five years technology. These kind of enhanced geothermal projects, technologies that we're seeing at the moment, just the first few pilots are really coming into service kind of now and over the next couple of years. I think before we see large-scale deployment, that's going to be into the 2030s. Yeah, that's my experience as well. Really exciting technology. Hopefully it'll help us in the long term, and there's a lot of potential in the United States for geothermal, but it'll take several years.

1:02:07So we do have an energy crunch right now. Ashijo, to answer your question, I think we're going to need an all-of-the-above approach in the near term in order to continue to deploy AI to meet consumers' growing demands through electrification and everything. So we need to do that to expand the grid. And for those of us who care about the emissions associated with that and the potential for natural gas to play a critical role in the next few years, I would say it's an investment that is going to pay off very, very quickly because AI's second order impacts on emissions, there's growing consensus that they are going to be, AI will be a net contributor to emissions reductions very, very soon.

1:02:45And I personally think that's within the next three years.

1:02:47Amy Myers Jaffe:And what about the question on international competition then? And I don't know if anyone wants to take that on, but this question of whether it's going to be possible to build data centers elsewhere in the world, outside the United States, that you can build and power at significantly lower cost. What did you say, like half the cost that you can? That was half the cost that you can in China, right? Was it? Which is already probably, I would guess, lower than the cost you can build in the U.S. I don't know. Yeah, I would just suggest, and I'll offer what I focus on, and that is how do we get this done in the U.S., right?

1:03:19Right. And that is, you know, I think some folks will call it energy dominance. Some will call it winning the AI race. For us, we call it meeting customer demand. Our customers want to be on the cloud and they want to innovate in AI. And we need to build data center infrastructure to make that happen. And at the end of the day, I think from an energy perspective, we see energy infrastructure as being the linchpin to making this all happen. Transmission infrastructure in particular, high voltage interregional transmission, to be even more specific, is going to be critical. goal, not only for serving new data center demand and other electrification sources like EVs and manufacturing and broader scale economic growth, but also for integrating new generation resources at scale.

1:04:02And that's what we're focused on every day at AWS.

1:04:06Ed Crooks:And ultimately, you know, five, 10 years ago, I would be on a panel and someone would ask me about, you know, the problem of equity in electricity prices or just the problem of electricity prices in general. And I used to say that this division between the wholesale market and the retail market is skewed up because in the end, people who are low income residential users should not be the interruptibles and they shouldn't be paying a higher price than the wholesale, you know, large load customer. And, you know, people would say, everybody would kind of like nod their head, and then they'd pull me aside after I was off the stage and say, you know what, that's not how electricity prices work in the United States, Mrs.

1:04:50Ed Crooks:Jaffe, right? But now we're seeing more discussion about companies and entities that have the larger loads, whether that's industrial load or data center load, wanting to really ensure that the electricity is there and so really willing to step up to the plate with a different pricing structure. And maybe that's, you know, the time has come, again, opportunity, because really, we need to go in that direction. On the international angle, I'll just say something very quickly. I was just in Thailand and Malaysia, I met with the power utilities in those countries too. There's a rush to meet data center demand in those countries as well.

1:05:34And you've heard about like a Singapore not being able to meet it, and you have data center companies or tech companies moving to Malaysia. So across the border in Johor, in Malaysia, there's a huge boom. But then you also have the question of many of the tech companies want zero carbon power. And when you have a nationally integrated utility, it's really hard to guarantee you're going to have zero carbon power because you're just buying from the grid. So for many countries, there are also local regulatory changes that need to happen to allow for either virtual PPAs or other types of green attributes to come to the tech company.

1:06:12So I don't know if the constraint is on data center building cost right now. I think it's much more on the energy provision side to make sure that the local power companies can actually meet the requirements of the data center companies, tech companies.

1:06:29Ed Crooks:Some of that, especially in the region you're mentioning, will eventually come from cross-border electricity trade. You're seeing a lot of that in Europe, but now Singapore and other countries are investigating, right? It's one line. It's very challenging. All you have in Southeast Asia is a hydro line from Laos going through Thailand, Malaysia to Singapore. The prices down in Singapore is exorbitantly high. You think New York has high power prices? Imagine transmitting power through three different countries and every utility along the way, charging you a willing charge to get the power to the end users.

1:07:08But people get desperate sometimes and that's what happens. We do see more need for regional grid integration, but when electricity and transmission line cross borders, the level of regulatory complexity is just unbelievable and takes a long time to resolve.

1:07:24Amy Myers Jaffe:I'm sorry, I think we should just about wrap it up there. Apologies if you had a question and we didn't get you. Go on then, go on. Just one very quickly, very quickly, and then we're going to wrap up. My question is, we've been talking about electricity,

1:07:37Ed Crooks:but one of the things that data center has is a lot of heat. We haven't talked about any thermal energy network yet. Now, certainly, if you are competing for electricity with residents, they may be angry at you, but if you provide them heat in the winter, they may be your friend. So just wondering if there is any discussion about thermal energy networks and basically data center becomes utility. Yeah, this is a really exciting opportunity. A lot of the energy that goes into data center does get converted to heat. It's used by the GPUs, CPUs, but then it gets converted to heat as well. So we do have partners.

1:08:16One of them I love to talk about is called Eco Data Center. They build their data centers out of timber. They power it with 100 % renewable energy, and they use district heating to take the heat from the data center and to heat the local communities. and I think we're only going to see more and more innovations here where we're able to reuse that and data centers will actually become not only assets to the grid but assets to the communities where they're at.

1:08:38Amy Myers Jaffe:Yeah, that is a really cool idea and definitely something we should be exploring further. Unfortunately, we do then definitely have to leave it there, but it's been great talking to you all. Thanks very much indeed, Josh. Thank you. Yeah, thank you. Great. Thanks, everybody. Thank you, Amy, and thanks very much indeed for hosting us here.

1:08:54Ed Crooks:Always wonderful to have a host, the Energy Gang at NYU.

1:08:57Amy Myers Jaffe:Absolutely, and thanks to everyone. at NYU for being such great hosts and, as I say, letting us use your wonderful room here. Thank you all very much indeed for coming. Thanks for your great questions, so please provide questions. And thanks to everyone listening to this podcast wherever you may be. As you know, we always value your feedback. Please do keep it coming. Leave us a review on YouTube if you want, or on your podcast platform. And we'll be back very soon with all the latest news and views on the Energy Transition. Until then, goodbye.