The Exchanges

Q that aren't necessarily designed to equip some of these, uh, peak loads. Others say AI is just different, you know, it's going to require exponentially more energy. Where do you, each of you fall on this spectrum of concern? Like, as people come to this and try to figure out just how substantial is this issue? Um, uh, Brian, we'll go to you. And then we'll go down the line.

A Yeah, as I, as I sort of alluded to before, I do think it's very real. And I do think that, you know, probably the most extreme projections are overblown. Um, I think on the other end, you know, the ones that I've seen on the low end, I think are, are pretty, pretty off base as well. Um, we've got a great indicator of investment and growth in this space, and it's the capex expenditure for all the big cloud companies. And if you look at that quarter over quarter, year over year, They are all accelerating investment. It's a great leading indicator of how much power, how many chips they're going to plug in in the future, um, and they're all betting really big on very large AI models, and so that's maybe one of the differences between AI and what we've seen in sort of native cloud is that there is very much an incentive to use more electricity. There's incentive to build bigger models to plug in More GPUs. And, you know, I had someone ask me the other day, well, what happens if, you know, if NVIDIA comes out with a chip that's twice as efficient? And my response to that is, well, then Meta or Microsoft or Google will plug in twice as many chips. If they have the power, they're going to build as big of a model as they can build. And so that's where I think AI is a little different than what we've seen. Not to say there's not incentives for efficiency, because there very much is, But…

AI assessment note: “I do think it's very real. And I do think that, you know, probably”

Q And do you have any numbers behind what we could potentially unlock?

A Uh, Well, yeah, in fact, there was a great RMI study that just came out, uh, very recently that looked at the PJM system and determined that you could unlock in excess of six gigawatts of extra capacity, and this is, I mean, PJM's a 90 gigawatt system, so six gigawatts is not nothing. That's pretty substantial, um, and, and this is for an investment of about a hundred million dollars that would return annually in excess of a billion dollars To rate payers. So, I mean, it's a huge return on investment without having to build a single new transmission line, right? And so when you think about the cost of doing that compared to the cost of simply putting better intelligence into the existing system, it's a massive return on investment.

AI assessment note: “determined that you could unlock in excess of six gigawatts of extra capacity”

Q And do you have any numbers behind what we could potentially unlock?

A Uh, Well, yeah, in fact, there was a great RMI study that just came out, uh, very recently that looked at the PJM system and determined that you could unlock in excess of six gigawatts of extra capacity, and this is, I mean, PJM's a 90 gigawatt system, so six gigawatts is not nothing. That's pretty substantial, um, and, and this is for an investment of about a hundred million dollars that would return annually in excess of a billion dollars To rate payers. So, I mean, it's a huge return on investment without having to build a single new transmission line, right? And so when you think about the cost of doing that compared to the cost of simply putting better intelligence into the existing system, it's a massive return on investment.

AI assessment note: “determined that you could unlock in excess of six gigawatts of extra capacity”

Q that aren't necessarily designed to equip some of these, uh, peak loads. Others say AI is just different, you know, it's going to require exponentially more energy. Where do you, each of you fall on this spectrum of concern? Like, as people come to this and try to figure out just how substantial is this issue? Um, uh, Brian, we'll go to you. And then we'll go down the line.

A Yeah, as I, as I sort of alluded to before, I do think it's very real. And I do think that, you know, probably the most extreme projections are overblown. Um, I think on the other end, you know, the ones that I've seen on the low end, I think are, are pretty, pretty off base as well. Um, we've got a great indicator of investment and growth in this space, and it's the capex expenditure for all the big cloud companies. And if you look at that quarter over quarter, year over year, They are all accelerating investment. It's a great leading indicator of how much power, how many chips they're going to plug in in the future, um, and they're all betting really big on very large AI models, and so that's maybe one of the differences between AI and what we've seen in sort of native cloud is that there is very much an incentive to use more electricity. There's incentive to build bigger models to plug in More GPUs. And, you know, I had someone ask me the other day, well, what happens if, you know, if NVIDIA comes out with a chip that's twice as efficient? And my response to that is, well, then Meta or Microsoft or Google will plug in twice as many chips. If they have the power, they're going to build as big of a model as they can build. And so that's where I think AI is a little different than what we've seen. Not to say there's not incentives for efficiency, because there very much is, But…

AI assessment note: “I do think it's very real. And I do think that, you know, probably the most extreme projections are overblown.”

Q of the spectrum, you have a lot of data center stalwarts who say, look at all the amazing innovations we've had in, in the efficiency of computing and the build out of these data centers and data center energy demand wasn't nearly what we thought it would be from the nineties, um, into the 2000, just because we got better at building data centers. Where are you in that spectrum?

A Yeah, I, I'm, I guess I'm, I'm probably a little bit in the middle. Um, I, I definitely think that we over-index on counting on, quote, energy miracles, because, again, going back to what we were talking about before, it takes a long time to build this stuff. So we are not going to have significant contribution of, say, nuclear energy, certainly not fusion, that's going to meet the challenge that we're talking about right now, which is we've got to build a lot of stuff, like, this year and next year and the year after that, right? So we have a lot that has to be done between now and then, right? I definitely am a huge proponent of that, and we should keep investing and focusing on that, because we will need that in the coming decades. I think we just have to be realistic about what we can achieve in the near term. I'm also skeptical of the idea that efficiency is going to save us. In fact, I, I've heard that a ton over the years, and, you know, my response to that has always been, the only time that efficiency is going to Reduce the input needed for any process is when you're starting to exhaust the demand for that resource, right? And, and when you think about something like data, we're nowhere close to satiating human demand for data, which means that if you make it more efficient, you just accelerate up that S-curve faster, and you actually use more of it, not less of it. An…

AI assessment note: “I'm probably a little bit in the middle. Um, I, I definitely think”

Q So then how are you actually mapping out projects?

A The opportunity set in front of us is sit down with the utility, understand, you know, where are the points on your system where you think there might be Opportunities to deliver a significant amount of power capacity. It's really the inverse of what you do when you're doing renewable development, right? You're looking for what are pots, the spots on the grid I can actually connect and inject power in a significant amount of power. We're doing the reverse to say, okay, what are the points on the grid where you can pull a significant amount of power out? And then to dive in further and say, okay, well, what are the constraints that we're seeing? Like you say, okay, we think there's 300 megawatts here. Okay, well, what would it take for that to be 500 megawatts Or a gigawatt. Like, what sort of investments would you have to undertake? You, thinking about things like grid-enhancing technologies. Thinking about how we use storage as a, as a flexible resource. Because the, the thing that I think people that work in this industry understand, but most people don't understand, is when a utility says that they don't have power, they don't mean that most of the time there's not power. They just mean there might be a handful of hours a year where there's not power. So they can't commit To, to delivering 24 seven, three 65 a year. And so the questions then become, okay, how can we take the…

AI assessment note: “sit down with the utility, understand, you know, where are the points on your system”

Q So then how are you actually mapping out projects?

A The opportunity set in front of us is sit down with the utility, understand, you know, where are the points on your system where you think there might be Opportunities to deliver a significant amount of power capacity. It's really the inverse of what you do when you're doing renewable development, right? You're looking for what are pots, the spots on the grid I can actually connect and inject power in a significant amount of power. We're doing the reverse to say, okay, what are the points on the grid where you can pull a significant amount of power out? And then to dive in further and say, okay, well, what are the constraints that we're seeing? Like you say, okay, we think there's 300 megawatts here. Okay, well, what would it take for that to be 500 megawatts Or a gigawatt. Like, what sort of investments would you have to undertake? You, thinking about things like grid-enhancing technologies. Thinking about how we use storage as a, as a flexible resource. Because the, the thing that I think people that work in this industry understand, but most people don't understand, is when a utility says that they don't have power, they don't mean that most of the time there's not power. They just mean there might be a handful of hours a year where there's not power. So they can't commit To, to delivering 24 seven, three 65 a year. And so the questions then become, okay, how can we take the…

AI assessment note: “sit down with the utility, understand, you know, where are the points on your system”

Q of the spectrum, you have a lot of data center stalwarts who say, look at all the amazing innovations we've had in, in the efficiency of computing and the build out of these data centers and data center energy demand wasn't nearly what we thought it would be from the nineties, um, into the 2000, just because we got better at building data centers. Where are you in that spectrum?

A Yeah, I, I'm, I guess I'm, I'm probably a little bit in the middle. Um, I, I definitely think that we over-index on counting on, quote, energy miracles, because, again, going back to what we were talking about before, it takes a long time to build this stuff. So we are not going to have significant contribution of, say, nuclear energy, certainly not fusion, that's going to meet the challenge that we're talking about right now, which is we've got to build a lot of stuff, like, this year and next year and the year after that, right? So we have a lot that has to be done between now and then, right? I definitely am a huge proponent of that, and we should keep investing and focusing on that, because we will need that in the coming decades. I think we just have to be realistic about what we can achieve in the near term. I'm also skeptical of the idea that efficiency is going to save us. In fact, I, I've heard that a ton over the years, and, you know, my response to that has always been, the only time that efficiency is going to Reduce the input needed for any process is when you're starting to exhaust the demand for that resource, right? And, and when you think about something like data, we're nowhere close to satiating human demand for data, which means that if you make it more efficient, you just accelerate up that S-curve faster, and you actually use more of it, not less of it. An…

AI assessment note: “I'm probably a little bit in the middle. Um, I definitely think that we over-index”

Q Brian, how are you seeing some of the supply agreements change? We have a question here about the Amazon Talon deal. We have a question also about onsite generation and how data centers are thinking about, uh, differently about procurement. Like, how are you seeing procurement and agreements change?

A Well, I mean, I think the Amazon Talon deals is, it's sort of a corollary to what John's talking about, where, you know, what you're doing is saying, Go, go to where the power is. And, and in John's case, he's talking about go to where the clean power is. Um, here you're just saying go to where the capacity is, right? And you could argue whether it's clean or not. I mean, it's, it's nuclear, it's zero carbon. It's not net new, so it's not bringing new capacity onto the grid. Um, but it's really about, you know, a shift in the industry from really a focus on, if you look at the history of the cloud industry, it started with a focus on fiber, right? Get as close to the network as possible. That's why you have concentrations in places like Northern Virginia and Amsterdam, because those were some of the biggest network hubs in the world. Um, then it really became about GDP and eyeballs. Get as close to the people as you can. Uh, now it's really about power. It's get to where the power is, because as John said, there's a convergence between, you know, energy and data. If you think about what data is, it's really just electricity turned into light. It is a form of just energy conversion, right? And so, There's this, been this realization over the last, you know, 18 to 24 months that this whole business is really about getting access to power. And of course, you know, access to clean …

AI assessment note: “what you're doing is saying, Go, go to where the power is.”

Q And if we look at grid enhancing technologies, like the portfolio of grid enhancing technologies, what are they generally? What are the other technologies that could help us expand existing capacity?

A Yeah, there's a, there's a few different things. I mean, dynamic line rating is one. Um, uh, you also have, uh, different forms of, of transmission optimization. Think more like, um, you know, like ways, for instance, where you can actually control traffic, you know, to, one thing is, you know, Even knowing that there's congestion in the system, but the other is, can we start to redirect power flows over different parts of the system, right? So there's, there's different ways that we can use these technologies and, and use them in a combined way as well, because they all sort of do different things, but as you put more intelligence on that system, you can incrementally start to eke out more capacity where, where otherwise you would think there's not any.

AI assessment note: “dynamic line rating is one. Um, uh, you also have, uh, different forms”

Q And if we look at grid enhancing technologies, like the portfolio of grid enhancing technologies, what are they generally? What are the other technologies that could help us expand existing capacity?

A Yeah, there's a, there's a few different things. I mean, dynamic line rating is one. Um, uh, you also have, uh, different forms of, of transmission optimization. Think more like, um, you know, like ways, for instance, where you can actually control traffic, you know, to, one thing is, you know, Even knowing that there's congestion in the system, but the other is, can we start to redirect power flows over different parts of the system, right? So there's, there's different ways that we can use these technologies and, and use them in a combined way as well, because they all sort of do different things, but as you put more intelligence on that system, you can incrementally start to eke out more capacity where, where otherwise you would think there's not any.

AI assessment note: “dynamic line rating is one. Um, uh, you also have, uh, different forms”

Q behind the meter, clean energy, um, grid enhancing technologies. Like, you know, how do you think about the mixing and matching of these technologies and like what's stand out as the most impactful? Brian, we'll go to you. I know it's like very site specific and how you're developing a particular site. How do you think about, like, the sets of technologies and what is most available and impactful today?

A Yeah, and, and it is very site-specific and location-specific, and it's, it's application-specific as well, as John was alluding to, with things like dispatchable, you know, workloads, which, which is not the sum total of all workloads, right, as he was noting. There are workloads that do still need twenty-four-seven uptime, but there's a lot of different ways to achieve that, and I think one of the challenges we have in the industry right now is You know, most of these data center operators are very desperate for capacity, so they're going to utilities. They're asking for firm service. They want it delivered twenty-four-seven, and then that's then defaulting utilities into thinking about, okay, well, I guess we can build some gas plants to solve that, rather than really peeling back the onion and looking at what are all the solutions that we have and all the tools that we have at our disposal, which includes things like grid-enhancing technologies, getting more out of the existing system, As Michelle was saying, um, how do we use storage more intelligently across the system to create more flexibility? I mean, ultimately, it's about how do we sort of even mimic flexible loads? Even if I'm a data center and I need twenty-four-seven uptime, there's a lot of resources that data centers already have on their sites. They have generators. They have batteries. Those, those can be opti…

AI assessment note: “tools that we have at our disposal, which includes things like grid-enhancing technologies”

Q behind the meter, clean energy, um, grid enhancing technologies. Like, you know, how do you think about the mixing and matching of these technologies and like what's stand out as the most impactful? Brian, we'll go to you. I know it's like very site specific and how you're developing a particular site. How do you think about, like, the sets of technologies and what is most available and impactful today?

A Yeah, and, and it is very site-specific and location-specific, and it's, it's application-specific as well, as John was alluding to, with things like dispatchable, you know, workloads, which, which is not the sum total of all workloads, right, as he was noting. There are workloads that do still need twenty-four-seven uptime, but there's a lot of different ways to achieve that, and I think one of the challenges we have in the industry right now is You know, most of these data center operators are very desperate for capacity, so they're going to utilities. They're asking for firm service. They want it delivered twenty-four-seven, and then that's then defaulting utilities into thinking about, okay, well, I guess we can build some gas plants to solve that, rather than really peeling back the onion and looking at what are all the solutions that we have and all the tools that we have at our disposal, which includes things like grid-enhancing technologies, getting more out of the existing system, As Michelle was saying, um, how do we use storage more intelligently across the system to create more flexibility? I mean, ultimately, it's about how do we sort of even mimic flexible loads? Even if I'm a data center and I need twenty-four-seven uptime, there's a lot of resources that data centers already have on their sites. They have generators. They have batteries. Those, those can be opti…

AI assessment note: “They have generators. They have batteries. Those, those can be optimized”

Q that's to come, so we can talk a little bit about that. Brian, over to you. You worked at Microsoft for more than a decade. You were procuring clean energy to meet corporate operations and data centers. How did that become more challenging after the public release of ChatGPT? Like, how dramatically did that increase computing needs, and A year and a half on, how has that challenge only increased?

A Yeah, so I think it happened at a very unique time, because if you look back to the prior decade, you know, when you were referencing all these stories, we were talking about how the internet was going to gobble up all the world's power. There was, there was something happening during that decade of sort of the founding of the cloud. If you go back to say the early days around 20 10, there was still a lot of computation that was happening on premise. So there was a period during the, around 20 10 to 20 20, where data centers On the whole had relatively flat load growth, but it wasn't because there wasn't significant load growth happening sort of under the hood. What was happening is you had a lot of businesses shutting down their own very inefficient data centers that were using multiples of the amount of electricity that a hyperscale data center would use. And the hyperscale data center could absorb that with, with significantly less, uh, electricity consumption. For, for more computation, right? So you had a significant increase in computation over that period of time with almost no increase in electricity, but a lot of people attribute that to efficiencies. Well, data centers are just getting more efficient, but really it was happening. We're just moving environments. So there was a form of efficiency happening there and that you're shutting down old, very inefficient, you k…

AI assessment note: “I think it happened at a very unique time, because if you look back”

Q And knowing what you know, Brian, about how to serve these pockets of demand, does it feel like we can serve them without building? A lot of this new gas?

A I think we can. I, I think there is a lot of, a lot of levers that we have at our disposal, you know, and I have, I've had empathy for my friends in the southeast. I know that they're under a lot of pressure, and I know that, you know, they, they're at the tip of the spear here because, you know, not only are they seeing a lot of demand from data center growth, they're actually where we're seeing a lot of demand for factories as well. So, you know, they're all under a tremendous amount of pressure to deliver, you know, for their shareholders, for their states, As much as they can, um, but there are other levers that we have, and there are ways to, to manage this, and it, it just, it's going to take a little bit of time to really work through what are these solutions, and how can we deliver them at scale, but I am optimistic that we will be able to create pathways where we can continue to grow the grid, and we can do it in a way that, that is truly sustainable.

AI assessment note: “I think we can. I, I think there is a lot of, a lot of levers”

Q And knowing what you know, Brian, about how to serve these pockets of demand, does it feel like we can serve them without building? A lot of this new gas?

A I think we can. I, I think there is a lot of, a lot of levers that we have at our disposal, you know, and I have, I've had empathy for my friends in the southeast. I know that they're under a lot of pressure, and I know that, you know, they, they're at the tip of the spear here because, you know, not only are they seeing a lot of demand from data center growth, they're actually where we're seeing a lot of demand for factories as well. So, you know, they're all under a tremendous amount of pressure to deliver, you know, for their shareholders, for their states, As much as they can, um, but there are other levers that we have, and there are ways to, to manage this, and it, it just, it's going to take a little bit of time to really work through what are these solutions, and how can we deliver them at scale, but I am optimistic that we will be able to create pathways where we can continue to grow the grid, and we can do it in a way that, that is truly sustainable.

AI assessment note: “I think we can. I, I think there is a lot of levers”

Q not adding tens of megawatts, but now gigawatts of demand, if manufacturing facilities are doing the same, we're seeing roughly the same type of demand in manufacturing, and we have all these EVs hitting the grid, um, We need to build quickly to service all this new demand, and what's that infrastructure challenge? How much do we need to expand the grid itself to accommodate integrating all those new resources?

A Well, we are going to have to build a lot more infrastructure. We're going to have to build a lot more transmission lines, a lot more power plants, but we also have to be realistic about how hard that, that stuff is. Uh, so I, I just started listening to a fantastic podcast called The Big Dig. Uh, a lot of people are familiar with the Boston tunnel project, um, that was, you know, took way too long and was way over budget. Um, but what I find so fascinating about it is, is so much of it has been about just the on the ground politics and the community engagement and, like, all the stuff that makes building stuff really hard. It's like, you look at these things and like, well, why did this take so long? It's like, well, there's like million reasons why it took so long, right? And, and that's the kind, that's the stuff that I think people vastly underestimate. And so, Yes, we need to be building new things, but we also have to be realistic about the fact that we're going to have to find growth inside the existing system too. We have to find ways to better optimize the system we have and get more out of the system we have rather than just counting on, well, we're just going to build a bunch more power plants because we will eventually, but we've got a lot we need to do between now and then in terms of meeting load growth, interconnecting Customer requests, et cetera.

AI assessment note: “We're going to have to build a lot more transmission lines, a lot more power plants”

Q that you, uh, have co-founded a company called Cloverleaf Infrastructure. You're taking a lot of your understanding of site development and procurement from your days at Microsoft, and, um, combining this with your understanding of all these different technologies, both on the generation side, on the grid enhancing technology side, and putting it together to help utilities figure out how to Manage this load growth. What are you guys doing?

A Yeah. So really excited about the announcement. Um, this is something that's been in the works for, for a little while for us. And it goes back to what you said is, you know, my passion is around how do we grow the grid? How do we meet these customer demands? Uh, and I, I saw firsthand all the challenges of doing that, especially with the burden of having to deliver so quickly, um, and being able to step out of that for a little while and start to think about how can we How can we really be more intentional in working with utilities with the, from a customer perspective to say, let's create a, a better runway for customers that, that come to you and are going to start asking you for increasingly large interconnects. You know, can, can we actually make that process easier where utilities can start to, um, better understand the investments they can make on their system, how those investments would be, will be received automatically. Buy those future customers, um, and really just provide them a signal into what's coming, um, because we've, we've seen a bit of a disconnect there, too, in that there's not always a great signal from these companies as to what they need. Oftentimes, they don't even know. Uh, going back to the conversation we were having about ChatGPT, you know, no one was fully prepared for the scale and the speed that that was happening, and again, if technology is …

AI assessment note: “let's create a, a better runway for customers that, that come to you”

Q not adding tens of megawatts, but now gigawatts of demand, if manufacturing facilities are doing the same, we're seeing roughly the same type of demand in manufacturing, and we have all these EVs hitting the grid, um, We need to build quickly to service all this new demand, and what's that infrastructure challenge? How much do we need to expand the grid itself to accommodate integrating all those new resources?

A Well, we are going to have to build a lot more infrastructure. We're going to have to build a lot more transmission lines, a lot more power plants, but we also have to be realistic about how hard that, that stuff is. Uh, so I, I just started listening to a fantastic podcast called The Big Dig. Uh, a lot of people are familiar with the Boston tunnel project, um, that was, you know, took way too long and was way over budget. Um, but what I find so fascinating about it is, is so much of it has been about just the on the ground politics and the community engagement and, like, all the stuff that makes building stuff really hard. It's like, you look at these things and like, well, why did this take so long? It's like, well, there's like million reasons why it took so long, right? And, and that's the kind, that's the stuff that I think people vastly underestimate. And so, Yes, we need to be building new things, but we also have to be realistic about the fact that we're going to have to find growth inside the existing system too. We have to find ways to better optimize the system we have and get more out of the system we have rather than just counting on, well, we're just going to build a bunch more power plants because we will eventually, but we've got a lot we need to do between now and then in terms of meeting load growth, interconnecting Customer requests, et cetera.

AI assessment note: “we're going to have to find growth inside the existing system too”

Q that you, uh, have co-founded a company called Cloverleaf Infrastructure. You're taking a lot of your understanding of site development and procurement from your days at Microsoft, and, um, combining this with your understanding of all these different technologies, both on the generation side, on the grid enhancing technology side, and putting it together to help utilities figure out how to Manage this load growth. What are you guys doing?

A Yeah. So really excited about the announcement. Um, this is something that's been in the works for, for a little while for us. And it goes back to what you said is, you know, my passion is around how do we grow the grid? How do we meet these customer demands? Uh, and I, I saw firsthand all the challenges of doing that, especially with the burden of having to deliver so quickly, um, and being able to step out of that for a little while and start to think about how can we How can we really be more intentional in working with utilities with the, from a customer perspective to say, let's create a, a better runway for customers that, that come to you and are going to start asking you for increasingly large interconnects. You know, can, can we actually make that process easier where utilities can start to, um, better understand the investments they can make on their system, how those investments would be, will be received automatically. Buy those future customers, um, and really just provide them a signal into what's coming, um, because we've, we've seen a bit of a disconnect there, too, in that there's not always a great signal from these companies as to what they need. Oftentimes, they don't even know. Uh, going back to the conversation we were having about ChatGPT, you know, no one was fully prepared for the scale and the speed that that was happening, and again, if technology is …

AI assessment note: “create a better runway for customers that come to you and are going to start asking”

Q that's to come, so we can talk a little bit about that. Brian, over to you. You worked at Microsoft for more than a decade. You were procuring clean energy to meet corporate operations and data centers. How did that become more challenging after the public release of ChatGPT? Like, how dramatically did that increase computing needs, and A year and a half on, how has that challenge only increased?

A Yeah, so I think it happened at a very unique time, because if you look back to the prior decade, you know, when you were referencing all these stories, we were talking about how the internet was going to gobble up all the world's power. There was, there was something happening during that decade of sort of the founding of the cloud. If you go back to say the early days around 20 10, there was still a lot of computation that was happening on premise. So there was a period during the, around 20 10 to 20 20, where data centers On the whole had relatively flat load growth, but it wasn't because there wasn't significant load growth happening sort of under the hood. What was happening is you had a lot of businesses shutting down their own very inefficient data centers that were using multiples of the amount of electricity that a hyperscale data center would use. And the hyperscale data center could absorb that with, with significantly less, uh, electricity consumption. For, for more computation, right? So you had a significant increase in computation over that period of time with almost no increase in electricity, but a lot of people attribute that to efficiencies. Well, data centers are just getting more efficient, but really it was happening. We're just moving environments. So there was a form of efficiency happening there and that you're shutting down old, very inefficient, you k…

AI assessment note: “it happened at a very unique time, because if you look back to the prior decade”