The Exchanges

Q high enough reliability to what you want, so you also put a UPS on site, uh, which just bridges seconds to minutes of power outages, basically, and then you generally put backup generators on site as well, which are supposed to You know, fill in the, the blanks where you have longer outages. So that architecture, grid connection, UPS, backup genset, that's the kind of basic, like, dominant paradigm, right?

A Correct. Yeah. Particularly for your traditional cloud data centers, you'll see that. Um, I think with some of the AI training sites, we've seen it's more of a move away from backup generators. Uh, some of that is in part because of that bad change. Like, they could handle an outage, um, if it, if it ever happened. And keep in mind, like, the, the outages we're talking about that the generators there to protect are pretty rare. Because we're talking about these sites being connected at very high voltages On the transmission system. So, you know, we're talking about like, you know, winter storm Yuri sort of events that you're really concerned about. Um, so in that case, both for that reason, and I think out of necessity, because especially if you're talking about these gigawatt scale sites we're seeing, you're not getting diesel generators permitted at that sort of scale anyway.

AI assessment note: “Correct. Yeah. Particularly for your traditional cloud data centers, you'll see that.”

Q Can we talk about a region, a region sort of an, I've, I've come to learn from you mostly like what a region means in the context of a A single hyperscaler or a single customer, but like, why do you think in regions, and what is a region?

A Sure, a region is just what the cloud applications look like to the outside world. So if you go on to AWS, you go on to Azure, and you say, I want to stand up an application, you're going to have an option to, to select a region, and it'll be called US East or US West too. Um, what that really is a designation of a cluster of data centers that are all within a certain latency Uh, envelope of one another. So Northern Virginia is a great example. AWS has their largest region, which is their U.S. East region. That is made up of, at this point, probably dozens of data centers. But the outside world, it looks like one big machine. Um, so that's what a region is. And so those regions kind of, and the reason that you have places like Northern Virginia, Amsterdam, those were where the biggest network hubs were. So everyone clustered around there initially as Everyone was launching some of their early cloud regions, and then over time, you know, Microsoft, Amazon, Google all have dozens of regions around the world, but that dozens of regions consists of hundreds of individual data centers.

AI assessment note: “what that really is a designation of a cluster of data centers”

Q Alright, so then now in, in the new world, um, which by the way still has all those same cloud data centers and cloud requirements, it's still growing, but what we're adding on to it is, is this AI world wherein there, it's important, I think, in the context of whether you can introduce any flexibility here to distinguish model training and inference, right?

A Yes, and that's one that comes up a lot is this idea of training models being curtailable, because they actually fit in that category of, uh, sort of indexing that I was talking about previously, which is that's sort of a batch process. So training models in and of themselves are batch processes, and so in theory they could actually shut down during certain periods. Um, that would not be true for inferencing, but so inferencing is, is a lot More akin to the search function. So if you go into a chat GPT and you want to, you want to make your picture, you want it to tell you a story, um, you want that to happen very quickly. So those, those applications are still going to require a very high availability similar to what, what you would expect in normal cloud applications. Um, so I, I think, I think it's a little bit overblown to say, and I think this is also true of Something like the conversation around crypto, but that these are highly curtailable loads that you could just, you know, attach a training model to a wind farm and only run it, you know, on average, 35% of the time. Nobody's going to do that, because the cost of that, of that infrastructure, the server is extraordinarily high. So you still want to get very high utilization out of those assets. So they're not, they're not, maybe you can avoid Significant contributions to things like system peak for, you know, a few ho…

AI assessment note: “training models in and of themselves are batch processes... that would not be true for inferencing”

Q Yeah. Okay. Wow. All right. So, 20 years. Um, paint me the really quick picture in your perspective of how we got here. Like, why did we land in this place where, where there is this absolutely enormous upwelling of opposition?

A Yeah. So, it's, it's interesting because, you know, we've, I mean, you know, we've been building data centers at, at some scale for 20 years in this country. Even a few years ago, you know, you were getting the red carpet rolled out for you when you'd go meet with Communities and cities, and we've somehow, in a relatively short time, gone from, you know, red carpet to pitchforks, you know, and it, it, and it wasn't, there wasn't even, like, a whole break in between, uh, you know, it was really fast, and so I, I think there's a number of things going on. I mean, one, of course, like, what we're building today is much bigger, you know, in one fell swoop than what we built 15 years ago. Now, if you look at places like Quincy, Washington, for instance, Microsoft probably has close to a gigawatt of data centers in Quincy Washington. I don't know the exact number, but it's, it's gotta be sort of in that ballpark, but it's happened over a period. I think they built that first data center in 2007, right? So almost 20 years, you know, they've gone to a gigawatt and there was a fantastic story just in the last week about Quincy Washington building a fifteen million dollar aquatic center and they have a hundred and fifty million dollar state of the art school and their unemployment rate has gone from like 29% to six percent. You know, so they've, they've been able over a period of 20 year…

AI assessment note: “what we're building today is much bigger, you know, in one fell swoop”

Q Or less, right? I mean, on a 24 seven basis, right? Like I just, I just was talking to a data center operator who said that their average actual utilization relative to nameplate capacity is like 40 to 50% over the course of a year.

A And so, you know, you can quickly do the math on if I'm paying 2700 dollars to KW, just roughly for that generation, and I'm having to overbuild it by, you know, maybe even two X, right? The, the per, Kilowatt hour cost of that system is extraordinarily high. I mean, extraordinarily high. Uh, and if you look at it in a place like Texas, for instance, where the, the average price for electricity on any given day is actually pretty low. Like the, the real time price may be sitting around 20 dollars a megawatt hour. So you go off grid in Texas and you're paying somewhere between a 150 to 200 dollars a megawatt hour, 24 seven, and your neighboring data center connected to the grid is paying 20 dollars for that same power. Now I'm leaving out the T and D You know, I mean, there's stuff on top of that, but you know, the, the average cost of electricity in the market in Texas is pretty cheap most of the time. Um, and the only argument, you know, you ever had for building something like a baseload generator in Texas is that sometimes the price would go to 5000 or 9000 dollars a megawatt hour, but with the massive amounts of solar and storage coming out of the grid, which you've probably talked about in another show, we're not seeing those spikes anymore. We're not seeing the scarcity of pricing, uh,

AI assessment note: “you can quickly do the math on if I'm paying 2700 dollars to KW”

Q concepts like this, there is a bunch of consternation from the customer set about the cost that they're, that's being proposed to To them. Do you think that's just, are they just positioning ultimately? And like, if you, if you forced it, they would cave or is, or have they not woken up? Like has the customer not woken up to the actual scale of the watt bit spread yet?

A I think this, I think a little bit of that is positioning. I think a little bit of is sort of maybe just a lack of, of really understanding the nature of, of building out large infrastructure. Again, There's a lot of players in this space that aren't particularly, uh, sophisticated when it comes to energy systems and utility rates and regulations. There are a lot that are. I mean, a lot of, especially the big tech companies have, you know, really extraordinarily talented energy teams and, you know, know how to work with utilities and regulators. Um, but, you know, again, you've got, if you've got some company that, you know, was going to operate a Bitcoin mine or, They were going to do green hydrogen project and now they decided, no, I'm, I'm actually a AI data center company now. It's like, okay, well, you know, when you start getting questions from utilities about, um, the, the cost of this infrastructure, sometimes people's eyes start to bug out because they see things that end in billions and they're like, whoa, I'm, I'm kind of getting in a lot deeper than I thought. But I think that just shows a, a lack of understanding, um, Of the overall value of that electricity. Yes, the cost is very high when we're talking about several hundred megawatts or multiple gigawatts of power. But again, the revenue on the other side, or even, even looking at the Overall CapEx deployment. Um…

AI assessment note: “I think a little bit of that is positioning. I think a little bit of is”

Q You also, when you, you put together this watt-bit spread piece, described it in terms of manufacturing theory in a way that I thought was useful to think about how different actors in this equation are thinking of it in different economic terms. So can you walk through the, like, what, who is operating how?

A Yeah, so Um, in, in terms of manufacturing, there's sort of two different models. Uh, one is sort of lean manufacturing, which says hold as little inventory as possible. You know, everything's about, you know, just in time. Um, and then if, if lead times extend to deliver, that's fine. The customers will just have to wait, but we're going to, you know, improve margins by not having a bunch of excess inventory. Um, there's another model, which is the theory of constraints, which says, no, what you need to focus on is throughput. That's how you create You know, enterprise value, and therefore you need to find areas of the supply chain, uh, or the manufacturing process that, that tend to become constrained and make sure that you, you're always building extra capacity. So more labor, more, you know, overhead, uh, for those particular points of the system. So if, if I'm downstream in the system, so I'm data center operators, I'm cloud providers, I'm NVIDIA, I'm open AI, uh, I'm very much thinking about the world in terms of theory of constraints, because again, the cost of the base infrastructure for my end product is actually really low. And when I say base infrastructure, I talk about land, you know, power, um, a data center shell, you know, those things are relatively inexpensive. In the grand scheme of operating, you know, my business. So I would be willing to pay a premium for …

AI assessment note: “I'm very much thinking about the world in terms of theory of constraints”

Q So you were there for the emergence of ChatGPT and, you know, the first kind of, I don't know, 18 months or something after that. So what was happening? Your, your job was to procure energy for Microsoft broadly, but including for data centers. Um, what was it like that last. 18 months, 24 months, like how did things change how quickly and like, what was the dynamic?

A Yeah, the change was pretty rapid. I think we all knew that Microsoft was working with OpenAI and that there was this whole movement towards AI potentially coming at some point in the future, but I don't think there was a realization, and if I'm going back to probably the summer of twenty-twenty-two, this was before ChatGPT-III was released, I don't think there was really a realization of the The scale of the magnitude of what the technology was going to do and how fast it was going to be adopted. And so we started hearing rumblings of it that summer, and I started getting some odd questions that I'd never gotten before about the scale of certain data centers and how big could you make a data center. And then it, once GPT-III was released in November, that's when it really started to sink in. And then I think the, the next level of, of realization was the next spring when 3.5 came out and you saw this massive leap When on really what, what I thought of as a sort of a half click, wasn't even the full click to GPT-IV, but just a half click. And that's the moment that I realized that the problem we were going to have was going to be whether we would have enough power to support this technology that was moving at a pace that, as you know, moves way faster than the electric utility industry.

AI assessment note: “the problem we were going to have was going to be whether we would have enough power”

Q some locations, right? Because Data centers are clustered in certain regions, and you also want to be able to get, ah, low latency to, like, highly populated areas, and so it was already, if you were in Singapore or London or something like that, it was, power was already a constraint, right? And so what you were realizing was that it was going to be a broader constraint than that?

A Yeah, it was really an accelerant, because I would say in the maybe two years prior to that, we had really shifted our strategy towards a power Sort of power first approach to how we thought about siting, because over that decade, we had gone from a baseline of, you know, a couple hundred megawatts, and we were growing at a, at a relatively rapid pace, but that's also kind of a small denominator. So the incremental tranches every year that we had to go procure were measured in the tens of, or maybe hundreds of megawatts. Towards the end of, of the last decade, That denominator kept growing, and now the denominator was in gigawatts. And the concern we started to realize is that the tranche size, even if we're still growing at the same rate, those tranche sizes are quite large. And so we already were starting to be concerned about the fact that we were out there looking for hundreds of megawatts, if not gigawatts, on a yearly basis, and then you add AI into the equation. And that's when we really started to realize that there's a, there's going to be A significant challenge, not just for Microsoft, but the entire industry, and not just the AI cloud industry, but the entire electric utility industry to support the continued growth of native cloud applications on top of now this new and, and somewhat uncertain trajectory of where AI was going to go.

AI assessment note: “Yeah, it was really an accelerant, because I would say in the maybe two years”

Q others as well. What do you make of all of that? Like, how much of that do you think might be speculative? How much is real? And just, what's your broad perspective on, on, like, we're, we're clearly in a moment, um, how should we be thinking about the actual expected load if, if, uh, power were not a constraint? And then we'll talk about what constraints power does add.

A Yeah, it's certainly the forecasts you've seen that are based on sort of those AEP numbers of requests they have in queue. If you add all those up, that's way above the actual demand, because there's a lot of zombie requests, just like in the generation queue. We're not going to build all the projects that are in generation queue right now. A bunch of those are just developers that, that hope to be able to build a project, but there wouldn't be enough demand to support every project in the queue that we have right now. Both on the supply side and the demand side. So the sum total of the, you know, the things you've seen that the sky is falling and AEP needs 90 gigawatts. There's no reason to believe that AEP is going to connect 90 gigawatts in the next 10 years. Um, There is still a very large demand inside, and I'll use AEP specifically, and there's a reason why we're talking about AEP, and I'll get to that in just a second, but there is still a extraordinarily large amount of demand inside of that system, and it is in excess of 10 gigawatts, almost certainly. Um, the reason AEP has gotten a lot of attention is they operate one of the highest voltage Systems in the country. So they operate a seven 65 KV system, uh, which is, you know, higher than what anyone else has. There's a little bit of seven 65 in New York as well. Um, but if you track what's been happening in terms of a…

AI assessment note: “that's way above the actual demand, because there's a lot of zombie requests”

Q others as well. What do you make of all of that? Like, how much of that do you think might be speculative? How much is real? And just, what's your broad perspective on, on, like, we're, we're clearly in a moment, um, how should we be thinking about the actual expected load if, if, uh, power were not a constraint? And then we'll talk about what constraints power does add.

A Yeah, it's certainly the forecasts you've seen that are based on sort of those AEP numbers of requests they have in queue. If you add all those up, that's way above the actual demand, because there's a lot of zombie requests, just like in the generation queue. We're not going to build all the projects that are in generation queue right now. A bunch of those are just developers that, that hope to be able to build a project, but there wouldn't be enough demand to support every project in the queue that we have right now. Both on the supply side and the demand side. So the sum total of the, you know, the things you've seen that the sky is falling and AEP needs 90 gigawatts. There's no reason to believe that AEP is going to connect 90 gigawatts in the next 10 years. Um, There is still a very large demand inside, and I'll use AEP specifically, and there's a reason why we're talking about AEP, and I'll get to that in just a second, but there is still a extraordinarily large amount of demand inside of that system, and it is in excess of 10 gigawatts, almost certainly. Um, the reason AEP has gotten a lot of attention is they operate one of the highest voltage Systems in the country. So they operate a seven 65 KV system, uh, which is, you know, higher than what anyone else has. There's a little bit of seven 65 in New York as well. Um, but if you track what's been happening in terms of a…

AI assessment note: “If you add all those up, that's way above the actual demand”

Q Alright, so then now in, in the new world, um, which by the way still has all those same cloud data centers and cloud requirements, it's still growing, but what we're adding on to it is, is this AI world wherein there, it's important, I think, in the context of whether you can introduce any flexibility here to distinguish model training and inference, right?

A Yes, and that's one that comes up a lot is this idea of training models being curtailable, because they actually fit in that category of, uh, sort of indexing that I was talking about previously, which is that's sort of a batch process. So training models in and of themselves are batch processes, and so in theory they could actually shut down during certain periods. Um, that would not be true for inferencing, but so inferencing is, is a lot More akin to the search function. So if you go into a chat GPT and you want to, you want to make your picture, you want it to tell you a story, um, you want that to happen very quickly. So those, those applications are still going to require a very high availability similar to what, what you would expect in normal cloud applications. Um, so I, I think, I think it's a little bit overblown to say, and I think this is also true of Something like the conversation around crypto, but that these are highly curtailable loads that you could just, you know, attach a training model to a wind farm and only run it, you know, on average, 35% of the time. Nobody's going to do that, because the cost of that, of that infrastructure, the server is extraordinarily high. So you still want to get very high utilization out of those assets. So they're not, they're not, maybe you can avoid Significant contributions to things like system peak for, you know, a few ho…

AI assessment note: “Yes, and that's one that comes up a lot is this idea of training models”

Q some locations, right? Because Data centers are clustered in certain regions, and you also want to be able to get, ah, low latency to, like, highly populated areas, and so it was already, if you were in Singapore or London or something like that, it was, power was already a constraint, right? And so what you were realizing was that it was going to be a broader constraint than that?

A Yeah, it was really an accelerant, because I would say in the maybe two years prior to that, we had really shifted our strategy towards a power Sort of power first approach to how we thought about siting, because over that decade, we had gone from a baseline of, you know, a couple hundred megawatts, and we were growing at a, at a relatively rapid pace, but that's also kind of a small denominator. So the incremental tranches every year that we had to go procure were measured in the tens of, or maybe hundreds of megawatts. Towards the end of, of the last decade, That denominator kept growing, and now the denominator was in gigawatts. And the concern we started to realize is that the tranche size, even if we're still growing at the same rate, those tranche sizes are quite large. And so we already were starting to be concerned about the fact that we were out there looking for hundreds of megawatts, if not gigawatts, on a yearly basis, and then you add AI into the equation. And that's when we really started to realize that there's a, there's going to be A significant challenge, not just for Microsoft, but the entire industry, and not just the AI cloud industry, but the entire electric utility industry to support the continued growth of native cloud applications on top of now this new and, and somewhat uncertain trajectory of where AI was going to go.

AI assessment note: “there's going to be A significant challenge, not just for Microsoft, but the entire industry”

Q it's just very lucrative for the hyperscalers to run these data centers. Do you really think it's not, maybe, perhaps not to the extreme where you just, like, connect your, your training model to, you know, off-grid solar or whatever and operate a 20% capacity factor, but, but just say, like, can you, can you avoid system peak daily, potentially? Can you shut down for a couple hours a day?

A It's, it's possible. I think it goes back to that sort of economic motivation argument and utilization of the asset, and it's, it's akin to the conversation You know, I'm sure you've probably had on your show about things like, um, you know, hydrogen, you know, electrolyzers, right? Electrolyzers are really expensive. So you want to run them at a very high capacity factor. Um, and, and it's similar to how you would think about a nuclear plant. You're not going to take a nuclear plant and turn it on and off every day because it's very expensive from a CapEx standpoint. So therefore you want high utilization. So anything that has very high upfront CapEx Usually needs very high utilization to justify that capex, and so that's why I think you're not going to see, at least from an economic standpoint, a significant push towards having a lot of curtailable workloads. Now, you also noted the capacity issue and the availability issue, so there's two different things at play. One is, could I curtail this workload and, you know, avoid, you know, really high power prices in the afternoon? You know, if it's a really hot day in Texas, for instance. But the second thing is, The thing utilities are really grappling with, and we'll go back to AEP, if AEP has to connect 10 gigawatts of firm load that needs twenty-four-seven power, then it needs 10 gigawatts of new peak capacity to support their…

AI assessment note: “It's, it's possible. I think it goes back to that sort of economic motivation”

Q Can we talk about a region, a region sort of an, I've, I've come to learn from you mostly like what a region means in the context of a A single hyperscaler or a single customer, but like, why do you think in regions, and what is a region?

A Sure, a region is just what the cloud applications look like to the outside world. So if you go on to AWS, you go on to Azure, and you say, I want to stand up an application, you're going to have an option to, to select a region, and it'll be called US East or US West too. Um, what that really is a designation of a cluster of data centers that are all within a certain latency Uh, envelope of one another. So Northern Virginia is a great example. AWS has their largest region, which is their U.S. East region. That is made up of, at this point, probably dozens of data centers. But the outside world, it looks like one big machine. Um, so that's what a region is. And so those regions kind of, and the reason that you have places like Northern Virginia, Amsterdam, those were where the biggest network hubs were. So everyone clustered around there initially as Everyone was launching some of their early cloud regions, and then over time, you know, Microsoft, Amazon, Google all have dozens of regions around the world, but that dozens of regions consists of hundreds of individual data centers.

AI assessment note: “a designation of a cluster of data centers that are all within a certain latency”

Q Okay. First thing, um, when you're developing a new data center, campus land, whatever, when you're developing a new data center, what are the basic requirements from a power perspective? And there are a bunch of other requirements. What are the basic requirements that you have from a power perspective? What has to be true?

A What has to be true is you still have to have a very high level of availability of power. I mean, outside of, of crypto operations, any sort of modern, whether it's an AI data center or a cloud data center still necessitates a significantly high availability of power, uh, in part because, you know, the, the capex costs associated with the infrastructure you're putting in there are so high. Uh, you want to have high utilization plus the services that you're serving out of that, whether it's AI inferencing or, Some sort of traditional cloud application still requires a high level of availability. The one area that comes up a lot in this discussion is training. And so does this training, can, can that act a little bit more as a batch workload? And it's true by definition it can. And at the same time, nobody wants to build a twenty billion dollar training model and just turn it on and off, you know, every time the electricity starts to cut out.

AI assessment note: “you still have to have a very high level of availability of power.”

Q entirely serve the load of the data center, though I think that might be, you tell me what you think, I think that's more of a mirage than anything else. More likely sitting there operating 24 seven, or as close to 24 seven as it can alongside a grid connection, ultimately. So can you straw man for me the argument for like when you might actually want to do that?

A Yeah, I mean, the argument is that if I go to utility and they tell me it's going to be five to seven years to get the connection at the scale I want, then maybe it's faster for me to just build my own generation. So that, that's the argument. And it's sort of further bolstered by the idea that, which I actually think is a false idea, but an idea that because I'm putting a 24 seven load on the grid, I need to match it with a 24 seven generation source. And you hear that a lot out of the CART administration about, well, wind and solar can't help us do what we need to do because they're intermittent, so we need to have lots of baseload generators because we need to connect these data centers and we need to keep the lights on. So that's the other part of the argument, that I need to match the, the output of this resource with what I need to input into my data center.

AI assessment note: “the argument is that if I go to utility and they tell me it's going to be five to seven years”

Q trying to build still. I mean, if you're a power provider, you're still getting inundated with large load interconnection requests, more so than ever, perhaps. But there is this demand planning challenge. Is that manifesting in, like, any change in strategy for those who are trying to build data centers? Are they, are they pulling back? Are they pairing the plans? Are they just, like, full steam ahead and hoping?

A Yeah, I, I think it's more probably, probably, Somewhere in the middle. I think there is some pullback, uh, not because, again, there's not conviction that the, uh, the opportunity is there, but rather, again, for any, any one individual player, you know, the ability to commit billions of dollars to electric utilities to build out more good infrastructure, um, is a hard pill to swallow if you're not fully convinced that you have a customer on the other side of that to receive it. So That, that's the challenge, I think, that the industry's in right now, is that there is still some hesitation when it, when it actually comes time to write that check. And utilities are getting a little more savvy around, you know, really holding feet to the fire for some of these, these companies. Um, and some of them are big tech companies, and some of them are just two guys with a truck that decided they were going to be data center developers, and they go get a queue position, and they have to pay 10,000 dollars, that's it, to, to get in the queue. Which is, which is shockingly low. Um, but, you know, utilities have just never had to deal with this before, you know, having this much large load, large load coming in at once. So I think it is, you know, in some ways sort of looking a little bit like a pullback when it's really just uncertainty. You know, I think the conviction is still very much t…

AI assessment note: “I think it's more probably, probably, Somewhere in the middle. I think there is some pullback”

Q everything is going to get solved here. Um, but as, as you know, as well as I do, like the, this market, the electricity market is not structured entirely that way. So one of the ways in which I feel like there's a disconnect is just on that timeline. So I wonder how you think about, how do you incentivize the construction of new long-term infrastructure for an uncertain time?

A Yeah, so when I say that the, the way I think about the spread is it's really about the value of capacity in a given year, right? So the value of a megawatt in twenty-twenty-seven is worth more than a value of a megawatt in twenty-thirty-two, because there's an assumption that by twenty-thirty-two, power will probably be more abundant, that will have sort of run through this cycle. So the real question of, like, how you, you monetize this in a way that's Rational for, for all actors is as a utility, I should look for ways to make investments in my system that allow me to accelerate the delivery of capacity and find entities that are willing to pay for that capacity in a given year at a higher price relative to delivery in, you know, five years later. That doesn't mean they're paying, you know, on a per megawatt hour basis more forever. But it means they're paying, in essence, higher demand charges to recover the cost of that infrastructure, you know, so that they can get plugged in sooner. Because that load's not going to go away. I mean, data centers don't really get turned off. Like, once you have the customer, they're going to be there. So that, that, that investment will be, in the utility parlance, used and useful, you know, for its useful life. Um, it's really about the timing of When I, I deliver that first electron. That's what's sort of being mispriced right now.

AI assessment note: “they're paying, in essence, higher demand charges to recover the cost of that infrastructure”

Q everybody wants, like, if you can get watts, you know, you'll, you'll pay a premium, and there's a lot of money sloshing around into that space, but is there some risk that the underlying watt bit spread is not as high as it seems today? Like, could it be artificially inflated today by just where we are in the cycle? Or do you think that's a, that's a sustainable thing?

A It's, it's possible, though I, I suspect that given the, my fundamental belief is that the demand for compute, let's just say through twenty-thirty, will exceed the available power in the market. And therefore, the marginal value of the next watt, uh, or gigawatt, frankly, that you can produce will remain quite high for some time. Because there will be A shortage of power available to plug in GPUs over that period. Now, at some point, We will get back to some level of equilibrium where the market starts to, to settle out. Um, but just given the time dimensions here of what it takes to build out energy infrastructure, you know, I feel pretty convinced that we're going to be in a period of, of shortage, and therefore every marginal, uh, gigawatt is going to have substantial value to, to some player in the market.

AI assessment note: “demand for compute, let's just say through twenty-thirty, will exceed the available power”

Q So you were there for the emergence of ChatGPT and, you know, the first kind of, I don't know, 18 months or something after that. So what was happening? Your, your job was to procure energy for Microsoft broadly, but including for data centers. Um, what was it like that last. 18 months, 24 months, like how did things change how quickly and like, what was the dynamic?

A Yeah, the change was pretty rapid. I think we all knew that Microsoft was working with OpenAI and that there was this whole movement towards AI potentially coming at some point in the future, but I don't think there was a realization, and if I'm going back to probably the summer of twenty-twenty-two, this was before ChatGPT-III was released, I don't think there was really a realization of the The scale of the magnitude of what the technology was going to do and how fast it was going to be adopted. And so we started hearing rumblings of it that summer, and I started getting some odd questions that I'd never gotten before about the scale of certain data centers and how big could you make a data center. And then it, once GPT-III was released in November, that's when it really started to sink in. And then I think the, the next level of, of realization was the next spring when 3.5 came out and you saw this massive leap When on really what, what I thought of as a sort of a half click, wasn't even the full click to GPT-IV, but just a half click. And that's the moment that I realized that the problem we were going to have was going to be whether we would have enough power to support this technology that was moving at a pace that, as you know, moves way faster than the electric utility industry.

AI assessment note: “I realized that the problem we were going to have was going to be whether we would have enough power”

Q Do you mean that that capacity has to be on site or not?

A It's not clear. Yeah, because the bring your own capacity concept, as we think about it, is you just have to make sure that you're funding all necessary capacity additions, and so we're doing a project right now, uh, in Oklahoma, and it's with a smaller municipal, and we're doing like a hundred percent of the sourcing of all the generation capacity of the wind, solar, storage, and everything that's going into that, um, and we are packaging it up and delivering it to the utility to then turn around and deliver it To the data center, right? So when I hear bring your own capacity, bring your own generation, that's the world that I think about is, okay, go, go solve it for the utility. Be financially responsible for backing it. But yeah, I think most people probably think about, oh, that just means you're going to have your own generation behind the meter. Um, and I've heard a number of people talk about that, and, and I think they even think One, it, it, it definitely doesn't solve the problem we're talking about, because if you're talking about the issue is, uh, you know, if, if the inflationary pressure on rates is not coming from, you know, more load on the system, but rather it's coming from higher cost for equipment, well, just buying that equipment and put it behind your meter doesn't do anything but provide more upward pressure for, for the, the, the other stuff that you're…

AI assessment note: “It's not clear. Yeah, because the bring your own capacity concept, as we think about it”

Q And why do you think people are missing that?

A A couple of things. First of all, and it only took us, what, 20 minutes maybe into this to, to mention Tyler Norris's name. So Tyler's paper about flexibility, which everyone's talking about, and rightfully so, I was actually just with Tyler this week talking about this. Um, Tyler's paper does a great job articulating my perspective, which is the problem we're trying to solve here is not that I need a 24 seven generation to match a 24 seven load. It's that I need to solve for The summer peaks and the winter system peaks, uh, in order to connect alone. That's what a utility does, and I think it's a, I, I think there's a misunderstanding that when you go to utility and say, okay, where's the power going to come from? The utility goes and solves for 8760 hours. Where's your power going to come from? That's not what they do. They look at would the incremental addition of this load on the system cause me to exceed what I can supply on the hottest summer day and the coldest winter morning. So first of all, it's, it's a capacity problem, not an energy problem. Um, and so flexibility Being able to identify sources where we can, whether it's on the customer side of the meter or the utility side of the meter, unlock more flexibility and unlock more capacity on that system is really the goal. The second part is, in addition to the time element, there's also this sort of space element. I m…

AI assessment note: “I think there's a misunderstanding that when you go to utility and say”

Q at the epicenter of this madness, and we've talked before on this podcast a little bit about it, but it's been a little while, and so I guess I want to start by bringing us up to speed on what you're seeing in the trenches, so to speak, today. Like, what's changed over the past, I don't know, six to nine months at this nexus of data centers and energy?

A Well, the market is still clearly very hot. I mean, there's huge demand for data center capacity, though there is a caveat in there in that I feel like a lot of the big players right now are struggling with something very similar that we struggled with, uh, back in, you know, the 20 tens when we were building out cloud infrastructure, which is exactly how much infrastructure should any one company build, uh, because you're, you're sort of building for your own stack. Um, cause I mean, keep in mind, like Microsoft and Meta, you know, they're not Equinix, right? They don't build data center capacity and then, then lease it out. I mean, they're, they're largely running first party platforms. Um, so That makes demand planning pretty tricky, because even if you have extraordinarily high conviction, as we did in the early twenty-tens, that the cloud market was going to be very large, we did not have incredible conviction on exactly what Azure's market share would be versus AWS versus GCP.

AI assessment note: “Well, the market is still clearly very hot. I mean, there's huge demand”

Q is anything that we can do to make the data centers More palatable to the grid from, from a grid perspective to enable some more of this growth and not like totally swamp the electricity system. So first question is in historic times when you were just building cloud data centers, like why do you need such, uh, insane resiliency? And then is there any difference in the new world?

A Yeah, the resiliency is a function of what we were going back talking about previously in terms of these regions. I think people have this sense of the cloud as really being this sort of ethereal thing, and applications can just move around, and if one data center goes down, then, well, you can just move everything to another data center. That's not really how it works. When someone's writing an application in the cloud, it's going to a physical location on, on the grid inside of a particular region. Um, that region Itself may have some redundancy. So there may be, as I said, multiple data centers inside that region. So if one data center goes down, you don't lose an application, but you still need very high availability for those data centers so that the customer experience is, is good so that your searches show up really fast. So when you go into retrieve a document, it comes up right away. If someone else is editing it, you're seeing it in real time. All of that requires very, very high availability. Um, There are a few exceptions. I mean, there are some applications that don't require, uh, as much, uh, availability. So one, one example might be, uh, the web crawlers that do sort of the indexing for search, right? So those, you know, if you drop your new podcast and put it on your, on your website and, you know, it doesn't show up in a Google search for a few minutes, that's…

AI assessment note: “you still need very high availability for those data centers so that the customer experience”

Q is anything that we can do to make the data centers More palatable to the grid from, from a grid perspective to enable some more of this growth and not like totally swamp the electricity system. So first question is in historic times when you were just building cloud data centers, like why do you need such, uh, insane resiliency? And then is there any difference in the new world?

A Yeah, the resiliency is a function of what we were going back talking about previously in terms of these regions. I think people have this sense of the cloud as really being this sort of ethereal thing, and applications can just move around, and if one data center goes down, then, well, you can just move everything to another data center. That's not really how it works. When someone's writing an application in the cloud, it's going to a physical location on, on the grid inside of a particular region. Um, that region Itself may have some redundancy. So there may be, as I said, multiple data centers inside that region. So if one data center goes down, you don't lose an application, but you still need very high availability for those data centers so that the customer experience is, is good so that your searches show up really fast. So when you go into retrieve a document, it comes up right away. If someone else is editing it, you're seeing it in real time. All of that requires very, very high availability. Um, There are a few exceptions. I mean, there are some applications that don't require, uh, as much, uh, availability. So one, one example might be, uh, the web crawlers that do sort of the indexing for search, right? So those, you know, if you drop your new podcast and put it on your, on your website and, you know, it doesn't show up in a Google search for a few minutes, that's…

AI assessment note: “you still need very high availability for those data centers so that the customer experience”

Q some share of that six hundred billion, how much of the revenue you need, six hundred billion plus in revenue, or ideally six hundred billion plus in earnings to make up for that CapEx. How much of that can they attain? And so there's like a, it's a classic You've got a land grab combined with a tragedy of the commons, right? And like figuring that out is super hard.

A It's really hard. And, and what happened during the 2010 is you had the co-location market sort of fill the gap because every company under-invested in their infrastructure, uh, which is why, you know, Microsoft and Amazon and Google leased capacity from the likes of Vantage and Cyrus One and, and Equinix. Um, but the difference between That era and this era is the skill set needed to fill that gap is not the skill set that was used in the prior era, which was being really good at real estate and fiber. And if you look at the makeup of those co-location companies, they are largely real estate and fiber, uh, dominated in terms of the, the talent and the skills. Um, because energy wasn't a challenge when we were building 50 megawatt data centers for a hundred megawatt data centers. Um, now the, the world has changed. And those companies that, that filled the gap in the last era are really not particularly well equipped, equipped to fill it in this era, um, because the, the depth of energy talent at those companies is actually quite thin. Uh, and so it's really changed the problem set to being not solving for, can I find real estate in some proximity to Northern Virginia, but where can I get a gigawatt of power, which is a completely different challenge.

AI assessment note: “what happened during the 2010 is you had the co-location market sort of fill”

Q enables a data center to get connected, that it doesn't, one, raise prices for other customers, and two, you know, limit the amount of other stuff that might have gotten connected, or limit the sort of, the productive use of clean generation that they might have been building. Like, is there, how much innovation is there to be had in the contractual structure, the tariffs, Between data centers and utilities?

A I think that's a really important piece of it in that if a utility is going to be making, well, one, you know, so it's not the question you asked, but I think it's important, which is, you know, historically, utilities have been motivated to spend as much money as possible on infrastructure because they want to put that in their rate base and make money on it. So there's already a change when you go start talking to utility and saying, hey, I want you to not spend a lot of money. They spend a little bit of money, Just to get more out of your existing system. So that's a pivot from the way utilities have tended to think about the world. Um, and I've had a lot of conversations with folks about this, and one of the questions that comes up is, well, why would a utility want to do that? Because, you know, it's not a lot of rate-based, say, you know, deploying grid-enhancing technologies versus building a new power plant. The reason they want to do it is the reason we're having this conversation is there's a massive demand that they want to go after. They want to attract those customers into their service territory. They want that economic development, job growth, and all the things to come with it. And so they are motivated to say yes, to move quickly, and to the extent that they can extract more out of the existing system, they then have an advantage relative to their peers to attr…

AI assessment note: “so it's not the question you asked, but I think it's important”

Q enables a data center to get connected, that it doesn't, one, raise prices for other customers, and two, you know, limit the amount of other stuff that might have gotten connected, or limit the sort of, the productive use of clean generation that they might have been building. Like, is there, how much innovation is there to be had in the contractual structure, the tariffs, Between data centers and utilities?

A I think that's a really important piece of it in that if a utility is going to be making, well, one, you know, so it's not the question you asked, but I think it's important, which is, you know, historically, utilities have been motivated to spend as much money as possible on infrastructure because they want to put that in their rate base and make money on it. So there's already a change when you go start talking to utility and saying, hey, I want you to not spend a lot of money. They spend a little bit of money, Just to get more out of your existing system. So that's a pivot from the way utilities have tended to think about the world. Um, and I've had a lot of conversations with folks about this, and one of the questions that comes up is, well, why would a utility want to do that? Because, you know, it's not a lot of rate-based, say, you know, deploying grid-enhancing technologies versus building a new power plant. The reason they want to do it is the reason we're having this conversation is there's a massive demand that they want to go after. They want to attract those customers into their service territory. They want that economic development, job growth, and all the things to come with it. And so they are motivated to say yes, to move quickly, and to the extent that they can extract more out of the existing system, they then have an advantage relative to their peers to attr…

AI assessment note: “so it's not the question you asked, but I think it's important”