WEBVTT - How Electric Utilities Will Handle Booming AI Datacenter Demand 0:00:02.520 --> 0:00:13.240 Bloomberg Audio Studios, Podcasts, Radio News. 0:00:17.880 --> 0:00:21.040 Hello and welcome to another episode of the Odd Lots podcast. 0:00:21.079 --> 0:00:22.480 I'm Joe Wisenthal and. 0:00:22.440 --> 0:00:23.440 I'm Tracy Alloway. 0:00:23.800 --> 0:00:27.120 So, Tracy, a thing that keeps coming up is that 0:00:27.360 --> 0:00:29.200 all these AI data centers are going to use a 0:00:29.200 --> 0:00:31.600 lot of electricity. I keep hearing that. 0:00:32.040 --> 0:00:35.680 Yes. Also, I just realized every time you use chat 0:00:35.720 --> 0:00:39.960 GPT to write like a satirical song, you're diverting energy 0:00:40.000 --> 0:00:42.680 away from someone turning on a light bulb or something 0:00:42.720 --> 0:00:45.280 like that, potentially something like zero sum game. 0:00:45.400 --> 0:00:46.000 Yeah, so be. 0:00:46.040 --> 0:00:50.560 Careful about your your random chat GPT queries. Although I 0:00:50.600 --> 0:00:53.360 think the training is like the more Yeah, like I think, 0:00:53.560 --> 0:00:56.240 but you know, like so maybe your song there, maybe 0:00:56.280 --> 0:00:57.880 it's okay. I don't think it's that bad. 0:00:58.280 --> 0:01:00.760 In the grand scheme of things, probably not. But there 0:01:00.840 --> 0:01:05.200 is this overarching conversation about AI's energy use. So what 0:01:05.319 --> 0:01:07.960 exactly it is? This is a big question I have. 0:01:08.120 --> 0:01:12.280 How do you disaggregate AI servers from your run of 0:01:12.319 --> 0:01:15.720 the mill software servers, How much it's going to consume, 0:01:16.240 --> 0:01:20.120 how that capacity is going to get allocated and built out? 0:01:20.520 --> 0:01:23.199 And I think there is this sense that we could 0:01:23.319 --> 0:01:28.240 end up going into very different, very extreme directions here. 0:01:28.319 --> 0:01:32.679 So you could have this great situation that because AI 0:01:33.040 --> 0:01:36.800 is a desirable activity, because it's profitable in many respects, 0:01:36.800 --> 0:01:41.280 that big tech ends up accelerating the energy capacity build out, 0:01:41.400 --> 0:01:46.000 maybe they even start building more green technology capabilities in 0:01:46.040 --> 0:01:49.040 an ideal world. But then you have the polar opposite 0:01:49.080 --> 0:01:52.680 scenario where you need all this power to develop this 0:01:52.800 --> 0:01:56.120 technology there isn't enough and it's sort of a race 0:01:56.160 --> 0:01:58.559 to the bottom where you have tech companies just trying 0:01:58.560 --> 0:02:01.640 to get energy wherever they can, maybe they even start 0:02:01.720 --> 0:02:04.960 using coal and things like that. So it feels like 0:02:05.880 --> 0:02:08.519 there's two very different paths that we could be going 0:02:08.560 --> 0:02:08.959 down here. 0:02:09.040 --> 0:02:10.639 Yeah, there's a lot here for us. 0:02:10.880 --> 0:02:11.040 You know. 0:02:11.040 --> 0:02:13.480 I remember Jigger shaw at Or when we interviewed him 0:02:13.480 --> 0:02:15.839 at the Texas Tribune conference like over a year ago, 0:02:16.560 --> 0:02:19.000 he brought this up. It's getting more and more attention. 0:02:19.120 --> 0:02:21.840 It keeps coming up. On the side of episode. Steve 0:02:21.880 --> 0:02:25.240 Eisman obviously recently talked about it. But I feel like 0:02:25.840 --> 0:02:28.120 it's time to like sort of make it a central 0:02:28.160 --> 0:02:31.240 part of the conversation and actually learn about numbers and 0:02:31.280 --> 0:02:33.480 where this power is generated from and like, yeah, like 0:02:33.520 --> 0:02:35.600 how much are we really talking about here. We know 0:02:35.680 --> 0:02:37.560 the you know, the tech company is highly aware of it. 0:02:37.560 --> 0:02:40.080 There was a headline recently about Microsoft maybe wanting to 0:02:40.120 --> 0:02:43.960 do something with on site nuclear development. They also, you know, 0:02:44.080 --> 0:02:45.320 they did do something. 0:02:45.480 --> 0:02:48.400 So I think they headline, well, didn't they buy a 0:02:48.480 --> 0:02:52.480 data center maybe next to a nuclear power plant? Susqahanna thing? 0:02:52.520 --> 0:02:53.120 I thought they did. 0:02:53.320 --> 0:02:55.000 Yeah, I think you're right about that. But then the 0:02:55.000 --> 0:02:58.639 other element too, is you mentioned that one solution here 0:02:58.760 --> 0:03:02.000 is just fossil fuels and dirty energy, except that all 0:03:02.000 --> 0:03:05.280 these tech companies are very like progressive minded and they 0:03:05.320 --> 0:03:07.600 all have these net zero commitments by you know, we're 0:03:07.600 --> 0:03:11.040 going to get it all from you know, windmills or sorry, 0:03:11.160 --> 0:03:13.640 wind turbines and solar and batteries and stores. 0:03:13.720 --> 0:03:17.200 Windmills next to AI servers would be an interesting one. 0:03:17.280 --> 0:03:19.480 But like you know, at some point, the rubber's got 0:03:19.480 --> 0:03:21.840 to hit the road with like how realistic are there 0:03:21.919 --> 0:03:24.680 net zero commitments or how can they achieve them if 0:03:24.680 --> 0:03:27.960 they're engaging in this investment activity that is highly energy intensive. 0:03:28.040 --> 0:03:28.079 No. 0:03:28.200 --> 0:03:30.960 Absolutely, And you are seeing a lot of this discussion 0:03:31.160 --> 0:03:34.920 reflected in the conversation around AI investment at this point. 0:03:34.960 --> 0:03:37.040 So I think a lot of people feel like they 0:03:37.080 --> 0:03:40.800 missed that first wave of chips around in video, so 0:03:40.840 --> 0:03:44.600 everyone's looking for the sort of second order investment play 0:03:44.680 --> 0:03:47.080 and a lot of people now are talking about energy 0:03:47.200 --> 0:03:50.280 or cooling and HVAC, so we need to talk about it. 0:03:50.560 --> 0:03:50.760 Well. 0:03:50.800 --> 0:03:53.520 I am really excited because I do believe we have 0:03:54.000 --> 0:03:58.120 the perfect guest for this topic, someone who spent twelve 0:03:58.240 --> 0:04:02.040 years at Microsoft, way before it was hot to talk 0:04:02.080 --> 0:04:06.560 about how tech companies needed all this electricity and energy. 0:04:06.800 --> 0:04:10.320 He was the first energy hire at Microsoft, and he 0:04:10.400 --> 0:04:13.119 recently left last year. He left to start his own 0:04:13.160 --> 0:04:15.800 firm to work on this problem specifically. So we were 0:04:15.840 --> 0:04:17.880 going to be speaking with Brian Janis. He is the 0:04:17.920 --> 0:04:22.120 co founder and chief strategy officer at Cloverleaf Infrastructure, a 0:04:22.120 --> 0:04:26.120 power development company that works closely with utilities on solving 0:04:26.200 --> 0:04:28.760 this problem. So, Brian, thank you so much for coming 0:04:28.760 --> 0:04:29.840 on out lots. 0:04:29.720 --> 0:04:31.400 Thank you for having me. Really excited to be here. 0:04:31.760 --> 0:04:35.080 So you got hired for Microsoft twelve years ago to 0:04:35.120 --> 0:04:38.520 do energy, and at the time, I don't think anyone 0:04:38.720 --> 0:04:42.400 was talking about energy as being like a particularly important 0:04:42.440 --> 0:04:47.719 aspect of these software companies or these big tech companies strategies. 0:04:48.000 --> 0:04:50.080 What was going on back then, or like, what did 0:04:50.120 --> 0:04:52.320 they see when they felt like, hey, we need to 0:04:52.400 --> 0:04:53.640 hire a VP of energy here. 0:04:54.040 --> 0:04:56.359 Yeah. It was actually funny because I had spent my 0:04:56.440 --> 0:04:59.679 career prior to that working with mainly large energy consumers 0:04:59.720 --> 0:05:01.400 who were the big who you'd expect it to be, 0:05:01.480 --> 0:05:05.400 the big industrial companies, And so when Microsoft came calling 0:05:05.480 --> 0:05:07.480 and said, hey, we need to get a full time 0:05:07.920 --> 0:05:10.640 energy person, I told them it sounded like a dead 0:05:10.720 --> 0:05:13.599 end job to be the energy person at a tech company, 0:05:13.600 --> 0:05:16.920 because why would they ever actually care about this issue. 0:05:17.200 --> 0:05:19.000 And the person that was recruiting me said, Hey, I 0:05:19.040 --> 0:05:21.560 think there is something to this, this whole cloud thing, 0:05:21.680 --> 0:05:23.479 and I think energy is going to start to be 0:05:24.000 --> 0:05:27.680 pretty central to what we're doing as a company. And 0:05:27.960 --> 0:05:31.640 you'll fast forward a decade and I remember having a 0:05:31.640 --> 0:05:33.719 conversation before I left the company. I was talking to 0:05:33.720 --> 0:05:36.160 the head of corporate strategy and he said to me, 0:05:36.240 --> 0:05:39.279 He's like, I don't think people quite realize the degree 0:05:39.279 --> 0:05:42.800 to which Microsoft is really just an energy company. We 0:05:42.839 --> 0:05:46.320 need power and we need silicon. We need chips. That's it, 0:05:46.400 --> 0:05:48.200 that's the business. If we don't have one of those 0:05:48.240 --> 0:05:50.120 two things, we're in a lot of trouble and so 0:05:50.200 --> 0:05:52.960 it was it was remarkable to see the shift over 0:05:53.000 --> 0:05:56.840 that decade plus of maybe one or two people at 0:05:56.839 --> 0:05:59.440 the company starting to think that energy might be important 0:05:59.440 --> 0:06:03.800 to us, when day to energy is actually absolutely central 0:06:04.279 --> 0:06:05.800 to everything the business does. 0:06:06.600 --> 0:06:08.440 So talk to us a little bit more about that 0:06:08.520 --> 0:06:12.799 cultural shift, because Joe and I heard from someone else recently, 0:06:12.839 --> 0:06:15.159 but they were saying that a bunch of the big 0:06:15.200 --> 0:06:18.960 tech companies that you would be very familiar with had 0:06:19.000 --> 0:06:23.359 representatives down in Houston for zero Week, so the annual 0:06:23.480 --> 0:06:26.039 energy conference, which they were describing as kind of a 0:06:26.080 --> 0:06:30.599 new development. But how familiar are tech companies nowadays with 0:06:30.880 --> 0:06:35.240 energy usage or needs and how much expertise have they 0:06:35.279 --> 0:06:37.640 actually built out in that capacity. 0:06:39.000 --> 0:06:41.680 Yeah, it's been a tremendous shift. And I mean if 0:06:41.680 --> 0:06:43.800 you would have gone to Seria Week even five years ago, 0:06:44.520 --> 0:06:47.120 you would not have seen a whole lot of engagement 0:06:47.160 --> 0:06:52.040 from the tech industry. But as their businesses have shifted 0:06:52.040 --> 0:06:56.240 to the cloud, and as the business opportunities that sit 0:06:56.279 --> 0:06:58.599 in front of them, particularly when it comes down to AI, 0:07:00.040 --> 0:07:03.760 as those have arisen, there's been a recognition that power 0:07:03.800 --> 0:07:06.760 really is central to what they're doing. And it was 0:07:06.800 --> 0:07:09.679 a slow shift. If you go back to the advent 0:07:09.800 --> 0:07:14.200 of the first cloud data centers. It really was about 0:07:14.400 --> 0:07:17.080 being close to the network, and so that was the 0:07:17.200 --> 0:07:20.440 driver of strategically, where do you put data centers, Well, 0:07:20.440 --> 0:07:22.320 you put them where the biggest network cubs are. So 0:07:22.600 --> 0:07:24.840 that's why we have lots of data centers in Northern Virginia. 0:07:24.960 --> 0:07:27.360 That's why we have lots of data centers in Amsterdam. 0:07:27.640 --> 0:07:31.720 Everyone was chasing network. Probably middle of last decade, there 0:07:31.760 --> 0:07:34.320 was a shift and it started to go, actually, we 0:07:34.320 --> 0:07:38.040 want to be close to eyeballs. So this started a 0:07:38.240 --> 0:07:41.440 sort of a land grab of all the cloud data 0:07:41.440 --> 0:07:44.160 centers starting to build lots of new data SISIs new 0:07:44.160 --> 0:07:46.480 countries because they wanted to be close to where the 0:07:46.520 --> 0:07:50.960 customers were. And so from about fall of twenty nineteen 0:07:51.200 --> 0:07:54.000 through probably spring of twenty twenty two, I think Microsoft 0:07:54.040 --> 0:07:56.720 added was adding close to a region a month in 0:07:56.800 --> 0:07:59.640 terms of new data center regions they were establishing around 0:07:59.640 --> 0:08:04.320 the world. And then in mid twenty twenty two that's 0:08:04.360 --> 0:08:08.560 when the realization started to sink in that wait a minute, 0:08:08.640 --> 0:08:11.360 this whole game is about power, because that's when we 0:08:11.360 --> 0:08:16.040 were first starting to hear rumblings of what open aye 0:08:16.160 --> 0:08:18.880 was working on and the scale of what chat gp 0:08:19.360 --> 0:08:21.600 T three was going to be, which was sort of 0:08:21.640 --> 0:08:23.920 the first big release where everyone was like, wait a minute, 0:08:23.960 --> 0:08:26.240 this is this is kind of a big deal what 0:08:26.280 --> 0:08:29.120 AI is doing and how fast it's moving. And then 0:08:29.240 --> 0:08:32.720 when we had that release of chat gpt T three 0:08:32.840 --> 0:08:35.600 in the fall of twenty two, and then shortly thereafter, 0:08:37.600 --> 0:08:39.960 three point five was released, and there was a massive 0:08:40.520 --> 0:08:44.040 increase in capability in that release if you recall, you know, 0:08:44.080 --> 0:08:46.360 in terms of what it could do on you know, 0:08:46.480 --> 0:08:51.000 getting scores on various tests and things. And it was 0:08:51.040 --> 0:08:55.720 that moment that I realized, this technology is moving way 0:08:55.800 --> 0:09:00.559 faster than the utility industry is moving. If we can 0:09:00.640 --> 0:09:04.880 make this much improvement in this technology in a six 0:09:04.920 --> 0:09:09.320 month time horizon, we're in a lot of trouble because 0:09:09.400 --> 0:09:11.600 the power industry does not move that fast. 0:09:12.160 --> 0:09:15.320 So I'm really fascinated by this idea that the release, 0:09:15.520 --> 0:09:17.840 like you know, like you were at Microsoft and so 0:09:17.960 --> 0:09:20.200 you had a front row seat to what open ai 0:09:20.400 --> 0:09:23.800 was doing with GPT one and GPD two, and there 0:09:23.800 --> 0:09:25.839 were a lot of people aware of this, and I'm 0:09:25.880 --> 0:09:28.760 sort of fascinated by this idea that it was that 0:09:28.880 --> 0:09:32.600 commercialization or sort of making easy in public CHATJEP. When 0:09:32.640 --> 0:09:35.800 it became chat GPT, they're like, oh, this is serious, 0:09:35.840 --> 0:09:38.400 and then we sell everyone rushing to buy in video 0:09:38.480 --> 0:09:41.559 chips and all these vcs pivoting to AI, et cetera. 0:09:41.720 --> 0:09:45.320 So talk to us about like the math there. It 0:09:45.400 --> 0:09:48.360 feels like there has been this sort of level shift 0:09:48.520 --> 0:09:52.760 up in sort of expectations of data center demand growth 0:09:53.160 --> 0:09:56.439 basically as a function of all of the excitement for AI. 0:09:57.960 --> 0:09:59.960 Yeah. I think then you're right. I mean, it's not 0:10:00.080 --> 0:10:03.520 like we didn't know that Microsoft had a partnership with 0:10:03.559 --> 0:10:06.120 open Ai and that that you know, AI was going 0:10:06.200 --> 0:10:10.320 to consume energy. I think everyone though, was a bit 0:10:10.400 --> 0:10:18.240 surprised at just how quickly what chat GPT could do 0:10:18.280 --> 0:10:21.200 just captured the collective consciousness. Yeah, and I think, I 0:10:21.200 --> 0:10:24.000 mean you probably remember when when that was released. I 0:10:24.040 --> 0:10:27.200 mean it it really sort of surprised everyone, and it 0:10:27.280 --> 0:10:30.360 became this thing where suddenly, even though we sort of 0:10:30.440 --> 0:10:32.520 knew what we were working on, it wasn't until you 0:10:32.600 --> 0:10:35.320 sort of put it out into the world that you 0:10:35.400 --> 0:10:39.360 realize maybe what you've created. And I mean that's where 0:10:39.520 --> 0:10:42.000 we realized we are we are running up this curve 0:10:42.160 --> 0:10:46.040 of capability a lot faster than we thought, uh, and 0:10:46.320 --> 0:10:48.720 the a number of applications that are getting built on this, 0:10:48.800 --> 0:10:51.000 in the number of different ways that it's being used, 0:10:51.000 --> 0:10:54.199 and how it's just become sort of common parlance. I mean, 0:10:54.200 --> 0:10:56.880 everyone knows what chat GPG three is, and no one 0:10:56.920 --> 0:10:59.720 what it was the month before that, right, So there 0:10:59.720 --> 0:11:01.480 there as a bit of a I think of a 0:11:01.960 --> 0:11:04.960 surprise in terms of just how quickly it was going 0:11:05.040 --> 0:11:09.040 to capture, you know, the collective consciousness and then you know, 0:11:09.360 --> 0:11:12.240 obviously lead to everything that's sort of being created as 0:11:12.240 --> 0:11:13.920 a result. And so we just we just moved up 0:11:13.920 --> 0:11:16.440 that curve so quickly, and I think that's that's where 0:11:16.480 --> 0:11:20.360 the industry maybe got you know, certainly the utilities were 0:11:20.400 --> 0:11:23.560 behind because as you may have seen there, a lot 0:11:23.600 --> 0:11:26.559 of them are starting to restate their load growth expectations 0:11:26.920 --> 0:11:30.480 and that was something that was not happening right before that. 0:11:30.520 --> 0:11:33.080 And so we've had massive changes just in the last 0:11:33.080 --> 0:11:35.560 two years of how utilities are one of the number 0:11:35.559 --> 0:11:39.240 cap So you know, if you take a look at 0:11:39.320 --> 0:11:43.959 a utility like Dominion in Virginia, so that's the largest 0:11:44.000 --> 0:11:46.120 concentration of data centers in the United States, so they're 0:11:46.160 --> 0:11:49.600 pretty good representative of what's happening. If you go back 0:11:49.600 --> 0:11:53.120 to twenty twenty one, they were forecasting load growth over 0:11:53.160 --> 0:11:56.080 a period of fifteen years of just a few percent. 0:11:56.240 --> 0:11:59.120 I mean it was about it was single digit growth 0:11:59.160 --> 0:12:03.120 over that entire period, so not yearly growth, but over 0:12:03.200 --> 0:12:07.000 fifteen years, single digit growth. By twenty twenty three, they 0:12:07.000 --> 0:12:10.880 were forecasting to grow two x over fifteen years. Wow. 0:12:10.960 --> 0:12:14.000 Now keep in mind this is electric utility. They do 0:12:14.080 --> 0:12:17.959 ten year planning cycles. So because they have very long 0:12:18.040 --> 0:12:21.680 lead times for equipment, for getting rights of way for 0:12:21.800 --> 0:12:27.920 transmission lines, they aren't companies that easily respond to a 0:12:28.080 --> 0:12:32.120 two x order of magnitude, you know, growth change over 0:12:32.120 --> 0:12:33.559 a period of fifteen years. I mean that is a 0:12:34.200 --> 0:12:37.679 that is a massive change for electric utility, particularly given 0:12:37.720 --> 0:12:41.200 the fact that the growth rate over the last fifteen 0:12:41.240 --> 0:12:44.520 to twenty years has been close to zero, so there's 0:12:44.679 --> 0:12:47.720 been relatively no load growth in fifteen to twenty years. 0:12:48.200 --> 0:12:51.400 Now suddenly you have utilities having to pivot to doubling 0:12:51.440 --> 0:12:54.280 the size of their system in that same horizon. 0:13:10.600 --> 0:13:13.240 I want to ask a very basic question, but I 0:13:13.280 --> 0:13:16.520 think it will probably inform the rest of this conversation. 0:13:16.960 --> 0:13:20.760 But when we say that AI consumes a lot of energy. 0:13:21.559 --> 0:13:25.120 Where is that consumption actually coming from? And Joe touched 0:13:25.160 --> 0:13:27.040 on this in the intro, but is it you know, 0:13:27.120 --> 0:13:30.839 the sheer scale of users on these platforms, is it 0:13:31.000 --> 0:13:34.440 I imagine the training that you need in order to 0:13:34.480 --> 0:13:39.120 develop these models, and then does that energy usage differ 0:13:39.280 --> 0:13:42.960 in any way from more traditional technologies. 0:13:43.320 --> 0:13:47.920 Yeah. So whenever I think about the consumption of electricity 0:13:47.960 --> 0:13:50.840 for AI or really any other application, I think you 0:13:50.840 --> 0:13:53.800 have to start at sort of the core of what 0:13:53.840 --> 0:13:56.559 we're talking about, which is really the human capacity for data. 0:13:57.120 --> 0:14:02.160 Like whether it's AI or cloud. Humans have a massive 0:14:02.280 --> 0:14:06.200 capacity to consume data. And if you think about where 0:14:06.240 --> 0:14:08.920 we are in this curve, I mean we're on some 0:14:09.080 --> 0:14:13.920 form of S curve right of human data consumption, which 0:14:14.040 --> 0:14:20.640 then directly ties to data centers, devices, energy consumption ultimately, 0:14:20.640 --> 0:14:22.840 because what we're doing is we're turning energy into data. 0:14:22.880 --> 0:14:26.200 We take electrons, we convert them to light, we move 0:14:26.240 --> 0:14:29.360 them around to your TV screens and your phones and 0:14:29.400 --> 0:14:34.800 your laptops, etc. So that's the uber trend that we're 0:14:34.880 --> 0:14:37.520 riding up right now. And so we're climbing this S curve. 0:14:38.000 --> 0:14:40.240 I don't know that anyone has a good sense of 0:14:41.320 --> 0:14:45.000 how steep or how long this curve will go. If 0:14:45.040 --> 0:14:47.480 you go back to look at something like electricity, it 0:14:47.520 --> 0:14:51.200 was roughly about one hundred year. S curve started in 0:14:51.240 --> 0:14:54.480 the beginning of last century, and it really started to flatline, 0:14:54.720 --> 0:14:57.560 as I mentioned before, towards the beginning of this century. 0:14:57.880 --> 0:15:00.440 Now we have this new trajectory that we're ringing, this 0:15:00.480 --> 0:15:02.320 new S curve that we're entering, that's going to sort 0:15:02.360 --> 0:15:04.760 of change that narrative. But you know that S curve 0:15:04.760 --> 0:15:07.200 for electricity took about one hundred years. No one knows 0:15:07.240 --> 0:15:10.880 where we are on that data curve today. So when 0:15:10.880 --> 0:15:14.760 you inject something like AI, you create a whole new 0:15:14.960 --> 0:15:18.320 opportunity for humans to consume data, to do new things 0:15:18.360 --> 0:15:21.000 with data that we couldn't do before, and so you 0:15:21.160 --> 0:15:24.040 accelerate us up this curve. Right, So, we were sitting 0:15:24.400 --> 0:15:27.280 somewhere along this curve. AI comes along, and now we're 0:15:27.320 --> 0:15:29.800 just moving up even further, And of course that means 0:15:30.240 --> 0:15:35.160 more energy consumption because the energy intensity of running an 0:15:35.160 --> 0:15:39.000 AI query versus a traditional search is much higher. Now 0:15:39.280 --> 0:15:42.480 what you can do with AI obviously is also much 0:15:42.480 --> 0:15:45.320 greater than what you can do with a traditional search, 0:15:45.360 --> 0:15:50.600 So there is a positive return on that invested energy. 0:15:51.080 --> 0:15:54.160 So that's you know when when oftentimes when this conversation 0:15:54.280 --> 0:15:58.000 comes up, there's a lot of consternation and panic over well, 0:15:58.000 --> 0:15:59.840 what are we going to do? You know, we're going 0:15:59.920 --> 0:16:02.400 to we're going to run out of energy. The nice 0:16:02.400 --> 0:16:04.960 thing about electricity is we can always make more. We're 0:16:05.160 --> 0:16:07.360 we're never going to run out of electricity. Not to 0:16:07.400 --> 0:16:10.000 say that there's not times where the grid is under 0:16:10.000 --> 0:16:12.760 constraint and there you know, you have risks of brownouts 0:16:12.760 --> 0:16:16.480 and blackouts. That's that's the reality. But we can invest 0:16:16.600 --> 0:16:19.200 we can invest more in transmission lines, we can invest 0:16:19.240 --> 0:16:23.239 more in power plants, and we can create enough electricity 0:16:23.680 --> 0:16:26.120 to to match that demand. 0:16:26.560 --> 0:16:29.400 Just to sort of clarify a point in adding on 0:16:29.440 --> 0:16:33.760 to Tracy's question, you mentioned that doing an AI query 0:16:34.160 --> 0:16:36.400 is more energy intensive than say, if I had just 0:16:36.480 --> 0:16:38.760 done a Google search, or if I had done a 0:16:38.800 --> 0:16:41.400 Being search or something like that. Like, what is it 0:16:41.560 --> 0:16:47.640 about the process of delivering these capabilities that makes it 0:16:47.720 --> 0:16:53.000 more computationally intensive or energy intensive then the previous generation 0:16:53.360 --> 0:16:56.360 of data usage or data querying online. 0:16:57.560 --> 0:17:00.840 There's two aspects to it, and we sort of alluded 0:17:00.880 --> 0:17:04.239 to it earlier. But the first is the training. So 0:17:04.240 --> 0:17:06.680 The first is the building of the large language model 0:17:07.000 --> 0:17:12.280 that itself is very energy intensive. These are extraordinarily large 0:17:12.400 --> 0:17:18.520 machines collections of machines that use very dense chips to 0:17:18.600 --> 0:17:22.440 create these language models that ultimately then get queried when 0:17:22.440 --> 0:17:26.040 you do an inference. So then you go to CHATGBT 0:17:26.200 --> 0:17:29.119 and you ask it to give you a menu or 0:17:29.840 --> 0:17:32.680 dinner party you want to have this weekend. It's then 0:17:33.280 --> 0:17:36.399 referencing that large language model and creating this response. And 0:17:36.440 --> 0:17:40.479 of course that process is more computationally intensive because it's 0:17:40.920 --> 0:17:43.160 it's doing a lot more things than a traditional searchces 0:17:43.400 --> 0:17:47.159 personal search just matched the words you put into a 0:17:48.160 --> 0:17:51.640 database of knowledge that it put together. But these large 0:17:51.680 --> 0:17:54.359 language models are much more complex, and then the therefore 0:17:54.480 --> 0:17:56.320 the things you're asking to do is more complex. So 0:17:56.720 --> 0:18:00.439 it will will almost by definition, be a more energy 0:18:00.480 --> 0:18:03.640 intensive process. Now, it's not to say that it can't 0:18:03.640 --> 0:18:08.360 get more efficient, and it will. And Nvidio just last 0:18:08.400 --> 0:18:11.720 week was releasing, you know, some data on some of 0:18:11.760 --> 0:18:14.840 its next generation chips that are going to be significantly 0:18:14.880 --> 0:18:18.159 more efficient than the prior generation. But one of the 0:18:18.160 --> 0:18:20.240 things that we need to be careful of is to 0:18:21.200 --> 0:18:24.680 think that because something becomes more efficient, then therefore we're 0:18:24.720 --> 0:18:28.080 going to use less of the input resource, in this 0:18:28.160 --> 0:18:32.560 case electricity. That's that's not how it works, because going 0:18:32.600 --> 0:18:37.760 back to the concept of human capacity for consuming data, 0:18:38.119 --> 0:18:40.480 all we do is we find more things to compute. 0:18:40.840 --> 0:18:43.760 And this is you've probably heard of Jabon's paradox, and 0:18:43.760 --> 0:18:46.520 this is the idea that, well, if we make more 0:18:46.640 --> 0:18:49.360 efficient steam engines. He was an economists in the eighteen 0:18:49.440 --> 0:18:50.959 hundreds and he said, well, if you make more efficient 0:18:51.000 --> 0:18:53.480 steam engines, then we'll use us coal. And he's like, no, 0:18:53.520 --> 0:18:54.959 that's not what's going to happen. We're going to use 0:18:54.960 --> 0:18:57.199 more coal because we're going to mechanize more things. And 0:18:57.240 --> 0:18:59.919 that's exactly what we do with data just because we 0:19:00.320 --> 0:19:02.439 because we've had more law for years and so chips 0:19:02.440 --> 0:19:06.640 have become incredibly more efficient than they were decades ago, 0:19:06.720 --> 0:19:10.120 but we didn't use less energy. We used much more 0:19:10.240 --> 0:19:14.480 energy because we could put chips in everything. So that's 0:19:14.640 --> 0:19:17.040 the trend line that we're on. It's still climbing that 0:19:17.160 --> 0:19:21.000 curve of consumption, and so no amount of efficiency is 0:19:21.040 --> 0:19:24.200 going to take us at this point at least, because 0:19:24.440 --> 0:19:27.119 I don't believe we're anywhere close to the bend in 0:19:27.160 --> 0:19:30.600 that s curve. No amount of efficiency is going to 0:19:30.640 --> 0:19:34.760 take us off of continuing to consume more electricity, and 0:19:34.760 --> 0:19:35.760 at least in the near term. 0:19:36.400 --> 0:19:39.800 So I have another basic building block kind of question. 0:19:40.000 --> 0:19:43.560 But when we say that technology companies are aware of 0:19:43.640 --> 0:19:48.679 the importance of energy usage or availability, and that this 0:19:48.760 --> 0:19:52.040 is something they happened working on, what exactly is the 0:19:52.119 --> 0:19:56.720 process by which a tech company gets its energy. So 0:19:56.840 --> 0:19:59.720 you know you have a big data center, I imagine 0:19:59.720 --> 0:20:03.760 you have some sort of agreement with whatever utility is 0:20:03.800 --> 0:20:06.879 in that area. But I also imagine that that agreement 0:20:07.000 --> 0:20:12.639 looks very very different to like my household energy bill 0:20:12.840 --> 0:20:13.720 or something like that. 0:20:14.560 --> 0:20:19.680 I'm certain it does hopefully significant orders of magnitude. Yes, 0:20:20.640 --> 0:20:24.560 So there's two components. I mean, one is, if you're 0:20:24.920 --> 0:20:26.840 building a data center, you have to plug it in somewhere. 0:20:26.880 --> 0:20:28.399 You've got to plug it into the grid, right, So 0:20:29.040 --> 0:20:33.400 there you're working with your local electric utility or transmission 0:20:33.400 --> 0:20:38.440 company and doing planning for how big is the facility 0:20:38.440 --> 0:20:40.080 to be, how much power is it going to pull 0:20:40.119 --> 0:20:42.080 off the grid at any given time, and then over 0:20:42.119 --> 0:20:44.440 a period of time, because these facilities just tend to 0:20:44.480 --> 0:20:48.679 grow forever, and so that's the physical nuts and bolts 0:20:48.680 --> 0:20:52.160 of connecting to the grid. Now, the second piece, of course, 0:20:52.200 --> 0:20:54.600 is there needs to be some generation source as well, 0:20:55.040 --> 0:20:57.679 like where's the power going to come from? And so 0:20:57.760 --> 0:21:01.359 those two things are related, but they could be somewhat disconnected. 0:21:01.400 --> 0:21:04.720 And so this is where you see you know, these 0:21:04.880 --> 0:21:07.160 especially the tech companies who've really been leaders in this space, 0:21:07.720 --> 0:21:11.960 entering into all these power purchase agreements for wind energy 0:21:12.000 --> 0:21:15.359 and for solar energy and in some cases nuclear. You 0:21:15.640 --> 0:21:19.600 mentioned the project earlier that's actually an AWS project where 0:21:19.600 --> 0:21:23.080 they cite it next to the Susquehanna Nuclear plant. Right, 0:21:23.160 --> 0:21:26.919 So all of that is around where are the electrons 0:21:26.960 --> 0:21:29.760 going to come from? And how can with that purchasing 0:21:29.840 --> 0:21:32.080 power of being some of the largest energy consumers in 0:21:32.119 --> 0:21:36.560 the planet, how can they begin to influence the mix 0:21:36.640 --> 0:21:40.000 of generation on the grid? Right? And that's the critical issue, 0:21:40.040 --> 0:21:43.159 is that you're trying to influence where that power is 0:21:43.200 --> 0:21:47.040 being generated from. It's not. And one thing just to 0:21:47.760 --> 0:21:50.600 keep in mind is that you know the electrons you get, 0:21:50.840 --> 0:21:53.320 you know, whether it's at your house or the data 0:21:53.320 --> 0:21:55.560 center down the street, they're all the same electrons. You're 0:21:55.560 --> 0:21:57.840 all pulling from the same grid. But what you're trying 0:21:57.880 --> 0:22:02.040 to do is influence how that that generation is being created, 0:22:02.440 --> 0:22:05.920 and that's where these purchase agreements come in for all 0:22:05.920 --> 0:22:07.200 these different sources of energy. 0:22:07.840 --> 0:22:10.440 All right, now, let's bring the question back to say 0:22:10.560 --> 0:22:11.919 the utility side or save. 0:22:11.840 --> 0:22:12.680 The dominion side. 0:22:12.720 --> 0:22:16.760