1
00:00:00,160 --> 00:00:02,719
Speaker 1: This is Tom Rowlands Reese and you're listening to Switched

2
00:00:02,759 --> 00:00:05,960
Speaker 1: on the podcast brought to you by BNF. The rapid

3
00:00:06,040 --> 00:00:09,400
Speaker 1: rise of energy intensive AI data centers is reshaping the

4
00:00:09,440 --> 00:00:13,000
Speaker 1: near term outlook for US power markets. Outpacing the energy

5
00:00:13,039 --> 00:00:16,720
Speaker 1: demand growth of EVS, hydrogen and all other demand classes

6
00:00:16,760 --> 00:00:19,599
Speaker 1: out to twenty thirty, data centers will account for eight

7
00:00:19,640 --> 00:00:22,959
Speaker 1: point six percent of US electricity demand by twenty thirty five.

8
00:00:23,120 --> 00:00:25,919
Speaker 1: That's almost twice as much as today. Largely owned and

9
00:00:25,960 --> 00:00:29,720
Speaker 1: operated by a few highly consolidated companies with very deep pockets,

10
00:00:29,920 --> 00:00:33,280
Speaker 1: this concentration of capital allows for rapid expansion and a

11
00:00:33,360 --> 00:00:37,760
Speaker 1: significant influence over future energy infrastructure investment. So what strategies

12
00:00:37,760 --> 00:00:40,839
Speaker 1: are these companies employing to optimize their data center rollout?

13
00:00:41,080 --> 00:00:43,520
Speaker 1: On today's show, I'm joined by BNF's head of US

14
00:00:43,560 --> 00:00:47,720
Speaker 1: Power Helen Co and US Power Senior Associate Natalie Lemandebrata,

15
00:00:47,960 --> 00:00:50,640
Speaker 1: and together we discuss findings from their note US data

16
00:00:50,640 --> 00:00:53,519
Speaker 1: center market outlook the Age of AI, which B and

17
00:00:53,560 --> 00:00:55,800
Speaker 1: EF clients can find at BNF go on the Bloomberg

18
00:00:55,880 --> 00:00:58,800
Speaker 1: Terminal or on BNF dot com. All right, let's get

19
00:00:58,800 --> 00:01:01,640
Speaker 1: to talking about the outlook AI data centers with Helen

20
00:01:01,720 --> 00:01:15,959
Speaker 1: and Natalie. Helen, thank you for being here. Thanks Tom

21
00:01:16,240 --> 00:01:18,039
Speaker 1: and Natalie thanks for being here as well.

22
00:01:18,200 --> 00:01:18,880
Speaker 2: Thank you Tom.

23
00:01:19,600 --> 00:01:23,240
Speaker 1: So Natalie reports up to Helen, and Helen reports up

24
00:01:23,280 --> 00:01:25,560
Speaker 1: to me, and I'm not saying that to flex. I'm

25
00:01:25,560 --> 00:01:27,960
Speaker 1: saying it for a couple of reasons. One is, I'm

26
00:01:28,000 --> 00:01:31,840
Speaker 1: like super proud to have such smart people on my team,

27
00:01:32,160 --> 00:01:34,559
Speaker 1: and I'm also particularly proud of the work they've done

28
00:01:34,640 --> 00:01:39,640
Speaker 1: around data centers. But also this situation of this reporting

29
00:01:39,720 --> 00:01:41,720
Speaker 1: line means that I get to have catch ups with

30
00:01:41,800 --> 00:01:44,880
Speaker 1: them regularly, which has been pretty useful to me personally

31
00:01:45,120 --> 00:01:48,520
Speaker 1: because this question around AI and data centers has had

32
00:01:48,560 --> 00:01:50,760
Speaker 1: a lot of people talking, a lot of people have opinions,

33
00:01:51,000 --> 00:01:54,040
Speaker 1: and so I have often found myself in situations where

34
00:01:54,200 --> 00:01:57,960
Speaker 1: people are expressing their opinions, and in those situations, I've

35
00:01:58,040 --> 00:02:01,520
Speaker 1: developed this tactic to different myself from the pack, which

36
00:02:01,600 --> 00:02:06,040
Speaker 1: is that in the situation, I just regurgitate whatever Helen

37
00:02:06,040 --> 00:02:08,400
Speaker 1: and Natalie last told me about data centers, and everyone

38
00:02:08,440 --> 00:02:11,200
Speaker 1: thinks that I'm really smart. So first off, let's start

39
00:02:11,200 --> 00:02:13,480
Speaker 1: just with the headline numbers. How much data center build

40
00:02:13,560 --> 00:02:16,680
Speaker 1: are we expecting in the US According to the report

41
00:02:16,680 --> 00:02:17,919
Speaker 1: that we just published, so.

42
00:02:17,919 --> 00:02:21,440
Speaker 2: BNF's latest outlook has data center and demand more than

43
00:02:21,480 --> 00:02:25,080
Speaker 2: doubling from thirty five gigatts today to close to eighty

44
00:02:25,120 --> 00:02:28,360
Speaker 2: gigawatts in twenty thirty five. This would account for close

45
00:02:28,360 --> 00:02:31,600
Speaker 2: to nine percent of total US electricity demand.

46
00:02:32,000 --> 00:02:34,480
Speaker 1: Wow, so we're expecting let me just do the mass

47
00:02:34,600 --> 00:02:38,440
Speaker 1: suddenly like forty five gigawatts ish, which for those of

48
00:02:38,480 --> 00:02:41,000
Speaker 1: you who are you know, maybe new to the energy space,

49
00:02:41,080 --> 00:02:45,000
Speaker 1: that's like twenty to thirty nuclear plants, and nuclear plants

50
00:02:45,000 --> 00:02:46,960
Speaker 1: are really really big and take a long time to build.

51
00:02:46,960 --> 00:02:50,480
Speaker 1: That's a lot of demand. So we are forecasting some

52
00:02:50,840 --> 00:02:53,680
Speaker 1: astronomical amount of data center build. How do we compare

53
00:02:53,880 --> 00:02:56,680
Speaker 1: to everyone else that has an opinion on this topic?

54
00:02:57,320 --> 00:03:03,760
Speaker 3: We're relatively conservative and relatively conservative. Yes, yeah, our overall

55
00:03:04,000 --> 00:03:07,800
Speaker 3: demand build is fairly low in terms of uptake relative

56
00:03:07,880 --> 00:03:09,160
Speaker 3: to our third parties.

57
00:03:09,639 --> 00:03:13,840
Speaker 1: Okay, so how come they are forecasting something so much

58
00:03:13,919 --> 00:03:16,120
Speaker 1: more aggressive than us, or how come we are so

59
00:03:16,240 --> 00:03:17,720
Speaker 1: much more conservative than them?

60
00:03:18,000 --> 00:03:22,239
Speaker 3: Well, we don't really know what our third party counterparts

61
00:03:22,320 --> 00:03:24,560
Speaker 3: do in terms of their forecast, but what we do

62
00:03:24,680 --> 00:03:27,960
Speaker 3: know is like how we forecast data centers, and our

63
00:03:28,040 --> 00:03:31,359
Speaker 3: focus was really to look at like how data centers

64
00:03:31,400 --> 00:03:34,280
Speaker 3: move from one stage to the next. So in our

65
00:03:34,320 --> 00:03:36,280
Speaker 3: project data base, what we know is that we can

66
00:03:36,320 --> 00:03:39,520
Speaker 3: see data center stages. So we have like early stage,

67
00:03:39,800 --> 00:03:43,160
Speaker 3: which is basically anything kind of just just got announced.

68
00:03:43,200 --> 00:03:45,840
Speaker 3: We have projects that are committed, which is anything that

69
00:03:45,880 --> 00:03:49,000
Speaker 3: has some type of like land or permitting agreement, things

70
00:03:49,000 --> 00:03:52,400
Speaker 3: that are under construction and then live. And what we

71
00:03:52,480 --> 00:03:55,080
Speaker 3: did was we tracked how these data centers moved from

72
00:03:55,080 --> 00:03:57,200
Speaker 3: one stage to the next, and we looked at like

73
00:03:57,280 --> 00:03:59,920
Speaker 3: the probability of how these data centers moved.

74
00:03:59,760 --> 00:04:02,520
Speaker 1: From and stay to the next. So typically, how long

75
00:04:02,560 --> 00:04:05,000
Speaker 1: does it take for a data center, you know, from

76
00:04:05,680 --> 00:04:08,960
Speaker 1: early on in this pipeline to being commissioned, how long

77
00:04:08,960 --> 00:04:09,600
Speaker 1: would that take?

78
00:04:10,040 --> 00:04:13,360
Speaker 3: Yeah, So what we've found based on data between twenty

79
00:04:13,440 --> 00:04:15,920
Speaker 3: twenty to twenty twenty four is that it takes seven

80
00:04:16,040 --> 00:04:18,719
Speaker 3: years to build a data center, which is a really,

81
00:04:18,880 --> 00:04:19,680
Speaker 3: really long time.

82
00:04:19,960 --> 00:04:23,240
Speaker 1: Okay, So that's really interesting. So in a way, our

83
00:04:23,320 --> 00:04:25,760
Speaker 1: forecast we're saying with a fair amount of confidence because

84
00:04:25,800 --> 00:04:27,440
Speaker 1: you know, twenty thirty five is only just over a

85
00:04:27,480 --> 00:04:31,160
Speaker 1: decade away, and we have data on everything that's getting built,

86
00:04:31,240 --> 00:04:33,400
Speaker 1: and we know that it takes most of that decade

87
00:04:33,440 --> 00:04:37,440
Speaker 1: for it to get built. So anyone who's forecasting something

88
00:04:37,440 --> 00:04:39,480
Speaker 1: more aggressive than us either has a different view on

89
00:04:39,520 --> 00:04:41,479
Speaker 1: how long it takes to build data centers, or they

90
00:04:41,560 --> 00:04:43,960
Speaker 1: must have different data, or maybe they're using a completely

91
00:04:43,960 --> 00:04:46,680
Speaker 1: different methodology. But it's good to know that we're the

92
00:04:46,680 --> 00:04:48,920
Speaker 1: ones doing it right. So who's building all of these

93
00:04:49,000 --> 00:04:53,919
Speaker 1: data centers this sort of colossal volume of new demand.

94
00:04:54,360 --> 00:04:57,720
Speaker 2: Yeah, so the data center market is pretty concentrated. There's

95
00:04:57,760 --> 00:05:01,800
Speaker 2: two main types of owners. There's colo data centers who

96
00:05:01,960 --> 00:05:05,719
Speaker 2: have buildings with multiple tenants, and you have these self

97
00:05:05,760 --> 00:05:09,360
Speaker 2: build companies, which are typically your large tech companies or

98
00:05:09,440 --> 00:05:14,440
Speaker 2: hyperscalers is what they're usually referred to. And the hyperscalers

99
00:05:14,480 --> 00:05:19,640
Speaker 2: of Google, Amazon, Microsoft, and Meta are close to fifty

100
00:05:19,640 --> 00:05:23,560
Speaker 2: percent of total operating capacity today and this is only

101
00:05:23,600 --> 00:05:26,960
Speaker 2: set to grow. They're building much larger campuses close to

102
00:05:27,040 --> 00:05:31,640
Speaker 2: gigawatts scale. Amazon has multiple gigawat data center campuses in

103
00:05:31,720 --> 00:05:36,680
Speaker 2: development in Virginia. Meta has another two gigawatts in Louisiana,

104
00:05:36,960 --> 00:05:40,039
Speaker 2: and just as a point of reference, in the last decade,

105
00:05:40,120 --> 00:05:43,880
Speaker 2: data centers have typically been in the tens of megawatts.

106
00:05:44,279 --> 00:05:46,680
Speaker 2: So really, as we're pushing through these hundreds of megawatts

107
00:05:46,680 --> 00:05:50,560
Speaker 2: and gigawat size, the rise of uptake will be much

108
00:05:50,720 --> 00:05:51,960
Speaker 2: faster and larger.

109
00:05:52,240 --> 00:05:54,680
Speaker 1: Okay, And so just just to make sure I've understood

110
00:05:54,800 --> 00:05:58,640
Speaker 1: overall correctly, those four companies are fifty percent of the

111
00:05:58,760 --> 00:06:02,120
Speaker 1: data center build today, but we think that there's going

112
00:06:02,160 --> 00:06:04,440
Speaker 1: to be even more because they're the companies that are

113
00:06:04,440 --> 00:06:07,400
Speaker 1: building these really big data centers that are so much

114
00:06:07,400 --> 00:06:10,360
Speaker 1: of what we're expecting. So you've already alluded to these

115
00:06:10,440 --> 00:06:12,840
Speaker 1: data centers are maybe bigger than the ones we've seen

116
00:06:12,880 --> 00:06:15,000
Speaker 1: in the past, that's the trend. But in what other

117
00:06:15,040 --> 00:06:17,719
Speaker 1: ways are the data centers that we're expecting different to

118
00:06:17,800 --> 00:06:19,960
Speaker 1: the ones that we've seen in the past.

119
00:06:20,680 --> 00:06:23,279
Speaker 3: Yeah, before we jump into that, I think it's important

120
00:06:23,279 --> 00:06:26,560
Speaker 3: to understand how be any of categorized as data centers.

121
00:06:26,800 --> 00:06:30,760
Speaker 3: So we categorize data centers in three different ways. Vers size,

122
00:06:30,800 --> 00:06:35,200
Speaker 3: which Natalie had alluded to, first retail, wholesale and hyperscaler

123
00:06:35,320 --> 00:06:37,359
Speaker 3: So that's just based on the project size of a

124
00:06:37,440 --> 00:06:40,640
Speaker 3: data center. And then there's operator type, which is the

125
00:06:40,680 --> 00:06:45,280
Speaker 3: ownership so either self build or co location, which Natalie

126
00:06:45,320 --> 00:06:50,120
Speaker 3: had already explained. And then there's workload, So workload is

127
00:06:50,200 --> 00:06:54,080
Speaker 3: based on just the computing process of a data center,

128
00:06:54,400 --> 00:06:58,040
Speaker 3: and there are many different types of workload from cloud

129
00:06:58,279 --> 00:07:02,520
Speaker 3: or enterprise, telecom, crypto mining, and that determines a lot

130
00:07:02,600 --> 00:07:06,960
Speaker 3: about the data center's overall infrastructure and then also their

131
00:07:07,040 --> 00:07:09,240
Speaker 3: overall load and power consumption.

132
00:07:09,760 --> 00:07:11,720
Speaker 1: So then, I mean one of the other things that

133
00:07:11,760 --> 00:07:13,720
Speaker 1: you've spoken to me about is when we're talking about

134
00:07:13,760 --> 00:07:16,840
Speaker 1: AI data centers, that there's two main flavors, So can

135
00:07:16,880 --> 00:07:18,920
Speaker 1: you just talk me through those as well.

136
00:07:19,560 --> 00:07:23,400
Speaker 2: Within AI, we mainly branch it out in two main workloads,

137
00:07:23,440 --> 00:07:27,280
Speaker 2: AI training and AI inference. AI training is processing a

138
00:07:27,360 --> 00:07:29,720
Speaker 2: large amount of data in order to train these large

139
00:07:29,800 --> 00:07:33,600
Speaker 2: language models, and AI inference is taking those already train

140
00:07:33,720 --> 00:07:36,520
Speaker 2: models for real time applications in use, like when you're

141
00:07:36,600 --> 00:07:40,160
Speaker 2: careering in chat GPT, And those two workloads have also

142
00:07:40,800 --> 00:07:46,400
Speaker 2: different location and constraints of the data center itself. AI inference,

143
00:07:46,560 --> 00:07:50,840
Speaker 2: since they're theoretically interacting with the end user in real time,

144
00:07:51,120 --> 00:07:54,600
Speaker 2: they'll care more about latency and location parameters to that

145
00:07:54,760 --> 00:07:57,960
Speaker 2: end user. AI training a lot of that processing happens

146
00:07:57,960 --> 00:08:01,040
Speaker 2: on site, so theoretically they could be more flexible on

147
00:08:01,080 --> 00:08:04,520
Speaker 2: where they locate, and they could follow where there's available

148
00:08:04,560 --> 00:08:06,480
Speaker 2: power or other constraints.

149
00:08:06,840 --> 00:08:09,480
Speaker 1: Okay, I mean you use the word theoretically there, which

150
00:08:09,480 --> 00:08:11,160
Speaker 1: maybe is doing a lot of work. And we'll come

151
00:08:11,160 --> 00:08:14,560
Speaker 1: back to where people are actually building data centers later on,

152
00:08:14,560 --> 00:08:17,160
Speaker 1: because I do have a question about that. But roughly,

153
00:08:17,240 --> 00:08:20,200
Speaker 1: what is do we have an idea of what proportion

154
00:08:20,320 --> 00:08:24,080
Speaker 1: of the data center building the pipeline is for AI

155
00:08:24,240 --> 00:08:26,400
Speaker 1: and what proportion of it is training and what proportion

156
00:08:26,520 --> 00:08:27,360
Speaker 1: of it is inference.

157
00:08:27,960 --> 00:08:32,240
Speaker 2: That's actually very tough to ascertain exactly what the split is.

158
00:08:32,280 --> 00:08:35,440
Speaker 2: If we look at a large Gigawat campus, some of

159
00:08:35,480 --> 00:08:37,800
Speaker 2: their buildings could be used for training today, but it

160
00:08:37,800 --> 00:08:40,200
Speaker 2: could be used for inference in the future. Similarly, we

161
00:08:40,280 --> 00:08:44,400
Speaker 2: talked about different owner types a co location building, they

162
00:08:44,440 --> 00:08:47,520
Speaker 2: could have, you know, multiple tenants, and unless you know

163
00:08:47,559 --> 00:08:50,320
Speaker 2: exactly who the tenant is and what type of workloads

164
00:08:50,320 --> 00:08:52,920
Speaker 2: they're running, it's also difficult to know if it is

165
00:08:52,960 --> 00:08:55,839
Speaker 2: for AI training at inference. But we can infer based

166
00:08:55,880 --> 00:08:57,280
Speaker 2: on where they're getting built.

167
00:08:57,679 --> 00:09:00,679
Speaker 1: But these distinctions, I mean, what I'm here here is

168
00:09:00,880 --> 00:09:03,640
Speaker 1: that it's not just that nobody tells us which it

169
00:09:03,720 --> 00:09:05,760
Speaker 1: is that makes it a little bit of a gray area.

170
00:09:05,800 --> 00:09:09,000
Speaker 1: It's that actually even a data center itself might sometimes

171
00:09:09,200 --> 00:09:11,000
Speaker 1: at one point in its life be doing one thing

172
00:09:11,040 --> 00:09:12,520
Speaker 1: and at a different point in it's life be doing

173
00:09:12,559 --> 00:09:13,040
Speaker 1: something else.

174
00:09:13,360 --> 00:09:15,880
Speaker 3: I think for power folks, I often try to like

175
00:09:16,000 --> 00:09:18,720
Speaker 3: frame it like a data center is very similar to

176
00:09:18,760 --> 00:09:21,840
Speaker 3: a battery. Like batteries can do multiple different types of

177
00:09:21,960 --> 00:09:25,720
Speaker 3: energy services or ancillary services, a data center can do

178
00:09:25,800 --> 00:09:29,000
Speaker 3: multiple different types of workload. It can be doing AI

179
00:09:29,040 --> 00:09:32,760
Speaker 3: training or cloud as long as the configuration is correct,

180
00:09:32,840 --> 00:09:34,560
Speaker 3: or it could do those types of workload.

181
00:09:34,679 --> 00:09:37,000
Speaker 1: Got it. So just because you've optimized one to one

182
00:09:37,000 --> 00:09:39,839
Speaker 1: thing doesn't mean that it can't do the other thing.

183
00:09:40,120 --> 00:09:45,360
Speaker 3: Yeah, And particularly in colocation data centers where basically these

184
00:09:45,360 --> 00:09:48,720
Speaker 3: companies are renting out their RT servers to tenants, those

185
00:09:48,760 --> 00:09:51,920
Speaker 3: tenants are probably doing different types of workload. There can

186
00:09:52,000 --> 00:09:54,360
Speaker 3: be multiple applications in a data center.

187
00:09:54,640 --> 00:09:56,800
Speaker 2: Yeah, a workload, which is often why you see in

188
00:09:56,840 --> 00:10:00,079
Speaker 2: colocation that they're optimizing for everything because they don't know

189
00:10:00,120 --> 00:10:03,720
Speaker 2: who their tenant is. I will add, though, in terms

190
00:10:03,720 --> 00:10:07,199
Speaker 2: of AI data centers that they are quite different from

191
00:10:07,280 --> 00:10:10,000
Speaker 2: data centers today. We already talked that these data centers

192
00:10:10,000 --> 00:10:12,280
Speaker 2: are getting larger. A lot of that has to do

193
00:10:12,320 --> 00:10:15,360
Speaker 2: with the GPUs in those servers being much more so.

194
00:10:15,400 --> 00:10:17,280
Speaker 1: Can you just what a GPUs.

195
00:10:17,040 --> 00:10:22,520
Speaker 2: Graphical processing units, which is they're similar to CPUs central

196
00:10:22,520 --> 00:10:26,640
Speaker 2: processing units, but they're very specialized in they're cooler.

197
00:10:28,080 --> 00:10:30,719
Speaker 1: I think that we've got to the level of understanding.

198
00:10:30,920 --> 00:10:32,839
Speaker 1: I'm good with it. So it's like it's like a chip,

199
00:10:32,960 --> 00:10:34,040
Speaker 1: but it's a different kind of chip.

200
00:10:34,120 --> 00:10:37,320
Speaker 2: Yeah, it's very good at parallel processing, which is optimized

201
00:10:37,360 --> 00:10:40,440
Speaker 2: for training large language models. I'm sure lots of people

202
00:10:40,440 --> 00:10:43,080
Speaker 2: have heard of Navidia and their GPUs. A lot of

203
00:10:43,080 --> 00:10:46,439
Speaker 2: tech companies are also building their AI accelerators, which are

204
00:10:46,520 --> 00:10:50,720
Speaker 2: specialized chips for these models. So a lot of the

205
00:10:50,760 --> 00:10:54,520
Speaker 2: design of the data center today in order to accommodate

206
00:10:54,559 --> 00:10:57,960
Speaker 2: AI training workloads are different. A lot of that has

207
00:10:58,000 --> 00:11:00,520
Speaker 2: to do with these GPUs and the rack that they're

208
00:11:00,520 --> 00:11:02,800
Speaker 2: on are going to be much more power dense, which

209
00:11:02,800 --> 00:11:06,520
Speaker 2: means they need a lot more sophisticated cooling technologies to

210
00:11:06,559 --> 00:11:09,640
Speaker 2: accommodate those sort of rack densities that could be tenext

211
00:11:09,720 --> 00:11:13,440
Speaker 2: to what typical data centers are today. We've seen, for example,

212
00:11:13,600 --> 00:11:16,960
Speaker 2: Metas scrapping data centers that were not ail ready, so

213
00:11:17,000 --> 00:11:20,120
Speaker 2: it's quite difficult to retrofit a data center from five

214
00:11:20,160 --> 00:11:22,040
Speaker 2: years ago to an AI workload.

215
00:11:22,400 --> 00:11:23,360
Speaker 1: So while a.

216
00:11:23,400 --> 00:11:26,079
Speaker 2: Data center in the future could have multiple uses, it's

217
00:11:26,120 --> 00:11:28,920
Speaker 2: also very specialized to what they're going to be running.

218
00:11:29,400 --> 00:11:31,560
Speaker 1: Got it is it? Would it be right for me

219
00:11:31,640 --> 00:11:34,360
Speaker 1: to say that, like, a data center designed for AI

220
00:11:34,760 --> 00:11:37,439
Speaker 1: can be used for other things too, but a data

221
00:11:37,440 --> 00:11:40,560
Speaker 1: center not designed for AI probably can't do AI? Is

222
00:11:40,559 --> 00:11:41,880
Speaker 1: that a fair statement?

223
00:11:42,200 --> 00:11:45,040
Speaker 3: I think that's a pretty fair statement. We're basically seeing

224
00:11:45,080 --> 00:11:48,199
Speaker 3: new data centers being designed in a way that allows

225
00:11:48,280 --> 00:11:51,800
Speaker 3: for AI training, and so there's an influence of like

226
00:11:51,880 --> 00:11:55,080
Speaker 3: AI and data center design to be larger and more

227
00:11:55,200 --> 00:11:56,000
Speaker 3: power tents.

228
00:11:56,440 --> 00:12:00,439
Speaker 1: So in these training data centers that building these models.

229
00:12:00,679 --> 00:12:02,800
Speaker 1: And one of the charts in the note that I

230
00:12:02,840 --> 00:12:06,240
Speaker 1: really loved but haven't quite fully digested is one showing

231
00:12:06,480 --> 00:12:09,040
Speaker 1: the amount of and you don't have to explain this

232
00:12:09,200 --> 00:12:15,120
Speaker 1: unit amount of terror flops required too. Terrorflop that's just

233
00:12:15,160 --> 00:12:17,520
Speaker 1: the unit of like computer work, isn't it.

234
00:12:17,600 --> 00:12:19,640
Speaker 2: Yeah, it's just a basic unit of computation.

235
00:12:20,040 --> 00:12:24,360
Speaker 1: Yeah, the amount of terror flops needed to design different

236
00:12:24,559 --> 00:12:28,320
Speaker 1: AI models as they've become more and more in sophisticated

237
00:12:28,600 --> 00:12:31,840
Speaker 1: and so obviously there's been an expectation that trend is

238
00:12:31,840 --> 00:12:34,680
Speaker 1: going to continue, and everyone was you know, freaking out

239
00:12:34,960 --> 00:12:36,920
Speaker 1: in both good ways and bad ways about all the

240
00:12:36,920 --> 00:12:38,920
Speaker 1: power demand that this will intel. And then I remember

241
00:12:38,960 --> 00:12:41,320
Speaker 1: deep seek came along and a lot of the people

242
00:12:41,440 --> 00:12:44,480
Speaker 1: were like, oh, this changes everything. Can you just provide

243
00:12:44,520 --> 00:12:46,120
Speaker 1: a bit of clarity on all of this.

244
00:12:46,720 --> 00:12:49,080
Speaker 3: There was a couple of things I really tried to

245
00:12:49,200 --> 00:12:52,160
Speaker 3: understand about all of this, just in terms of like

246
00:12:52,360 --> 00:12:55,040
Speaker 3: power market fundamentals. I think when it comes to like

247
00:12:55,240 --> 00:12:59,480
Speaker 3: forecasting for power demand, all things come down to some

248
00:13:00,160 --> 00:13:04,280
Speaker 3: very basic constructs. It's usually like how much quantity of

249
00:13:04,280 --> 00:13:08,800
Speaker 3: something and then the energy intensity of that something. And

250
00:13:08,960 --> 00:13:14,200
Speaker 3: for data centers it's a very similar process. And when

251
00:13:14,200 --> 00:13:18,240
Speaker 3: we think about like the energy efficiency of large language

252
00:13:18,280 --> 00:13:21,360
Speaker 3: model AI training data centers, there's like a couple things

253
00:13:21,400 --> 00:13:24,760
Speaker 3: to think through. First is around the energy intensity of

254
00:13:24,800 --> 00:13:28,920
Speaker 3: power consumption from a chips perspective, and chip innovation is

255
00:13:28,960 --> 00:13:33,960
Speaker 3: often like confused with like increasing energy efficiency. That's not

256
00:13:34,080 --> 00:13:37,840
Speaker 3: necessarily the case. Typically when we think about chip innovation,

257
00:13:38,320 --> 00:13:42,400
Speaker 3: it's often optimized for like operations per second, which means

258
00:13:42,440 --> 00:13:46,280
Speaker 3: that like typically more like every generation of chips tend

259
00:13:46,320 --> 00:13:49,880
Speaker 3: to have draw more power and therefore, like it increases

260
00:13:49,920 --> 00:13:53,920
Speaker 3: the energy intensity of a data center. So there are

261
00:13:54,480 --> 00:13:57,800
Speaker 3: new chips that are getting invented, like the Nvidia Blackwell

262
00:13:57,800 --> 00:14:01,400
Speaker 3: that focuses on energy efficiency. But in general what we've

263
00:14:01,400 --> 00:14:04,400
Speaker 3: seen is like the more advanced chips tend to draw

264
00:14:04,480 --> 00:14:08,240
Speaker 3: on more power. The other thing that is really important

265
00:14:08,280 --> 00:14:11,640
Speaker 3: for AI training is then like the number of parameters

266
00:14:12,000 --> 00:14:15,560
Speaker 3: that an AI training model focuses on. So parameters is

267
00:14:15,640 --> 00:14:20,360
Speaker 3: just like points of active information for a model to

268
00:14:20,440 --> 00:14:21,000
Speaker 3: think through.

269
00:14:21,240 --> 00:14:23,360
Speaker 1: So it's just the number of different things it thinks

270
00:14:23,400 --> 00:14:26,520
Speaker 1: about sort of like so if I was thinking, like

271
00:14:27,000 --> 00:14:29,120
Speaker 1: should I go to work today or should I walk

272
00:14:29,160 --> 00:14:32,440
Speaker 1: to work today? Like, if I had one parameter, it

273
00:14:32,560 --> 00:14:34,760
Speaker 1: might be what's the weather like outside? And if the

274
00:14:34,840 --> 00:14:37,720
Speaker 1: two parameter might be what's the weather like outside and

275
00:14:37,760 --> 00:14:39,360
Speaker 1: what day of the week is it? Yeah, you might

276
00:14:39,400 --> 00:14:42,240
Speaker 1: determine whether or not I go to work. So that's

277
00:14:42,280 --> 00:14:44,520
Speaker 1: like what each of those is a parameter.

278
00:14:44,280 --> 00:14:49,160
Speaker 3: Exactly, exactly precisely and in general, like in the large

279
00:14:49,240 --> 00:14:51,760
Speaker 3: language model community historically, or at least in the power

280
00:14:51,800 --> 00:14:55,440
Speaker 3: industry historically, what people had thought through was that like

281
00:14:56,040 --> 00:15:01,600
Speaker 3: more sophisticated models required more parameters. So with every generation

282
00:15:01,800 --> 00:15:05,600
Speaker 3: of like chatchipt or Gemina or cloud. If you look

283
00:15:05,640 --> 00:15:08,960
Speaker 3: at their like technical reports, what you'll see is that

284
00:15:09,000 --> 00:15:12,000
Speaker 3: there's increasing amount of parameters that is used to like

285
00:15:12,080 --> 00:15:14,960
Speaker 3: train these large language models, and so in general, like

286
00:15:15,240 --> 00:15:19,160
Speaker 3: there was this overall consensus that the energy intensity of

287
00:15:19,200 --> 00:15:23,320
Speaker 3: air training is an upward trend. You use more powerful chips,

288
00:15:23,360 --> 00:15:26,600
Speaker 3: and you're using your training on more and more parameters,

289
00:15:26,840 --> 00:15:30,280
Speaker 3: and so that's all driving more and more electricity consumption.

290
00:15:30,680 --> 00:15:34,320
Speaker 3: Deep Seek came out in December twenty twenty four, and

291
00:15:34,400 --> 00:15:37,920
Speaker 3: in their technical report, what was really cool was that

292
00:15:38,000 --> 00:15:41,800
Speaker 3: they had a different training process. So it uses something

293
00:15:41,840 --> 00:15:45,800
Speaker 3: called a mixture of experts training process, where instead of

294
00:15:45,880 --> 00:15:49,320
Speaker 3: just like plugging in all of their parameters all at once,

295
00:15:49,720 --> 00:15:53,920
Speaker 3: what they did was they pre categorized their parameters into experts.

296
00:15:54,240 --> 00:15:58,720
Speaker 3: So like certain parameters that specialize in math or like colors,

297
00:15:58,960 --> 00:16:01,520
Speaker 3: they would like pre categze them so that when you

298
00:16:01,560 --> 00:16:04,560
Speaker 3: put in a query of a question to ask the

299
00:16:04,600 --> 00:16:08,080
Speaker 3: model to train, it would know which specialization to pull

300
00:16:08,160 --> 00:16:12,040
Speaker 3: the parameters from, which then drew on less power. So

301
00:16:12,400 --> 00:16:15,760
Speaker 3: the other thing that the technical report published was that

302
00:16:15,880 --> 00:16:19,840
Speaker 3: deep Seek performed at a very high level with chat

303
00:16:19,880 --> 00:16:23,000
Speaker 3: GPT or like other types of large language models that

304
00:16:23,160 --> 00:16:26,320
Speaker 3: used a lot more parameters, and so it kind of

305
00:16:26,320 --> 00:16:30,520
Speaker 3: broke the assumption that you needed a whole bunch of

306
00:16:30,560 --> 00:16:33,760
Speaker 3: parameters to make really sophisticated large language models.

307
00:16:34,280 --> 00:16:37,240
Speaker 1: And so do you think that that will massively change

308
00:16:37,440 --> 00:16:39,480
Speaker 1: the outlook for power demand from AI?

309
00:16:40,160 --> 00:16:42,280
Speaker 3: So to answer your question from like a data center

310
00:16:42,320 --> 00:16:47,520
Speaker 3: demand perspective, not necessarily in our near term forecast on

311
00:16:47,880 --> 00:16:50,920
Speaker 3: in our data center outlook, we use like a project

312
00:16:50,920 --> 00:16:54,680
Speaker 3: by project level forecasts, right. But in our new energy outlook,

313
00:16:54,920 --> 00:16:58,440
Speaker 3: which focused more on like long term forecasts for data

314
00:16:58,440 --> 00:17:02,920
Speaker 3: center demand, it focused on that fundamentals based way of forecasting,

315
00:17:02,960 --> 00:17:06,800
Speaker 3: which it looks at long term in any given like market,

316
00:17:07,240 --> 00:17:11,240
Speaker 3: what data generation and data usage looks like relative to

317
00:17:11,280 --> 00:17:14,320
Speaker 3: the energy intensity of that data generation.

318
00:17:14,880 --> 00:17:15,040
Speaker 2: Right.

319
00:17:15,200 --> 00:17:17,280
Speaker 1: And there's a point you make in the report I

320
00:17:17,320 --> 00:17:19,359
Speaker 1: recall and I don't think you were talking about deep

321
00:17:19,400 --> 00:17:22,160
Speaker 1: seek here. I think you were talking about data center efficiency.

322
00:17:22,200 --> 00:17:24,760
Speaker 1: But bring up Jevins paradox, which is this idea that

323
00:17:24,760 --> 00:17:27,320
Speaker 1: if you make something more efficient, it doesn't mean necessarily

324
00:17:27,359 --> 00:17:30,440
Speaker 1: that we save energy, it's that we just do more

325
00:17:30,480 --> 00:17:33,160
Speaker 1: with what we were going to consume anyway, And could

326
00:17:33,200 --> 00:17:35,240
Speaker 1: the same logic be applied for deep seek is if

327
00:17:35,280 --> 00:17:38,280
Speaker 1: it is more efficient with parameters and therefore energy and

328
00:17:38,440 --> 00:17:41,320
Speaker 1: also computer usage, then that just opens the door to

329
00:17:41,359 --> 00:17:44,360
Speaker 1: do cooler things with AI than would have previously been possible,

330
00:17:44,680 --> 00:17:45,879
Speaker 1: rather than to save energy.

331
00:17:46,280 --> 00:17:50,240
Speaker 3: Yeah, the energy intensity curve of AI training data centers

332
00:17:50,280 --> 00:17:53,080
Speaker 3: like instead of it being like an upward swing, it

333
00:17:53,800 --> 00:17:56,919
Speaker 3: goes down. But we also then know it opens up

334
00:17:57,000 --> 00:17:59,480
Speaker 3: a lot more opportunities for a lot of different types

335
00:17:59,480 --> 00:18:03,920
Speaker 3: of business to maybe do AI training right, which means

336
00:18:03,920 --> 00:18:07,120
Speaker 3: that you have more companies that may be doing this.

337
00:18:07,520 --> 00:18:10,080
Speaker 3: So got it runs paradox.

338
00:18:09,920 --> 00:18:12,480
Speaker 1: Very interesting and actually I think that brings some real

339
00:18:12,480 --> 00:18:15,840
Speaker 1: clarity into what this whole deep seat thing means for

340
00:18:15,880 --> 00:18:20,920
Speaker 1: power demand, which my main takeaway is like not that much. Ultimately,

341
00:18:21,240 --> 00:18:24,080
Speaker 1: it's very difficult to say, but we shouldn't be saying, oh,

342
00:18:24,119 --> 00:18:26,159
Speaker 1: this means all this data center demand growth isn't going

343
00:18:26,200 --> 00:18:29,560
Speaker 1: to happen. Yes, So pulling out again, there's all of

344
00:18:29,600 --> 00:18:32,560
Speaker 1: this data center build that's going to happen in the US,

345
00:18:32,800 --> 00:18:35,040
Speaker 1: four companies are going to be behind a little bit

346
00:18:35,040 --> 00:18:36,960
Speaker 1: more than half of it. Where are they going to

347
00:18:37,000 --> 00:18:39,640
Speaker 1: be building all of this? And why are they going

348
00:18:39,680 --> 00:18:40,960
Speaker 1: to be building in those places?

349
00:18:41,440 --> 00:18:44,800
Speaker 2: Yeah? I'll also add on those four companies that are

350
00:18:45,119 --> 00:18:48,159
Speaker 2: most of the data center market, they're also the companies

351
00:18:48,440 --> 00:18:51,560
Speaker 2: that you know aren't building these AI training models and

352
00:18:51,640 --> 00:18:55,120
Speaker 2: have the capacity to train large scale models because it's

353
00:18:55,160 --> 00:18:59,280
Speaker 2: a very costly exercise that is only set to grow.

354
00:18:59,480 --> 00:19:02,080
Speaker 2: They're also forty of the data center market, but they're

355
00:19:02,119 --> 00:19:05,560
Speaker 2: also the ones training AI models because actually, not many

356
00:19:05,560 --> 00:19:08,040
Speaker 2: companies can train A models, right, right.

357
00:19:07,880 --> 00:19:10,360
Speaker 1: So a lot of the new big demand is becoming

358
00:19:10,440 --> 00:19:13,359
Speaker 1: from these four companies. So a lot of the demand

359
00:19:13,440 --> 00:19:14,680
Speaker 1: that we're talking about.

360
00:19:14,880 --> 00:19:17,000
Speaker 2: Yeah, I guess today when we look at the data

361
00:19:17,000 --> 00:19:20,480
Speaker 2: center fleet, they're mostly for cloud. But going forward, the

362
00:19:20,520 --> 00:19:24,040
Speaker 2: companies that can actually train AI models is less than

363
00:19:24,080 --> 00:19:26,720
Speaker 2: ten companies training like frontier models, and that those are

364
00:19:26,800 --> 00:19:28,240
Speaker 2: going to be the big tech companies.

365
00:19:28,520 --> 00:19:31,760
Speaker 1: And so then where is this happening and why in

366
00:19:31,800 --> 00:19:32,800
Speaker 1: those locations.

367
00:19:33,119 --> 00:19:36,160
Speaker 2: So in our forecast we see three main markets emerge

368
00:19:36,200 --> 00:19:39,280
Speaker 2: through twenty thirty five. We break them down by power

369
00:19:39,320 --> 00:19:42,680
Speaker 2: region for power forecasting purposes.

370
00:19:42,200 --> 00:19:45,960
Speaker 1: And that's just because that's how we think. I don't

371
00:19:46,000 --> 00:19:47,560
Speaker 1: see states, I see power region.

372
00:19:47,880 --> 00:19:51,280
Speaker 2: Yeah, so we see PGM or Coat and Southeast, and

373
00:19:51,359 --> 00:19:57,720
Speaker 2: within PGM, which spans fourteen states, we see Virginia continuing

374
00:19:57,760 --> 00:20:01,320
Speaker 2: being one of the biggest markets so than Virginia has

375
00:20:01,400 --> 00:20:05,800
Speaker 2: been data center capital for the last decade. If Virginia

376
00:20:05,840 --> 00:20:09,000
Speaker 2: was a country, it would follow the US and China

377
00:20:09,040 --> 00:20:11,360
Speaker 2: as having the largest data center market.

378
00:20:11,320 --> 00:20:12,480
Speaker 4: In the world.

379
00:20:12,680 --> 00:20:14,680
Speaker 1: Wow. Can I say that again? Wow?

380
00:20:14,840 --> 00:20:15,080
Speaker 3: Yeah.

381
00:20:15,119 --> 00:20:19,119
Speaker 2: So Northern Virginia has been kind of at the center

382
00:20:19,320 --> 00:20:21,760
Speaker 2: of data center build out. A lot of this has

383
00:20:21,800 --> 00:20:24,399
Speaker 2: to do with a bit of history. They had the

384
00:20:24,440 --> 00:20:28,760
Speaker 2: first Internet exchange point in the nineties, which was kind

385
00:20:28,760 --> 00:20:31,480
Speaker 2: of the beginning of small data center build out, and

386
00:20:31,720 --> 00:20:35,679
Speaker 2: data centers typically continue to cluster an existing markets, so

387
00:20:36,000 --> 00:20:42,000
Speaker 2: as data centers grow, they'll have supporting infrastructure like fiber optics,

388
00:20:42,600 --> 00:20:48,120
Speaker 2: utility relationships, and workforce availability that allows more data centers

389
00:20:48,160 --> 00:20:51,240
Speaker 2: to continue to grow. So over time, Northern Virginia just

390
00:20:51,280 --> 00:20:53,879
Speaker 2: has gotten hotter and hotter, and we see in our

391
00:20:53,880 --> 00:20:57,320
Speaker 2: project Pipeline a lot of that continuing to grow. We

392
00:20:57,440 --> 00:21:00,320
Speaker 2: do know that a lot of this could be more

393
00:21:00,400 --> 00:21:03,920
Speaker 2: AI inference rather than AI training, but still a lot

394
00:21:03,920 --> 00:21:07,840
Speaker 2: of data center demand happening. There. Another state in PJM

395
00:21:07,960 --> 00:21:11,440
Speaker 2: is Ohio, which is emerging as one of the main

396
00:21:11,680 --> 00:21:16,000
Speaker 2: hubs in the Midwest. Google is building data centers in

397
00:21:16,080 --> 00:21:19,359
Speaker 2: New Albany, which is just outside Columbus, one of the

398
00:21:19,440 --> 00:21:22,600
Speaker 2: largest cities in Ohio, as well as other co location

399
00:21:22,960 --> 00:21:24,520
Speaker 2: and hyperscular companies.

400
00:21:24,800 --> 00:21:27,719
Speaker 1: You know, you've highlighted some a couple of major markets

401
00:21:27,720 --> 00:21:29,560
Speaker 1: with PGM. I think I think I thought was really

402
00:21:29,560 --> 00:21:33,160
Speaker 1: interesting and maybe slightly paradoxical in the report you wrote

403
00:21:33,359 --> 00:21:35,639
Speaker 1: was there's this chart from that has a survey of

404
00:21:35,760 --> 00:21:39,160
Speaker 1: data center developers saying what do you prioritize when you're

405
00:21:39,160 --> 00:21:41,000
Speaker 1: thinking about where to build a data center? And I

406
00:21:41,000 --> 00:21:43,439
Speaker 1: think we said that the top three survey results it

407
00:21:43,480 --> 00:21:45,680
Speaker 1: wasn't our survey, it was a third parties. The top

408
00:21:45,680 --> 00:21:48,120
Speaker 1: three survey results all related to energy. It was something

409
00:21:48,160 --> 00:21:51,280
Speaker 1: like security of supply, how cheap the energy is, how

410
00:21:51,440 --> 00:21:53,480
Speaker 1: green the energy is. I can't remember, don't quote me

411
00:21:53,480 --> 00:21:56,240
Speaker 1: on them, but it was energy related. When you look

412
00:21:56,359 --> 00:22:00,240
Speaker 1: at the data of where they're currently building and have been, well,

413
00:22:00,320 --> 00:22:04,240
Speaker 1: it kind of completely contradicts that thesis. You know, PGM

414
00:22:04,320 --> 00:22:07,600
Speaker 1: doesn't have the cheapest or cleanest electricity in the US.

415
00:22:07,680 --> 00:22:09,600
Speaker 1: I mean you could say it has good security of supply,

416
00:22:09,880 --> 00:22:11,760
Speaker 1: So what is behind this paradox?

417
00:22:12,240 --> 00:22:14,560
Speaker 2: I think we a lot of the hype right now

418
00:22:14,600 --> 00:22:18,280
Speaker 2: in data center built out is AI related, and we

419
00:22:18,359 --> 00:22:21,800
Speaker 2: did talk about how AI training could be more flexible.

420
00:22:21,960 --> 00:22:25,080
Speaker 2: And we do see a lot of our merging innovations

421
00:22:25,119 --> 00:22:29,760
Speaker 2: of siting near stranded renewable assets and going against the

422
00:22:29,800 --> 00:22:34,960
Speaker 2: grain of traditional sighting. But most as we said, most

423
00:22:35,000 --> 00:22:37,720
Speaker 2: of them are continuing to build out an existing data

424
00:22:37,720 --> 00:22:42,080
Speaker 2: center market. So market's pre AI one theory is basically,

425
00:22:42,280 --> 00:22:45,720
Speaker 2: they're investing billions of dollars in a data center for

426
00:22:45,920 --> 00:22:48,880
Speaker 2: the next decade or so. While in the near term

427
00:22:48,880 --> 00:22:51,280
Speaker 2: they can plan for AI training. In the future, it

428
00:22:51,280 --> 00:22:54,480
Speaker 2: could be AI inference, or they could you know, retrofit

429
00:22:54,520 --> 00:22:57,240
Speaker 2: and sell it to a co location company altogether, and

430
00:22:57,600 --> 00:23:01,879
Speaker 2: they need to plan for those latency and redidancy requirements today.

431
00:23:02,119 --> 00:23:05,800
Speaker 2: So even though in the near term they could cite

432
00:23:05,840 --> 00:23:08,040
Speaker 2: it in you know, West Texas where there's a lot

433
00:23:08,040 --> 00:23:11,080
Speaker 2: of renewables, we still see like Dallas Fort Worth and

434
00:23:11,280 --> 00:23:14,560
Speaker 2: like northern Virginia as being hotspots, although we are seeing

435
00:23:14,880 --> 00:23:18,199
Speaker 2: a trend of going a bit outside of urban locations

436
00:23:18,200 --> 00:23:21,199
Speaker 2: and going more to where there is power supply.

437
00:23:21,520 --> 00:23:23,640
Speaker 1: Got it. I remember earlier in the podcast you said

438
00:23:23,840 --> 00:23:27,280
Speaker 1: in theory about you know where you could build training

439
00:23:27,359 --> 00:23:29,680
Speaker 1: data cents and this is the this is what you're

440
00:23:29,680 --> 00:23:31,399
Speaker 1: saying is in practice, it's like this, but we are

441
00:23:31,440 --> 00:23:34,000
Speaker 1: seeing a bit of a trend, but just maybe not

442
00:23:34,040 --> 00:23:36,080
Speaker 1: as much as you would think to the sort of

443
00:23:36,440 --> 00:23:37,960
Speaker 1: energy ideal locations.

444
00:23:38,240 --> 00:23:38,680
Speaker 4: Yeah.

445
00:23:38,880 --> 00:23:42,119
Speaker 3: I guess one thing that we did also notice, like

446
00:23:42,400 --> 00:23:46,240
Speaker 3: to your point on that little contradiction, is that hyperscalers

447
00:23:46,240 --> 00:23:50,359
Speaker 3: do take a dual strategy. They're both building in existing

448
00:23:50,520 --> 00:23:55,080
Speaker 3: regions and also trialing in new locations. If you look

449
00:23:55,240 --> 00:23:59,000
Speaker 3: in our report, you're seeing that the large four companies

450
00:23:59,160 --> 00:24:02,639
Speaker 3: they're building with in their own data center clusters and

451
00:24:02,720 --> 00:24:05,960
Speaker 3: also like looking for new markets at the same time.

452
00:24:06,200 --> 00:24:09,360
Speaker 3: And they're doing that because for them, speed and scale

453
00:24:09,520 --> 00:24:14,879
Speaker 3: of development is really critical, particularly because like their AI

454
00:24:15,040 --> 00:24:18,080
Speaker 3: business requires them to kind of take a like a.

455
00:24:18,040 --> 00:24:20,800
Speaker 1: Winner is that it's like a there's a computing power

456
00:24:20,920 --> 00:24:24,440
Speaker 1: arms race happening, yeah now, and so it's all very

457
00:24:24,480 --> 00:24:26,760
Speaker 1: well saying, oh, we'd ideally build it here, but it's

458
00:24:26,760 --> 00:24:28,439
Speaker 1: like we just need this right now, and we're going

459
00:24:28,520 --> 00:24:29,080
Speaker 1: to do what's tried.

460
00:24:29,480 --> 00:24:29,960
Speaker 4: Yeah.

461
00:24:30,000 --> 00:24:32,600
Speaker 3: Like, I guess it's like a winner takes all game

462
00:24:32,880 --> 00:24:36,320
Speaker 3: in the AI business. So they want to build data

463
00:24:36,320 --> 00:24:39,520
Speaker 3: centers as quickly as possible, So they're taking all options.

464
00:24:39,680 --> 00:24:42,439
Speaker 1: They want their their particular AI model to be the

465
00:24:42,440 --> 00:24:45,080
Speaker 1: Coca Cola of AI. So what does all this mean

466
00:24:45,119 --> 00:24:47,480
Speaker 1: for the power sector? I mean, how is this going

467
00:24:47,520 --> 00:24:50,000
Speaker 1: to affect the supply mix just keeping up with all

468
00:24:50,000 --> 00:24:53,000
Speaker 1: this demand? How's it going to affect the regulatory model?

469
00:24:53,720 --> 00:24:56,480
Speaker 1: You know, it's not designed for just suddenly having loads

470
00:24:56,520 --> 00:24:59,720
Speaker 1: of new demand dropped on a in very concentrated regions,

471
00:24:59,760 --> 00:25:02,040
Speaker 1: as we I've just learned. And then you know what

472
00:25:02,040 --> 00:25:04,520
Speaker 1: innovations are we seeing to try and cope with all

473
00:25:04,520 --> 00:25:06,359
Speaker 1: of this new demand in the power sector.

474
00:25:06,880 --> 00:25:10,760
Speaker 3: Yeah, So what we're currently seeing is a very strong

475
00:25:10,840 --> 00:25:15,199
Speaker 3: reaction towards all of this new data center demand, particularly

476
00:25:15,240 --> 00:25:19,199
Speaker 3: from utilities. So in Ohio, where I guess there's a

477
00:25:19,200 --> 00:25:22,760
Speaker 3: lot of new investment in data centers, what we're seeing

478
00:25:22,880 --> 00:25:25,480
Speaker 3: is ap Ohio, which is the utility in the region.

479
00:25:25,720 --> 00:25:29,560
Speaker 3: They've proposed a new tariff which in that tariff they've

480
00:25:29,680 --> 00:25:32,439
Speaker 3: acquired or requested that new data centers pay up to

481
00:25:32,520 --> 00:25:37,000
Speaker 3: eighty five percent of their projected energy usage, which makes

482
00:25:37,040 --> 00:25:40,240
Speaker 3: it less attractive of a market for these data centers

483
00:25:40,280 --> 00:25:43,160
Speaker 3: to want to build in that region. In a way,

484
00:25:43,240 --> 00:25:47,520
Speaker 3: it's to prevent increasing retail rates from a utilities perspective,

485
00:25:47,680 --> 00:25:50,399
Speaker 3: but there's a lot of pushback for installment.

486
00:25:50,760 --> 00:25:54,000
Speaker 1: Yeah, that's I mean, because that's controversial because the entire

487
00:25:54,040 --> 00:25:58,560
Speaker 1: basis of the regulated monopoly is providing equal access to

488
00:25:58,680 --> 00:26:02,480
Speaker 1: all consumers, even if those consumers that are being discriminated

489
00:26:02,520 --> 00:26:07,040
Speaker 1: against our massive corporations, still does undermine the philosophical basis.

490
00:26:07,119 --> 00:26:08,920
Speaker 1: So I'm kind of interested to see what's going to

491
00:26:08,960 --> 00:26:11,240
Speaker 1: happen there. What do you think's going to mean for

492
00:26:11,720 --> 00:26:15,560
Speaker 1: the supply mix, you know, wind, solar, gas, what's going

493
00:26:15,600 --> 00:26:15,919
Speaker 1: to do it?

494
00:26:16,280 --> 00:26:19,520
Speaker 2: So I think one of the emerging innovations in moving

495
00:26:19,560 --> 00:26:22,439
Speaker 2: from you know, this hyperload growth environment and there's not

496
00:26:22,560 --> 00:26:26,720
Speaker 2: much supply available is this trend of colocation or having

497
00:26:26,800 --> 00:26:31,240
Speaker 2: on site generation. A lot of traditional grid planning was

498
00:26:31,320 --> 00:26:33,919
Speaker 2: for that peak power, and I think we saw this

499
00:26:34,160 --> 00:26:36,640
Speaker 2: a couple of years ago, and you know a lot

500
00:26:36,640 --> 00:26:40,040
Speaker 2: of taxes and how they integrated crypto mining is having

501
00:26:40,040 --> 00:26:43,920
Speaker 2: these as flexible loads and curtailing during hours of peak demand.

502
00:26:44,320 --> 00:26:48,520
Speaker 2: We know that non crypto workloads like AI training or

503
00:26:48,560 --> 00:26:53,080
Speaker 2: inference may not necessarily curtail, but we do see one

504
00:26:53,440 --> 00:26:56,720
Speaker 2: model in which they'll have some sort of on site

505
00:26:56,800 --> 00:27:01,880
Speaker 2: generation or colocation supply where they could as a whole

506
00:27:01,880 --> 00:27:06,040
Speaker 2: campus interact with grid needs in terms of the total

507
00:27:06,080 --> 00:27:09,960
Speaker 2: supply mix. A lot of the short term needs means

508
00:27:10,040 --> 00:27:13,479
Speaker 2: that they'll build whatever technology is fastest, and in our

509
00:27:13,600 --> 00:27:17,560
Speaker 2: data most of that is when solar batteries, but reliability

510
00:27:17,640 --> 00:27:20,120
Speaker 2: is also a huge part of data centers. They're known

511
00:27:20,160 --> 00:27:23,400
Speaker 2: for having five nines, which is ninety nine point nine

512
00:27:23,520 --> 00:27:27,840
Speaker 2: nine nine percent uptime, which means that you know, firm

513
00:27:27,880 --> 00:27:31,080
Speaker 2: capacity like natural gas generation or diesel gent chats that

514
00:27:31,080 --> 00:27:34,919
Speaker 2: they've used traditionally will also be a large part of

515
00:27:34,960 --> 00:27:36,680
Speaker 2: the solution in the short term.

516
00:27:36,800 --> 00:27:38,880
Speaker 1: So it's a real bit of an open question there.

517
00:27:39,040 --> 00:27:41,760
Speaker 1: It's either how quickly things can get built and what

518
00:27:41,880 --> 00:27:45,520
Speaker 1: truly is the fastest solution versus the long term needs. Helen,

519
00:27:45,520 --> 00:27:47,080
Speaker 1: thank you very much for joining us today.

520
00:27:47,320 --> 00:27:49,760
Speaker 4: Thank you, Ton and Natalie, thank you so much for

521
00:27:49,840 --> 00:27:59,639
Speaker 4: joining Thanks Sam.

522
00:28:00,040 --> 00:28:02,960
Speaker 1: Day's episode of Switched On was produced by Cam Gray

523
00:28:03,160 --> 00:28:05,520
Speaker 1: with production assistance from Kamala Shelling.

524
00:28:05,680 --> 00:28:08,840
Speaker 4: Bloomberg NIF is a service provided by Bloomberg Finance LP

525
00:28:09,000 --> 00:28:09,879
Speaker 4: and its affiliates.

526
00:28:09,920 --> 00:28:12,640
Speaker 1: This recording does not constitute, nor should it be construed

527
00:28:12,640 --> 00:28:16,440
Speaker 1: as investment in vice investment recommendations, or a recommendation as

528
00:28:16,480 --> 00:28:19,320
Speaker 1: to an investment or other strategy. Bloomberg ANIF should not

529
00:28:19,359 --> 00:28:22,399
Speaker 1: be considered as information sufficient upon which to base an

530
00:28:22,440 --> 00:28:25,600
Speaker 1: investment decision. Neither Bloomberg Finance LP nor any of its

531
00:28:25,600 --> 00:28:29,320
Speaker 1: affiliates makes any representation or warranty as to the accuracy

532
00:28:29,400 --> 00:28:32,199
Speaker 1: or completeness of the information contained in this recording, and

533
00:28:32,320 --> 00:28:32,560
Speaker 1: any

534
00:28:32,600 --> 00:28:35,879
Speaker 2: Liability as a result of this recording is expressly disclaimed