1
00:00:02,720 --> 00:00:16,439
Speaker 1: Bloomberg Audio Studios, Podcasts, Radio News.

2
00:00:18,520 --> 00:00:22,400
Speaker 2: Hello and welcome to another episode of the Odd Lots podcast.

3
00:00:22,480 --> 00:00:24,800
Speaker 3: I'm Joe Wisenthal and I'm Tracy Alloway.

4
00:00:25,120 --> 00:00:28,320
Speaker 2: Tracy covering the AI boom is actually reminding me a

5
00:00:28,360 --> 00:00:31,240
Speaker 2: little bit of the tariff boom in April, simply because

6
00:00:31,280 --> 00:00:33,880
Speaker 2: every day they are new headlines, like they're just today

7
00:00:34,040 --> 00:00:38,519
Speaker 2: we're recording this November twelfth, Anthropic commits fifty billion dollars

8
00:00:38,520 --> 00:00:41,160
Speaker 2: to build AI data centers in the US. So the

9
00:00:41,240 --> 00:00:44,479
Speaker 2: advanced model companies are vertically integrating more to build their

10
00:00:44,479 --> 00:00:46,879
Speaker 2: own data centers. Every day some new development.

11
00:00:47,080 --> 00:00:49,760
Speaker 3: Yeah, it's becoming pretty hard to keep up. So I

12
00:00:49,760 --> 00:00:52,400
Speaker 3: think we're probably just going to talk in terms of

13
00:00:52,440 --> 00:00:54,520
Speaker 3: billions and trillions. We're just going to say lots and

14
00:00:54,560 --> 00:00:57,280
Speaker 3: lots of money is going into the space. But the

15
00:00:57,320 --> 00:00:59,840
Speaker 3: way I've been thinking about it is, Okay, at this

16
00:01:00,240 --> 00:01:04,640
Speaker 3: everyone agrees that the AI buildout is super expensive, and

17
00:01:04,760 --> 00:01:08,240
Speaker 3: all these companies are spending massive amounts of capex to

18
00:01:08,360 --> 00:01:12,200
Speaker 3: do this, and I'm starting to think that AI capex

19
00:01:12,319 --> 00:01:16,399
Speaker 3: is kind of like the Schrodinger's Cat of markets in

20
00:01:16,440 --> 00:01:20,400
Speaker 3: the sense that it could either be a massive strength

21
00:01:20,600 --> 00:01:23,920
Speaker 3: for these companies because the capex is so expensive and

22
00:01:23,959 --> 00:01:26,200
Speaker 3: it takes so much money to build out, and so

23
00:01:26,319 --> 00:01:28,320
Speaker 3: anyone who manages to do it kind of builds a

24
00:01:28,360 --> 00:01:32,400
Speaker 3: moat around their business. Or it could be a massive weakness,

25
00:01:32,480 --> 00:01:34,720
Speaker 3: right if you're spending all this money and then that

26
00:01:34,760 --> 00:01:37,640
Speaker 3: doesn't end up generating the revenues that you actually need

27
00:01:37,680 --> 00:01:41,640
Speaker 3: to justify it. And going back to the Schrodinger's analogy,

28
00:01:42,560 --> 00:01:45,000
Speaker 3: it seems like we just don't know what's going to

29
00:01:45,080 --> 00:01:47,720
Speaker 3: come out of the box, right, Like it's simultaneously a

30
00:01:47,760 --> 00:01:51,240
Speaker 3: strength and a weakness, and until we build out AGI

31
00:01:51,400 --> 00:01:53,000
Speaker 3: or whatever, like, we're just not going to know.

32
00:01:53,720 --> 00:01:56,080
Speaker 2: I told her, right, there's so much at stake here,

33
00:01:56,280 --> 00:01:59,520
Speaker 2: and obviously we know the numbers are absolutely enormous. They're staggering,

34
00:01:59,600 --> 00:02:02,720
Speaker 2: and we could talk about them too. The financing structures

35
00:02:02,760 --> 00:02:04,240
Speaker 2: are also very interesting.

36
00:02:04,520 --> 00:02:04,680
Speaker 4: You know.

37
00:02:04,760 --> 00:02:09,320
Speaker 2: It's one thing if you just have Meta or Alphabet

38
00:02:09,440 --> 00:02:11,440
Speaker 2: and they make a ton of money already and they're

39
00:02:11,480 --> 00:02:14,160
Speaker 2: spending money on data centers whatever. That's one thing. It's

40
00:02:14,200 --> 00:02:18,320
Speaker 2: another thing when you start seeing these SPVs where the

41
00:02:18,440 --> 00:02:20,560
Speaker 2: hyperscaler puts in this amount of money and then the

42
00:02:20,600 --> 00:02:22,800
Speaker 2: private credit puts in this equity and then they borrow

43
00:02:22,880 --> 00:02:25,840
Speaker 2: a bunch and then there's all these questions about the payback.

44
00:02:25,880 --> 00:02:28,440
Speaker 2: And we think of tech as from years and years

45
00:02:28,480 --> 00:02:31,280
Speaker 2: as basically being this equity story, and when it becomes

46
00:02:31,320 --> 00:02:33,960
Speaker 2: a credit story. Yeah, and when you know people are

47
00:02:33,960 --> 00:02:37,440
Speaker 2: talking about quoting Oracle CDs, I always forget these companies

48
00:02:37,480 --> 00:02:40,520
Speaker 2: even have CDs because I'm so unused to thinking of

49
00:02:40,520 --> 00:02:43,760
Speaker 2: big tech companies as credits. So when I see people

50
00:02:43,800 --> 00:02:47,440
Speaker 2: starting to tweet Oracle CDs charts or core Weave CDs charts,

51
00:02:47,440 --> 00:02:50,040
Speaker 2: It's like, Okay, we are in a different level of

52
00:02:50,120 --> 00:02:51,560
Speaker 2: capital intensity.

53
00:02:51,160 --> 00:02:53,840
Speaker 3: Right, and some of those swaps have been going up lately.

54
00:02:54,160 --> 00:02:56,960
Speaker 3: I'm going to say one more thing, thinking back to

55
00:02:57,000 --> 00:02:59,880
Speaker 3: the two thousand and eight financial crisis. I remember the

56
00:03:00,040 --> 00:03:03,320
Speaker 3: economist at Raymond James I it is Jeff's out who

57
00:03:03,400 --> 00:03:06,440
Speaker 3: went on to become a very big name. Yeah, we

58
00:03:06,480 --> 00:03:08,600
Speaker 3: should have him on the podcast. But he made the

59
00:03:08,639 --> 00:03:13,200
Speaker 3: point that historically when you had real estate crashes property crashes,

60
00:03:13,600 --> 00:03:16,960
Speaker 3: it was usually because of a problem in the economy.

61
00:03:17,160 --> 00:03:19,080
Speaker 3: But then what happened in the run up to two

62
00:03:19,120 --> 00:03:21,720
Speaker 3: thousand and seven two thousand and eight is the housing

63
00:03:21,760 --> 00:03:25,960
Speaker 3: market crash became the proximate cause of the troubles in

64
00:03:26,000 --> 00:03:28,520
Speaker 3: the economy. And if you think about how much money

65
00:03:28,600 --> 00:03:33,000
Speaker 3: is being spent on AI right now again billions, trillions

66
00:03:33,040 --> 00:03:37,800
Speaker 3: possibly of dollars, it's very easy to see how AI

67
00:03:37,880 --> 00:03:40,760
Speaker 3: could borph into a problem for the wider real economy.

68
00:03:40,800 --> 00:03:43,280
Speaker 2: Totally just on this note, and then we'll get into

69
00:03:43,360 --> 00:03:46,160
Speaker 2: our conversation. The Center for Public Enterprises out with a

70
00:03:46,200 --> 00:03:49,600
Speaker 2: great report today called Bubble or Nothing by Ed vat Aarun,

71
00:03:50,000 --> 00:03:52,440
Speaker 2: pointing out one of the things that makes data centers

72
00:03:52,480 --> 00:03:55,480
Speaker 2: interesting is how they sit at this intersection of essentially

73
00:03:55,880 --> 00:04:00,119
Speaker 2: industrial spending and real estate. It's an interesting ascid class

74
00:04:00,360 --> 00:04:02,280
Speaker 2: for its own right. So much to talk about. We

75
00:04:02,320 --> 00:04:04,600
Speaker 2: could never do a justice in one episode, but that

76
00:04:04,680 --> 00:04:06,640
Speaker 2: means we got to do more. Anyway. I'm very excited

77
00:04:06,640 --> 00:04:08,680
Speaker 2: for today's episode. We really do have the perfect guest.

78
00:04:08,920 --> 00:04:11,480
Speaker 2: Someone who's been writing about this for a long time,

79
00:04:11,640 --> 00:04:14,000
Speaker 2: someone who's just been writing about the Internet and all

80
00:04:14,040 --> 00:04:16,600
Speaker 2: things for longer than any of us, someone who's been

81
00:04:16,640 --> 00:04:20,160
Speaker 2: blogging and investing for far longer than either of us

82
00:04:20,240 --> 00:04:22,560
Speaker 2: or anything like that. Way more knowledgeable about how these

83
00:04:22,600 --> 00:04:25,840
Speaker 2: businesses worked, and most very focused on the data center

84
00:04:25,880 --> 00:04:28,200
Speaker 2: buildout we're going to be speaking with Paul Kadrowski. He

85
00:04:28,279 --> 00:04:31,240
Speaker 2: is a fellow at the MIT Institute for the Digital Economy,

86
00:04:31,480 --> 00:04:38,159
Speaker 2: also a partner at sk Ventures, and longtime internet blogger, writer, newsletter, yapper, etc.

87
00:04:39,000 --> 00:04:42,360
Speaker 2: Someone we've never never had on the podcast before. So Paul,

88
00:04:42,440 --> 00:04:44,320
Speaker 2: thank you so much for joining us.

89
00:04:44,440 --> 00:04:46,880
Speaker 5: Hey, guys, thanks good to be here. Other than the

90
00:04:46,920 --> 00:04:47,599
Speaker 5: blogging part, but.

91
00:04:47,800 --> 00:04:50,839
Speaker 2: No, it's all. It's all. You're a true pioneer in

92
00:04:50,920 --> 00:04:53,920
Speaker 2: that and it's impressive that you still write with the

93
00:04:53,960 --> 00:04:57,320
Speaker 2: output that you do. At some point in the last year,

94
00:04:58,000 --> 00:05:00,240
Speaker 2: I feel like you really got laser focused, maybe in

95
00:05:00,240 --> 00:05:02,880
Speaker 2: the last two years, really got laser focused on the

96
00:05:02,960 --> 00:05:05,560
Speaker 2: data center's story is this is where the action is.

97
00:05:05,880 --> 00:05:07,960
Speaker 6: Yeah, I did, and in part just because I caught

98
00:05:08,000 --> 00:05:09,719
Speaker 6: myself by surprise with it.

99
00:05:09,720 --> 00:05:10,200
Speaker 5: It was weird.

100
00:05:10,240 --> 00:05:12,160
Speaker 6: I was looking at first half GDP day it actually

101
00:05:12,200 --> 00:05:15,719
Speaker 6: first quarter GDP data earlier in the year, and you know,

102
00:05:15,760 --> 00:05:17,760
Speaker 6: this has become a commonplace that people know this, but

103
00:05:17,800 --> 00:05:20,480
Speaker 6: I hadn't realized what a large fraction of GDP growth

104
00:05:20,640 --> 00:05:22,720
Speaker 6: in the first quarter data centers were was on the

105
00:05:22,800 --> 00:05:25,680
Speaker 6: order of fifty percent, much larger if you included all

106
00:05:25,680 --> 00:05:27,760
Speaker 6: sort of externalities all the other things that data center

107
00:05:27,800 --> 00:05:31,240
Speaker 6: spending in turn kind of accelerates. And then obviously the

108
00:05:31,240 --> 00:05:33,080
Speaker 6: same thing was true in the second quarter, and it

109
00:05:33,120 --> 00:05:35,360
Speaker 6: was I got back to thinking about my dog, and

110
00:05:35,400 --> 00:05:36,960
Speaker 6: I was my analogy is that.

111
00:05:36,800 --> 00:05:38,680
Speaker 3: As one does, as one does.

112
00:05:38,960 --> 00:05:41,359
Speaker 6: I got to get like my dog barks when the

113
00:05:41,400 --> 00:05:44,200
Speaker 6: mailman comes to the house and keeps barking, and then

114
00:05:44,240 --> 00:05:47,200
Speaker 6: the mailman goes away. And I'm convinced he thinks he

115
00:05:47,279 --> 00:05:49,600
Speaker 6: makes the mailman go away, right, he has this really

116
00:05:49,640 --> 00:05:52,800
Speaker 6: screw causality, and it's like, dude, if you don't bark,

117
00:05:52,839 --> 00:05:54,720
Speaker 6: it goes away. Anyway, this is part of the job.

118
00:05:54,760 --> 00:05:58,280
Speaker 6: They just go away. And I think about macro policy

119
00:05:58,320 --> 00:06:00,160
Speaker 6: in the same way that if you don't understand and

120
00:06:00,279 --> 00:06:02,960
Speaker 6: the drivers of GDP growth, you're likely to think to

121
00:06:03,000 --> 00:06:05,200
Speaker 6: whatever it is you would most like to be causing

122
00:06:05,240 --> 00:06:07,200
Speaker 6: GDP growth is doing that. So in the case of

123
00:06:07,200 --> 00:06:09,520
Speaker 6: the US in the first half of the year, you know,

124
00:06:09,680 --> 00:06:11,920
Speaker 6: was this puzzle was, well, maybe it's terroifts, maybe tariffs

125
00:06:11,960 --> 00:06:14,120
Speaker 6: are actually contributing to it, maybe consumers are much.

126
00:06:14,040 --> 00:06:15,400
Speaker 5: More resilient than we expected.

127
00:06:15,880 --> 00:06:18,080
Speaker 6: And as it turns out, a huge factor, probably the

128
00:06:18,160 --> 00:06:21,920
Speaker 6: largest factor, was this sort of unintentional private sector stimulus

129
00:06:21,960 --> 00:06:26,000
Speaker 6: program otherwise known as data centers, and for me that

130
00:06:26,080 --> 00:06:28,840
Speaker 6: I'll start it. So that started this puzzle of understanding

131
00:06:29,400 --> 00:06:33,080
Speaker 6: this sort of disconmisserate size, the consequences of that size,

132
00:06:33,440 --> 00:06:36,880
Speaker 6: and the acceleration's consequences in terms of where where the

133
00:06:36,920 --> 00:06:37,960
Speaker 6: money is coming from, and all.

134
00:06:37,920 --> 00:06:38,719
Speaker 5: Sorts of other things.

135
00:06:38,760 --> 00:06:41,760
Speaker 6: But just to reframe in terms of something you guys

136
00:06:41,800 --> 00:06:44,320
Speaker 6: were already talking about, and this I think is super important,

137
00:06:44,320 --> 00:06:47,960
Speaker 6: and understanding why this particular episode is likely to turn

138
00:06:47,960 --> 00:06:49,640
Speaker 6: out to be historically really important.

139
00:06:50,200 --> 00:06:53,679
Speaker 2: Wait, when you say you're referred to this podcast episode,

140
00:06:53,720 --> 00:06:57,040
Speaker 2: you're not referring to the broader episode of AI data Center.

141
00:06:57,160 --> 00:06:58,960
Speaker 5: Entirely, just the podcast.

142
00:07:01,080 --> 00:07:03,480
Speaker 6: Who Cares about data centers at the ten year anniversary

143
00:07:03,480 --> 00:07:06,680
Speaker 6: of bad Law. So the reason why sort of it's

144
00:07:06,680 --> 00:07:08,800
Speaker 6: going to be historically important is because, for the first time,

145
00:07:08,839 --> 00:07:12,160
Speaker 6: we combine all the major ingredients of every historical bubbles

146
00:07:12,160 --> 00:07:14,280
Speaker 6: in a single bubble. We have a metabubble no pun

147
00:07:14,320 --> 00:07:17,400
Speaker 6: intended for meta. We have real estate. You guys just

148
00:07:17,400 --> 00:07:19,640
Speaker 6: talked about this, right, Some of the largest bubbles in

149
00:07:19,720 --> 00:07:22,640
Speaker 6: US history had some relationship to real estate. We have

150
00:07:22,680 --> 00:07:25,880
Speaker 6: a great technology story. Almost all the large modern bubbles

151
00:07:25,880 --> 00:07:27,200
Speaker 6: has something to do with technology.

152
00:07:27,200 --> 00:07:28,120
Speaker 5: We have loose credit.

153
00:07:28,480 --> 00:07:30,880
Speaker 6: Most of the major bubbles in some sense have a

154
00:07:30,920 --> 00:07:34,480
Speaker 6: loose credit aspect. And one of the other exacerbating pieces

155
00:07:34,520 --> 00:07:36,800
Speaker 6: that some of the largest bubbles, thinking about even the

156
00:07:36,800 --> 00:07:40,240
Speaker 6: financial crisis, is some kind of notional government backstop. You know,

157
00:07:40,280 --> 00:07:43,520
Speaker 6: think about the role in terms of broadening home ownership

158
00:07:43,520 --> 00:07:45,320
Speaker 6: in the context of the real estate bubble, and the

159
00:07:45,400 --> 00:07:48,160
Speaker 6: role that Fanny and Freddie played and loosening credit standards

160
00:07:48,160 --> 00:07:50,560
Speaker 6: and all of those things. This is the first bubble

161
00:07:50,600 --> 00:07:52,760
Speaker 6: that has all of that. It's like, we said, you

162
00:07:52,800 --> 00:07:56,040
Speaker 6: know what would be great, Let's create a bubble that

163
00:07:56,160 --> 00:07:58,040
Speaker 6: takes everything that ever worked and put it.

164
00:07:58,000 --> 00:07:59,720
Speaker 5: All in one. And this is what we've done.

165
00:08:00,120 --> 00:08:02,480
Speaker 6: Got a speculative real estate component is probably one of

166
00:08:02,520 --> 00:08:07,440
Speaker 6: the strongest technology stories we ever had back to rural electrification.

167
00:08:07,560 --> 00:08:09,840
Speaker 6: In terms of a technology story, we have loose credit.

168
00:08:09,880 --> 00:08:12,800
Speaker 6: You guys talked about what's happening with respect to not

169
00:08:12,880 --> 00:08:14,840
Speaker 6: just the role of private credit, but how private credit

170
00:08:14,880 --> 00:08:17,840
Speaker 6: is largely supplanted commercial banks with respect to being lenders here.

171
00:08:18,000 --> 00:08:19,560
Speaker 6: So we have all of these pieces that have all

172
00:08:19,560 --> 00:08:21,200
Speaker 6: come together at once, and I think in terms of

173
00:08:21,240 --> 00:08:24,360
Speaker 6: framing what's going on right now. It's really important to

174
00:08:24,480 --> 00:08:28,040
Speaker 6: understand that it brings together all of these components and

175
00:08:28,080 --> 00:08:29,720
Speaker 6: ways we've never seen before, which is one of the

176
00:08:29,720 --> 00:08:31,720
Speaker 6: reasons why the notion that we can land this thing

177
00:08:31,840 --> 00:08:33,560
Speaker 6: on the runway gently is nonsense.

178
00:08:34,360 --> 00:08:37,160
Speaker 3: I love that framing the metal babble is perfect. Also,

179
00:08:37,520 --> 00:08:40,120
Speaker 3: I had an epiphany earlier. I already told Joe, so

180
00:08:40,160 --> 00:08:43,040
Speaker 3: you can attest to this, but I realized private credit

181
00:08:43,640 --> 00:08:47,000
Speaker 3: kind of supplanted shadow banking as the term. Right like

182
00:08:47,240 --> 00:08:49,800
Speaker 3: after two thousand and eight, we called it shadow banking,

183
00:08:50,240 --> 00:08:52,480
Speaker 3: and then at some point it flipped to I guess

184
00:08:52,559 --> 00:08:54,480
Speaker 3: the couplier private credit.

185
00:08:54,559 --> 00:08:57,120
Speaker 2: Shadow bank always owned it sinister right away that private

186
00:08:57,120 --> 00:08:57,560
Speaker 2: credit is.

187
00:08:57,559 --> 00:08:59,560
Speaker 3: Well, someone figured that out and they're like, well, now

188
00:08:59,559 --> 00:09:00,439
Speaker 3: it's private credit.

189
00:09:00,520 --> 00:09:01,640
Speaker 5: I like to think of it as a kind of

190
00:09:01,679 --> 00:09:05,880
Speaker 5: financial witness protection program. It was like, oh, you're those guys.

191
00:09:06,200 --> 00:09:07,480
Speaker 5: That's great, now who you are?

192
00:09:07,920 --> 00:09:10,120
Speaker 6: Yeah, it's kind of like that, And it's now like

193
00:09:10,200 --> 00:09:13,000
Speaker 6: one point whatever. It is one point seven trillion dollars

194
00:09:13,080 --> 00:09:16,560
Speaker 6: is the size of which is larger than many components

195
00:09:16,559 --> 00:09:19,040
Speaker 6: of the orthodox lending market combined. In terms of the

196
00:09:19,080 --> 00:09:21,839
Speaker 6: private credit industry itself, so that's a huge new piece

197
00:09:21,840 --> 00:09:24,520
Speaker 6: of this that sometimes escapes notice how big it is

198
00:09:24,520 --> 00:09:26,520
Speaker 6: and why it emerged, So all of those pieces.

199
00:09:26,640 --> 00:09:28,880
Speaker 3: Yeah, it's stunning the growth that we've seen. Let me

200
00:09:28,920 --> 00:09:31,840
Speaker 3: ask a very basic question before we go further. But

201
00:09:32,160 --> 00:09:35,719
Speaker 3: one thing I've been wondering is Joe mentioned that anthropic

202
00:09:35,760 --> 00:09:39,160
Speaker 3: headline that we heard before. We've seen Meta raising financing

203
00:09:39,200 --> 00:09:42,720
Speaker 3: for data center builds, all that stuff. Why do these

204
00:09:42,880 --> 00:09:47,679
Speaker 3: massively profitable and cash rich companies have to raise financing

205
00:09:48,120 --> 00:09:49,040
Speaker 3: at all?

206
00:09:49,240 --> 00:09:51,880
Speaker 6: Well, they don't, but there's these irritating shareholders out there

207
00:09:52,480 --> 00:09:55,560
Speaker 6: get all pissy whenever you start diluting earnings pre shared

208
00:09:55,559 --> 00:09:57,840
Speaker 6: too much and diverting it towards a single source. Now

209
00:09:57,840 --> 00:10:00,280
Speaker 6: that's not the case with private companies obviously, but by

210
00:10:00,320 --> 00:10:03,000
Speaker 6: the same token, open ai doesn't have the luxury of

211
00:10:03,000 --> 00:10:05,600
Speaker 6: having cash flows via which they can do any of

212
00:10:05,640 --> 00:10:09,200
Speaker 6: the things we're describing, so anthropic open Ai and everyone

213
00:10:09,200 --> 00:10:11,280
Speaker 6: else they have no option other than to do exactly

214
00:10:11,280 --> 00:10:13,560
Speaker 6: what we're describing. It's a different story with respect to

215
00:10:14,360 --> 00:10:16,840
Speaker 6: how what percentage of Google's free cash flow or Amazon

216
00:10:16,880 --> 00:10:19,240
Speaker 6: free cash flow that they want to continue to divert

217
00:10:19,440 --> 00:10:22,040
Speaker 6: towards data centers. So in terms of the privates, this

218
00:10:22,080 --> 00:10:25,000
Speaker 6: is the only option that they have. The public's obviously

219
00:10:25,000 --> 00:10:27,959
Speaker 6: increasing the hyperscalers increasingly. We've got up to the point

220
00:10:27,960 --> 00:10:30,520
Speaker 6: where around five hundred billion dollars or fifty percent of

221
00:10:30,520 --> 00:10:33,600
Speaker 6: their free cash flow is going directly towards spending on

222
00:10:33,760 --> 00:10:36,400
Speaker 6: data centers, and that's obviously a point at which you know,

223
00:10:36,480 --> 00:10:38,120
Speaker 6: we have other things we have to do with free

224
00:10:38,200 --> 00:10:41,440
Speaker 6: cash flow, and including having some of it be earnings

225
00:10:41,440 --> 00:10:44,160
Speaker 6: per share, and so we increasingly it's become the option.

226
00:10:44,320 --> 00:10:46,240
Speaker 5: You see what METT is doing recently with respect it

227
00:10:46,360 --> 00:10:47,119
Speaker 5: is SPVs.

228
00:10:47,400 --> 00:10:50,760
Speaker 6: We bring in other participants, create new financing vehicles, and

229
00:10:50,800 --> 00:10:52,920
Speaker 6: then we play this entertaining game of it's not really

230
00:10:53,000 --> 00:10:53,440
Speaker 6: our debt.

231
00:10:53,480 --> 00:10:55,320
Speaker 5: It's in an SPV. I don't have to roll it

232
00:10:55,360 --> 00:10:56,280
Speaker 5: back onto my own.

233
00:10:56,160 --> 00:10:58,839
Speaker 6: Balance sheet and then bring in new lenders, new private

234
00:10:58,840 --> 00:10:59,760
Speaker 6: credit firms and others.

235
00:11:00,000 --> 00:11:02,200
Speaker 5: So that's the reason. Obviously it's partly because of the scale.

236
00:11:02,240 --> 00:11:04,400
Speaker 6: It's probably because the privates who have no other option,

237
00:11:04,760 --> 00:11:07,400
Speaker 6: and it's probably we've kind of tapped out the public

238
00:11:07,440 --> 00:11:09,520
Speaker 6: companies in terms of the fraction of free cash flow

239
00:11:10,080 --> 00:11:10,920
Speaker 6: that they.

240
00:11:10,760 --> 00:11:14,920
Speaker 5: Feel as if they can spend with impunity on these projects.

241
00:11:14,640 --> 00:11:16,959
Speaker 2: Explain to us for those who don't know. You know, again,

242
00:11:17,160 --> 00:11:19,880
Speaker 2: SPV one of these terms that we really haven't heard

243
00:11:19,960 --> 00:11:22,400
Speaker 2: in a while. And there's nothing inherently bad about an

244
00:11:22,480 --> 00:11:25,720
Speaker 2: SPV except that you only hear about them typically after

245
00:11:25,840 --> 00:11:27,800
Speaker 2: there's something, you know, some sort of crazy.

246
00:11:27,679 --> 00:11:29,640
Speaker 5: Ride, which is weird obviously, But yes, tell.

247
00:11:29,559 --> 00:11:31,679
Speaker 2: How would you U say in the broad strokes, how

248
00:11:31,720 --> 00:11:34,800
Speaker 2: would you characterize what these financing vehicles are?

249
00:11:35,280 --> 00:11:37,959
Speaker 6: So Mechanically, it's just a way of making sure that

250
00:11:38,000 --> 00:11:39,720
Speaker 6: I don't have to roll data onto my balance sheet.

251
00:11:39,760 --> 00:11:42,960
Speaker 6: But legally it's a structure into which I and my

252
00:11:43,040 --> 00:11:46,439
Speaker 6: partners contribute capital that in exchange for which they retain

253
00:11:46,559 --> 00:11:49,600
Speaker 6: legal title to the project that we've created, which allows

254
00:11:49,679 --> 00:11:52,280
Speaker 6: us to all contribute capitalists but not have to put

255
00:11:52,280 --> 00:11:54,160
Speaker 6: it back on my balance sheet and therefore not to

256
00:11:54,200 --> 00:11:55,359
Speaker 6: have that debt rated.

257
00:11:55,640 --> 00:11:56,920
Speaker 5: Which is really the key.

258
00:11:57,040 --> 00:11:59,520
Speaker 6: Now, if you look at the actual intrinstics, say, for example,

259
00:11:59,559 --> 00:12:01,640
Speaker 6: the reason that a project that they did in conjunction

260
00:12:01,720 --> 00:12:04,720
Speaker 6: with blue Out, it's wild and byzantine. It looks like

261
00:12:04,760 --> 00:12:06,400
Speaker 6: something you might have seen and what was that in

262
00:12:06,440 --> 00:12:08,240
Speaker 6: Harry Potter or the forest with all the spider webs.

263
00:12:08,240 --> 00:12:09,720
Speaker 5: It looks a little like that, right where.

264
00:12:09,559 --> 00:12:11,640
Speaker 6: Everything's connected to everything and all I know is something

265
00:12:11,640 --> 00:12:14,160
Speaker 6: and here's going to get me. So there's incredible complexity,

266
00:12:14,200 --> 00:12:16,680
Speaker 6: but at the core, it's a mechanism via which I

267
00:12:16,679 --> 00:12:18,720
Speaker 6: can raise more capital and keep it off my balance

268
00:12:18,800 --> 00:12:21,280
Speaker 6: sheet by creating a legal entity that controls the actual

269
00:12:21,360 --> 00:12:23,559
Speaker 6: data center and I don't. Therefore I have to put

270
00:12:23,559 --> 00:12:26,400
Speaker 6: it back, roll it all back onto my balance sheet, navierated.

271
00:12:26,920 --> 00:12:31,320
Speaker 6: Now there's weird intricacies obviously, So for example, what happens

272
00:12:31,440 --> 00:12:34,400
Speaker 6: if at some period in the future this thing isn't

273
00:12:34,440 --> 00:12:37,120
Speaker 6: performing the way we expect who owns it at that point?

274
00:12:37,320 --> 00:12:38,960
Speaker 5: Is there a payment exchange, does.

275
00:12:38,840 --> 00:12:41,200
Speaker 6: It become metas, does it become blue ouls, does it

276
00:12:41,200 --> 00:12:45,000
Speaker 6: become someone else? And these things will turn out to matter.

277
00:12:45,080 --> 00:12:46,840
Speaker 6: Right now, no one cares. If you go through some

278
00:12:46,880 --> 00:12:49,040
Speaker 6: of the documents on these things, it's not entirely clear

279
00:12:49,480 --> 00:12:51,600
Speaker 6: what the recourse payment will be when it ever, if

280
00:12:51,600 --> 00:12:54,120
Speaker 6: and when it ever has to revert back to another owner,

281
00:12:54,160 --> 00:12:55,760
Speaker 6: and it's not going to be held on to by

282
00:12:55,760 --> 00:12:57,440
Speaker 6: the SPV. And I think this will turn out to

283
00:12:57,440 --> 00:12:59,480
Speaker 6: be really important four or five years down the road,

284
00:13:00,080 --> 00:13:14,200
Speaker 6: but right now nobody cares.

285
00:13:16,960 --> 00:13:21,199
Speaker 3: So Number one, the lifespan of data centers is actually

286
00:13:21,360 --> 00:13:24,440
Speaker 3: not that long. I can't remember the exact estimate, but

287
00:13:24,480 --> 00:13:27,000
Speaker 3: maybe like three or four years something like that. And

288
00:13:27,040 --> 00:13:30,160
Speaker 3: then also you have this risk that tenants are sort

289
00:13:30,200 --> 00:13:33,240
Speaker 3: of rolling through and no one knows what that actually

290
00:13:33,280 --> 00:13:35,200
Speaker 3: means for the structure of the debt, and you kind

291
00:13:35,200 --> 00:13:37,080
Speaker 3: of get this asset liability mismatch.

292
00:13:37,760 --> 00:13:40,280
Speaker 6: Yeah, so I'll start with the first one first. So

293
00:13:40,520 --> 00:13:42,760
Speaker 6: this gets into something Michael Berry was tweeting about the

294
00:13:42,800 --> 00:13:46,319
Speaker 6: other day, which was sort of entertaining that back about

295
00:13:46,360 --> 00:13:51,120
Speaker 6: four years ago, tech companies changed the appreciation schedule or

296
00:13:51,160 --> 00:13:53,040
Speaker 6: the assets inside of data centers.

297
00:13:53,240 --> 00:13:57,200
Speaker 5: They extended them somewhat. Now, that wasn't an error.

298
00:13:57,240 --> 00:14:00,520
Speaker 6: The reality is that data centers used for the purposes

299
00:14:00,559 --> 00:14:02,400
Speaker 6: like at aws, where You've got a big S three

300
00:14:02,440 --> 00:14:05,160
Speaker 6: bucket and I'm storing data inside of it. Those things

301
00:14:05,280 --> 00:14:08,280
Speaker 6: generally speaking, the assets are long lived. I'm not running

302
00:14:08,280 --> 00:14:10,880
Speaker 6: them flat out, it's not. These are not streetcar racers

303
00:14:10,880 --> 00:14:13,200
Speaker 6: that I'm running around inside of a data center. These

304
00:14:13,200 --> 00:14:16,640
Speaker 6: are relatively inexpensive chips that I'm using for really mundane

305
00:14:16,679 --> 00:14:20,680
Speaker 6: purposes like storing large amounts terabytes exhibites of data inside

306
00:14:20,680 --> 00:14:23,080
Speaker 6: of s three buckets, so it's not unreasonable to say

307
00:14:23,080 --> 00:14:26,040
Speaker 6: their lifespans fairly long. They're not being taxed that heavily,

308
00:14:26,160 --> 00:14:29,160
Speaker 6: so pushing out the depreciation schedule makes a lot of sense.

309
00:14:29,240 --> 00:14:32,640
Speaker 6: But that was coincident with the emergence of GPU driven

310
00:14:32,720 --> 00:14:36,000
Speaker 6: data centers using products like the chips from Nvidia, and

311
00:14:36,040 --> 00:14:39,200
Speaker 6: those have much shorter lifespans, so depending on the usage.

312
00:14:39,200 --> 00:14:39,720
Speaker 5: So there's two.

313
00:14:39,640 --> 00:14:44,400
Speaker 6: Different reasons why the lifespan and therefore the depreciation schedule

314
00:14:44,440 --> 00:14:46,760
Speaker 6: of a GPU inside of a data center is very different.

315
00:14:46,880 --> 00:14:49,560
Speaker 6: So the reason most people think about is, oh, well,

316
00:14:49,640 --> 00:14:51,720
Speaker 6: technology changes really quickly and I want to have the

317
00:14:51,800 --> 00:14:53,400
Speaker 6: latest and greatest, and therefore I'm going to have to

318
00:14:53,480 --> 00:14:58,320
Speaker 6: upgrade all the time. That's important, but it's probably about equal,

319
00:14:58,360 --> 00:15:01,600
Speaker 6: if not maybe slightly less important the nature of how

320
00:15:01,640 --> 00:15:04,200
Speaker 6: the chip is used inside the data center. So when

321
00:15:04,280 --> 00:15:07,560
Speaker 6: you run using like the latest, say a Nvidia chip

322
00:15:07,600 --> 00:15:10,160
Speaker 6: for training a model, those things are being run flat

323
00:15:10,160 --> 00:15:12,480
Speaker 6: out twenty four hours a day, seven days a week,

324
00:15:12,480 --> 00:15:14,720
Speaker 6: which is why they're liquid cool. They're inside of these

325
00:15:15,080 --> 00:15:17,520
Speaker 6: giant centers where one of your primary problems is keeping

326
00:15:17,560 --> 00:15:18,120
Speaker 6: them all cool.

327
00:15:18,600 --> 00:15:20,880
Speaker 5: It's like saying I bought a used car and.

328
00:15:20,840 --> 00:15:22,720
Speaker 6: I don't care what it was used for. Well, if

329
00:15:22,760 --> 00:15:24,440
Speaker 6: it turns out it was used by someone who was

330
00:15:24,480 --> 00:15:27,240
Speaker 6: doing like Laman's twenty four hours of endurance with it,

331
00:15:27,480 --> 00:15:30,160
Speaker 6: that's very different. Even if the mileage is the same

332
00:15:30,200 --> 00:15:31,280
Speaker 6: as someone who only drove to.

333
00:15:31,320 --> 00:15:32,200
Speaker 5: Church on Sundays.

334
00:15:32,360 --> 00:15:36,720
Speaker 6: Right, these are very different consequences with respect to what's

335
00:15:36,760 --> 00:15:38,920
Speaker 6: called the thermal degradation of the chip. The chip's been

336
00:15:39,000 --> 00:15:42,440
Speaker 6: run hot and flat out, so it probably it's useful.

337
00:15:42,480 --> 00:15:45,400
Speaker 6: Lifespan might be on the order of two years, maybe

338
00:15:45,400 --> 00:15:48,680
Speaker 6: even eighteen months. So there's a huge difference in terms

339
00:15:48,720 --> 00:15:51,800
Speaker 6: of how the chip was used, leaving aside whether or

340
00:15:51,840 --> 00:15:53,600
Speaker 6: not there's a new generation of what's come along. So

341
00:15:53,760 --> 00:15:56,920
Speaker 6: that takes us back to these depreciation schedules. So these

342
00:15:56,960 --> 00:16:00,920
Speaker 6: depreciation schedules change, just as the nature of how the

343
00:16:00,920 --> 00:16:04,360
Speaker 6: lifespan of the chips changed dramatically, because I can use

344
00:16:04,360 --> 00:16:06,560
Speaker 6: something for you know, storing things in s three buckets

345
00:16:06,600 --> 00:16:08,720
Speaker 6: for a long time, six to eight years isn't unreasonable.

346
00:16:09,200 --> 00:16:12,760
Speaker 6: But if I'm doing the the Laman's endurance equivalent with

347
00:16:12,880 --> 00:16:16,200
Speaker 6: a GPU, it might be eighteen months. That's a huge

348
00:16:16,240 --> 00:16:19,000
Speaker 6: difference in terms of the likely lifespan of a product

349
00:16:19,040 --> 00:16:21,840
Speaker 6: that I'm depreciating over a very different period. And so

350
00:16:21,920 --> 00:16:24,280
Speaker 6: that's a huge part of the problem here with respect

351
00:16:24,320 --> 00:16:28,440
Speaker 6: to understanding the intrinsics in terms of how data centers

352
00:16:29,040 --> 00:16:31,360
Speaker 6: can and can't make money. How you have to think

353
00:16:31,400 --> 00:16:34,560
Speaker 6: about the likely capex requirements because of this much shorter

354
00:16:34,640 --> 00:16:37,000
Speaker 6: life span of the underlying technology, and then.

355
00:16:36,840 --> 00:16:40,400
Speaker 3: Talk about the tenancy rollover risk. I guess we might

356
00:16:40,400 --> 00:16:40,760
Speaker 3: call it.

357
00:16:41,200 --> 00:16:44,760
Speaker 6: Yeah, it's really interesting. So one way to think about

358
00:16:45,160 --> 00:16:48,160
Speaker 6: data centers is as giant apartment buildings. Right, They're essentially

359
00:16:48,200 --> 00:16:50,560
Speaker 6: gigantic commercial pieces of commercial real estate with a bunch

360
00:16:50,600 --> 00:16:53,560
Speaker 6: of tenants. Sometimes there's a lot of tenants, sometimes there's

361
00:16:53,560 --> 00:16:55,920
Speaker 6: only one. Sometimes Google bought the whole apartment building and

362
00:16:55,960 --> 00:16:57,760
Speaker 6: just moved in, Or it's a giant office building they

363
00:16:57,800 --> 00:16:59,840
Speaker 6: just moved in. It's all theirs, right, So think about

364
00:16:59,880 --> 00:17:02,320
Speaker 6: it in those sorts of terms. And the reason why

365
00:17:02,400 --> 00:17:04,960
Speaker 6: as a sponsor of a data center I might take

366
00:17:05,000 --> 00:17:07,159
Speaker 6: a different view on how many tenants I want is

367
00:17:07,200 --> 00:17:09,720
Speaker 6: again you think about it in terms of what can

368
00:17:09,760 --> 00:17:11,919
Speaker 6: I get Google to pay? But whereasus what can I

369
00:17:12,000 --> 00:17:15,000
Speaker 6: get someone who's a much flightier tenant to pay? Well,

370
00:17:15,040 --> 00:17:17,399
Speaker 6: I can get the flightier tenants, more of them and

371
00:17:17,480 --> 00:17:21,440
Speaker 6: diversified as all leasing inside the data center, paying higher

372
00:17:21,560 --> 00:17:24,800
Speaker 6: lease rates for GPUs over the period of tendency than

373
00:17:24,800 --> 00:17:26,840
Speaker 6: I can get a Google to pay. Why because Google's

374
00:17:26,840 --> 00:17:28,880
Speaker 6: got great credit, they don't have to pay very much

375
00:17:28,880 --> 00:17:29,600
Speaker 6: and they know they don't.

376
00:17:29,600 --> 00:17:31,280
Speaker 5: So if you look at the commercial real estate.

377
00:17:31,200 --> 00:17:34,520
Speaker 6: Data, the cap rate, the blended cap rate for these

378
00:17:34,560 --> 00:17:37,320
Speaker 6: for the largest data centers that are tenanted by hyperscalers

379
00:17:37,920 --> 00:17:41,000
Speaker 6: is horrible. It's like four point eight five point three percent.

380
00:17:41,119 --> 00:17:43,920
Speaker 6: It's like, why don't you just buy a treasure you're doing.

381
00:17:44,160 --> 00:17:47,440
Speaker 6: So what happens then is people start blending in more

382
00:17:47,440 --> 00:17:50,359
Speaker 6: different kinds of tenants to Tracy's point, as an effort

383
00:17:50,400 --> 00:17:53,240
Speaker 6: to try and improve the yield the cap rate on

384
00:17:53,600 --> 00:17:56,240
Speaker 6: the underlying instrument, which is the data center. So you

385
00:17:56,280 --> 00:17:58,440
Speaker 6: could do all of this should start to sound familiar

386
00:17:58,520 --> 00:18:01,040
Speaker 6: because it's this idea of a blend together all of

387
00:18:01,040 --> 00:18:03,359
Speaker 6: these different tendencies. I can increase the yield of the

388
00:18:03,400 --> 00:18:07,479
Speaker 6: securitized instrument, but that also changes the risk profile of

389
00:18:07,520 --> 00:18:09,400
Speaker 6: what comes out at the other end, which just takes

390
00:18:09,480 --> 00:18:12,920
Speaker 6: us to things like the increasing usage of these things

391
00:18:12,920 --> 00:18:15,639
Speaker 6: in asset backed securities, which are these trench securities that

392
00:18:15,720 --> 00:18:18,359
Speaker 6: have all the different pieces, We have different layers associated

393
00:18:18,400 --> 00:18:21,399
Speaker 6: with it, and that's a reflection of well, there's different

394
00:18:21,440 --> 00:18:24,280
Speaker 6: tenants inside these data centers, and people want different exposures

395
00:18:24,280 --> 00:18:26,240
Speaker 6: to risks. So I may only want to buy the

396
00:18:26,280 --> 00:18:29,440
Speaker 6: senior tranch. You may want to buy the mezzanine and trace.

397
00:18:29,520 --> 00:18:31,560
Speaker 6: He may want to buy the equity charge.

398
00:18:31,800 --> 00:18:33,959
Speaker 3: Can I just say, I know we already said this,

399
00:18:34,359 --> 00:18:39,719
Speaker 3: but Paul is truly, truly the perfect guest. I remember

400
00:18:39,720 --> 00:18:43,359
Speaker 3: reading his coverage of subprime and securitization in like two

401
00:18:43,400 --> 00:18:45,560
Speaker 3: thousand and eight, and so having someone who's able to

402
00:18:45,640 --> 00:18:49,880
Speaker 3: synthesize that experience with what's going on now is just fantastic.

403
00:18:50,040 --> 00:18:51,760
Speaker 2: I kind of can't believe we're doing this again. I know,

404
00:18:51,800 --> 00:18:54,280
Speaker 2: I mean, look, I mean again, there's nothing inherently wrong

405
00:18:54,320 --> 00:18:58,960
Speaker 2: with SPVs. There's nothing inherently wrong tranching, right, Like a

406
00:18:58,960 --> 00:19:02,159
Speaker 2: lot of these things are very intuitive, etc. But it

407
00:19:02,240 --> 00:19:05,360
Speaker 2: is still a little weird how central this is and

408
00:19:05,400 --> 00:19:08,639
Speaker 2: how it's the same old There's nothing I mean, on

409
00:19:08,680 --> 00:19:10,520
Speaker 2: some financial level, it feels very familiar.

410
00:19:10,680 --> 00:19:12,520
Speaker 5: No, there's nothing new un to the sun.

411
00:19:13,160 --> 00:19:15,199
Speaker 6: But I think that point is really important It's not

412
00:19:15,280 --> 00:19:18,200
Speaker 6: that tranches are evil. It's not the securitization is evil,

413
00:19:18,280 --> 00:19:20,840
Speaker 6: or that asset backed security your project finance is evil.

414
00:19:21,200 --> 00:19:26,000
Speaker 5: No, all of these things are terrific pieces of the arsenal.

415
00:19:26,040 --> 00:19:28,920
Speaker 6: Whenever you're actually raising money for projects, the issues start

416
00:19:28,960 --> 00:19:30,800
Speaker 6: to arise at the scale, which is what you guys

417
00:19:30,840 --> 00:19:34,200
Speaker 6: have already alluded to. But the secondary piece, which again

418
00:19:34,280 --> 00:19:37,560
Speaker 6: will sound painfully familiar to the financial crisis, is there's

419
00:19:37,560 --> 00:19:40,119
Speaker 6: a flywheel that gets created at the back end of this.

420
00:19:40,680 --> 00:19:44,600
Speaker 6: So once you start securitizing the yield producing assets in

421
00:19:44,640 --> 00:19:47,720
Speaker 6: the form of these tranch securities, the people who are

422
00:19:47,720 --> 00:19:50,640
Speaker 6: purchasing those things don't give a rats ask what's going

423
00:19:50,680 --> 00:19:53,919
Speaker 6: on inside this AI. I joke all the time that

424
00:19:53,960 --> 00:19:55,720
Speaker 6: a lot of these people can't spell AI. They don't

425
00:19:55,720 --> 00:19:57,359
Speaker 6: care what's going on inside the.

426
00:19:57,400 --> 00:19:58,520
Speaker 5: Data center, right.

427
00:19:59,119 --> 00:20:00,840
Speaker 6: It could be you know, the world Hide and Go

428
00:20:00,920 --> 00:20:03,359
Speaker 6: Seek Championships had going on in there. I don't care

429
00:20:03,600 --> 00:20:05,240
Speaker 6: as long as it generates heels and I.

430
00:20:05,160 --> 00:20:06,280
Speaker 5: Can securitize it.

431
00:20:06,600 --> 00:20:06,719
Speaker 3: Well.

432
00:20:06,720 --> 00:20:09,000
Speaker 6: It's very much an analogous to what's happened in prior

433
00:20:09,080 --> 00:20:12,440
Speaker 6: periods like this, where again you get this secondary flywheel

434
00:20:12,480 --> 00:20:15,840
Speaker 6: effect of let's just create more of these things because

435
00:20:15,840 --> 00:20:18,680
Speaker 6: our customers want more and they're really easy to securitize

436
00:20:18,720 --> 00:20:20,520
Speaker 6: and look gets backst up by Meta and Google or

437
00:20:20,520 --> 00:20:21,040
Speaker 6: whoever else.

438
00:20:21,160 --> 00:20:24,200
Speaker 2: Well, so this actually brings important point. I mentioned this

439
00:20:24,359 --> 00:20:27,040
Speaker 2: great report out from the Center for Public Enterprise. One

440
00:20:27,080 --> 00:20:31,120
Speaker 2: of the things that they pointed out is in this

441
00:20:31,280 --> 00:20:34,720
Speaker 2: market environment where everyone is just you know, there's this

442
00:20:34,800 --> 00:20:38,399
Speaker 2: sort of AI pixi us, but also just the reality

443
00:20:38,440 --> 00:20:41,040
Speaker 2: if your revenues are surging, the market probably loves you,

444
00:20:41,920 --> 00:20:45,399
Speaker 2: like talk to us about the unit economics. Here is

445
00:20:45,400 --> 00:20:48,960
Speaker 2: the incentive for all the players essentially to just grow

446
00:20:49,080 --> 00:20:52,920
Speaker 2: the top line as much as possible, even if these

447
00:20:52,960 --> 00:20:56,320
Speaker 2: aren't whether we're talking about inference on a per token basis,

448
00:20:56,720 --> 00:20:59,119
Speaker 2: even if these aren't particularly profitable, how do you think

449
00:20:59,160 --> 00:21:01,880
Speaker 2: about the union economics of some of these businesses and

450
00:21:01,960 --> 00:21:05,200
Speaker 2: how that could eventually perhaps sort of you know, come

451
00:21:05,240 --> 00:21:06,280
Speaker 2: home to Ruster to speak.

452
00:21:06,600 --> 00:21:07,359
Speaker 5: Yeah, So.

453
00:21:09,160 --> 00:21:11,159
Speaker 6: The term of art obviously is these things have negative

454
00:21:11,240 --> 00:21:13,840
Speaker 6: unit economics, which is a fancy way of saying that

455
00:21:13,880 --> 00:21:15,639
Speaker 6: we lose money on every sale and try to make

456
00:21:15,640 --> 00:21:18,200
Speaker 6: it up on volume. Right, I mean, that's the that's

457
00:21:18,200 --> 00:21:20,320
Speaker 6: the problem here. So but that's okay, I mean, we've

458
00:21:20,320 --> 00:21:21,960
Speaker 6: had lots of Amazon.

459
00:21:21,520 --> 00:21:24,119
Speaker 5: In its early days that negative unit economics. You can

460
00:21:24,119 --> 00:21:24,760
Speaker 5: get past that.

461
00:21:25,080 --> 00:21:27,000
Speaker 6: And as an aside, I'll say right here, all of

462
00:21:27,000 --> 00:21:28,639
Speaker 6: the things that I'm saying is and to say that

463
00:21:28,720 --> 00:21:31,760
Speaker 6: you know AI is some kind of you know, free

464
00:21:31,960 --> 00:21:36,040
Speaker 6: tamagatche thing, that's just a fad as an incredibly important technology.

465
00:21:36,240 --> 00:21:38,439
Speaker 6: What we're talking about is how it's funded and the

466
00:21:38,440 --> 00:21:40,600
Speaker 6: consequences of doing that in terms of what's going to

467
00:21:40,600 --> 00:21:42,639
Speaker 6: happen with respect to the businesses and the return on

468
00:21:42,680 --> 00:21:46,680
Speaker 6: those businesses. Right, So, the unit economics are dire for

469
00:21:46,800 --> 00:21:49,280
Speaker 6: a bunch of reasons, have mostly having to do with

470
00:21:49,960 --> 00:21:52,840
Speaker 6: the more tokens you have to produce. The costs rise

471
00:21:53,040 --> 00:21:56,280
Speaker 6: more or less linearly with the demand on the system.

472
00:21:56,320 --> 00:21:59,600
Speaker 6: As opposed to an orthodox software business where the more

473
00:21:59,640 --> 00:22:02,280
Speaker 6: people use my service, the more people across which I

474
00:22:02,280 --> 00:22:05,800
Speaker 6: can spread my relatively fixed costs. That's not the way

475
00:22:05,960 --> 00:22:09,800
Speaker 6: that for the most part, current generation large language models

476
00:22:09,840 --> 00:22:14,200
Speaker 6: were costs rise linearly or sublinearly with the number of users,

477
00:22:14,240 --> 00:22:16,680
Speaker 6: which makes for really crappy unit.

478
00:22:16,480 --> 00:22:18,280
Speaker 5: Economics, and that's a big part of the problem.

479
00:22:18,359 --> 00:22:20,600
Speaker 6: So from there you get to the question of Okay,

480
00:22:20,680 --> 00:22:22,280
Speaker 6: so what does it have to look like in terms

481
00:22:22,280 --> 00:22:25,159
Speaker 6: of making it look profitable. There's lots of ways to

482
00:22:25,200 --> 00:22:27,000
Speaker 6: back into this. You can do bottoms up models. It

483
00:22:27,040 --> 00:22:29,560
Speaker 6: would suggest that like if every iPhone newsrun earth paid

484
00:22:29,560 --> 00:22:33,160
Speaker 6: fifty bucks do at work, we could have around a

485
00:22:33,160 --> 00:22:35,760
Speaker 6: four hundred billion dollar, five hundred billion dollar annual stream

486
00:22:35,760 --> 00:22:37,480
Speaker 6: of revenue flowing. And well, that's not going to happen,

487
00:22:37,560 --> 00:22:39,080
Speaker 6: but it's worth pointing out like that would do it.

488
00:22:39,119 --> 00:22:40,480
Speaker 6: But it gives you a sense of the kind of

489
00:22:40,520 --> 00:22:44,280
Speaker 6: scale of what at a consumer level, for example, it

490
00:22:44,359 --> 00:22:45,159
Speaker 6: might have to look like.

491
00:22:45,560 --> 00:22:46,800
Speaker 5: People come out it from the other end.

492
00:22:46,800 --> 00:22:48,320
Speaker 6: One of my favorite ways that people come out is

493
00:22:48,320 --> 00:22:50,520
Speaker 6: to say, well, we could create a viable model here.

494
00:22:50,520 --> 00:22:53,200
Speaker 6: If you think this was in the JPM call last week.

495
00:22:53,200 --> 00:22:54,600
Speaker 6: I don't know if you guys saw the summary of it,

496
00:22:54,640 --> 00:22:57,040
Speaker 6: but it was huge fun for the whole family listening.

497
00:22:57,080 --> 00:22:59,560
Speaker 6: And so one of the ways they backed into it

498
00:22:59,560 --> 00:23:02,400
Speaker 6: was a top model where they said, well, the global

499
00:23:02,520 --> 00:23:04,639
Speaker 6: TAM for human labor.

500
00:23:05,480 --> 00:23:08,560
Speaker 5: I love the five trillion dollars. I love the global TAM.

501
00:23:08,600 --> 00:23:08,880
Speaker 5: I said.

502
00:23:08,920 --> 00:23:10,720
Speaker 6: That was right up there with saying like if I

503
00:23:10,800 --> 00:23:13,240
Speaker 6: reduce humans to their chemical components, here's what.

504
00:23:13,200 --> 00:23:13,879
Speaker 5: I can get for you.

505
00:23:14,240 --> 00:23:17,600
Speaker 3: Well, this was this was Steve Eisman's line, which was like,

506
00:23:17,800 --> 00:23:21,080
Speaker 3: beware of anyone that mentions tam right.

507
00:23:21,040 --> 00:23:23,200
Speaker 6: Right, right, no exactly, and so then and then they play.

508
00:23:23,280 --> 00:23:26,240
Speaker 6: The next step is of course to say, well, imagine

509
00:23:26,280 --> 00:23:28,600
Speaker 6: we can get ten percent of that, right, which is

510
00:23:28,880 --> 00:23:31,240
Speaker 6: which is obviously one of the oldest cliches. It's like saying,

511
00:23:31,240 --> 00:23:32,600
Speaker 6: you know, I'm going to get five percent of the

512
00:23:32,680 --> 00:23:34,560
Speaker 6: Chinese market. No one ever gets five percent of the

513
00:23:34,600 --> 00:23:35,160
Speaker 6: Chinese market.

514
00:23:35,240 --> 00:23:35,800
Speaker 5: This doesn't happen.

515
00:23:35,880 --> 00:23:38,480
Speaker 6: So the same thing won't happen with global labor. But

516
00:23:38,560 --> 00:23:40,200
Speaker 6: if you were to do that, you do the math

517
00:23:40,280 --> 00:23:42,600
Speaker 6: on that that call those kinds of numbers gets you

518
00:23:42,680 --> 00:23:45,480
Speaker 6: to a weighted average cost of capital basis to a

519
00:23:45,480 --> 00:23:48,680
Speaker 6: reasonable return on current and planned expenditures with respect to

520
00:23:48,760 --> 00:23:51,320
Speaker 6: AI data centers. If you assume we're heading to about

521
00:23:51,320 --> 00:23:54,120
Speaker 6: a three or four trillion dollar a number, which is

522
00:23:54,359 --> 00:23:57,119
Speaker 6: kind of the I think it's around the number that

523
00:23:57,160 --> 00:23:58,600
Speaker 6: most people put out there, which I think is a

524
00:23:58,600 --> 00:24:00,880
Speaker 6: completely wrong number, but nevertheles that's the kind of number

525
00:24:00,880 --> 00:24:01,919
Speaker 6: and what you'd have to do to get there.

526
00:24:01,960 --> 00:24:02,800
Speaker 5: So you can get there from.

527
00:24:02,640 --> 00:24:05,600
Speaker 6: A bottoms up model by making some really unreasonable assumptions

528
00:24:05,640 --> 00:24:07,920
Speaker 6: about the total numbers of subscribers and what they pay.

529
00:24:08,400 --> 00:24:10,639
Speaker 6: You can get there from a top down model. You

530
00:24:10,680 --> 00:24:12,600
Speaker 6: can also get there by thinking about it purely in

531
00:24:12,680 --> 00:24:16,400
Speaker 6: terms of industrial users. I think about purely API users

532
00:24:16,480 --> 00:24:19,879
Speaker 6: just for end retail users of AI don't exist. And say,

533
00:24:20,200 --> 00:24:22,720
Speaker 6: you know, Andthropics projecting seventy billion dollars in revenue in

534
00:24:22,760 --> 00:24:25,320
Speaker 6: twenty twenty eight, something like thirty five percent of their

535
00:24:25,320 --> 00:24:28,800
Speaker 6: current revenues. Most of their revenues today are from their API.

536
00:24:29,080 --> 00:24:31,960
Speaker 6: Thirty five percent of that is from software developers that

537
00:24:32,080 --> 00:24:37,000
Speaker 6: split between two large users, Copilot and Cursor. And so

538
00:24:37,119 --> 00:24:38,920
Speaker 6: you know, we can model that out. Everybody has to

539
00:24:39,000 --> 00:24:40,560
Speaker 6: become a software developer.

540
00:24:40,160 --> 00:24:41,480
Speaker 5: And we can make the math work.

541
00:24:41,720 --> 00:24:44,400
Speaker 6: The problem is it's got huge fragility right in customer

542
00:24:44,440 --> 00:24:48,159
Speaker 6: concentration risk. So a Cursor disappears as a user of

543
00:24:48,280 --> 00:24:51,399
Speaker 6: Entropics API, and you just blew out fifteen percent of

544
00:24:51,920 --> 00:24:54,280
Speaker 6: your revenues because they're gone and they've done something else.

545
00:24:54,840 --> 00:24:56,640
Speaker 6: And as it turns out, Cursor a two weeks ago

546
00:24:56,680 --> 00:24:58,800
Speaker 6: announced that they were trading their own internal model that

547
00:24:58,840 --> 00:25:00,600
Speaker 6: you could use for software developed and you wouldn't have

548
00:25:00,680 --> 00:25:03,920
Speaker 6: to call the Anthropic API so you can think about

549
00:25:03,920 --> 00:25:05,800
Speaker 6: all these different ways to get there, but they all

550
00:25:05,840 --> 00:25:08,000
Speaker 6: have a lot of built in fragility with respect to

551
00:25:08,800 --> 00:25:11,040
Speaker 6: so we all become software developers and we all subscribe

552
00:25:11,080 --> 00:25:11,520
Speaker 6: to Cursor.

553
00:25:12,840 --> 00:25:15,000
Speaker 3: Just going back to the used car analogy that you

554
00:25:15,040 --> 00:25:18,160
Speaker 3: mentioned before, when we're thinking about all this financing of

555
00:25:18,440 --> 00:25:21,439
Speaker 3: the AI capex spen, is it useful to think of

556
00:25:22,000 --> 00:25:26,880
Speaker 3: GPUs essentially as the collateral the problem?

557
00:25:26,960 --> 00:25:28,160
Speaker 5: Yes, or what would you.

558
00:25:28,080 --> 00:25:29,720
Speaker 3: Call the collateral in this case?

559
00:25:29,920 --> 00:25:31,160
Speaker 5: So what ends up happening.

560
00:25:31,359 --> 00:25:33,040
Speaker 6: The collateral in this case is the gp There's no

561
00:25:33,119 --> 00:25:35,800
Speaker 6: question it is the GPA. The issue is this disconnect,

562
00:25:35,800 --> 00:25:38,600
Speaker 6: this temporal mismatch that you alluded to earlier with respect

563
00:25:38,600 --> 00:25:41,520
Speaker 6: to the duration of the underlying debt and the assets

564
00:25:41,520 --> 00:25:42,640
Speaker 6: that are producing.

565
00:25:42,400 --> 00:25:44,320
Speaker 5: The income that allows me to pay for the debt.

566
00:25:44,600 --> 00:25:48,159
Speaker 6: Right, so we've got this probably unprecedented temporal messmatch with

567
00:25:48,200 --> 00:25:52,120
Speaker 6: thirty year loans and two year depreciation on the underlying collateral,

568
00:25:52,160 --> 00:25:55,000
Speaker 6: which is essentially the GPUs that are the income producing assets.

569
00:25:55,840 --> 00:25:59,560
Speaker 6: And so that creates this constant refinancing risk because I'm

570
00:25:59,560 --> 00:26:01,040
Speaker 6: going to can you only have to turn over the

571
00:26:01,040 --> 00:26:03,720
Speaker 6: base And we've seen this many many times right now,

572
00:26:03,760 --> 00:26:05,480
Speaker 6: it's easy to turn it over, but in two years

573
00:26:05,480 --> 00:26:07,800
Speaker 6: it may not be possible. There's a wave of refinancings

574
00:26:07,840 --> 00:26:09,640
Speaker 6: coming in twenty twenty eight in many of the more.

575
00:26:09,560 --> 00:26:10,719
Speaker 5: Speculative data centers.

576
00:26:11,000 --> 00:26:12,639
Speaker 6: Will they be able to turn over their debt and

577
00:26:12,680 --> 00:26:14,399
Speaker 6: refinance all the GPUs today?

578
00:26:14,400 --> 00:26:16,320
Speaker 5: They could? This today is in twenty twenty eight.

579
00:26:16,800 --> 00:26:21,240
Speaker 6: So that's the inherent problem, is this structural temporal mismatch

580
00:26:21,320 --> 00:26:24,200
Speaker 6: between the income producing assets and the duration of the lungs.

581
00:26:24,200 --> 00:26:26,480
Speaker 6: And it gets worse if you think about it in

582
00:26:26,520 --> 00:26:29,320
Speaker 6: more realistic terms, think about it in terms of one

583
00:26:29,359 --> 00:26:31,720
Speaker 6: of the other gating factors here that's driving all.

584
00:26:31,600 --> 00:26:34,240
Speaker 5: Of this is the scarcity of energy supply. It's really difficult.

585
00:26:35,000 --> 00:26:36,680
Speaker 6: You can hook them up to the well. It's actually

586
00:26:36,720 --> 00:26:38,119
Speaker 6: kind of turned into a bit of a joke. I

587
00:26:38,119 --> 00:26:39,520
Speaker 6: can hook you up to the grid, but I can't

588
00:26:39,520 --> 00:26:40,800
Speaker 6: give you power. I don't know if you saw the

589
00:26:40,840 --> 00:26:43,800
Speaker 6: recent episode with the Oregon Public Utilities Commission, Amazon had

590
00:26:43,840 --> 00:26:46,840
Speaker 6: three data centers that they connected to the grid, and

591
00:26:46,880 --> 00:26:49,080
Speaker 6: it was kind of like the Oregon PUC said, Oh,

592
00:26:49,119 --> 00:26:51,840
Speaker 6: you want power too, Oh, I can't help you with that.

593
00:26:52,200 --> 00:26:53,040
Speaker 5: We can't help you with that.

594
00:26:53,160 --> 00:26:55,560
Speaker 6: So now there's a complaint in it the Oregon PUC

595
00:26:55,640 --> 00:26:59,480
Speaker 6: from ADS, Amazon's the digital services group that runs aws,

596
00:26:59,480 --> 00:27:01,440
Speaker 6: complaining we now have data centers, but.

597
00:27:01,400 --> 00:27:02,240
Speaker 5: We have no power.

598
00:27:02,800 --> 00:27:04,680
Speaker 6: Right it sounds a little bit like like a winter

599
00:27:04,800 --> 00:27:07,320
Speaker 6: storm hazard or something, but it's the structural problem with

600
00:27:07,359 --> 00:27:10,639
Speaker 6: respect to the inability. We can connect people, but we

601
00:27:10,680 --> 00:27:13,240
Speaker 6: can't provide them with power. So the next stage is

602
00:27:13,280 --> 00:27:15,639
Speaker 6: and this takes bets back to the collateral problem in

603
00:27:15,680 --> 00:27:18,560
Speaker 6: the temporal mismatch, is that people are doing behind the

604
00:27:18,600 --> 00:27:22,119
Speaker 6: meter power. They're building natural gas or if you're fair me,

605
00:27:22,320 --> 00:27:26,040
Speaker 6: you're saying wild things about nuclear power and you're saying, Okay,

606
00:27:26,080 --> 00:27:27,840
Speaker 6: I'm coming with my own power. You don't need to

607
00:27:27,880 --> 00:27:30,119
Speaker 6: connect me to the grid. I'm going to power this myself.

608
00:27:30,880 --> 00:27:33,199
Speaker 6: That creates two or three different issues, but among the

609
00:27:33,240 --> 00:27:36,360
Speaker 6: more important is think about how long lived an asset

610
00:27:36,400 --> 00:27:38,920
Speaker 6: a natural gas plant is. This is not something that's

611
00:27:38,960 --> 00:27:41,680
Speaker 6: got a five year lifespan and we just truly wave goodbye.

612
00:27:41,760 --> 00:27:44,480
Speaker 6: This is going to be running probably twenty five to

613
00:27:44,520 --> 00:27:47,919
Speaker 6: thirty years. And the only thing your ability to forecast.

614
00:27:48,560 --> 00:27:50,399
Speaker 6: We know the cost of the natural gas plant, but

615
00:27:50,440 --> 00:27:51,920
Speaker 6: in terms of the cost of the center, and it's

616
00:27:51,920 --> 00:27:54,840
Speaker 6: incompability to generate enough income to pay off the loan

617
00:27:54,880 --> 00:27:58,080
Speaker 6: associated with the natural gas plant. God help you if

618
00:27:58,119 --> 00:28:00,560
Speaker 6: you think you can sort that out, because what you've

619
00:28:00,560 --> 00:28:03,760
Speaker 6: really got is a huge likelihood of a stranded asset

620
00:28:03,800 --> 00:28:06,240
Speaker 6: of their natural gas plants that are longer useful for

621
00:28:06,359 --> 00:28:07,919
Speaker 6: powering these things that they were built for.

622
00:28:23,880 --> 00:28:26,840
Speaker 2: The good news is that Daniel Jurgen said this on

623
00:28:27,040 --> 00:28:31,359
Speaker 2: the show. You know the back orders for natural gas turbines,

624
00:28:31,359 --> 00:28:33,080
Speaker 2: like you probably if you ordered one today, you would

625
00:28:33,080 --> 00:28:35,440
Speaker 2: probably get it in twenty thirty. So the good news

626
00:28:35,480 --> 00:28:37,960
Speaker 2: that I suppose ten years is that at least you

627
00:28:38,000 --> 00:28:40,560
Speaker 2: don't have to have the turbines sitting there for years.

628
00:28:40,560 --> 00:28:42,160
Speaker 2: Like I don't know, Maybe I don't know if that's

629
00:28:42,200 --> 00:28:44,280
Speaker 2: good news at all, but there are se I may

630
00:28:44,320 --> 00:28:46,160
Speaker 2: never get it in, You may never get the gas

631
00:28:46,200 --> 00:28:48,440
Speaker 2: plant built. Anyway, someone will be stuck with the book.

632
00:28:48,760 --> 00:28:51,680
Speaker 6: It kind of raises this goes back to Tracy's question earlier.

633
00:28:51,720 --> 00:28:55,040
Speaker 6: This raises a really interesting thing. So like, honestly, what

634
00:28:55,120 --> 00:28:57,719
Speaker 6: the f are all these people doing who are announcing

635
00:28:58,920 --> 00:29:02,200
Speaker 6: the giant unding translation. I think of it like people

636
00:29:02,320 --> 00:29:04,560
Speaker 6: all showing up with the OK Corral at once and

637
00:29:04,600 --> 00:29:07,240
Speaker 6: It's like, dude over there has one gun, I got two.

638
00:29:07,280 --> 00:29:09,959
Speaker 3: Yeah, I got Oh that's not a nice this is anie.

639
00:29:10,160 --> 00:29:10,320
Speaker 5: Yeah.

640
00:29:10,600 --> 00:29:13,560
Speaker 6: But it's this deterrence. It's this deterrence program that's going on.

641
00:29:13,640 --> 00:29:16,560
Speaker 6: Don't even imagine spending fifty because I'm spending one hundred.

642
00:29:17,080 --> 00:29:19,360
Speaker 5: No point in you doing any of those. That's very

643
00:29:19,600 --> 00:29:20,520
Speaker 5: game theoretic.

644
00:29:20,800 --> 00:29:23,040
Speaker 3: Well, this also worries me because you hear so many

645
00:29:23,120 --> 00:29:27,200
Speaker 3: people framing this as like an existential competition. Right, and

646
00:29:27,240 --> 00:29:31,600
Speaker 3: once you start calling something existential, the limit on spend,

647
00:29:31,760 --> 00:29:33,000
Speaker 3: well it becomes unlimited.

648
00:29:33,080 --> 00:29:33,240
Speaker 5: Right.

649
00:29:33,360 --> 00:29:35,400
Speaker 3: It's about survival, so you'll spend anything.

650
00:29:35,520 --> 00:29:39,120
Speaker 2: That's why the conversation has turned in recent weeks to

651
00:29:39,240 --> 00:29:42,120
Speaker 2: the one entity that actually, at least in theory, can

652
00:29:42,160 --> 00:29:43,560
Speaker 2: print as much money as possible.

653
00:29:43,840 --> 00:29:47,720
Speaker 6: Right, that's the you know, the Sarah Friar's accidental foot

654
00:29:47,720 --> 00:29:49,040
Speaker 6: in mouth the thing earlier in the week.

655
00:29:49,240 --> 00:29:49,880
Speaker 5: But that's right.

656
00:29:49,920 --> 00:29:52,120
Speaker 6: But that's again goes back to my original point about

657
00:29:52,120 --> 00:29:56,560
Speaker 6: what makes this bubble unusual. It's this element that not

658
00:29:56,600 --> 00:29:58,760
Speaker 6: only is there a kind of bagstock, but there's actually

659
00:29:59,000 --> 00:30:01,960
Speaker 6: a notion of wrapping in the flag. We have to

660
00:30:02,000 --> 00:30:04,520
Speaker 6: win this competition, we have to do what it takes.

661
00:30:04,560 --> 00:30:08,080
Speaker 6: This is existential. It's US versus China, and it's not

662
00:30:08,160 --> 00:30:10,600
Speaker 6: just the US doing this. I was talking to some

663
00:30:10,640 --> 00:30:13,840
Speaker 6: Canadian policymakers just earlier this morning, exact same thing going

664
00:30:13,840 --> 00:30:15,800
Speaker 6: on there. We have to build a domestic in the

665
00:30:15,880 --> 00:30:18,360
Speaker 6: same thing in the UK, same thing in Germany. And

666
00:30:18,440 --> 00:30:21,520
Speaker 6: so there's this idea around the world that sovereign ai

667
00:30:21,640 --> 00:30:24,720
Speaker 6: is something that's incredibly important. So this this government backstop

668
00:30:24,760 --> 00:30:27,240
Speaker 6: isn't just mythic, it's it's global. It's this idea that

669
00:30:27,280 --> 00:30:28,960
Speaker 6: we all have to win, we all have to win,

670
00:30:28,960 --> 00:30:32,520
Speaker 6: which obviously can't happen, but that the government's playing a

671
00:30:32,600 --> 00:30:34,120
Speaker 6: role and that that be can trace this kind of

672
00:30:34,160 --> 00:30:35,680
Speaker 6: limitless course of capital.

673
00:30:35,720 --> 00:30:38,000
Speaker 2: You know. So one of the things that's going on,

674
00:30:38,200 --> 00:30:40,480
Speaker 2: and maybe it's part of the same the sort of

675
00:30:40,640 --> 00:30:44,880
Speaker 2: maximalist strategy mentioned Anthropic wants to get into data centers,

676
00:30:44,920 --> 00:30:48,800
Speaker 2: so everyone's sort of looking at how they can expand vertically.

677
00:30:48,840 --> 00:30:50,920
Speaker 2: Can I own the data centers? I think? You know,

678
00:30:51,000 --> 00:30:53,760
Speaker 2: Sam Altman has talked about owning chips or owning a

679
00:30:53,760 --> 00:30:57,240
Speaker 2: semiconductor fab at some point, like maybe that'll be part

680
00:30:57,240 --> 00:30:59,960
Speaker 2: of the story. Who knows. There's one thing that I don't.

681
00:31:00,040 --> 00:31:01,920
Speaker 2: I'm sort of curious. I'd love to have your take

682
00:31:02,000 --> 00:31:05,760
Speaker 2: on there was. At the end of September, Meta announced

683
00:31:05,760 --> 00:31:08,280
Speaker 2: the deal to buy Compute from core Weave, one of

684
00:31:08,280 --> 00:31:12,280
Speaker 2: these neo clouds. I don't totally get that because Meta

685
00:31:12,320 --> 00:31:14,440
Speaker 2: has its own data centers, et cetera. Do you have

686
00:31:14,480 --> 00:31:18,960
Speaker 2: some intuitive sense about what an established hyperscaler needs a

687
00:31:19,040 --> 00:31:22,520
Speaker 2: neo cloud for in this arrangement, what core Weave can

688
00:31:22,560 --> 00:31:25,680
Speaker 2: supply that Meta can't build on its own or buy

689
00:31:25,680 --> 00:31:26,160
Speaker 2: on its own.

690
00:31:26,840 --> 00:31:30,440
Speaker 5: Nothing.

691
00:31:29,480 --> 00:31:32,840
Speaker 6: So here's what's going on. This is what's going on

692
00:31:32,960 --> 00:31:36,000
Speaker 6: is that there's this form of hoarding going on. So

693
00:31:36,800 --> 00:31:39,640
Speaker 6: what's happening is is people saying, you have capacity, I

694
00:31:39,680 --> 00:31:40,600
Speaker 6: can lock that up.

695
00:31:40,760 --> 00:31:41,640
Speaker 5: I'll lock that up.

696
00:31:42,000 --> 00:31:44,000
Speaker 6: And because I can't lock it up yet by building

697
00:31:44,000 --> 00:31:45,960
Speaker 6: a data center quickly enough, I'll lock it up in

698
00:31:45,960 --> 00:31:49,400
Speaker 6: the marketplace. So once you start thinking of compute as

699
00:31:49,400 --> 00:31:52,920
Speaker 6: a hordable commodity, and what people are doing is trying

700
00:31:52,960 --> 00:31:56,479
Speaker 6: to hoard it, control it before someone else can do it,

701
00:31:56,480 --> 00:31:59,120
Speaker 6: because until they bring on their own access capacity. That's

702
00:31:59,160 --> 00:32:01,560
Speaker 6: really what's going on in a lot of these transactions.

703
00:32:01,600 --> 00:32:03,720
Speaker 6: This is a way of making sure that I may

704
00:32:03,720 --> 00:32:05,800
Speaker 6: not need this but you sure can have it. And

705
00:32:05,880 --> 00:32:08,400
Speaker 6: so there's there's an element of compute hoarding going on

706
00:32:08,560 --> 00:32:12,080
Speaker 6: across the map because of you know, this backlog and building.

707
00:32:11,880 --> 00:32:13,600
Speaker 5: Data centers that may or may not ever get built.

708
00:32:13,640 --> 00:32:14,280
Speaker 5: So that's the answer.

709
00:32:14,360 --> 00:32:16,160
Speaker 6: The answer isn't that they care at all about whether

710
00:32:16,200 --> 00:32:19,120
Speaker 6: or not they can run giant workloads on any particular

711
00:32:19,280 --> 00:32:22,600
Speaker 6: neo clouds provider. It's the idea of hoarding capacity and

712
00:32:22,640 --> 00:32:24,840
Speaker 6: making sure that no one else can have it, like

713
00:32:24,960 --> 00:32:27,120
Speaker 6: trying to have like the Hunt Brothers and the getting

714
00:32:27,120 --> 00:32:28,360
Speaker 6: a corner on the silver market.

715
00:32:29,320 --> 00:32:31,040
Speaker 3: You know, I want to go back to China because

716
00:32:31,080 --> 00:32:33,840
Speaker 3: it is true that the US and China seem locked

717
00:32:33,880 --> 00:32:37,320
Speaker 3: in this existential race for AI supremacy, but they seem

718
00:32:37,360 --> 00:32:39,680
Speaker 3: to be taking very different approaches to it. And in

719
00:32:39,720 --> 00:32:42,240
Speaker 3: the US, it's all about spending as much money as

720
00:32:42,280 --> 00:32:45,720
Speaker 3: you can developing these you know, state of the art,

721
00:32:46,000 --> 00:32:50,040
Speaker 3: mostly closed source models, whereas in China it seems to

722
00:32:50,080 --> 00:32:53,880
Speaker 3: be much more about rapid adoption and creating open source

723
00:32:53,920 --> 00:32:56,800
Speaker 3: models that just get out into the market much faster

724
00:32:57,400 --> 00:33:01,960
Speaker 3: and much more cheaply. And so I'm curious, like, which

725
00:33:02,000 --> 00:33:04,520
Speaker 3: of those approaches do you think it's going to win?

726
00:33:04,560 --> 00:33:04,840
Speaker 5: Here.

727
00:33:05,600 --> 00:33:09,920
Speaker 6: Yeah, so that's a really good question. So I think

728
00:33:10,280 --> 00:33:11,920
Speaker 6: it's going to be something closer to.

729
00:33:11,880 --> 00:33:14,120
Speaker 5: The Chinese approach, but not for the reasons they expect.

730
00:33:14,440 --> 00:33:17,840
Speaker 6: So the reason is because, so what, let's I'll reframe

731
00:33:17,840 --> 00:33:19,800
Speaker 6: what the Chinese are doing slightly, so I'll say that

732
00:33:19,840 --> 00:33:21,800
Speaker 6: instead of it just being a sort of an example

733
00:33:21,840 --> 00:33:24,160
Speaker 6: of open source, I don't think that's right. The right

734
00:33:24,200 --> 00:33:25,840
Speaker 6: way to think about it is they're using this kind

735
00:33:25,880 --> 00:33:29,000
Speaker 6: of distillation approach increasingly where there's kind of a you

736
00:33:29,040 --> 00:33:30,920
Speaker 6: think about it like, Okay, I'm a sales manager. I

737
00:33:30,920 --> 00:33:32,760
Speaker 6: don't want to train all my salespeople. I'm going to

738
00:33:32,760 --> 00:33:33,760
Speaker 6: train this dude.

739
00:33:33,640 --> 00:33:35,840
Speaker 5: And they're going to train all the sales But that's distillation, right.

740
00:33:35,880 --> 00:33:38,560
Speaker 6: You train the trainer, I train somebody who trains something else,

741
00:33:38,600 --> 00:33:41,360
Speaker 6: and something else in this case are these smaller models.

742
00:33:41,360 --> 00:33:44,960
Speaker 6: So that approach of kind of training the trainer really

743
00:33:44,960 --> 00:33:47,880
Speaker 6: speeds up the process of creating new models because I

744
00:33:47,960 --> 00:33:50,880
Speaker 6: distill them, I train them out of out of other

745
00:33:50,960 --> 00:33:54,200
Speaker 6: models that are really compute intensive, like anthropics or opening

746
00:33:54,200 --> 00:33:56,520
Speaker 6: eyes or whoever else is right. So the notion is,

747
00:33:57,440 --> 00:34:00,120
Speaker 6: so is there are huge efficiency gains to be had

748
00:34:00,200 --> 00:34:03,440
Speaker 6: in training and the Chinese are showing the huge efficiency

749
00:34:03,480 --> 00:34:05,520
Speaker 6: gains to be had, and the one way to think

750
00:34:05,560 --> 00:34:09,600
Speaker 6: about it is that the transformer models that underlie large

751
00:34:09,640 --> 00:34:12,920
Speaker 6: language models that are so computationally intensive, went from the

752
00:34:13,040 --> 00:34:16,400
Speaker 6: lab to the market faster than any product in technology history.

753
00:34:16,680 --> 00:34:19,560
Speaker 6: So they're absolutely bloated and full of crap. Right, So

754
00:34:19,680 --> 00:34:22,719
Speaker 6: these things are wildly inefficient. There's all kinds of other

755
00:34:22,760 --> 00:34:25,319
Speaker 6: ways to do the same sorts of things, one of

756
00:34:25,360 --> 00:34:27,480
Speaker 6: which is distillations. So what you're really seeing is a

757
00:34:27,560 --> 00:34:30,879
Speaker 6: kind of an accident of history that we've came down.

758
00:34:30,960 --> 00:34:33,839
Speaker 6: The US came down this path that led directly out

759
00:34:33,840 --> 00:34:36,880
Speaker 6: of the original transformer paper in twenty seventeen, and the

760
00:34:36,920 --> 00:34:38,319
Speaker 6: Chinese have said, yeah, we're not going to be able

761
00:34:38,320 --> 00:34:39,200
Speaker 6: to do that for a bunch.

762
00:34:39,000 --> 00:34:41,200
Speaker 5: Of different reasons. But we don't have to do.

763
00:34:41,120 --> 00:34:43,520
Speaker 6: That because I can take this approach of distillation, which

764
00:34:43,600 --> 00:34:44,640
Speaker 6: lets us get you.

765
00:34:44,719 --> 00:34:46,320
Speaker 5: If you look at Kimmy, this sort.

766
00:34:46,160 --> 00:34:49,520
Speaker 6: Of relatively recent open source these things are actually really

767
00:34:49,520 --> 00:34:52,479
Speaker 6: effective in benchmark very well, and it's not surprising because

768
00:34:52,480 --> 00:34:54,440
Speaker 6: they've been trained by really good trainers, which.

769
00:34:54,280 --> 00:34:55,680
Speaker 5: Is to say some of the other models that are

770
00:34:55,719 --> 00:34:56,200
Speaker 5: out there.

771
00:34:56,640 --> 00:34:59,520
Speaker 6: But these are about efficiency games, which should then ask

772
00:34:59,600 --> 00:35:02,279
Speaker 6: the question is whoa wait a minute, if there's all

773
00:35:02,280 --> 00:35:05,200
Speaker 6: these efficiency gains ahead from training, and training is seventy

774
00:35:05,200 --> 00:35:08,160
Speaker 6: percent of the workload on data centers? Hang on a second,

775
00:35:08,239 --> 00:35:11,640
Speaker 6: aren't we completely misforecasting the likely future the arc of

776
00:35:11,680 --> 00:35:14,640
Speaker 6: demand for compute And the answer is yes. And this

777
00:35:14,760 --> 00:35:17,840
Speaker 6: is rather than looking at it as an example of

778
00:35:17,840 --> 00:35:20,120
Speaker 6: why China is doing something better for worse, another way

779
00:35:20,120 --> 00:35:23,320
Speaker 6: of looking at is saying, just just refuted the approach

780
00:35:23,360 --> 00:35:26,239
Speaker 6: that we're taking to training altogether, because it shows how

781
00:35:26,280 --> 00:35:30,000
Speaker 6: blowdd and inefficient the approach we're taking is, and yet

782
00:35:30,040 --> 00:35:32,839
Speaker 6: we're projecting on that basis what future data center needs are.

783
00:35:33,280 --> 00:35:35,520
Speaker 2: Part of the question, it seems to me, and this

784
00:35:35,560 --> 00:35:38,360
Speaker 2: is where it gets a little bit philosophical, is what

785
00:35:38,520 --> 00:35:42,440
Speaker 2: do these AI companies think they're building? Because one theory

786
00:35:42,520 --> 00:35:46,200
Speaker 2: is like, well, maybe they're building business tools, right, maybe

787
00:35:46,239 --> 00:35:48,879
Speaker 2: they're building business tools of various sorts. And if they're

788
00:35:48,920 --> 00:35:52,279
Speaker 2: building business tools of various sorts, that implies the possibility

789
00:35:52,280 --> 00:35:55,200
Speaker 2: that eventually they get good enough. This does the job right,

790
00:35:55,320 --> 00:35:59,000
Speaker 2: This makes it easier for this website. You can use

791
00:35:59,000 --> 00:36:02,760
Speaker 2: an agent to book your travel, and the technology works,

792
00:36:02,760 --> 00:36:04,480
Speaker 2: and we don't have to keep building it because we

793
00:36:04,520 --> 00:36:06,839
Speaker 2: got to the point where it works. And then there

794
00:36:06,880 --> 00:36:09,239
Speaker 2: is this other question of like, well, maybe they want

795
00:36:09,280 --> 00:36:12,520
Speaker 2: to build something called AGI or ASI that's like so

796
00:36:12,760 --> 00:36:16,040
Speaker 2: sci fi et cetera, in which case you could never

797
00:36:16,080 --> 00:36:19,040
Speaker 2: get enough, or simply having built the thing that allows

798
00:36:19,080 --> 00:36:21,480
Speaker 2: you to book your travel or book a dinner reservation

799
00:36:21,800 --> 00:36:25,320
Speaker 2: or translated text or whatever, that's not nearly enough. You

800
00:36:25,400 --> 00:36:28,040
Speaker 2: you hear different things. But what do you think the

801
00:36:28,120 --> 00:36:31,080
Speaker 2: builders at the cutting edge of these labs are going for?

802
00:36:31,280 --> 00:36:33,840
Speaker 2: Is it really the sort of sci fi building god

803
00:36:33,920 --> 00:36:37,320
Speaker 2: cliche or do they want to build profitable business tools?

804
00:36:38,680 --> 00:36:39,719
Speaker 5: So it's the first.

805
00:36:39,440 --> 00:36:41,440
Speaker 6: Thing until you challenge them, and then it's the second.

806
00:36:41,560 --> 00:36:44,960
Speaker 6: So what happens is if you have the conversation internally,

807
00:36:44,960 --> 00:36:47,520
Speaker 6: they'll say, yeah, no, no, no, we're building this really effective

808
00:36:47,560 --> 00:36:51,120
Speaker 6: productivity enhancing tools that'll be used across a host of businesses,

809
00:36:51,160 --> 00:36:52,600
Speaker 6: and these all sounds really good.

810
00:36:52,640 --> 00:36:54,840
Speaker 5: But then when you walk through some of the math.

811
00:36:54,680 --> 00:36:58,160
Speaker 6: In terms of justifying the ROI on the spend, all

812
00:36:58,200 --> 00:37:00,000
Speaker 6: of a sudden, then it turns into what I call faith.

813
00:37:00,280 --> 00:37:02,960
Speaker 5: Argumentation about AGI, and they.

814
00:37:02,800 --> 00:37:06,080
Speaker 6: Say it's like the greatest call option ever, Like what

815
00:37:06,120 --> 00:37:08,120
Speaker 6: would you pay for a call option that could get

816
00:37:08,160 --> 00:37:10,520
Speaker 6: you anything, and it's like, well, wait a minute, this

817
00:37:10,560 --> 00:37:13,000
Speaker 6: isn't a way of justifying any particular expenditure.

818
00:37:13,040 --> 00:37:14,520
Speaker 5: This is just faith based argumentation.

819
00:37:14,680 --> 00:37:17,960
Speaker 6: We're saying, you know, with the uber call option for anything,

820
00:37:18,000 --> 00:37:19,600
Speaker 6: you should be willing to pay anything for it. And

821
00:37:19,640 --> 00:37:22,520
Speaker 6: obviously that that kind of justification doesn't get you anywhere.

822
00:37:22,520 --> 00:37:26,560
Speaker 6: So in house they'll arm wave a lot about these

823
00:37:26,560 --> 00:37:27,840
Speaker 6: different models that will emerge.

824
00:37:27,840 --> 00:37:28,319
Speaker 5: Who knows.

825
00:37:28,360 --> 00:37:29,839
Speaker 6: I had someone at inn Vidia tell me the other

826
00:37:29,920 --> 00:37:31,760
Speaker 6: day that we really are just waiting for the uber

827
00:37:31,800 --> 00:37:33,440
Speaker 6: of ai to come along and show.

828
00:37:33,320 --> 00:37:37,040
Speaker 5: Us the future. And I'm like, okay, so that's it's

829
00:37:37,080 --> 00:37:38,120
Speaker 5: not an answer, right.

830
00:37:38,239 --> 00:37:42,440
Speaker 2: So because in theory, if you're building a business productivity tool,

831
00:37:42,840 --> 00:37:46,680
Speaker 2: then eventually you could solve your unit economics problem. Right,

832
00:37:46,760 --> 00:37:49,080
Speaker 2: If you're just trying to build a really great business opportunity,

833
00:37:49,120 --> 00:37:50,360
Speaker 2: then as simply you know what, we don't have to

834
00:37:50,400 --> 00:37:53,200
Speaker 2: build anymore. It works, and then the cash flow just

835
00:37:53,200 --> 00:37:57,200
Speaker 2: starts pouring in and the cost per token goes down can.

836
00:37:57,040 --> 00:37:59,280
Speaker 6: And there's a bunch of that already happening. It's really interesting.

837
00:37:59,400 --> 00:38:02,440
Speaker 6: But what's incre thing happening is the problems they're solving

838
00:38:02,440 --> 00:38:05,600
Speaker 6: are really mundane, and so it's things like I'm trying

839
00:38:05,600 --> 00:38:07,960
Speaker 6: to onboard a bunch of new suppliers right now that

840
00:38:08,000 --> 00:38:10,480
Speaker 6: people have weird zip codes and they sometimes don't match up.

841
00:38:10,520 --> 00:38:12,520
Speaker 6: I have a dude in the back who fixes that.

842
00:38:12,840 --> 00:38:14,560
Speaker 6: I'd rather have someone who could do it faster so

843
00:38:14,640 --> 00:38:17,360
Speaker 6: they could onboard a lot more suppliers. Oh, it turns

844
00:38:17,440 --> 00:38:19,759
Speaker 6: out these small language models are really good at that.

845
00:38:19,800 --> 00:38:23,200
Speaker 6: These micro models like IBM's granted and whatever else, But

846
00:38:23,320 --> 00:38:27,960
Speaker 6: those things require a fraction of the training, are very cheap,

847
00:38:28,120 --> 00:38:31,520
Speaker 6: are not going to justify anywhere near the economics needed

848
00:38:31,600 --> 00:38:34,439
Speaker 6: to pay for the current spend. And yet those things

849
00:38:34,480 --> 00:38:37,360
Speaker 6: are almost likely very likely the future because it'll be

850
00:38:37,400 --> 00:38:40,920
Speaker 6: profitably get token used from micro models often hosted internally

851
00:38:41,440 --> 00:38:45,040
Speaker 6: to do really mundane background tasks, not very glamorous onboarding

852
00:38:45,080 --> 00:38:49,640
Speaker 6: new suppliers, matching records, great stuff, just not really very exciting.

853
00:38:49,680 --> 00:38:52,399
Speaker 6: But large language models are amazing at it, and small

854
00:38:52,480 --> 00:38:54,719
Speaker 6: language models are amazing at it, and almost.

855
00:38:54,440 --> 00:38:58,839
Speaker 3: Free and writing songs, right, Joe, I'm actually I'm still

856
00:38:58,840 --> 00:39:02,480
Speaker 3: annoyed that AI is like getting into art and music

857
00:39:02,520 --> 00:39:05,480
Speaker 3: writing and all the fun stuff versus the stuff that

858
00:39:05,520 --> 00:39:08,680
Speaker 3: I don't want to do like folding launchy to your classic.

859
00:39:08,320 --> 00:39:10,560
Speaker 5: Example or matching customer records. Are that?

860
00:39:11,200 --> 00:39:14,120
Speaker 3: So, going back to the beginning of this conversation when

861
00:39:14,160 --> 00:39:16,640
Speaker 3: we were just talking about the scale of AI investment

862
00:39:16,719 --> 00:39:19,920
Speaker 3: and its impact on the US economy, I'm pretty sure

863
00:39:20,080 --> 00:39:23,520
Speaker 3: you are one of the ones who's described AI capex

864
00:39:23,600 --> 00:39:27,560
Speaker 3: as like a private sector stimulus program for the US economy.

865
00:39:28,040 --> 00:39:31,840
Speaker 3: What are the actual consequences, either positive or negative, of

866
00:39:31,960 --> 00:39:36,400
Speaker 3: having this massive private sector spend in the economy versus

867
00:39:36,640 --> 00:39:39,720
Speaker 3: something I guess more typical, which would be a government

868
00:39:39,760 --> 00:39:43,680
Speaker 3: stimulus or maybe growth driven by consumer spending or something

869
00:39:43,719 --> 00:39:44,000
Speaker 3: like that.

870
00:39:44,600 --> 00:39:48,000
Speaker 6: Yeah, So to an orthodox economist, the old line is like,

871
00:39:48,040 --> 00:39:49,879
Speaker 6: it really doesn't matter what we pay people to do as.

872
00:39:49,840 --> 00:39:51,840
Speaker 5: Long as we pay them, right. It's the idea of I.

873
00:39:51,840 --> 00:39:53,560
Speaker 6: Should be, I should be you should be willing to

874
00:39:53,600 --> 00:39:55,920
Speaker 6: pay people to dig holes in the ground and people.

875
00:39:55,640 --> 00:39:56,960
Speaker 5: Over there to fill the holes back in.

876
00:39:57,040 --> 00:39:59,439
Speaker 6: Again, it really doesn't matter as long as the money

877
00:39:59,440 --> 00:40:03,200
Speaker 6: he's out there circulation, right, It's just it's all just stimulus.

878
00:40:03,280 --> 00:40:06,920
Speaker 5: Right. So, to that way of thinking, it doesn't matter

879
00:40:07,320 --> 00:40:09,560
Speaker 5: because the money's all finding its way back into the economy.

880
00:40:09,600 --> 00:40:12,680
Speaker 6: But I think that's obviously hugely misleading, because in this context,

881
00:40:12,719 --> 00:40:15,760
Speaker 6: these are investments created with an expectation of a return.

882
00:40:16,239 --> 00:40:19,160
Speaker 6: If they can't, then that flows backwards into all the

883
00:40:19,239 --> 00:40:21,360
Speaker 6: entities that are built on that basis, whether it's private

884
00:40:21,400 --> 00:40:24,440
Speaker 6: credit firms and their returns, the S and P five hundred,

885
00:40:24,440 --> 00:40:26,160
Speaker 6: what is it like? Thirty five percent now is AI

886
00:40:26,239 --> 00:40:30,160
Speaker 6: related mag seven meg ten whatever? Fifty percent now the

887
00:40:30,239 --> 00:40:32,680
Speaker 6: last two years return. So this is a massive negative

888
00:40:32,719 --> 00:40:34,919
Speaker 6: wealth effect when you unwind it, not just in terms

889
00:40:34,920 --> 00:40:36,640
Speaker 6: of the direct spending, but in terms of the wealth

890
00:40:36,640 --> 00:40:39,480
Speaker 6: effect with respect to what people's holdings are. So this

891
00:40:39,560 --> 00:40:41,120
Speaker 6: is not as simple as saying this has just been

892
00:40:41,160 --> 00:40:42,440
Speaker 6: a wonderful stimulus program.

893
00:40:42,560 --> 00:40:44,880
Speaker 5: We're paying people to dig holes and filling them back in. Again,

894
00:40:45,320 --> 00:40:45,719
Speaker 5: this is.

895
00:40:45,680 --> 00:40:48,520
Speaker 6: A wasting asset on something that's likely to be produced

896
00:40:48,560 --> 00:40:51,279
Speaker 6: in quantities that we can never earn an economic return from,

897
00:40:51,360 --> 00:40:55,439
Speaker 6: in part because of wildly flawed assumptions and projections about

898
00:40:55,440 --> 00:40:58,240
Speaker 6: the future of demand for those units. And so that's

899
00:40:58,280 --> 00:41:00,560
Speaker 6: that's the deep structural problem, and can get into this

900
00:41:00,600 --> 00:41:03,320
Speaker 6: whole question of like, well it was just private equity

901
00:41:03,360 --> 00:41:06,560
Speaker 6: guys get hurt, you know, cares Screw those guys, right,

902
00:41:07,040 --> 00:41:08,879
Speaker 6: And it's not, of course, because as we just talked

903
00:41:08,880 --> 00:41:10,760
Speaker 6: about it, it's it's in equity funds.

904
00:41:10,520 --> 00:41:12,040
Speaker 2: It's firefighters and teachers money.

905
00:41:12,120 --> 00:41:14,720
Speaker 6: Yeah, and it's in reeds now look at the larger

906
00:41:14,760 --> 00:41:16,840
Speaker 6: holdings and reads now increasingly our data centers.

907
00:41:16,920 --> 00:41:17,759
Speaker 5: Yeah. And it's even in.

908
00:41:17,760 --> 00:41:20,000
Speaker 6: Sort of sneaky backdoor ways like we're seeing increasing I

909
00:41:20,000 --> 00:41:21,880
Speaker 6: don't if you guys are familiar with these new interval funds.

910
00:41:21,880 --> 00:41:23,520
Speaker 5: They're appearing there all over.

911
00:41:23,400 --> 00:41:26,080
Speaker 2: Now, Paul Kadrowski, we could I have a million more

912
00:41:26,160 --> 00:41:28,359
Speaker 2: questions you could ask you, But much like the race

913
00:41:28,360 --> 00:41:31,160
Speaker 2: towards a GI itself, that would imply that we'll ever

914
00:41:31,320 --> 00:41:34,200
Speaker 2: actually get to the end of this conversation. So how

915
00:41:34,239 --> 00:41:37,239
Speaker 2: about we wrap here and then just plan on, you know,

916
00:41:37,360 --> 00:41:41,160
Speaker 2: revisiting the com six months, maybe three years. We just

917
00:41:41,239 --> 00:41:44,000
Speaker 2: keep revisiting down the line where we are in the cycle.

918
00:41:44,080 --> 00:41:45,960
Speaker 5: As long as we haven't been turned into paper clips.

919
00:41:45,960 --> 00:41:46,279
Speaker 1: I'm good.

920
00:41:46,880 --> 00:41:49,920
Speaker 2: Yeah, that's the no one talks about the nightmare. I

921
00:41:49,960 --> 00:41:52,680
Speaker 2: feel like that was a no one talks about the

922
00:41:52,719 --> 00:41:56,480
Speaker 2: old school paper clip maximizer stuff. Everyone's onto more esoteric fears.

923
00:41:56,520 --> 00:41:58,319
Speaker 5: I know people have moved on. We need to worry.

924
00:41:58,480 --> 00:42:01,560
Speaker 3: Does anyone wait, did anyone ever try to securitize Clippy?

925
00:42:01,680 --> 00:42:03,200
Speaker 5: They didn't, right, I don't think so.

926
00:42:03,440 --> 00:42:05,920
Speaker 2: No, thanks Paul.

927
00:42:06,040 --> 00:42:07,720
Speaker 6: Hey, thanks guys.

928
00:42:19,440 --> 00:42:21,319
Speaker 2: Paul's so good. That's a lot of fun. He's so good.

929
00:42:21,400 --> 00:42:24,440
Speaker 3: Here's my highest form of praise for an odd thought's guest.

930
00:42:24,640 --> 00:42:26,880
Speaker 3: I am going to go back and read that transcript

931
00:42:26,920 --> 00:42:27,839
Speaker 3: from beginning to end.

932
00:42:27,880 --> 00:42:30,600
Speaker 2: It is a very good that is a very good

933
00:42:30,840 --> 00:42:32,920
Speaker 2: practice to do. You're not going to listen to it.

934
00:42:33,800 --> 00:42:34,520
Speaker 5: I'm going to read it.

935
00:42:34,760 --> 00:42:36,360
Speaker 2: Yeah, I can read it. I can't listen to it.

936
00:42:36,520 --> 00:42:37,520
Speaker 3: I just listened to it.

937
00:42:37,600 --> 00:42:39,200
Speaker 2: I can need to read it. I can't listen to

938
00:42:39,280 --> 00:42:41,640
Speaker 2: our episodes. No, I just you know, I think there's

939
00:42:41,680 --> 00:42:43,480
Speaker 2: a lot, there's a lot more to do on all

940
00:42:43,520 --> 00:42:47,000
Speaker 2: this topic, but the financing in particular and some of

941
00:42:47,040 --> 00:42:52,239
Speaker 2: these arrangements. It's just incredible how the speed with which

942
00:42:52,719 --> 00:42:55,680
Speaker 2: I guess I would say the financing has gotten interesting.

943
00:42:55,840 --> 00:42:57,360
Speaker 2: Do you know what I'm saying that? I think like

944
00:42:57,400 --> 00:43:00,880
Speaker 2: a data center project ten years ago, Microsoft AWS thing

945
00:43:01,160 --> 00:43:05,120
Speaker 2: just seemed like a fairly straightforward is probably more complicated

946
00:43:05,160 --> 00:43:07,880
Speaker 2: than I appreciate at the time, but basically straightforward. We

947
00:43:07,920 --> 00:43:09,600
Speaker 2: make this money and part of it is going to

948
00:43:09,600 --> 00:43:13,160
Speaker 2: go to building more data centers to you know, serve

949
00:43:14,000 --> 00:43:16,759
Speaker 2: you know, Amazon Prime Streaming or whatever it is, or

950
00:43:16,800 --> 00:43:19,080
Speaker 2: some client thing or whatever. And then the degree of

951
00:43:19,120 --> 00:43:23,960
Speaker 2: complexity with these SPVs and rollover risk and depreciation schedules

952
00:43:24,000 --> 00:43:27,480
Speaker 2: and changing of who it's gotten very interesting, very fast.

953
00:43:27,760 --> 00:43:31,480
Speaker 3: Life Uh finds a way life finds. Yeah, that was

954
00:43:31,520 --> 00:43:35,560
Speaker 3: my terrible, terrible impression. I think that's absolutely right. One

955
00:43:35,560 --> 00:43:37,440
Speaker 3: thing I would say is the fact that a lot

956
00:43:37,480 --> 00:43:41,040
Speaker 3: of these big, supposedly cash rich companies are doing this

957
00:43:41,120 --> 00:43:44,399
Speaker 3: through SPVs that effectively preserve their balance sheet and their

958
00:43:44,440 --> 00:43:46,160
Speaker 3: cash flow so they can do something else with it.

959
00:43:46,239 --> 00:43:49,960
Speaker 3: I mean a lot of companies use SPVs. Sure, yeah,

960
00:43:50,480 --> 00:43:54,560
Speaker 3: But I do think it says something about the scale, yes, right,

961
00:43:54,640 --> 00:43:57,399
Speaker 3: Like there's a scale problem here where if all you're

962
00:43:57,480 --> 00:44:01,120
Speaker 3: spending was appearing on balance sheet investment might think very

963
00:44:01,200 --> 00:44:04,399
Speaker 3: very differently about your company. And then the other thing

964
00:44:04,400 --> 00:44:07,600
Speaker 3: I would say is I still think the comparing contrast

965
00:44:07,640 --> 00:44:11,080
Speaker 3: between the US and China and their approaches to AI.

966
00:44:11,800 --> 00:44:13,480
Speaker 3: You know, both of them, I think would agree that

967
00:44:13,520 --> 00:44:16,160
Speaker 3: this is an existential problem of some sort or an

968
00:44:16,200 --> 00:44:21,839
Speaker 3: existential competition. But they're following very different paths, and it

969
00:44:21,880 --> 00:44:24,279
Speaker 3: does seem to me like the arc of history kind

970
00:44:24,320 --> 00:44:26,960
Speaker 3: of leans towards stuff becoming cheaper.

971
00:44:28,160 --> 00:44:30,520
Speaker 2: The artifactory bend towards China.

972
00:44:31,400 --> 00:44:34,200
Speaker 3: Well that's that too, but it bends towards you know,

973
00:44:34,560 --> 00:44:36,960
Speaker 3: people generally want the cheaper thing, and they want the

974
00:44:36,960 --> 00:44:40,520
Speaker 3: thing that's like available now, and China seems to be

975
00:44:40,640 --> 00:44:41,319
Speaker 3: going for that.

976
00:44:41,440 --> 00:44:43,960
Speaker 2: The counter argument is that if you're going to use

977
00:44:44,000 --> 00:44:47,640
Speaker 2: an open source model for some purposes, you have to

978
00:44:47,640 --> 00:44:50,200
Speaker 2: supply your own electricity, right, you have to supply your

979
00:44:50,239 --> 00:44:52,279
Speaker 2: own inference. You've got to host on your service, like,

980
00:44:52,280 --> 00:44:56,239
Speaker 2: you still run into some constraints, and so rather than

981
00:44:56,320 --> 00:45:00,160
Speaker 2: having it beyond whatever whoever else is data center, you

982
00:45:00,160 --> 00:45:01,800
Speaker 2: gotta find a way to run it yourself.

983
00:45:01,880 --> 00:45:05,800
Speaker 3: Yeah, okay, but China has a leg an electricity.

984
00:45:05,080 --> 00:45:07,319
Speaker 2: Which was the point that Jensen Wong made. I mean,

985
00:45:07,360 --> 00:45:09,880
Speaker 2: part of the reason, like there's so much talk about

986
00:45:09,880 --> 00:45:13,400
Speaker 2: this these days right now, is that the industry insiders

987
00:45:13,400 --> 00:45:15,319
Speaker 2: are saying a bunch of weird things. Paul mentioned the

988
00:45:15,320 --> 00:45:18,120
Speaker 2: Sarah Friar comment yea, and she she sort of had

989
00:45:18,160 --> 00:45:20,640
Speaker 2: to walk back, but then she said there was the

990
00:45:20,680 --> 00:45:22,680
Speaker 2: Sam Altman thing where he was asked how are you

991
00:45:22,719 --> 00:45:24,040
Speaker 2: going to pay for all this? And he said, look,

992
00:45:24,040 --> 00:45:26,560
Speaker 2: you want to sell your shares or not, which is

993
00:45:26,600 --> 00:45:28,000
Speaker 2: like the interviewer probably thought he.

994
00:45:27,960 --> 00:45:29,160
Speaker 3: Was little defensive.

995
00:45:29,360 --> 00:45:32,880
Speaker 2: Obviously, Jensen Wong talking at a recently about how China

996
00:45:32,960 --> 00:45:35,480
Speaker 2: was going to win. Maybe he was saying that because

997
00:45:35,960 --> 00:45:38,759
Speaker 2: he wanted to catalyze more action on solving some of

998
00:45:38,800 --> 00:45:41,600
Speaker 2: the electricity problems in the US. But you know, the

999
00:45:41,800 --> 00:45:44,920
Speaker 2: very people at the center of this are saying things

1000
00:45:45,080 --> 00:45:48,960
Speaker 2: right now that you know. What's interesting too, is you

1001
00:45:49,000 --> 00:45:52,279
Speaker 2: know this bullwhip phenomenon everyone as Paul described it, he

1002
00:45:52,280 --> 00:45:54,600
Speaker 2: didn't use the word bullwhip, but when everyone is trying

1003
00:45:54,600 --> 00:45:56,640
Speaker 2: to get their hands on the same gear, you gotta

1004
00:45:56,680 --> 00:45:59,160
Speaker 2: wonder how sustaint what's the other side of a bullwep

1005
00:45:59,200 --> 00:46:01,360
Speaker 2: could look like? We just got to do more episodes

1006
00:46:01,360 --> 00:46:01,520
Speaker 2: on this.

1007
00:46:01,719 --> 00:46:03,680
Speaker 3: Yeah, we have to. Shall we leave it there for now?

1008
00:46:03,760 --> 00:46:04,759
Speaker 2: Let's leave it there all right?

1009
00:46:04,800 --> 00:46:07,439
Speaker 3: This has been another episode of the Audthots podcast. I'm

1010
00:46:07,440 --> 00:46:10,359
Speaker 3: Tracy Alloway. You can follow me at Tracy Alloway and.

1011
00:46:10,320 --> 00:46:13,000
Speaker 2: I'm Jill Wisenthal. You can follow me at The Stalwart.

1012
00:46:13,200 --> 00:46:16,440
Speaker 2: Check out Paul Kadrowski's writing at Paul Kadrowski dot com,

1013
00:46:16,480 --> 00:46:19,719
Speaker 2: follow our producers Carmen Rodriguez at Carman Arman, dash Ol

1014
00:46:19,719 --> 00:46:22,759
Speaker 2: Bennett at dashbod and Kilbrooks at Kilbrooks. And for more

1015
00:46:22,800 --> 00:46:25,279
Speaker 2: odd Lots content, go to Bloomberg dot com slash odd

1016
00:46:25,360 --> 00:46:27,840
Speaker 2: Lots with the daily newsletter and all of our episodes,

1017
00:46:28,040 --> 00:46:29,960
Speaker 2: and you can shout about all of these topics twenty

1018
00:46:29,960 --> 00:46:33,160
Speaker 2: four to seven in our discord Discord dot gg slash

1019
00:46:33,160 --> 00:46:33,680
Speaker 2: od Lots.

1020
00:46:33,880 --> 00:46:36,000
Speaker 3: And if you enjoy odd Lots, if you like it

1021
00:46:36,080 --> 00:46:40,160
Speaker 3: when we talk about the AI private credit leverage, subprime

1022
00:46:40,320 --> 00:46:43,960
Speaker 3: economy nexus, then please leave us a positive review on

1023
00:46:44,000 --> 00:46:46,960
Speaker 3: your favorite podcast platform. And remember, if you are a

1024
00:46:47,000 --> 00:46:50,040
Speaker 3: Bloomberg subscriber, you can listen to all of our episodes

1025
00:46:50,120 --> 00:46:51,400
Speaker 3: absolutely ad free.

1026
00:46:51,760 --> 00:46:52,759
Speaker 4: All you have to do is.

1027
00:46:52,719 --> 00:46:55,440
Speaker 3: Find the Bloomberg channel on Apple Podcasts and follow the

1028
00:46:55,440 --> 00:46:56,360
Speaker 3: instructions there.

1029
00:46:56,640 --> 00:47:14,200
Speaker 4: Thanks for listening in

1030
00:47:22,600 --> 00:47:23,239
Speaker 5: In