1
00:00:03,120 --> 00:00:17,400
Speaker 1: Bloomberg Audio Studios, Podcasts, Radio News.

2
00:00:20,040 --> 00:00:23,280
Speaker 2: Hello and welcome to another episode of the All Thoughts podcast.

3
00:00:23,480 --> 00:00:24,880
Speaker 2: I'm Tracy Alloway.

4
00:00:24,560 --> 00:00:25,759
Speaker 3: And I'm Joe Wisenthal.

5
00:00:26,600 --> 00:00:29,640
Speaker 2: Joe, AI is so hot right now, in the immortal

6
00:00:29,720 --> 00:00:32,440
Speaker 2: words of Mugatu, AI is so hot.

7
00:00:33,159 --> 00:00:33,479
Speaker 4: It is.

8
00:00:33,800 --> 00:00:36,640
Speaker 3: Yes, it is really hot. You know, you hear something.

9
00:00:36,960 --> 00:00:38,720
Speaker 3: There's a little bit of slowing down in some of

10
00:00:38,800 --> 00:00:42,960
Speaker 3: the progress on the models, but the recent in video

11
00:00:43,080 --> 00:00:46,880
Speaker 3: results speak for themselves. There is nothing that I've seen

12
00:00:47,240 --> 00:00:50,680
Speaker 3: yet that would suggest that this macro trend, at least

13
00:00:50,960 --> 00:00:53,120
Speaker 3: as an investment trend, and I'm not talking about stocks

14
00:00:53,159 --> 00:00:56,800
Speaker 3: per se, is anywhere close to quote slowing down.

15
00:00:57,160 --> 00:00:59,760
Speaker 2: Yeah, And the interesting thing is we seem to be

16
00:01:00,320 --> 00:01:03,040
Speaker 2: more and more players, some new types of players that

17
00:01:03,080 --> 00:01:05,679
Speaker 2: are getting into the space. So, you know, we have

18
00:01:05,840 --> 00:01:09,319
Speaker 2: AI funds kind of launching left and right. And one

19
00:01:09,319 --> 00:01:12,440
Speaker 2: of the newest players is a hedge fund called Magnetar

20
00:01:12,800 --> 00:01:16,400
Speaker 2: and I know them like primarily for credit stuff. I

21
00:01:16,440 --> 00:01:19,759
Speaker 2: think they were big in redcap trades for a while. Yeah,

22
00:01:19,800 --> 00:01:22,680
Speaker 2: and now they're launching an AI fund, a VC fund,

23
00:01:22,720 --> 00:01:24,839
Speaker 2: which is kind of unusual for this type of hedge

24
00:01:24,840 --> 00:01:25,240
Speaker 2: fund to.

25
00:01:25,160 --> 00:01:28,399
Speaker 3: Do totally, I mean I've heard of magnetar for a

26
00:01:28,440 --> 00:01:32,720
Speaker 3: long time, obviously, going back to the early twenty tens

27
00:01:33,040 --> 00:01:37,840
Speaker 3: at least, And look, I'm not surprised that various investors

28
00:01:37,920 --> 00:01:41,720
Speaker 3: are looking for what is their distinct way into this space?

29
00:01:41,800 --> 00:01:44,959
Speaker 3: And of course, look, we've done interviews with vcs of

30
00:01:45,040 --> 00:01:48,280
Speaker 3: various nature and positions in the past, and so I

31
00:01:48,320 --> 00:01:50,840
Speaker 3: guess you know, there's sort of two questions to my

32
00:01:51,000 --> 00:01:54,840
Speaker 3: mind anytime we're gonna be talking to someone investing in

33
00:01:54,920 --> 00:01:58,680
Speaker 3: early stage or any stage of AI, which is obviously

34
00:01:59,160 --> 00:02:01,320
Speaker 3: what is the thesis is what's going to win out

35
00:02:01,400 --> 00:02:04,880
Speaker 3: where we'll value a crew. But then from an investor perspective,

36
00:02:05,800 --> 00:02:09,600
Speaker 3: given so many entrants into this space, specifically whether on

37
00:02:09,639 --> 00:02:12,120
Speaker 3: the public equity side, whether on the private side, whether

38
00:02:12,160 --> 00:02:14,880
Speaker 3: on the VC side or early stage, late stage, what

39
00:02:15,040 --> 00:02:17,480
Speaker 3: do they, as a fund or an investor bring to

40
00:02:17,520 --> 00:02:20,480
Speaker 3: the table or will be able to see that the

41
00:02:20,639 --> 00:02:24,400
Speaker 3: other billions of dollars competing for AI profits do not see.

42
00:02:24,680 --> 00:02:27,040
Speaker 2: I have a slightly different question, which is for these

43
00:02:27,080 --> 00:02:30,520
Speaker 2: types of investors, like how much is it about how

44
00:02:30,560 --> 00:02:34,200
Speaker 2: good the technology is that they're investing in versus how

45
00:02:34,440 --> 00:02:36,960
Speaker 2: much is it about getting in the right position in

46
00:02:37,000 --> 00:02:37,919
Speaker 2: the capital stack.

47
00:02:38,040 --> 00:02:39,080
Speaker 3: So that's a great question.

48
00:02:39,200 --> 00:02:41,440
Speaker 2: I think it's going to be really interesting to talk

49
00:02:41,480 --> 00:02:45,720
Speaker 2: to someone who's coming from this perspective. And without further ado,

50
00:02:46,080 --> 00:02:48,160
Speaker 2: we have the perfect guest we're going to be speaking with,

51
00:02:48,280 --> 00:02:52,160
Speaker 2: Jim Prosco. He is a partner and senior portfolio manager

52
00:02:52,520 --> 00:02:57,400
Speaker 2: on Magnetar's Alternative credit and fixed income team. Jim, Welcome

53
00:02:57,440 --> 00:02:57,959
Speaker 2: to the show.

54
00:02:58,400 --> 00:02:59,480
Speaker 5: Thank you, great to be here.

55
00:02:59,800 --> 00:03:03,400
Speaker 2: So how does someone on a hedge funds fixed income

56
00:03:03,440 --> 00:03:04,520
Speaker 2: team get into AI.

57
00:03:05,800 --> 00:03:09,679
Speaker 4: Well, we have a long history of investments in private companies,

58
00:03:10,080 --> 00:03:13,040
Speaker 4: really dating back to an increased focus after the Financial

59
00:03:13,080 --> 00:03:16,400
Speaker 4: Crisis when spreads and yields got tighter and the private

60
00:03:16,440 --> 00:03:21,360
Speaker 4: markets seem more interesting. And we've often partnered with platforms

61
00:03:21,880 --> 00:03:24,760
Speaker 4: where we thought we could grow the platform and generate

62
00:03:24,840 --> 00:03:28,040
Speaker 4: an interesting asset, either a pool of cash flowing assets,

63
00:03:28,200 --> 00:03:31,320
Speaker 4: or help grow the company and participate in that growth

64
00:03:32,040 --> 00:03:36,760
Speaker 4: and support them through financing and other things like we

65
00:03:36,760 --> 00:03:39,760
Speaker 4: can support them through helping them with hiring or accounting

66
00:03:39,960 --> 00:03:42,360
Speaker 4: or other systems they need, and just to help them

67
00:03:42,400 --> 00:03:44,760
Speaker 4: grow generally. And so you know, I've been doing that

68
00:03:44,880 --> 00:03:47,040
Speaker 4: a long time and we've been a number of areas

69
00:03:47,320 --> 00:03:51,640
Speaker 4: like auto lending in Ireland, and then we've moved into

70
00:03:51,920 --> 00:03:55,520
Speaker 4: various fintech companies. We were one of the first institutional

71
00:03:55,520 --> 00:04:00,040
Speaker 4: investors in open Door before they went public. We're supporting

72
00:04:00,080 --> 00:04:04,040
Speaker 4: and investing in a very interesting fintech that is financing

73
00:04:04,080 --> 00:04:07,760
Speaker 4: restaurants right now, and so we felt we had experience

74
00:04:07,840 --> 00:04:11,880
Speaker 4: in that space, and then that sort of overlapped with

75
00:04:12,680 --> 00:04:15,920
Speaker 4: our relationship and our investment in core Weave, where we

76
00:04:15,920 --> 00:04:19,080
Speaker 4: were the first institutional investor in core Weave in twenty

77
00:04:19,279 --> 00:04:23,640
Speaker 4: twenty one. So we're very early in the trend of

78
00:04:23,680 --> 00:04:27,680
Speaker 4: putting capital into the AI infrastructure space and that's just

79
00:04:27,760 --> 00:04:30,039
Speaker 4: sort of grown as this whole market has grown to

80
00:04:30,160 --> 00:04:34,279
Speaker 4: encompass literally everything. Now, you know, we continue to look

81
00:04:34,320 --> 00:04:37,120
Speaker 4: for smart ways to invest, and you know, one of

82
00:04:37,120 --> 00:04:40,920
Speaker 4: those ways we felt was what can we provide that's

83
00:04:40,920 --> 00:04:43,320
Speaker 4: a value And one of the things we can provide

84
00:04:44,000 --> 00:04:47,080
Speaker 4: besides the general help we can give a growth stage

85
00:04:47,080 --> 00:04:52,280
Speaker 4: company is compute because that is the scarce resource right now,

86
00:04:52,320 --> 00:04:54,599
Speaker 4: and that's where all the capital is going to the

87
00:04:54,680 --> 00:04:58,480
Speaker 4: various parts of the value chain to deliver compute, and

88
00:04:58,560 --> 00:05:01,440
Speaker 4: so there's a competition to get compute, and if you're

89
00:05:01,440 --> 00:05:06,120
Speaker 4: a smaller company with limited capital or limited access to capital,

90
00:05:06,600 --> 00:05:08,839
Speaker 4: it can be difficult to get that, and so that

91
00:05:09,000 --> 00:05:11,039
Speaker 4: was sort of the value proposition we thought we could

92
00:05:11,040 --> 00:05:11,560
Speaker 4: bring to bear.

93
00:05:12,040 --> 00:05:14,680
Speaker 2: Joe, I have this vision in my head of vcs,

94
00:05:14,760 --> 00:05:18,120
Speaker 2: like going into startups bearing baskets full of chips.

95
00:05:18,279 --> 00:05:19,960
Speaker 3: Ah yeah, instead of just saying.

96
00:05:19,800 --> 00:05:22,680
Speaker 2: That, like our pitch is the relationship and the coaching as.

97
00:05:22,839 --> 00:05:25,720
Speaker 3: We have access to the chips or the energy plus chips.

98
00:05:26,160 --> 00:05:29,839
Speaker 3: Just for point of clarification, listeners should know we've talked

99
00:05:29,839 --> 00:05:33,280
Speaker 3: to Core. We've at least twice on the show, and

100
00:05:33,360 --> 00:05:38,240
Speaker 3: it feels like in the AI space specifically, this is

101
00:05:38,279 --> 00:05:40,720
Speaker 3: one of those names that's a very big deal, but

102
00:05:40,880 --> 00:05:43,479
Speaker 3: not many people don't know it the way they know

103
00:05:43,640 --> 00:05:47,919
Speaker 3: sayan in Nvidia at the very back end or Chatgypt

104
00:05:48,160 --> 00:05:50,920
Speaker 3: at the very front end, but they build a lot

105
00:05:50,960 --> 00:05:54,480
Speaker 3: of the data centers that are filled with Nvidia chips.

106
00:05:54,680 --> 00:05:56,920
Speaker 3: I want to get more into the business model there

107
00:05:56,960 --> 00:05:59,000
Speaker 3: because I have a lot of questions in the business

108
00:05:59,000 --> 00:06:02,240
Speaker 3: of selling compute, etc. But talk a little bit more

109
00:06:02,520 --> 00:06:06,120
Speaker 3: about you said your experience in the private side is

110
00:06:06,160 --> 00:06:10,480
Speaker 3: like this expertise with platforms per se. And when I

111
00:06:10,520 --> 00:06:12,880
Speaker 3: think of platforms, I think of companies that can acquire

112
00:06:12,920 --> 00:06:15,400
Speaker 3: lots of other companies or a lot can be built

113
00:06:15,480 --> 00:06:18,760
Speaker 3: onto them. Talk to us about how the platform specific

114
00:06:18,920 --> 00:06:23,760
Speaker 3: expertise informs you're thinking with a core weave or any

115
00:06:23,800 --> 00:06:26,320
Speaker 3: other AI investment that you're making now.

116
00:06:27,480 --> 00:06:33,640
Speaker 4: So we've tried to put capital into companies that are

117
00:06:34,200 --> 00:06:37,240
Speaker 4: trying to build their business in a particular space, and

118
00:06:37,880 --> 00:06:40,760
Speaker 4: oftentimes that could be a space where they generate a

119
00:06:40,800 --> 00:06:45,160
Speaker 4: cash flowing asset, like in the auto loan example, in

120
00:06:45,320 --> 00:06:48,880
Speaker 4: the open door example, they were acquiring real estate, which

121
00:06:48,920 --> 00:06:53,240
Speaker 4: was a hard asset. In that restaurant fintech example, they're

122
00:06:53,279 --> 00:06:57,560
Speaker 4: acquiring restaurant credit. And so we've tried to support businesses

123
00:06:57,680 --> 00:07:01,560
Speaker 4: that had some kind of asset or flow and work

124
00:07:01,600 --> 00:07:04,039
Speaker 4: with them on a number of ways that we can

125
00:07:04,120 --> 00:07:06,400
Speaker 4: add value. I think first and foremost is all these

126
00:07:06,440 --> 00:07:09,359
Speaker 4: growth stage companies need financing, and I think we have

127
00:07:09,440 --> 00:07:13,440
Speaker 4: great expertise from debt to equity, private to public, and

128
00:07:13,520 --> 00:07:16,560
Speaker 4: we can be innovative in trying to bring you the best,

129
00:07:17,440 --> 00:07:21,280
Speaker 4: most appropriate, lowest cost capital to these growth stage companies.

130
00:07:21,320 --> 00:07:22,800
Speaker 4: And like I said, as well as.

131
00:07:22,920 --> 00:07:25,880
Speaker 3: So just to be clear, just to understand in this context,

132
00:07:26,280 --> 00:07:30,239
Speaker 3: what makes AI distinct, say from other waves of tech

133
00:07:30,760 --> 00:07:34,040
Speaker 3: or what makes it distinct for say a magnetar is

134
00:07:34,080 --> 00:07:38,560
Speaker 3: in part this distinct capital demand that was not perhaps

135
00:07:39,040 --> 00:07:41,000
Speaker 3: as big of a deal during the SaaS wave of

136
00:07:41,040 --> 00:07:41,840
Speaker 3: the twenty tens.

137
00:07:43,040 --> 00:07:46,080
Speaker 4: Yes, so not not only a general capital demand, but

138
00:07:46,280 --> 00:07:49,440
Speaker 4: in many cases, for many of these companies, a very

139
00:07:49,480 --> 00:07:54,040
Speaker 4: specific demand to have capital to deploy with compute, and

140
00:07:54,120 --> 00:07:58,800
Speaker 4: because they need this very specific scarce resource, helping to

141
00:07:58,840 --> 00:08:03,240
Speaker 4: deliver that resource, and in particular helping to deliver that

142
00:08:03,320 --> 00:08:05,840
Speaker 4: resource in a high quality way. Where you have a

143
00:08:05,880 --> 00:08:09,080
Speaker 4: partner like core Weave that has I think there's a

144
00:08:09,080 --> 00:08:11,840
Speaker 4: lot of evidence that they have the highest performing AI

145
00:08:11,960 --> 00:08:16,360
Speaker 4: training cluster, and so that is really valuable to these

146
00:08:16,400 --> 00:08:20,040
Speaker 4: companies that might otherwise struggle to get enough compute to

147
00:08:20,200 --> 00:08:21,360
Speaker 4: further their business model.

148
00:08:21,720 --> 00:08:24,920
Speaker 2: Speaking of Core We've I'm really curious how that conversation

149
00:08:25,160 --> 00:08:28,640
Speaker 2: actually started because this was a new and novel thing.

150
00:08:28,680 --> 00:08:31,960
Speaker 2: I don't think we had chip based loans before to

151
00:08:32,040 --> 00:08:35,319
Speaker 2: my knowledge, and I keep hearing that asset based financing

152
00:08:35,400 --> 00:08:37,680
Speaker 2: is going to be like this next big thing in

153
00:08:37,800 --> 00:08:42,560
Speaker 2: private credit or it's the last real frontier in private credit.

154
00:08:42,800 --> 00:08:45,200
Speaker 2: How did you come up with this idea this deal?

155
00:08:46,840 --> 00:08:50,680
Speaker 4: Well, acid based financing is really a classic private credit

156
00:08:50,760 --> 00:08:54,480
Speaker 4: tool and there's a number of examples. Just if you

157
00:08:54,559 --> 00:08:59,160
Speaker 4: think about my example with the Irish auto lender. If

158
00:08:59,200 --> 00:09:02,480
Speaker 4: you buy a loan for a car, so the Irish

159
00:09:02,520 --> 00:09:06,080
Speaker 4: auto lenders generating car loans and those go and you

160
00:09:06,160 --> 00:09:10,119
Speaker 4: buy them in a vehicle, you have primarily the security

161
00:09:10,120 --> 00:09:12,960
Speaker 4: of the people paying on those loans, and so you

162
00:09:13,000 --> 00:09:16,800
Speaker 4: get paid back by the cash flow of the borrowers

163
00:09:17,040 --> 00:09:19,760
Speaker 4: paying their car loans back, but there's credit risk to

164
00:09:19,800 --> 00:09:22,720
Speaker 4: that they could potentially stop paying, and in the case

165
00:09:22,720 --> 00:09:26,239
Speaker 4: where they stop paying, then you have the cars collateral.

166
00:09:26,840 --> 00:09:32,880
Speaker 4: And really that metaphor applies almost directly to GPUs, where

167
00:09:33,400 --> 00:09:37,360
Speaker 4: if you're a company delivering high performance compute like Core

168
00:09:37,440 --> 00:09:42,600
Speaker 4: we've has, you're contractually selling that compute to some counterparty

169
00:09:42,600 --> 00:09:45,040
Speaker 4: that's going to use it in their case. You know,

170
00:09:45,120 --> 00:09:48,800
Speaker 4: that's often a very large, very credit worthy hyperscaler, but

171
00:09:48,920 --> 00:09:52,480
Speaker 4: not always. There could be smaller startups that have riskier

172
00:09:52,520 --> 00:09:58,079
Speaker 4: business models, and in that case, primarily by funding the GPU,

173
00:09:58,160 --> 00:10:01,280
Speaker 4: you're getting paid back with those controls actual cash flows

174
00:10:01,360 --> 00:10:04,439
Speaker 4: on the use of the GPU. But in the case

175
00:10:04,480 --> 00:10:09,480
Speaker 4: that company fails, then as backup you have the GPU itself. Now,

176
00:10:10,240 --> 00:10:12,560
Speaker 4: the GPU isn't really like the car where you'll probably

177
00:10:12,559 --> 00:10:14,920
Speaker 4: go out and sell it, but you get the time

178
00:10:15,080 --> 00:10:17,920
Speaker 4: back on the GPU, which you can then resell to

179
00:10:18,000 --> 00:10:21,120
Speaker 4: somebody else, and being a scarce asset, you can think

180
00:10:21,120 --> 00:10:23,920
Speaker 4: about what value that would have in a future time.

181
00:10:24,840 --> 00:10:29,160
Speaker 3: One difference that I could imagine with the GPU versus

182
00:10:29,240 --> 00:10:33,120
Speaker 3: other forms of assets, say whether it's a car or

183
00:10:33,160 --> 00:10:35,680
Speaker 3: say whether it's a house, is a certain here in

184
00:10:35,760 --> 00:10:40,319
Speaker 3: twenty twenty four, still unpredictability about many things in the future.

185
00:10:40,720 --> 00:10:45,360
Speaker 3: Will in video always be the gold standard so to speak?

186
00:10:45,520 --> 00:10:47,880
Speaker 3: In AI chips maybe it looks like it, yes, but

187
00:10:47,960 --> 00:10:52,120
Speaker 3: it doesn't seem guaranteed. How fast will the current generation

188
00:10:52,240 --> 00:10:56,120
Speaker 3: of chips that are deployed degrade in value? I imagine

189
00:10:56,120 --> 00:10:59,960
Speaker 3: there are fairly predictable sort of depreciation curves for cars

190
00:11:00,280 --> 00:11:03,920
Speaker 3: that perhaps are more uncertain for chips. And then also

191
00:11:04,200 --> 00:11:09,840
Speaker 3: the uncertainty of actual deployment given permitting and challenges with

192
00:11:10,040 --> 00:11:12,560
Speaker 3: energy and the other operational things that have to do

193
00:11:12,600 --> 00:11:15,600
Speaker 3: with a new company building a data center. Talk to

194
00:11:15,679 --> 00:11:19,200
Speaker 3: us about modeling or at least thinking through some of

195
00:11:19,240 --> 00:11:23,079
Speaker 3: the uncertainties with chips specifically, Well, depending.

196
00:11:22,679 --> 00:11:26,240
Speaker 4: What stage you get involved, you have the breadth of

197
00:11:26,320 --> 00:11:30,040
Speaker 4: all those different risks potentially. So if you're investing in

198
00:11:30,160 --> 00:11:34,600
Speaker 4: high performance compute but it's a greenfield data center, then

199
00:11:34,640 --> 00:11:36,360
Speaker 4: you have to think about all those things. You have

200
00:11:36,440 --> 00:11:39,560
Speaker 4: to think about the delivering of the power. You have

201
00:11:39,600 --> 00:11:42,319
Speaker 4: to think about the timing on all the components to

202
00:11:42,360 --> 00:11:45,680
Speaker 4: get to the data center. If you're making what we've

203
00:11:45,679 --> 00:11:48,520
Speaker 4: been talking about, which is sort of a GPU based loan,

204
00:11:48,920 --> 00:11:53,760
Speaker 4: then usually that loan is based upon a running GPU

205
00:11:53,960 --> 00:11:57,320
Speaker 4: and an existing high performance compute data center, so you

206
00:11:57,360 --> 00:12:00,559
Speaker 4: don't really have to think about some of the earlier

207
00:12:00,600 --> 00:12:03,520
Speaker 4: stage issues. You more have to think about how long

208
00:12:03,640 --> 00:12:07,120
Speaker 4: is my contract, how good is my contract, What do

209
00:12:07,240 --> 00:12:10,719
Speaker 4: I think the value of renting that chip out will

210
00:12:10,760 --> 00:12:14,560
Speaker 4: be at the end of that contract. How much rent

211
00:12:14,600 --> 00:12:16,520
Speaker 4: on that chip could I get if I had to

212
00:12:16,559 --> 00:12:19,040
Speaker 4: re rent that in the middle of the contract. So

213
00:12:19,080 --> 00:12:22,400
Speaker 4: it's more near term things on actually having a functioning

214
00:12:22,480 --> 00:12:25,560
Speaker 4: GPU in the data center, But all those other things

215
00:12:25,559 --> 00:12:27,679
Speaker 4: have to be financed too, and there's going to be

216
00:12:27,880 --> 00:12:32,200
Speaker 4: innovative and large amounts of capital dedicated to financing those things.

217
00:12:32,559 --> 00:12:35,640
Speaker 2: Setting aside the financing for a second, how hard has

218
00:12:35,679 --> 00:12:41,200
Speaker 2: it been just to find physical space in data centers, well, it's.

219
00:12:41,040 --> 00:12:44,559
Speaker 4: Been extremely scarce, and a lot of that is driven

220
00:12:44,720 --> 00:12:50,160
Speaker 4: by the search for power. The data centers required for

221
00:12:50,280 --> 00:12:53,760
Speaker 4: the new AI chips are much different than the old

222
00:12:53,840 --> 00:12:57,880
Speaker 4: data center. So it isn't really cost efficient in most

223
00:12:57,920 --> 00:13:00,120
Speaker 4: cases to go and take an old data center and

224
00:13:00,160 --> 00:13:03,840
Speaker 4: try to retrofit it because the amount of power just

225
00:13:03,920 --> 00:13:07,000
Speaker 4: a loan that has to go there is you know,

226
00:13:07,360 --> 00:13:11,120
Speaker 4: transcending an order of magnitude more per rack of GPUs now,

227
00:13:11,160 --> 00:13:15,040
Speaker 4: and so that's just you just can't really retrofit that efficiently.

228
00:13:15,080 --> 00:13:17,520
Speaker 4: It's better to build your own building. And so it's

229
00:13:17,600 --> 00:13:22,760
Speaker 4: really come down to things like permitting availability of power

230
00:13:23,559 --> 00:13:26,200
Speaker 4: and time to get all your components, and you know,

231
00:13:26,240 --> 00:13:28,960
Speaker 4: all these things have their own lead time. So it

232
00:13:29,000 --> 00:13:32,479
Speaker 4: had an interesting back and forth to Brian on curing transformers.

233
00:13:32,520 --> 00:13:35,839
Speaker 4: You know, all these little you know, nuances come into

234
00:13:35,920 --> 00:13:37,920
Speaker 4: play when you have to build a data center. And

235
00:13:38,000 --> 00:13:43,280
Speaker 4: so because power is really the limiting factor most of all,

236
00:13:43,679 --> 00:13:47,200
Speaker 4: you're seeing a lot of moves towards where the power is.

237
00:13:47,760 --> 00:13:50,600
Speaker 4: And it was recently an article on Bloomberg. I think

238
00:13:50,640 --> 00:13:53,000
Speaker 4: about a company in Texas that owns a bunch of

239
00:13:53,080 --> 00:13:55,560
Speaker 4: land that's now worth forty billion dollars, right, And that's

240
00:13:55,600 --> 00:13:59,960
Speaker 4: because they're near all this renewable power. But that isn't

241
00:14:00,000 --> 00:14:03,520
Speaker 4: the only thing. It's incredibly complex to operate this high

242
00:14:03,559 --> 00:14:06,679
Speaker 4: performance compute. So then you have to think about if

243
00:14:06,720 --> 00:14:08,760
Speaker 4: I try to build my data center out there where

244
00:14:08,800 --> 00:14:12,559
Speaker 4: the power is. Can I get everything out there, including

245
00:14:13,600 --> 00:14:15,000
Speaker 4: operational expertise?

246
00:14:15,080 --> 00:14:15,240
Speaker 5: Right?

247
00:14:15,280 --> 00:14:17,360
Speaker 4: Can I staff my data center with the kind of

248
00:14:17,360 --> 00:14:20,680
Speaker 4: experts I need to run this kind of highly technical,

249
00:14:20,760 --> 00:14:24,880
Speaker 4: high performance compute. And each generation is just getting more complicated.

250
00:14:25,160 --> 00:14:28,160
Speaker 4: We're going to have liquid cooling on the next generation

251
00:14:28,240 --> 00:14:31,840
Speaker 4: of Nvidia chips, probably immersion cooling right after that. It's

252
00:14:32,000 --> 00:14:36,120
Speaker 4: very complicated, very expensive, and very difficult to scale. Much

253
00:14:36,160 --> 00:14:39,280
Speaker 4: harder to do in a large size than it is

254
00:14:39,320 --> 00:14:39,640
Speaker 4: to do in.

255
00:14:39,640 --> 00:14:40,560
Speaker 5: A small size.

256
00:14:40,800 --> 00:14:45,800
Speaker 2: Maybe Magnetar can finance a small modular nuclear reactor. No, seriously,

257
00:14:45,840 --> 00:14:49,280
Speaker 2: because if you're financing the compute and securing that on

258
00:14:49,400 --> 00:14:51,920
Speaker 2: behalf of companies that you want to invest in, you

259
00:14:51,960 --> 00:14:54,520
Speaker 2: could go one layer down finance the energy.

260
00:14:55,480 --> 00:14:57,520
Speaker 4: And we're certainly interested in that, and we have a

261
00:14:57,640 --> 00:15:00,760
Speaker 4: history in investing in energy. We have investment right now

262
00:15:01,080 --> 00:15:04,040
Speaker 4: and a developer of utility scale solar power in the

263
00:15:04,120 --> 00:15:07,720
Speaker 4: US who has least some of that solar power to

264
00:15:08,080 --> 00:15:12,160
Speaker 4: various hyperscalers. So that is certainly a space we're interested in.

265
00:15:12,640 --> 00:15:16,120
Speaker 4: I was just in Miami meeting with a company that

266
00:15:16,240 --> 00:15:19,880
Speaker 4: has a novel heat sink battery technology that they want

267
00:15:19,880 --> 00:15:22,000
Speaker 4: to deploy to data centers that they're talking to a

268
00:15:22,040 --> 00:15:26,720
Speaker 4: bunch of data center type companies about launching that product there.

269
00:15:26,800 --> 00:15:29,160
Speaker 4: So there's a ton of interesting things, and just like

270
00:15:29,240 --> 00:15:32,160
Speaker 4: every other part of this ecosystem, it's going to require

271
00:15:32,160 --> 00:15:33,400
Speaker 4: an immense amount of capital.

272
00:15:33,760 --> 00:15:37,320
Speaker 3: I guess, just since we're sidetracked on the energy component

273
00:15:37,440 --> 00:15:42,680
Speaker 3: for now while we're here novel battery technologies, there's a

274
00:15:42,720 --> 00:15:44,880
Speaker 3: lot of them out there. There's a lot of startups

275
00:15:44,880 --> 00:15:48,600
Speaker 3: that have something novel and energy, and often one of

276
00:15:48,640 --> 00:15:51,320
Speaker 3: the things that they talk about is this chicken and

277
00:15:51,360 --> 00:15:55,280
Speaker 3: egg problem where they need capital, They need sort of

278
00:15:55,360 --> 00:15:58,600
Speaker 3: financing of some sort or another to build this stuff,

279
00:15:58,840 --> 00:16:01,120
Speaker 3: but the lenders don't really want to give it until

280
00:16:01,120 --> 00:16:03,280
Speaker 3: there's demand, and no one's just going to promise to

281
00:16:03,280 --> 00:16:06,720
Speaker 3: buy it until it's shown that it can work. Can

282
00:16:06,760 --> 00:16:08,240
Speaker 3: you talk a little bit, I mean again, I know

283
00:16:08,280 --> 00:16:10,720
Speaker 3: there's a little bit off track from GPUs themselves. But

284
00:16:10,840 --> 00:16:14,840
Speaker 3: since you were talking about similar yeah, talk about the batteries.

285
00:16:15,000 --> 00:16:17,160
Speaker 3: Can you talk a little bit about that dynamic as

286
00:16:17,200 --> 00:16:19,120
Speaker 3: it affects solving the energy side of the equation?

287
00:16:19,800 --> 00:16:20,440
Speaker 5: Yeah, for sure.

288
00:16:20,480 --> 00:16:22,760
Speaker 4: And it has some overlap with the way you look

289
00:16:22,840 --> 00:16:25,440
Speaker 4: at an AI company too. You know, if you think

290
00:16:25,480 --> 00:16:29,200
Speaker 4: about the core things that we really want to look at,

291
00:16:29,760 --> 00:16:36,000
Speaker 4: it's technology team and traction. So does their technology really work?

292
00:16:36,080 --> 00:16:38,440
Speaker 4: That's first and foremost. You know, what is this product?

293
00:16:38,480 --> 00:16:42,320
Speaker 4: Does it have some kind of advantage? And then traction

294
00:16:42,800 --> 00:16:44,400
Speaker 4: like time to market.

295
00:16:44,200 --> 00:16:45,560
Speaker 5: That's super important.

296
00:16:45,840 --> 00:16:49,200
Speaker 4: I was just talking to isokon pool side and like

297
00:16:49,520 --> 00:16:52,000
Speaker 4: to him, like those are the two most important things.

298
00:16:52,200 --> 00:16:55,440
Speaker 4: Speed to product, speed to market, because it's a race,

299
00:16:55,920 --> 00:16:58,960
Speaker 4: and even if you have the greatest technology, if you

300
00:16:59,040 --> 00:17:01,560
Speaker 4: take too long, someone's going to be using something else.

301
00:17:01,600 --> 00:17:05,240
Speaker 4: And that's certainly true in the energy space where energy

302
00:17:05,320 --> 00:17:09,200
Speaker 4: is of critical importance. So I think that for these

303
00:17:09,240 --> 00:17:13,160
Speaker 4: startups on the traction side, they really need some strategic

304
00:17:13,280 --> 00:17:17,520
Speaker 4: partnerships because their cost of capital is very high.

305
00:17:17,560 --> 00:17:20,919
Speaker 3: Strategic partnership is kind of like an existing company that

306
00:17:21,040 --> 00:17:23,760
Speaker 3: has a demand. It also has a lot of cash

307
00:17:23,920 --> 00:17:26,080
Speaker 3: and could theoretically be a buyer of their.

308
00:17:25,880 --> 00:17:29,679
Speaker 4: Solution, yes, and really on the other side too, So

309
00:17:29,840 --> 00:17:34,800
Speaker 4: for example, because their cost of capital is so high,

310
00:17:34,880 --> 00:17:36,679
Speaker 4: there's certain things that it's hard for him to do.

311
00:17:36,880 --> 00:17:39,000
Speaker 4: And one of the things that it's really hard for

312
00:17:39,080 --> 00:17:41,199
Speaker 4: all these startups to do, and this was true and

313
00:17:41,320 --> 00:17:44,680
Speaker 4: the recycling industry and other industries, is build a plant.

314
00:17:45,440 --> 00:17:48,760
Speaker 4: Like very expensive, time consuming to build a plant. You

315
00:17:48,800 --> 00:17:52,160
Speaker 4: don't really want to raise bc capital to build a plant,

316
00:17:52,440 --> 00:17:54,840
Speaker 4: and so it's important to have a partnership on the

317
00:17:54,880 --> 00:17:57,920
Speaker 4: manufacturing side too. And that was really like the first

318
00:17:57,920 --> 00:18:01,119
Speaker 4: thing this battery startup that I just visited talked about

319
00:18:01,520 --> 00:18:04,000
Speaker 4: is like getting that because you've got to be able

320
00:18:04,000 --> 00:18:05,960
Speaker 4: to deliver your product and you have to deliver it

321
00:18:06,000 --> 00:18:09,360
Speaker 4: on scale, and ideally you don't want to be wasting

322
00:18:09,400 --> 00:18:12,080
Speaker 4: time building your own plant on that and then like

323
00:18:12,119 --> 00:18:14,440
Speaker 4: you said, on the other end, you want to have

324
00:18:14,480 --> 00:18:18,080
Speaker 4: a partnership with the users of the energy, which is

325
00:18:18,400 --> 00:18:20,920
Speaker 4: all the people that either have data centers or use

326
00:18:21,040 --> 00:18:24,960
Speaker 4: data centers or customers of data centers, and you want

327
00:18:25,000 --> 00:18:29,560
Speaker 4: them to ideally put together an attract a financing relationship

328
00:18:29,600 --> 00:18:33,600
Speaker 4: where you know, in some form or fashion they're front

329
00:18:33,640 --> 00:18:37,280
Speaker 4: loading their payments to use so that you can use

330
00:18:37,320 --> 00:18:40,120
Speaker 4: that capital to actually build a product that they meet.

331
00:18:56,480 --> 00:18:58,879
Speaker 2: So Joe and I went to San Francisco a little

332
00:18:58,880 --> 00:19:01,680
Speaker 2: while ago and we saw some cool things. I had

333
00:19:01,680 --> 00:19:03,960
Speaker 2: my first ride in a way Moo, and we saw

334
00:19:04,000 --> 00:19:07,480
Speaker 2: some cool battery related technology. We also saw a lot

335
00:19:07,480 --> 00:19:11,560
Speaker 2: of vcs. Everyone very excited about AI. Obviously, they were

336
00:19:11,560 --> 00:19:15,160
Speaker 2: also talking about the difficulty of chasing deals right now,

337
00:19:15,560 --> 00:19:20,200
Speaker 2: how do you compete with those traditional vcs or are

338
00:19:20,240 --> 00:19:23,160
Speaker 2: you just not competing with them directly because you're taking

339
00:19:23,200 --> 00:19:25,760
Speaker 2: the slightly different GPU backed approach.

340
00:19:26,680 --> 00:19:27,760
Speaker 5: You know, I think it's both.

341
00:19:27,840 --> 00:19:30,560
Speaker 4: I think you're competing with them and to an extent,

342
00:19:30,760 --> 00:19:33,879
Speaker 4: partnering with them. And that's the thing we had to

343
00:19:33,920 --> 00:19:37,000
Speaker 4: ask ourselves before launching the fund, is what are we

344
00:19:37,080 --> 00:19:40,399
Speaker 4: bringing to the bear that's value added? And in this case,

345
00:19:40,440 --> 00:19:44,680
Speaker 4: we're bringing to bear the compute. And so often these startups,

346
00:19:44,800 --> 00:19:47,720
Speaker 4: even if they're backed by a strong VC, can have

347
00:19:47,800 --> 00:19:49,840
Speaker 4: a bit of a chicken and egg problem, which is

348
00:19:50,720 --> 00:19:53,600
Speaker 4: they need compute to develop their product, and they need

349
00:19:53,640 --> 00:19:56,600
Speaker 4: capital to buy that compute. But if they don't have

350
00:19:56,640 --> 00:20:00,240
Speaker 4: the compute lined up and the price locked in, then

351
00:20:00,280 --> 00:20:02,600
Speaker 4: the capital might be hesitant to go in because they'd

352
00:20:02,600 --> 00:20:05,159
Speaker 4: be like, we could put our capital into you, and

353
00:20:05,200 --> 00:20:07,159
Speaker 4: then it could take you an extra six months to

354
00:20:07,200 --> 00:20:10,440
Speaker 4: get your compute, and by that time some competitors passed

355
00:20:10,480 --> 00:20:14,240
Speaker 4: you by or the technology has changed. And on the

356
00:20:14,320 --> 00:20:16,800
Speaker 4: other hand, because they're a startup, they don't really have

357
00:20:16,840 --> 00:20:20,240
Speaker 4: the credit worthiness to just contract the compute. They most

358
00:20:20,359 --> 00:20:23,160
Speaker 4: likely have to pay up front, and so we bridge

359
00:20:23,200 --> 00:20:26,720
Speaker 4: that gap. And so if we go into a fundraising

360
00:20:26,840 --> 00:20:29,280
Speaker 4: round where there's a bunch of vcs putting cash in,

361
00:20:29,800 --> 00:20:34,240
Speaker 4: if they know that we're putting compute in alongside them

362
00:20:34,600 --> 00:20:37,119
Speaker 4: and that the second the round closes that compute will

363
00:20:37,160 --> 00:20:39,760
Speaker 4: be available to the company, that makes it easier to

364
00:20:39,880 --> 00:20:42,880
Speaker 4: raise the cash part of it. So we are competing

365
00:20:42,920 --> 00:20:45,480
Speaker 4: and we need that value added to be part of

366
00:20:45,520 --> 00:20:48,359
Speaker 4: the equation. But also I think it helps them to

367
00:20:48,480 --> 00:20:51,840
Speaker 4: raise from traditional vcs because we take that one risk

368
00:20:51,880 --> 00:20:52,520
Speaker 4: off the table.

369
00:20:52,840 --> 00:20:56,480
Speaker 3: How big is the market of companies that need compute,

370
00:20:56,520 --> 00:20:59,760
Speaker 3: because there are plenty of AI companies that just build

371
00:20:59,800 --> 00:21:05,720
Speaker 3: on top of an existing model like GPT or anthropics model,

372
00:21:05,720 --> 00:21:10,439
Speaker 3: et cetera. How many companies are actually out there and

373
00:21:10,520 --> 00:21:12,919
Speaker 3: who like not who are they specifically, but what are

374
00:21:12,920 --> 00:21:16,639
Speaker 3: the types of companies for whom actual access to compute

375
00:21:17,080 --> 00:21:19,240
Speaker 3: is an important part of their business.

376
00:21:20,320 --> 00:21:22,520
Speaker 4: Yes, well, you know it starts, of course with the

377
00:21:23,400 --> 00:21:28,439
Speaker 4: LM companies. You're using massive, huge, huge amounts of compute.

378
00:21:28,760 --> 00:21:30,200
Speaker 5: But then if you look.

379
00:21:30,080 --> 00:21:34,840
Speaker 4: At the rest of sort of the AI stack, there's

380
00:21:34,880 --> 00:21:37,040
Speaker 4: a couple areas where you're going to need compute, and

381
00:21:37,119 --> 00:21:43,240
Speaker 4: one is all the small model custom model companies, and

382
00:21:43,560 --> 00:21:45,479
Speaker 4: small commute a lot of different things. So you can

383
00:21:45,840 --> 00:21:49,000
Speaker 4: have some very small companies that are using a very

384
00:21:49,040 --> 00:21:52,000
Speaker 4: targeted model, like say in a vertical stack, you might

385
00:21:52,040 --> 00:21:56,560
Speaker 4: have a robotics company that is specifically training a model

386
00:21:56,760 --> 00:22:00,520
Speaker 4: to run a robot in a particular situation, and that

387
00:22:00,520 --> 00:22:03,520
Speaker 4: could be anything from a warehouse to doing surgery, right,

388
00:22:04,080 --> 00:22:09,679
Speaker 4: and they need compute to train that model or another

389
00:22:09,720 --> 00:22:12,360
Speaker 4: one which is huge and dominated by an existing big

390
00:22:12,359 --> 00:22:17,280
Speaker 4: players autonomous driving, but there are other autonomous driving companies

391
00:22:17,320 --> 00:22:21,480
Speaker 4: that are trying to be deployed at other automakers that

392
00:22:21,680 --> 00:22:23,320
Speaker 4: need compute to train those models.

393
00:22:23,520 --> 00:22:24,600
Speaker 5: Or weather models.

394
00:22:25,280 --> 00:22:28,200
Speaker 4: There's some really good companies that we've talked to doing

395
00:22:28,280 --> 00:22:31,879
Speaker 4: weather models. They need compute to train their model, and

396
00:22:32,000 --> 00:22:35,560
Speaker 4: so that whole model layer, and then even on the

397
00:22:35,640 --> 00:22:40,720
Speaker 4: app layer, they might be custom elements of small models

398
00:22:40,720 --> 00:22:42,480
Speaker 4: that they have that sit on top of the big

399
00:22:42,720 --> 00:22:44,760
Speaker 4: lms that they need some amount of compute for.

400
00:22:45,560 --> 00:22:46,639
Speaker 5: So there's quite a range.

401
00:22:46,680 --> 00:22:49,680
Speaker 4: You know, it's not everyone, you know, it's more in

402
00:22:49,720 --> 00:22:52,520
Speaker 4: that model application layer, and you know, less in the

403
00:22:52,760 --> 00:22:54,879
Speaker 4: infrastructure layer that need compute.

404
00:22:55,160 --> 00:22:58,280
Speaker 2: So this is one thing I always wonder about AI investment,

405
00:22:58,359 --> 00:23:00,400
Speaker 2: which is you have a lot of companies that are

406
00:23:00,400 --> 00:23:04,000
Speaker 2: building on top of existing models, as Joe mentioned, And

407
00:23:04,200 --> 00:23:07,240
Speaker 2: to some extent that makes sense because they can save

408
00:23:07,400 --> 00:23:10,080
Speaker 2: a lot of money by doing it, and realistically, are

409
00:23:10,080 --> 00:23:13,359
Speaker 2: you going to compete with Google or Microsoft? Probably not.

410
00:23:14,119 --> 00:23:16,440
Speaker 2: But on the other hand, I always wonder if you're

411
00:23:16,440 --> 00:23:19,679
Speaker 2: building on top of an existing model, how do you

412
00:23:19,800 --> 00:23:23,480
Speaker 2: ring fence that business? Because my assumption is if AI

413
00:23:23,800 --> 00:23:27,200
Speaker 2: gets better, maybe at some point the AI can replicate

414
00:23:27,280 --> 00:23:29,520
Speaker 2: any AI model basically.

415
00:23:30,920 --> 00:23:34,640
Speaker 4: So this is the first thing we always worry about

416
00:23:35,080 --> 00:23:39,040
Speaker 4: is does some giant company already have this product in

417
00:23:39,080 --> 00:23:41,840
Speaker 4: a closet with like twenty PhDs working on this and

418
00:23:41,920 --> 00:23:44,080
Speaker 4: somebody I was just at this conference and somebody coined

419
00:23:44,080 --> 00:23:48,200
Speaker 4: the phrase incumbent maximalist, And that's the man. You think

420
00:23:48,240 --> 00:23:50,320
Speaker 4: the incumbents are going to do everything and no one

421
00:23:50,320 --> 00:23:53,439
Speaker 4: else will ever succeed. And I think there's a few

422
00:23:53,680 --> 00:23:58,479
Speaker 4: use cases. There's things where it's a very specific task

423
00:23:59,400 --> 00:24:02,879
Speaker 4: that is hard to do well with a giant general

424
00:24:02,960 --> 00:24:06,240
Speaker 4: model and probably isn't worth doing well. Like if you're

425
00:24:06,480 --> 00:24:10,960
Speaker 4: focused on growing tens to hundreds of billions of dollars

426
00:24:10,960 --> 00:24:13,639
Speaker 4: of revenue, you can't be distracted by trying to do

427
00:24:13,720 --> 00:24:16,560
Speaker 4: every little thing. And we've seen this in previous tech

428
00:24:16,920 --> 00:24:20,480
Speaker 4: revolutions as well, and so it can be something that's

429
00:24:20,600 --> 00:24:26,439
Speaker 4: very focused on a space. We've seen legal accounting, sales.

430
00:24:27,080 --> 00:24:31,359
Speaker 4: There's some great companies that have virtual employees that they're

431
00:24:31,359 --> 00:24:35,280
Speaker 4: doing things that are very task specific. There's some companies

432
00:24:35,320 --> 00:24:39,000
Speaker 4: doing text of language and language to text and other

433
00:24:39,080 --> 00:24:43,240
Speaker 4: things for very specific applications. So you know that's one way.

434
00:24:43,680 --> 00:24:47,520
Speaker 4: The other way is data. The greatest ring fence is

435
00:24:47,680 --> 00:24:52,680
Speaker 4: to any AI company or business is data. Because you've

436
00:24:52,760 --> 00:24:56,280
Speaker 4: seen as the performance of some of the lms has

437
00:24:56,359 --> 00:24:59,000
Speaker 4: supposedly flattened out, a lot of that is because they've

438
00:24:59,000 --> 00:25:01,480
Speaker 4: just used all the data, like they've trained on the

439
00:25:01,520 --> 00:25:04,280
Speaker 4: whole Internet, there's nothing left and so now you have

440
00:25:04,320 --> 00:25:07,200
Speaker 4: to have other ways to train or novel sources of data.

441
00:25:07,280 --> 00:25:10,800
Speaker 4: So proprietary data is super valuable. And then there just

442
00:25:10,800 --> 00:25:14,119
Speaker 4: could be areas where they're conflicted. They don't want to

443
00:25:14,400 --> 00:25:17,440
Speaker 4: compete with their customers right now, although you know, competing

444
00:25:17,440 --> 00:25:20,159
Speaker 4: with your customers is a great tradition in the tech space,

445
00:25:20,560 --> 00:25:22,919
Speaker 4: but there could be situations where it's not worth it

446
00:25:22,960 --> 00:25:24,879
Speaker 4: to them yet to compete with their customers. And so

447
00:25:25,359 --> 00:25:27,679
Speaker 4: I think there's those different use cases where you know

448
00:25:27,680 --> 00:25:31,720
Speaker 4: you're going to see a small number of companies succeed.

449
00:25:32,119 --> 00:25:34,560
Speaker 3: I have a very stupid question, and actually I shouldn't

450
00:25:34,560 --> 00:25:36,440
Speaker 3: even be asking you. I should have asked it the

451
00:25:36,520 --> 00:25:39,200
Speaker 3: last time we talked to core Weave, but since you're here,

452
00:25:39,760 --> 00:25:42,399
Speaker 3: I'm gonna take them all again, or on the question

453
00:25:42,440 --> 00:25:45,320
Speaker 3: I didn't ask them. I know that Nvidia is an

454
00:25:45,320 --> 00:25:50,320
Speaker 3: investor in core Weave, but even setting aside that specific relationship,

455
00:25:50,680 --> 00:25:55,720
Speaker 3: the actual purchasing of chips, how does the pricing work

456
00:25:55,760 --> 00:25:58,320
Speaker 3: and how much is it a de facto auction? Where

457
00:25:58,359 --> 00:26:02,000
Speaker 3: As demand for chips boomed, in Vidia can expand its

458
00:26:02,080 --> 00:26:07,200
Speaker 3: margin versus in Vidia aims for a stable margin over time,

459
00:26:07,240 --> 00:26:10,320
Speaker 3: And I imagine this enters into your calculation to somewhat

460
00:26:10,520 --> 00:26:14,560
Speaker 3: thinking about a core Weaves future capital requirements. How does

461
00:26:14,600 --> 00:26:15,680
Speaker 3: that market for chips work?

462
00:26:16,960 --> 00:26:20,400
Speaker 4: Well, I can't comment on the internal workings of Nvidia

463
00:26:20,800 --> 00:26:22,280
Speaker 4: setting their prices.

464
00:26:21,960 --> 00:26:25,160
Speaker 3: But is an investor in a buyer whatever you I'm

465
00:26:25,200 --> 00:26:28,680
Speaker 3: a buyer of chips, how do I want to buy

466
00:26:28,680 --> 00:26:29,040
Speaker 3: some chips?

467
00:26:29,040 --> 00:26:31,800
Speaker 2: And now imagine it's like the container industry where you

468
00:26:31,880 --> 00:26:34,960
Speaker 2: have to have a specific relationship and there's a shipping

469
00:26:35,000 --> 00:26:38,800
Speaker 2: manager called Lars somewhere in northern Europe who holds the

470
00:26:38,880 --> 00:26:39,919
Speaker 2: keys to the chips.

471
00:26:40,640 --> 00:26:43,800
Speaker 4: Well, for any company using a resource, and it's certainly

472
00:26:43,840 --> 00:26:47,879
Speaker 4: true of companies using compute right, it's always a cost

473
00:26:48,040 --> 00:26:54,080
Speaker 4: benefit example. So there's great benefits to running your AI

474
00:26:54,240 --> 00:26:59,120
Speaker 4: training on an Nvidia ecosystem on a network like Core

475
00:26:59,200 --> 00:27:02,920
Speaker 4: weaves that's very fast and very reliable because you know,

476
00:27:02,960 --> 00:27:06,840
Speaker 4: when you train a model, you stop every fifteen or

477
00:27:06,880 --> 00:27:09,719
Speaker 4: thirty minutes to save your work, and if there's a

478
00:27:09,720 --> 00:27:11,520
Speaker 4: failure in there, you have to go back to the

479
00:27:11,600 --> 00:27:13,639
Speaker 4: last time you save your work and there's a huge

480
00:27:13,680 --> 00:27:18,680
Speaker 4: loss on that. So there's benefits to using the best technology,

481
00:27:19,280 --> 00:27:23,600
Speaker 4: but those are quantifiable, and if you're a particular kind

482
00:27:23,640 --> 00:27:28,080
Speaker 4: of technology becomes too expensive, you'll see people diversify out right.

483
00:27:28,119 --> 00:27:30,439
Speaker 4: I mean, there was just news the last two days

484
00:27:30,480 --> 00:27:35,880
Speaker 4: about Anthropic and AWS and aws's new chips, So there's

485
00:27:35,880 --> 00:27:38,119
Speaker 4: always some form of competition. I mean, in Vida is

486
00:27:38,119 --> 00:27:40,639
Speaker 4: sitting in a unique place where they've really had a

487
00:27:40,680 --> 00:27:45,200
Speaker 4: de facto monopoly on this, and I think their pricing

488
00:27:46,040 --> 00:27:48,679
Speaker 4: is being set in a way to grow the market,

489
00:27:48,800 --> 00:27:51,280
Speaker 4: right Like, they want to grow the market. I can't

490
00:27:51,320 --> 00:27:53,560
Speaker 4: speak for them, but you wouldn't want to set the

491
00:27:53,600 --> 00:27:57,399
Speaker 4: price of your product so high that you stifle the

492
00:27:57,440 --> 00:28:01,000
Speaker 4: market's growth, right Like, growth is more than making an

493
00:28:01,000 --> 00:28:03,960
Speaker 4: extra dollar on every widget, And so I think that's

494
00:28:04,000 --> 00:28:07,960
Speaker 4: got to be a calculation, and certainly to date it's

495
00:28:08,160 --> 00:28:10,919
Speaker 4: been fruitful in that this market has taken off like

496
00:28:11,000 --> 00:28:12,120
Speaker 4: almost no market ever.

497
00:28:28,320 --> 00:28:31,800
Speaker 2: I want to go back to the capital question, and

498
00:28:32,280 --> 00:28:36,359
Speaker 2: most venture capital comes in the form of equity. You're

499
00:28:36,440 --> 00:28:40,640
Speaker 2: doing something slightly different in my understanding. You're primarily going

500
00:28:40,800 --> 00:28:44,760
Speaker 2: down the debt and sort of fixed income route. That

501
00:28:44,880 --> 00:28:47,440
Speaker 2: seems so different because in my mind, when I think

502
00:28:47,480 --> 00:28:50,360
Speaker 2: about bond investing, and we've said this a number of

503
00:28:50,400 --> 00:28:53,880
Speaker 2: times on the show, it's all about avoiding losers, right Like,

504
00:28:53,920 --> 00:28:57,200
Speaker 2: there's limited upside, but you don't want a bankruptcy that

505
00:28:57,240 --> 00:29:01,480
Speaker 2: wipes out your investment, whereas equity the upside is basically uncapped.

506
00:29:01,520 --> 00:29:05,040
Speaker 2: So it's about finding that one stellar out performer or

507
00:29:05,080 --> 00:29:08,520
Speaker 2: that one lottery ticket. How do you square I guess

508
00:29:08,600 --> 00:29:11,800
Speaker 2: the risk averseness of some of this debt financing with

509
00:29:12,040 --> 00:29:16,240
Speaker 2: getting the huge upside that is potentially there from AI.

510
00:29:17,400 --> 00:29:22,200
Speaker 4: Well, the amount of financing required for this whole AI buildout,

511
00:29:22,280 --> 00:29:25,360
Speaker 4: which is on some immense scale of you know, people

512
00:29:25,360 --> 00:29:29,280
Speaker 4: have talked about the Manhattan Project, the building of the Interstates.

513
00:29:30,000 --> 00:29:34,080
Speaker 4: It's going to require capital in many forms for many things,

514
00:29:34,160 --> 00:29:37,200
Speaker 4: and I think there's a lot of thinking going on,

515
00:29:37,400 --> 00:29:41,680
Speaker 4: and you know, certainly we're part of that in deploying

516
00:29:41,720 --> 00:29:47,080
Speaker 4: the most efficient capital to the different layers of this buildout.

517
00:29:47,240 --> 00:29:50,040
Speaker 4: And so we've talked about a couple different things here.

518
00:29:50,040 --> 00:29:54,320
Speaker 4: We've talked about financing GPUs. So if you're financing GPUs

519
00:29:54,360 --> 00:29:58,040
Speaker 4: with debt, then you can really think through your downside protection,

520
00:29:58,800 --> 00:30:00,600
Speaker 4: just like in the audio metaphor.

521
00:30:00,800 --> 00:30:01,960
Speaker 2: Right, you have the collateral.

522
00:30:02,120 --> 00:30:04,600
Speaker 4: You have the collateral, you have the contract. You can

523
00:30:04,640 --> 00:30:08,080
Speaker 4: analyze the credit worthiness of the contract. You can look

524
00:30:08,160 --> 00:30:13,440
Speaker 4: at how the leasing curves of prior chip generations have decayed.

525
00:30:14,160 --> 00:30:17,080
Speaker 4: You have some real information there, you have a real asset,

526
00:30:17,800 --> 00:30:22,200
Speaker 4: you have real contracted cash flows. Now in the VC fund,

527
00:30:22,920 --> 00:30:25,240
Speaker 4: that's a lot different. In this case, this is true

528
00:30:25,480 --> 00:30:29,240
Speaker 4: venture equity, and it's just that it's being deployed in

529
00:30:29,280 --> 00:30:34,520
Speaker 4: a unique way where instead of cash, the compute has

530
00:30:34,600 --> 00:30:37,920
Speaker 4: been contractually secured and it's just being exchanged for the

531
00:30:37,960 --> 00:30:42,400
Speaker 4: equity directly, as I talked about before, saving that step

532
00:30:42,520 --> 00:30:45,920
Speaker 4: and de risking the process of acquiring compute for these

533
00:30:45,920 --> 00:30:47,000
Speaker 4: grow stage companies.

534
00:30:47,080 --> 00:30:49,640
Speaker 2: So you are doing equity through the VC fund.

535
00:30:49,800 --> 00:30:51,400
Speaker 5: The VC fund is equity.

536
00:30:51,480 --> 00:30:55,560
Speaker 4: Yes, it would be part of typically but not always

537
00:30:55,600 --> 00:30:58,960
Speaker 4: a part of a round that a growth stage company

538
00:30:59,040 --> 00:30:59,800
Speaker 4: might be doing.

539
00:31:00,080 --> 00:31:01,680
Speaker 2: Doing convertibles.

540
00:31:01,760 --> 00:31:06,160
Speaker 4: So we can do virtually anything across the debt equity

541
00:31:06,640 --> 00:31:10,640
Speaker 4: private public spectrum, and have in many cases in the

542
00:31:11,280 --> 00:31:15,560
Speaker 4: AI fund itself. Most of the companies being gross stage

543
00:31:16,040 --> 00:31:18,920
Speaker 4: are not really in a position to do debt, so

544
00:31:19,040 --> 00:31:21,640
Speaker 4: I think for the most part, I would expect that

545
00:31:21,720 --> 00:31:24,560
Speaker 4: those would all be venture equity investments.

546
00:31:25,160 --> 00:31:27,680
Speaker 3: I gotta chuckle when you're like, oh, we've been in

547
00:31:27,680 --> 00:31:29,880
Speaker 3: this space, it's way back, and then you said twenty

548
00:31:29,920 --> 00:31:32,040
Speaker 3: twenty one, But it does really sort of.

549
00:31:32,040 --> 00:31:32,640
Speaker 5: Speak to hell.

550
00:31:32,720 --> 00:31:33,760
Speaker 2: It feels like a long time.

551
00:31:33,880 --> 00:31:36,200
Speaker 3: Yeah, well, you know, I mean ched GBT, I think

552
00:31:36,240 --> 00:31:38,440
Speaker 3: came out at the very end of twenty twenty two

553
00:31:38,600 --> 00:31:40,640
Speaker 3: or maybe early twenty twenty three, and that was the

554
00:31:40,640 --> 00:31:42,520
Speaker 3: big light bulb moment for a lot of people. So

555
00:31:42,640 --> 00:31:45,400
Speaker 3: even being that active in a lot of this stuff

556
00:31:45,400 --> 00:31:49,880
Speaker 3: a year earlier truly is early. That being said, things

557
00:31:49,960 --> 00:31:52,720
Speaker 3: like core weave, things like data centers. The need for

558
00:31:52,800 --> 00:31:55,760
Speaker 3: compute is very well understood right now in a way

559
00:31:55,760 --> 00:31:58,440
Speaker 3: that perhaps in three years ago many people in the

560
00:31:58,600 --> 00:32:03,240
Speaker 3: credit and financing space weren't thinking of is that a

561
00:32:03,320 --> 00:32:07,320
Speaker 3: margin compressor for you? The fact that other entities, probably

562
00:32:07,320 --> 00:32:11,400
Speaker 3: many with much more capital than Magnetar has everyone has

563
00:32:11,480 --> 00:32:15,680
Speaker 3: now woken up to this opportunity of yes, there's going

564
00:32:15,720 --> 00:32:17,959
Speaker 3: to be a lot of financing needs in AI. And

565
00:32:18,000 --> 00:32:21,640
Speaker 3: do you see change in competition or spreads or anything

566
00:32:21,680 --> 00:32:22,040
Speaker 3: like that.

567
00:32:23,000 --> 00:32:27,440
Speaker 4: Well, I think it really depends on what you're financing.

568
00:32:27,680 --> 00:32:30,760
Speaker 4: So there's a lot of capital that's gone into all

569
00:32:30,800 --> 00:32:35,600
Speaker 4: these spaces, and certainly all across the stack.

570
00:32:35,320 --> 00:32:36,560
Speaker 5: Of financing compute.

571
00:32:36,600 --> 00:32:39,000
Speaker 4: You've seen a huge amount of capital come in, and

572
00:32:39,040 --> 00:32:44,320
Speaker 4: you've seen all the giant investment companies providers of capital.

573
00:32:44,360 --> 00:32:45,560
Speaker 5: Get involved and so.

574
00:32:47,200 --> 00:32:49,760
Speaker 4: There's a lot of capital in there, but there's also

575
00:32:50,320 --> 00:32:55,200
Speaker 4: like a huge need for capital, and it's very complex

576
00:32:55,320 --> 00:32:59,680
Speaker 4: thinking about the structuring and getting the right capital and

577
00:32:59,680 --> 00:33:02,080
Speaker 4: the right space. And so I think there's room to

578
00:33:02,120 --> 00:33:05,360
Speaker 4: be innovative. And I've spent the last twenty years at

579
00:33:05,360 --> 00:33:11,160
Speaker 4: Magnetar thinking about unique ways to source investments and deploy capital,

580
00:33:11,480 --> 00:33:13,160
Speaker 4: and I think that really comes to bear on this.

581
00:33:13,280 --> 00:33:16,360
Speaker 4: And because this whole market, like you said, is so new,

582
00:33:16,400 --> 00:33:18,920
Speaker 4: and we've only had chat GPT for a couple of years,

583
00:33:19,320 --> 00:33:22,760
Speaker 4: you know, you're seeing companies with all different ways of working.

584
00:33:22,840 --> 00:33:25,840
Speaker 4: You know, we I talked to a company in the

585
00:33:26,560 --> 00:33:29,760
Speaker 4: text of voice space at a conference last week and

586
00:33:29,840 --> 00:33:34,040
Speaker 4: they actually were buying their own DGX servers themselves and

587
00:33:34,120 --> 00:33:38,000
Speaker 4: just running on themselves in their own on premp site.

588
00:33:38,480 --> 00:33:41,720
Speaker 4: And We're like, sure, like that's something we can finance.

589
00:33:41,840 --> 00:33:43,000
Speaker 4: That's like a hard asset.

590
00:33:43,080 --> 00:33:44,720
Speaker 5: But no one's really looking at that yet.

591
00:33:44,760 --> 00:33:48,080
Speaker 4: Because most of the capital is so big, it has

592
00:33:48,160 --> 00:33:51,160
Speaker 4: to go to the biggest thing. So you have your

593
00:33:52,040 --> 00:33:55,800
Speaker 4: trillion dollar investment firm, which was a couple you're not

594
00:33:55,920 --> 00:34:00,000
Speaker 4: going to want to deploy twenty to fifty million dollars

595
00:34:00,280 --> 00:34:02,200
Speaker 4: in a one off thing. You're going to want to

596
00:34:02,240 --> 00:34:05,320
Speaker 4: deploy tens of billions of dollars in the biggest thing,

597
00:34:05,400 --> 00:34:09,440
Speaker 4: whether that's power, physical data centers, or GPUs.

598
00:34:10,520 --> 00:34:15,280
Speaker 2: What's the pitch to your investors, to Magnetars investors, Because again,

599
00:34:15,320 --> 00:34:17,319
Speaker 2: this is something I know you said you've been in

600
00:34:17,360 --> 00:34:19,840
Speaker 2: the tech space for a while, but it's still something

601
00:34:20,239 --> 00:34:23,360
Speaker 2: that feels fairly new. And when I think about AI,

602
00:34:24,320 --> 00:34:27,160
Speaker 2: there's been so much excitement over it. Some people have

603
00:34:27,239 --> 00:34:29,520
Speaker 2: been talking about whether or not it's in a bubble,

604
00:34:30,000 --> 00:34:31,920
Speaker 2: and I think about a hedge fund, and that's all

605
00:34:31,960 --> 00:34:38,120
Speaker 2: about uncorrelated returns and investing profitably through the cycle. I

606
00:34:38,160 --> 00:34:43,040
Speaker 2: get that you might be promising very large upside to investors,

607
00:34:43,200 --> 00:34:46,239
Speaker 2: but what is the hedge aspect of this.

608
00:34:47,680 --> 00:34:47,919
Speaker 5: Well.

609
00:34:48,080 --> 00:34:51,080
Speaker 4: As a firm, we've done many different products and many

610
00:34:51,080 --> 00:34:55,319
Speaker 4: different strategies for many different investors over the years, and

611
00:34:55,960 --> 00:34:59,680
Speaker 4: we've really been flexible in trying to deploy capital in

612
00:34:59,680 --> 00:35:01,880
Speaker 4: the most most interesting areas that are going to have

613
00:35:01,920 --> 00:35:05,200
Speaker 4: the best risk adjusted returns. And many of our investors

614
00:35:05,200 --> 00:35:07,080
Speaker 4: have been with us through the whole life of the

615
00:35:07,080 --> 00:35:10,839
Speaker 4: firm since two thousand and five and appreciate that. And

616
00:35:10,880 --> 00:35:17,239
Speaker 4: so we've done both diversified investment strategies where we just

617
00:35:17,320 --> 00:35:21,360
Speaker 4: thought the general pipeline of deploying structured capital has been great,

618
00:35:21,680 --> 00:35:24,640
Speaker 4: and then we've also done things targeted at a particular

619
00:35:24,719 --> 00:35:27,839
Speaker 4: asset when we thought that opportunity was great. And so

620
00:35:28,400 --> 00:35:32,200
Speaker 4: in the case of the VC fund, the value proposition

621
00:35:32,360 --> 00:35:37,080
Speaker 4: really is for the investor what it is for the company,

622
00:35:37,120 --> 00:35:41,440
Speaker 4: which is, we're bringing something unique to these gross stage

623
00:35:41,480 --> 00:35:46,040
Speaker 4: AI companies which will get us access to making investments

624
00:35:46,560 --> 00:35:49,600
Speaker 4: and what we hope will be the best best of

625
00:35:49,600 --> 00:35:53,200
Speaker 4: those companies with the best business models and the best teams.

626
00:35:53,920 --> 00:35:58,040
Speaker 4: And so we're going to use the unique compute that

627
00:35:58,080 --> 00:36:00,440
Speaker 4: we have and the way that we're going to exchange

628
00:36:00,480 --> 00:36:03,520
Speaker 4: that for equity and deliver that to these companies as

629
00:36:03,560 --> 00:36:06,640
Speaker 4: a way of getting access to investments in what's a

630
00:36:06,800 --> 00:36:09,400
Speaker 4: very as you mentioned, very competitive environment where there's a

631
00:36:09,440 --> 00:36:12,640
Speaker 4: lot of capital going into the space. And so I

632
00:36:12,680 --> 00:36:17,680
Speaker 4: think for investors that want to participate in that kind

633
00:36:17,719 --> 00:36:22,960
Speaker 4: of investment, in getting capital deployed into growth stage AI companies,

634
00:36:23,000 --> 00:36:25,160
Speaker 4: you know, this is a very unique opportunity, and so

635
00:36:25,239 --> 00:36:26,880
Speaker 4: we saw a lot of traction with that.

636
00:36:27,160 --> 00:36:30,400
Speaker 3: When you come in as a VC investor in some

637
00:36:30,440 --> 00:36:34,400
Speaker 3: of these startups, do you have to supply dollars or

638
00:36:34,560 --> 00:36:38,279
Speaker 3: in some cases or all cases, is your ability to

639
00:36:38,320 --> 00:36:40,560
Speaker 3: promise compute from day one enough for equity?

640
00:36:41,840 --> 00:36:46,520
Speaker 4: It really varies, and there's investments we've made both inside

641
00:36:46,520 --> 00:36:50,560
Speaker 4: and outside the fund, and it just depends on the situation.

642
00:36:50,760 --> 00:36:54,799
Speaker 4: So there can be companies that we find super interesting

643
00:36:55,000 --> 00:36:57,759
Speaker 4: but don't need compute, and in that case we could

644
00:36:57,800 --> 00:37:01,000
Speaker 4: invest in those companies directly outside of the fund. For

645
00:37:01,040 --> 00:37:04,600
Speaker 4: the fund itself, the proposition is equity for compute, and

646
00:37:04,680 --> 00:37:08,480
Speaker 4: so the fund itself is focused on companies that really

647
00:37:08,719 --> 00:37:11,920
Speaker 4: do need equity and are interested in equity. And I

648
00:37:11,960 --> 00:37:14,279
Speaker 4: really do need compute and are interested in compute on

649
00:37:14,320 --> 00:37:18,759
Speaker 4: corewaves network, and so that's the kind of companies that

650
00:37:18,800 --> 00:37:21,919
Speaker 4: will invest in from the fund. But as Magnetar as

651
00:37:21,960 --> 00:37:24,640
Speaker 4: a whole, we've been focused, like we talked about, on

652
00:37:24,680 --> 00:37:29,799
Speaker 4: everything from energy, through infrastructure, through other AI companies that

653
00:37:29,960 --> 00:37:32,400
Speaker 4: just don't happen to me compute right now.

654
00:37:32,960 --> 00:37:38,719
Speaker 3: Then, just to this point, your ability to promise or

655
00:37:38,800 --> 00:37:44,239
Speaker 3: give AI startups compute, this access to compute emerged via

656
00:37:44,280 --> 00:37:46,960
Speaker 3: that initial relationship as a financier.

657
00:37:47,120 --> 00:37:48,880
Speaker 2: This is what I was going to ask, which is

658
00:37:48,960 --> 00:37:52,759
Speaker 2: how worried. Are you about competitors doing the same thing

659
00:37:53,400 --> 00:37:56,839
Speaker 2: and providing GPU back debt or is it the case

660
00:37:56,880 --> 00:37:59,480
Speaker 2: that because of your first mover advantage with core Weave,

661
00:37:59,640 --> 00:38:01,960
Speaker 2: you can hold onto that advantage for a while.

662
00:38:02,840 --> 00:38:07,520
Speaker 4: So for the fund itself, it was the unique relationship

663
00:38:07,560 --> 00:38:10,239
Speaker 4: we had with coreweve where we felt they were the

664
00:38:10,280 --> 00:38:14,840
Speaker 4: best provider of AI training compute and we were able

665
00:38:14,920 --> 00:38:19,640
Speaker 4: to work with them to contract some of the very

666
00:38:19,680 --> 00:38:23,239
Speaker 4: scarce resource of that and then have that available to

667
00:38:23,320 --> 00:38:27,200
Speaker 4: deliver to these AI growth companies. And so that was

668
00:38:27,320 --> 00:38:32,280
Speaker 4: really where we were able to put together something unique because.

669
00:38:31,960 --> 00:38:34,560
Speaker 3: Day one that was understood to be part of the

670
00:38:34,600 --> 00:38:37,880
Speaker 3: payoff of being a financing partner to Corewave.

671
00:38:38,840 --> 00:38:40,520
Speaker 5: I wouldn't say from day one.

672
00:38:40,719 --> 00:38:44,080
Speaker 4: I would just say it's part of the natural growth

673
00:38:44,239 --> 00:38:49,239
Speaker 4: in their business and our growth in investing in the

674
00:38:49,320 --> 00:38:53,000
Speaker 4: AI market and in being a partner with them. Everyone

675
00:38:53,200 --> 00:38:55,360
Speaker 4: is both a partner and a competitor in this space,

676
00:38:55,480 --> 00:38:59,160
Speaker 4: and you know, Nvidia has multiple ways that they invest

677
00:38:59,160 --> 00:39:02,360
Speaker 4: in their customers, as do all the hyper scalers for example,

678
00:39:02,800 --> 00:39:06,520
Speaker 4: And so it's really about are you providing something unique,

679
00:39:07,160 --> 00:39:10,360
Speaker 4: something that's different, And you know, right now this moment

680
00:39:10,400 --> 00:39:14,320
Speaker 4: in time. We feel like the size of the compute

681
00:39:14,400 --> 00:39:18,600
Speaker 4: we're providing and the network we're providing it on and

682
00:39:18,680 --> 00:39:21,280
Speaker 4: the way that we can provide it in real time

683
00:39:22,040 --> 00:39:25,439
Speaker 4: is unique and is valuable to many companies. Now, look,

684
00:39:25,480 --> 00:39:27,520
Speaker 4: there could be some companies that are getting their compute

685
00:39:27,520 --> 00:39:30,120
Speaker 4: from somewhere else and it's just not a fit that's

686
00:39:30,160 --> 00:39:33,160
Speaker 4: certainly going to happen. But I think there's many, many

687
00:39:33,360 --> 00:39:36,439
Speaker 4: AI growth companies where this is very valuable to them

688
00:39:36,480 --> 00:39:39,120
Speaker 4: to get the compute on Quorwy's network, and that's going

689
00:39:39,200 --> 00:39:41,600
Speaker 4: to lead to a relationship with them.

690
00:39:42,200 --> 00:39:46,680
Speaker 3: When Amazon makes a VC investment, it's in large part

691
00:39:46,840 --> 00:39:49,680
Speaker 3: understood that it's the same sort of premise that they're

692
00:39:49,680 --> 00:39:51,920
Speaker 3: going to invest in some software company and the money

693
00:39:51,920 --> 00:39:55,560
Speaker 3: comes right back in because that company has AWS needs

694
00:39:55,640 --> 00:39:58,880
Speaker 3: and so it comes back. Obviously, we know that the

695
00:39:59,239 --> 00:40:03,120
Speaker 3: not only to the large legacy hyper scalers. Not only

696
00:40:03,120 --> 00:40:05,280
Speaker 3: they're building their own models, many of them they're building

697
00:40:05,280 --> 00:40:09,160
Speaker 3: their own silicon and Facebook has its own chips and

698
00:40:09,320 --> 00:40:12,799
Speaker 3: talked about Amazon and Google has I forget what their

699
00:40:12,840 --> 00:40:15,160
Speaker 3: whole thing is called. How do you think about them

700
00:40:15,239 --> 00:40:19,120
Speaker 3: as competitors to core weave in these sort of pure

701
00:40:19,239 --> 00:40:21,560
Speaker 3: chips and data center side. I know their partners, I

702
00:40:21,600 --> 00:40:24,120
Speaker 3: know their customers, et cetera. But they are also pure

703
00:40:24,160 --> 00:40:26,439
Speaker 3: competitors both to say a core weave and to say

704
00:40:26,440 --> 00:40:27,000
Speaker 3: an n video.

705
00:40:28,040 --> 00:40:28,279
Speaker 5: Yeah.

706
00:40:28,320 --> 00:40:32,400
Speaker 4: Again, everyone's a partner and a competitor, you know. I

707
00:40:32,440 --> 00:40:33,800
Speaker 4: think the difference.

708
00:40:33,480 --> 00:40:36,000
Speaker 3: Google's as TPUs is their thing. Anyway, Sorry, keep going,

709
00:40:36,040 --> 00:40:36,600
Speaker 3: I just couldn't.

710
00:40:36,920 --> 00:40:39,560
Speaker 4: Yeah, I mean the difference, as Brian talked about, is

711
00:40:39,840 --> 00:40:42,680
Speaker 4: the core Weave network was built for the ground up

712
00:40:43,000 --> 00:40:47,480
Speaker 4: to be hyper efficient at running AI solutions, and so

713
00:40:47,520 --> 00:40:50,600
Speaker 4: I think it's unique in that way, and I think

714
00:40:50,640 --> 00:40:55,000
Speaker 4: that's why it's grown so fast. But certainly everyone else

715
00:40:55,200 --> 00:40:58,000
Speaker 4: is trying to build their own out and there will

716
00:40:58,040 --> 00:41:01,360
Speaker 4: be other people that will have in Vida GPU chips

717
00:41:01,880 --> 00:41:05,440
Speaker 4: and that will include the hyperscalers. But you know, one

718
00:41:05,480 --> 00:41:06,520
Speaker 4: of the things we've.

719
00:41:06,320 --> 00:41:07,920
Speaker 5: Seen is that.

720
00:41:09,480 --> 00:41:14,920
Speaker 4: This is very hard technology. So it's particularly hard to

721
00:41:14,960 --> 00:41:19,920
Speaker 4: deploy at scale because you run into like real physics issues,

722
00:41:19,960 --> 00:41:24,080
Speaker 4: you know, surface area to volume type issues of getting

723
00:41:24,280 --> 00:41:27,840
Speaker 4: this much power to IRAQ with like how much cable

724
00:41:27,880 --> 00:41:31,279
Speaker 4: does that take? How much cooling does that take? How

725
00:41:31,280 --> 00:41:34,840
Speaker 4: do you run the software layer, like the software layer

726
00:41:34,840 --> 00:41:38,239
Speaker 4: to control you know, a node of eight GPUs is

727
00:41:38,280 --> 00:41:40,040
Speaker 4: going to be a lot different than if you're trying

728
00:41:40,120 --> 00:41:43,640
Speaker 4: to run one hundred and twenty eight thousand GPUs. And

729
00:41:43,680 --> 00:41:46,880
Speaker 4: so this problem gets more and more difficult, and you

730
00:41:47,000 --> 00:41:51,359
Speaker 4: need better technology and you need highly skilled people, and

731
00:41:51,520 --> 00:41:54,760
Speaker 4: so the bar is always moving. You know, there's always

732
00:41:54,760 --> 00:41:59,400
Speaker 4: a next generation chips that's going to be super complicated. Certainly,

733
00:42:00,080 --> 00:42:05,279
Speaker 4: the Blackwell deployments and the incremental new Blackwell generations are

734
00:42:05,320 --> 00:42:08,280
Speaker 4: going to be ever more complicated and trigger to deploy.

735
00:42:08,880 --> 00:42:13,080
Speaker 4: And you've seen issues already, right, You've seen hyperscalers and

736
00:42:13,239 --> 00:42:17,919
Speaker 4: other competitors in the space have reliability problems or be

737
00:42:17,920 --> 00:42:22,359
Speaker 4: behind schedule. Like it's not easy. It's a very complicated technology.

738
00:42:22,400 --> 00:42:25,239
Speaker 4: You're not plugging your GPU into the wall and it's

739
00:42:25,280 --> 00:42:28,279
Speaker 4: ready to run an AI model, and so like, I

740
00:42:28,320 --> 00:42:33,120
Speaker 4: think there's going to be value accrewing to skill an

741
00:42:33,160 --> 00:42:36,640
Speaker 4: efficiency and execution in the space, and you know that's

742
00:42:36,680 --> 00:42:37,760
Speaker 4: going to last for a while.

743
00:42:38,320 --> 00:42:43,040
Speaker 2: So some people draw an analogy between the current enthusiastic

744
00:42:43,239 --> 00:42:47,480
Speaker 2: cycle for AI and the early two thousands period where

745
00:42:47,520 --> 00:42:50,360
Speaker 2: we had a lot of enthusiasm for internet companies and

746
00:42:50,440 --> 00:42:54,120
Speaker 2: telecoms and things like that. Do you see evidence of

747
00:42:54,440 --> 00:42:57,040
Speaker 2: froth out there, or is it the case that because

748
00:42:57,200 --> 00:43:01,040
Speaker 2: of the huge amount of initial capital invents that's needed,

749
00:43:01,480 --> 00:43:04,960
Speaker 2: it's difficult to get I guess enough new entrance that

750
00:43:05,000 --> 00:43:06,440
Speaker 2: this would become a bubble.

751
00:43:07,560 --> 00:43:11,520
Speaker 4: Yeah, everything can become a bubble eventually in almost any

752
00:43:11,800 --> 00:43:17,440
Speaker 4: industry that's highly capital intensive. Usually if there's excess returns,

753
00:43:17,480 --> 00:43:20,160
Speaker 4: you'll see capital go into it until those returns aren't

754
00:43:20,160 --> 00:43:22,920
Speaker 4: good anymore, and a lot of capital will go in

755
00:43:23,040 --> 00:43:24,160
Speaker 4: before you figure out.

756
00:43:24,040 --> 00:43:25,000
Speaker 5: That last part.

757
00:43:25,520 --> 00:43:29,480
Speaker 4: But this is extremely early. Like if you look at

758
00:43:29,880 --> 00:43:33,120
Speaker 4: the capital that went into the Internet and then how

759
00:43:33,160 --> 00:43:37,640
Speaker 4: that value accrued to both the big tech companies and

760
00:43:37,680 --> 00:43:41,600
Speaker 4: the startups. People have looked at numbers like three trillion

761
00:43:41,680 --> 00:43:46,239
Speaker 4: dollars of equity value created with the large incumbents, but

762
00:43:46,280 --> 00:43:50,239
Speaker 4: there was another five hundred billion created for the new startups.

763
00:43:50,239 --> 00:43:53,800
Speaker 4: And we're just getting going here, right. We're just building

764
00:43:53,880 --> 00:43:57,480
Speaker 4: out the kind of data centers, the kind of energy infrastructure.

765
00:43:58,040 --> 00:44:00,839
Speaker 5: We're just starting to deploy products. If you talked to.

766
00:44:01,760 --> 00:44:07,439
Speaker 4: Enterprises, they're just starting to implement the most obvious use

767
00:44:07,480 --> 00:44:11,120
Speaker 4: cases for AI. So I think we're much too early

768
00:44:11,480 --> 00:44:14,480
Speaker 4: to worry about a bubble. I talked to somebody at

769
00:44:14,480 --> 00:44:17,799
Speaker 4: a hyperscaler and they were like, the last thing we're

770
00:44:17,840 --> 00:44:19,960
Speaker 4: worried about right now is having too much compute.

771
00:44:20,440 --> 00:44:24,080
Speaker 3: Last question for me, you say, we're early. There's still

772
00:44:24,200 --> 00:44:28,040
Speaker 3: no signs of too much compute. Earlier in the conversation,

773
00:44:28,160 --> 00:44:33,080
Speaker 3: you're like, this is a Manhattan project, scale project. Give

774
00:44:33,160 --> 00:44:35,719
Speaker 3: us some flashy number. How much has been deployed in

775
00:44:35,760 --> 00:44:38,120
Speaker 3: this area, and you know over the next ten years

776
00:44:38,480 --> 00:44:41,560
Speaker 3: how much capital is going to be demanded for this

777
00:44:41,680 --> 00:44:43,200
Speaker 3: space and how much will be needed.

778
00:44:44,000 --> 00:44:46,920
Speaker 4: So one one number I saw was that in twenty

779
00:44:47,239 --> 00:44:53,680
Speaker 4: twenty three, thirty seven billion dollars was deployed into AI infrastructure,

780
00:44:54,480 --> 00:44:57,960
Speaker 4: and in two thousand and thirty three that number is

781
00:44:58,000 --> 00:45:00,680
Speaker 4: going to be like four hundred and thirty billion in

782
00:45:00,760 --> 00:45:04,800
Speaker 4: that year. So this is trillion dollar scale investment.

783
00:45:05,880 --> 00:45:09,359
Speaker 2: Cool, You're cool, all right, Jim Presco, Thank you so

784
00:45:09,440 --> 00:45:11,120
Speaker 2: much for coming on all thoughts. That was great.

785
00:45:11,360 --> 00:45:13,000
Speaker 5: Thank you for having me.

786
00:45:13,080 --> 00:45:13,759
Speaker 3: Thank you so much.

787
00:45:26,680 --> 00:45:26,919
Speaker 4: Joe.

788
00:45:26,960 --> 00:45:30,759
Speaker 2: There's two things that I hear consistently about AI, and

789
00:45:30,880 --> 00:45:33,560
Speaker 2: one is it's going to need a lot of capital, yeah,

790
00:45:33,560 --> 00:45:35,640
Speaker 2: which Jim spoke to. And then the other thing I

791
00:45:35,680 --> 00:45:38,719
Speaker 2: always hear is well, at some point AI companies have

792
00:45:38,800 --> 00:45:42,840
Speaker 2: to actually produce revenue, and I guess the question is, like,

793
00:45:43,400 --> 00:45:46,440
Speaker 2: are they going to start producing revenue in time to

794
00:45:46,600 --> 00:45:48,560
Speaker 2: pay back that massive capital need.

795
00:45:49,440 --> 00:45:53,759
Speaker 3: Yes, it's very interesting because, look, I believe that there

796
00:45:53,800 --> 00:45:58,880
Speaker 3: are companies that are getting productive value out of AI models.

797
00:45:58,719 --> 00:45:59,920
Speaker 5: Like I believe that exists.

798
00:46:00,440 --> 00:46:02,960
Speaker 3: But you know, you talk about hundreds of billions over

799
00:46:03,000 --> 00:46:06,040
Speaker 3: the coming years and financing in the end that is

800
00:46:06,080 --> 00:46:10,120
Speaker 3: going to have to come from profitable deployment to customers,

801
00:46:10,360 --> 00:46:12,360
Speaker 3: and so like this to me is like, you know,

802
00:46:12,840 --> 00:46:15,960
Speaker 3: still a bit uncertain. I do think the financing that

803
00:46:16,000 --> 00:46:19,120
Speaker 3: we talked about is extremely interesting just in the context

804
00:46:19,160 --> 00:46:19,920
Speaker 3: of this conversation.

805
00:46:20,560 --> 00:46:20,880
Speaker 5: Yeah.

806
00:46:20,920 --> 00:46:24,000
Speaker 2: Absolutely, the GPU backed loans, Yeah.

807
00:46:23,800 --> 00:46:27,279
Speaker 3: Well, both the GPU backed loans and the opportunity that

808
00:46:27,280 --> 00:46:33,319
Speaker 3: that affords company like Magnetar to make GPU capacity in

809
00:46:33,400 --> 00:46:37,360
Speaker 3: lieu of cash for equity investments is extremely interesting. And

810
00:46:37,480 --> 00:46:39,880
Speaker 3: so and then you get this second order effect. So A,

811
00:46:40,480 --> 00:46:43,719
Speaker 3: you're providing something that other vcs can't because you are

812
00:46:43,719 --> 00:46:46,279
Speaker 3: giving them access to compute on day one. And then

813
00:46:46,440 --> 00:46:49,879
Speaker 3: b other vcs want to enter that deal because they

814
00:46:49,920 --> 00:46:52,600
Speaker 3: know that they're going to be investing in a company

815
00:46:52,880 --> 00:46:54,840
Speaker 3: that is not going to be have to scrambling for

816
00:46:55,000 --> 00:46:56,960
Speaker 3: compute once they get that VC cash.

817
00:46:57,280 --> 00:46:59,880
Speaker 2: It's a very sort of middle way approach because I

818
00:47:00,080 --> 00:47:03,799
Speaker 2: think so far the way we've seen AI investment unfold

819
00:47:04,040 --> 00:47:06,480
Speaker 2: is either it's the sort of picks and shovels approach

820
00:47:06,520 --> 00:47:09,719
Speaker 2: where you invest in the chip companies themselves and the

821
00:47:09,800 --> 00:47:13,440
Speaker 2: data centers, or it's you invest in the AI companies

822
00:47:13,440 --> 00:47:15,839
Speaker 2: that are doing cool things. But this is kind of both.

823
00:47:16,080 --> 00:47:18,799
Speaker 3: It is exactly both, and it sort of sounds like

824
00:47:18,880 --> 00:47:23,360
Speaker 3: some combination of foresightful planning and also stumbling into a

825
00:47:23,440 --> 00:47:27,799
Speaker 3: very good situation by which the firm's relationship with core Weave,

826
00:47:27,920 --> 00:47:32,120
Speaker 3: dating all the way back to twenty twenty one, does

827
00:47:32,239 --> 00:47:36,160
Speaker 3: now give them this a certain edge in the VCR.

828
00:47:36,320 --> 00:47:39,359
Speaker 3: It's just a really it's this is a fascinating sort

829
00:47:39,360 --> 00:47:41,360
Speaker 3: of open frontier in many respects.

830
00:47:41,440 --> 00:47:43,040
Speaker 2: I still want to know who came up with the

831
00:47:43,120 --> 00:47:46,560
Speaker 2: idea for chip based financing. Jim kind of evaded that

832
00:47:46,680 --> 00:47:48,440
Speaker 2: part of the question, but I want to know what

833
00:47:48,480 --> 00:47:50,080
Speaker 2: those initial conversations were Like.

834
00:47:50,320 --> 00:47:52,879
Speaker 3: Yeah, it's also just interesting to think about that on

835
00:47:52,920 --> 00:47:56,920
Speaker 3: some level, the analogies are like an Irish car lender, right,

836
00:47:57,040 --> 00:47:59,640
Speaker 3: So it's like, on some level this is a very

837
00:47:59,760 --> 00:48:04,480
Speaker 3: none and with technology that is highly uncertain. And then

838
00:48:04,640 --> 00:48:07,400
Speaker 3: on the other hand, if you're invested in a Carlo

839
00:48:07,480 --> 00:48:08,759
Speaker 3: and Company, you could sort of get it.

840
00:48:09,000 --> 00:48:10,480
Speaker 2: Yeah, all right, shall we leave it there.

841
00:48:10,600 --> 00:48:11,319
Speaker 5: Let's leave it there.

842
00:48:11,480 --> 00:48:14,560
Speaker 2: This has been another episode of the Oudlots podcast. I'm

843
00:48:14,600 --> 00:48:17,560
Speaker 2: Tracy Alloway. You can follow me at Tracy Alloway and.

844
00:48:17,520 --> 00:48:20,080
Speaker 3: I'm Joe Wisenthal. You can follow me at the Stalwart.

845
00:48:20,280 --> 00:48:23,640
Speaker 3: Follow our producers Kerman Rodriguez at Kerman armand dash Ol

846
00:48:23,640 --> 00:48:26,880
Speaker 3: Bennett at Dashbot and kill Brooks at Kilbrooks. Thank you

847
00:48:26,920 --> 00:48:29,960
Speaker 3: to our producer Moses Onam. From our Oddlots content, go

848
00:48:30,000 --> 00:48:32,799
Speaker 3: to Bloomberg dot com slash odd Lots, where we have transcripts,

849
00:48:32,800 --> 00:48:35,399
Speaker 3: a blog, and a daily newsletter and you can chet

850
00:48:35,400 --> 00:48:38,840
Speaker 3: about all of these topics, including AI twenty four seven

851
00:48:38,920 --> 00:48:41,720
Speaker 3: in our discord. Go there and check it out Discord

852
00:48:41,760 --> 00:48:42,640
Speaker 3: dot gg.

853
00:48:42,640 --> 00:48:46,040
Speaker 2: Slash odlocks And if you enjoy ad Blots, if you

854
00:48:46,160 --> 00:48:49,080
Speaker 2: like it when we dig into the capital structure of

855
00:48:49,160 --> 00:48:52,520
Speaker 2: AI investments, then please leave us a positive review on

856
00:48:52,600 --> 00:48:56,080
Speaker 2: your favorite podcast platform. And remember, if you are a

857
00:48:56,120 --> 00:49:00,000
Speaker 2: Bloomberg subscriber, you can listen to all of our episodes absolutely.

858
00:49:00,080 --> 00:49:02,200
Speaker 2: The ad free. All you need to do is find

859
00:49:02,200 --> 00:49:06,000
Speaker 2: the Bloomberg channel on Apple Podcasts and follow the instructions there.

860
00:49:06,480 --> 00:49:07,320
Speaker 2: Thanks for listening.