1
00:00:03,120 --> 00:00:12,680
Speaker 1: Bloomberg Audio Studios, Podcasts, radio News.

2
00:00:20,040 --> 00:00:24,040
Speaker 2: Hello and welcome to another episode of the Odd Lots podcast.

3
00:00:24,120 --> 00:00:25,880
Speaker 2: I'm Joe Wisenthal.

4
00:00:25,400 --> 00:00:26,520
Speaker 1: And I'm Tracy Alloway.

5
00:00:26,920 --> 00:00:29,640
Speaker 2: Tracy, you know, we've done tons of course on like

6
00:00:29,800 --> 00:00:34,559
Speaker 2: electricity and AI and data centers and all that stuff,

7
00:00:34,880 --> 00:00:38,320
Speaker 2: but we've never actually done like a well, we've never

8
00:00:38,440 --> 00:00:41,560
Speaker 2: talked to someone who is building data centers.

9
00:00:42,120 --> 00:00:43,479
Speaker 1: Putting it all together, you mean.

10
00:00:43,440 --> 00:00:45,800
Speaker 2: Yeah, putting it all together like what you know, just

11
00:00:45,800 --> 00:00:47,880
Speaker 2: a bunch of you know, I've had consultants, so we

12
00:00:47,960 --> 00:00:50,920
Speaker 2: talked to energy people, but like, how does this business

13
00:00:51,560 --> 00:00:54,920
Speaker 2: of essentially, I guess, building a building, putting a bunch

14
00:00:54,960 --> 00:00:57,880
Speaker 2: of chips in there, getting the electricity, and then in theory,

15
00:00:58,440 --> 00:01:00,440
Speaker 2: selling all of that at a markup? Like, how does

16
00:01:00,480 --> 00:01:01,240
Speaker 2: it actually work?

17
00:01:01,520 --> 00:01:01,720
Speaker 3: You know?

18
00:01:01,760 --> 00:01:04,120
Speaker 1: What I was reading recently This is kind of a tangent,

19
00:01:04,240 --> 00:01:07,399
Speaker 1: but not really because we're talking about the physical and

20
00:01:07,760 --> 00:01:12,920
Speaker 1: financial process of building these things. But I saw this

21
00:01:12,959 --> 00:01:17,240
Speaker 1: is online. There's a guide to the like physical Planning

22
00:01:18,080 --> 00:01:23,200
Speaker 1: around an IBM system three sixty from like nineteen sixty

23
00:01:23,280 --> 00:01:27,399
Speaker 1: three or something, and it's two hundred and thirteen pages long.

24
00:01:27,560 --> 00:01:28,480
Speaker 2: Have you read it yet?

25
00:01:28,840 --> 00:01:32,160
Speaker 1: I did flip through it, there's like there's guidance on

26
00:01:32,240 --> 00:01:36,760
Speaker 1: minimizing vibrations obviously, like temperature and humidity and stuff like that.

27
00:01:36,800 --> 00:01:39,000
Speaker 1: I did not read the full two hundred pages, but

28
00:01:39,040 --> 00:01:42,560
Speaker 1: I'm kind of thinking like if this is what if

29
00:01:42,560 --> 00:01:45,640
Speaker 1: this is all the thinking that had to go into

30
00:01:45,720 --> 00:01:49,800
Speaker 1: like one computer, albeit a supercomputer in the nineteen sixties,

31
00:01:49,840 --> 00:01:52,120
Speaker 1: but like a pretty basic machine. When we look back

32
00:01:52,160 --> 00:01:54,960
Speaker 1: on it now, how much planning and thinking has to

33
00:01:55,000 --> 00:01:58,720
Speaker 1: go into building like these huge cloud servers and all

34
00:01:58,760 --> 00:02:02,760
Speaker 1: their associated infrastry structure, both physical and software as well.

35
00:02:03,080 --> 00:02:05,600
Speaker 2: No, totally, and you know, we you know, one of

36
00:02:05,640 --> 00:02:07,880
Speaker 2: the ways that we've touched on this subject a little

37
00:02:07,880 --> 00:02:11,200
Speaker 2: bit is in our conversations with Steve Eisman, who's been

38
00:02:11,360 --> 00:02:13,600
Speaker 2: investing at least as far as we know, in a

39
00:02:13,600 --> 00:02:19,079
Speaker 2: lot of these like industrial HVAC companies and electricity gear

40
00:02:19,240 --> 00:02:22,280
Speaker 2: companies and stuff like that. So like companies that have

41
00:02:22,320 --> 00:02:25,280
Speaker 2: actually been around for a really long time, sort of

42
00:02:25,280 --> 00:02:29,000
Speaker 2: standard cyclical businesses, and then they've like caught the secular

43
00:02:29,160 --> 00:02:32,960
Speaker 2: tailwind because with this boom in AI data center construction,

44
00:02:33,560 --> 00:02:36,520
Speaker 2: suddenly there's this sort of continuous bid for all their

45
00:02:36,520 --> 00:02:37,640
Speaker 2: gear and services.

46
00:02:37,760 --> 00:02:41,600
Speaker 1: I'm going to start an anti vibration floor maker or something.

47
00:02:41,600 --> 00:02:43,560
Speaker 1: Do you think that's a viable business? Does anyone care

48
00:02:43,560 --> 00:02:44,840
Speaker 1: about vibrations anymore?

49
00:02:44,880 --> 00:02:48,600
Speaker 2: I am certain that in various high tech environments you

50
00:02:48,680 --> 00:02:50,960
Speaker 2: do not want to have vibrations. You know, you have

51
00:02:51,040 --> 00:02:53,240
Speaker 2: like a valuable chips, you don't want them to be

52
00:02:53,280 --> 00:02:54,639
Speaker 2: like degrading.

53
00:02:54,280 --> 00:02:56,080
Speaker 1: Because people are walking around.

54
00:02:55,919 --> 00:02:58,280
Speaker 2: Yeah, or just you know what, all the machine and

55
00:02:58,320 --> 00:03:01,560
Speaker 2: all your air conditioners and equipment and all that stuff,

56
00:03:01,600 --> 00:03:03,080
Speaker 2: you can't be having that stuff degrade.

57
00:03:03,200 --> 00:03:06,160
Speaker 1: Well, the other interesting thing that's happening in the space now,

58
00:03:06,240 --> 00:03:10,200
Speaker 1: So in addition to the physical challenge of building a

59
00:03:10,200 --> 00:03:13,600
Speaker 1: bunch of this stuff, there's also the financial aspect of it.

60
00:03:13,720 --> 00:03:17,400
Speaker 1: And I guess as AI becomes more and more of

61
00:03:17,440 --> 00:03:19,600
Speaker 1: a thing, and clearly, as you laid out, there's a

62
00:03:19,639 --> 00:03:22,400
Speaker 1: lot of enthusiasm around the space. At the moment, you

63
00:03:22,480 --> 00:03:26,520
Speaker 1: are seeing a bunch of financial entities get interested as well.

64
00:03:26,600 --> 00:03:29,799
Speaker 1: So obviously venture capital has been pouring money into the space,

65
00:03:29,840 --> 00:03:33,440
Speaker 1: but we're starting to see some new types of financial

66
00:03:33,520 --> 00:03:37,560
Speaker 1: investments in AI. And I'm thinking about one thing in particular,

67
00:03:37,960 --> 00:03:42,360
Speaker 1: and it is the recent GPU or chip backed loan

68
00:03:42,840 --> 00:03:46,320
Speaker 1: that was reported by the Wall Street Journal and I

69
00:03:46,360 --> 00:03:48,080
Speaker 1: think we should talk about that aspect.

70
00:03:47,760 --> 00:03:50,080
Speaker 2: Of it too, totally, because one of the things that's

71
00:03:50,200 --> 00:03:53,960
Speaker 2: happening in tech is this big sort of shift from like, okay,

72
00:03:54,000 --> 00:03:56,200
Speaker 2: we're all of your costs in the past, where a

73
00:03:56,200 --> 00:03:58,720
Speaker 2: lot of them were sort of op x, the cost

74
00:03:58,720 --> 00:04:02,200
Speaker 2: of engineers, et cetera. And now suddenly tech companies have

75
00:04:02,240 --> 00:04:04,680
Speaker 2: to think about CAPEX for the first time, these big

76
00:04:04,800 --> 00:04:06,720
Speaker 2: upfront costs that are in theory going to pay off

77
00:04:06,760 --> 00:04:09,480
Speaker 2: for a long time, which in theory then changes how

78
00:04:09,520 --> 00:04:11,080
Speaker 2: you should think about the financing model.

79
00:04:11,160 --> 00:04:12,880
Speaker 1: Absolutely well, I am.

80
00:04:12,840 --> 00:04:16,640
Speaker 2: Excited to say because we literally do have the perfect

81
00:04:16,720 --> 00:04:20,000
Speaker 2: guest we're going to be speaking with, Brian Venturo. He

82
00:04:20,080 --> 00:04:24,120
Speaker 2: is the chief strategy officer at core Weave. Corewave. For

83
00:04:24,160 --> 00:04:27,480
Speaker 2: those who don't know, it's probably the company right now

84
00:04:27,760 --> 00:04:31,120
Speaker 2: that people most associate with being at the heart of

85
00:04:31,200 --> 00:04:35,359
Speaker 2: the AI data center boom. They have a bunch of

86
00:04:35,360 --> 00:04:38,440
Speaker 2: in video chips, they have investments from in Nvidio right

87
00:04:38,560 --> 00:04:41,279
Speaker 2: here in the sweet spot. As you mentioned, one of

88
00:04:41,320 --> 00:04:44,200
Speaker 2: the interesting things that's going on is they not long

89
00:04:44,240 --> 00:04:49,440
Speaker 2: ago announced a debt financing facility sit back basically by

90
00:04:49,480 --> 00:04:53,679
Speaker 2: the GPUs that they would acquire, so literally the perfect

91
00:04:53,680 --> 00:04:58,920
Speaker 2: person to understand like the business of these AI cloud

92
00:04:59,360 --> 00:05:02,080
Speaker 2: data center So Brian, thank you so much for coming in.

93
00:05:02,360 --> 00:05:04,080
Speaker 3: Thanks for having me. It's the second time I've been

94
00:05:04,080 --> 00:05:04,760
Speaker 3: on the podcast.

95
00:05:05,040 --> 00:05:08,200
Speaker 2: That's right. We talked to Brian years ago. It's interesting

96
00:05:08,440 --> 00:05:10,920
Speaker 2: to think about at that time because I think that

97
00:05:11,040 --> 00:05:13,320
Speaker 2: may have been like twenty twenty or twenty one, and

98
00:05:13,360 --> 00:05:16,000
Speaker 2: the excitement then was that these chips could be used

99
00:05:16,200 --> 00:05:19,520
Speaker 2: for crypto mining and other things like sort of distributed

100
00:05:19,600 --> 00:05:23,520
Speaker 2: video editing and stuff like that, and then Ethereum stopped

101
00:05:23,600 --> 00:05:26,320
Speaker 2: using mining. But it was sort of fortuitous timing because

102
00:05:26,400 --> 00:05:30,040
Speaker 2: right around then AI went crazy and that's probably I

103
00:05:30,080 --> 00:05:31,960
Speaker 2: don't know, in my view, maybe a higher use of

104
00:05:32,000 --> 00:05:34,479
Speaker 2: these chips before we get to that. Do you worry

105
00:05:34,520 --> 00:05:37,920
Speaker 2: about vibration in your data center?

106
00:05:38,120 --> 00:05:42,760
Speaker 3: So everywhere that's close to a fault line is designed

107
00:05:42,760 --> 00:05:45,440
Speaker 3: around that and is part of code. So you know,

108
00:05:45,440 --> 00:05:47,800
Speaker 3: the engineering firms that help us build these data centers

109
00:05:47,800 --> 00:05:49,680
Speaker 3: have taken all of that into account, and all of

110
00:05:49,720 --> 00:05:54,600
Speaker 3: our racks are you know, seismically tuned to make sure

111
00:05:54,600 --> 00:05:57,239
Speaker 3: that we can withstand the normal vibration from the Earth.

112
00:05:57,640 --> 00:06:00,360
Speaker 3: So yeah, it's been something that's been in those annuals

113
00:06:00,360 --> 00:06:03,360
Speaker 3: for a long time. Some of our hardwer manufacturers actually

114
00:06:03,360 --> 00:06:06,479
Speaker 3: have vibration testing labs where they put the racks on

115
00:06:06,520 --> 00:06:08,839
Speaker 3: top of a big kind of platform that shakes, and

116
00:06:08,880 --> 00:06:12,240
Speaker 3: it's pretty dangerous and uncontrollable and hard to watch. But

117
00:06:12,720 --> 00:06:14,239
Speaker 3: you know, there's people out there that have been solving

118
00:06:14,240 --> 00:06:15,080
Speaker 3: this problem for decades.

119
00:06:15,120 --> 00:06:17,800
Speaker 1: Now I missed the boat on that business choir. It

120
00:06:17,920 --> 00:06:22,400
Speaker 1: sounds like it's been dealt with decades ago. Okay, well, actually,

121
00:06:22,760 --> 00:06:25,760
Speaker 1: why didn't I start with a very simple question, which

122
00:06:25,839 --> 00:06:29,400
Speaker 1: is when when you're looking at the business of core Weave,

123
00:06:29,880 --> 00:06:34,560
Speaker 1: so a specialized cloud service provider, let's put it that way,

124
00:06:34,960 --> 00:06:38,440
Speaker 1: what are the different components that you have to think about?

125
00:06:38,480 --> 00:06:41,040
Speaker 1: You know, Joe kind of alluded to all these different

126
00:06:41,240 --> 00:06:44,200
Speaker 1: ingredients that go into the business, but walk us through

127
00:06:44,360 --> 00:06:45,680
Speaker 1: what those actually are.

128
00:06:46,120 --> 00:06:49,120
Speaker 3: Sure, so, there's there's three pieces that as a management team,

129
00:06:49,120 --> 00:06:52,680
Speaker 3: we think are incredibly critical to the business. The first is,

130
00:06:53,240 --> 00:06:55,479
Speaker 3: you know, our technology services that we provide on top

131
00:06:55,520 --> 00:06:57,880
Speaker 3: of the hardware, right and this is everything from the

132
00:06:57,920 --> 00:07:00,480
Speaker 3: software layer through the support organization to you know, how

133
00:07:00,480 --> 00:07:03,000
Speaker 3: we work with our customers. This isn't the type of

134
00:07:03,040 --> 00:07:04,960
Speaker 3: thing that you just go plug in and it works.

135
00:07:05,600 --> 00:07:08,279
Speaker 3: In these large supercomputer clusters, there may be two hundred

136
00:07:08,360 --> 00:07:12,080
Speaker 3: thousand infinibank connections that connect all the GPUs together, and

137
00:07:12,120 --> 00:07:15,120
Speaker 3: if one of those connections fails for whatever reason, the

138
00:07:15,280 --> 00:07:18,040
Speaker 3: job will completely stop and have to restart from its

139
00:07:18,080 --> 00:07:20,720
Speaker 3: previous checkpoints. So, you know, everything that we do on

140
00:07:20,760 --> 00:07:22,800
Speaker 3: the software side and engineering side is to make sure

141
00:07:22,800 --> 00:07:25,800
Speaker 3: these clusters are as resilient and performant as they possibly

142
00:07:25,840 --> 00:07:28,680
Speaker 3: can be to ensure you know, our customers can run

143
00:07:28,720 --> 00:07:33,240
Speaker 3: their jobs, you know, increase efficiency and get all of

144
00:07:33,320 --> 00:07:36,000
Speaker 3: the kind of monetary value they can out of the chips.

145
00:07:36,920 --> 00:07:39,520
Speaker 3: So technology piece is really hard. It's something that I

146
00:07:39,560 --> 00:07:42,440
Speaker 3: think is very overlooked by the market, but it's just

147
00:07:42,480 --> 00:07:44,520
Speaker 3: as hard as the two other kind of pieces that

148
00:07:44,520 --> 00:07:47,560
Speaker 3: this business stands on. The second is, you know, the

149
00:07:47,600 --> 00:07:49,680
Speaker 3: physical nature of the business in that you have to

150
00:07:49,720 --> 00:07:52,680
Speaker 3: actually build and run these data centers and those hundreds

151
00:07:52,680 --> 00:07:56,160
Speaker 3: of thousands connections inside the supercomputers. Like somebody has to

152
00:07:56,160 --> 00:07:58,640
Speaker 3: go put those together and make sure they're clean and

153
00:07:58,680 --> 00:08:01,200
Speaker 3: make sure they're labeled correctly to be able to remediate failures.

154
00:08:01,720 --> 00:08:05,600
Speaker 3: And when you're building a thirty two thousand GPU supercomputer

155
00:08:05,880 --> 00:08:08,800
Speaker 3: that is one of the fastest three computers in the planet.

156
00:08:09,800 --> 00:08:13,360
Speaker 3: You know, you're running thousands of miles of cable inside

157
00:08:13,400 --> 00:08:16,520
Speaker 3: a very dense space, right. These data centers are built

158
00:08:16,640 --> 00:08:19,480
Speaker 3: very tiny to make sure that you can connect everything together,

159
00:08:20,160 --> 00:08:22,880
Speaker 3: and that becomes a huge logistical challenge. So, you know,

160
00:08:22,920 --> 00:08:25,120
Speaker 3: the data centerpiece, which we're going to talk more about today,

161
00:08:25,880 --> 00:08:28,440
Speaker 3: is very challenging to design for the use case. And

162
00:08:28,520 --> 00:08:30,080
Speaker 3: then the third piece is how the hell do you

163
00:08:30,200 --> 00:08:32,959
Speaker 3: finance the whole thing? Right, And you know, we've been

164
00:08:33,480 --> 00:08:36,800
Speaker 3: very successful in the financing aspect of this, but you know,

165
00:08:36,840 --> 00:08:39,880
Speaker 3: whether you're financing technology operations or the physical build of

166
00:08:39,920 --> 00:08:43,880
Speaker 3: these things, it is an incredibly capital intensive business and

167
00:08:44,240 --> 00:08:48,320
Speaker 3: constructing those financial instruments to back our business is very hard,

168
00:08:48,360 --> 00:08:51,120
Speaker 3: and we have to be very very thoughtful around who

169
00:08:51,120 --> 00:08:53,360
Speaker 3: the counterparties are, how do we think about credit risk,

170
00:08:53,679 --> 00:08:56,160
Speaker 3: how do our investors think about that credit risk, How

171
00:08:56,160 --> 00:08:59,319
Speaker 3: do we deal with contingencies inside the contracts to make

172
00:08:59,360 --> 00:09:01,160
Speaker 3: sure that they are financeable on the scale that we've

173
00:09:01,160 --> 00:09:02,400
Speaker 3: done over the last eighteen months.

174
00:09:02,640 --> 00:09:04,880
Speaker 2: Talk to us a little bit more. We could probably

175
00:09:05,120 --> 00:09:09,319
Speaker 2: talk about data center financing credit and have have that

176
00:09:09,440 --> 00:09:12,520
Speaker 2: be a whole episode, but when you think about you

177
00:09:12,640 --> 00:09:16,079
Speaker 2: have to think about your counter party's credit risk. Talk

178
00:09:16,080 --> 00:09:18,160
Speaker 2: to us a little bit about what you're who those are,

179
00:09:18,320 --> 00:09:20,080
Speaker 2: what the type of entity is.

180
00:09:20,400 --> 00:09:22,800
Speaker 3: Sure, so I'll get myself in trouble if I just

181
00:09:22,840 --> 00:09:25,400
Speaker 3: start naming them off. Yeah, some of them are more

182
00:09:25,400 --> 00:09:28,559
Speaker 3: public than others. You know, I'm going to refer to

183
00:09:28,600 --> 00:09:33,040
Speaker 3: them as you know, hyperscale customers. We have AI lab customers,

184
00:09:33,679 --> 00:09:37,800
Speaker 3: we have large enterprise customers. We've really constructed our portfolio

185
00:09:37,800 --> 00:09:40,600
Speaker 3: of business around the idea that you know, if we're

186
00:09:40,600 --> 00:09:42,960
Speaker 3: going to build ten billion dollars of infrastructure for somebody,

187
00:09:43,000 --> 00:09:44,640
Speaker 3: we have to know there's a balance sheet we can

188
00:09:44,720 --> 00:09:48,800
Speaker 3: lean into behind it, right, and we're the pace at

189
00:09:48,840 --> 00:09:53,320
Speaker 3: which we've grown. You know, our customers are demanding scale

190
00:09:53,440 --> 00:09:57,320
Speaker 3: so quickly that the credit of the counterparty is incredibly

191
00:09:57,360 --> 00:09:59,400
Speaker 3: important to find the low cost of capital we have

192
00:09:59,480 --> 00:10:02,120
Speaker 3: with these ADIT facilities we've announced, right, So you know,

193
00:10:02,160 --> 00:10:04,520
Speaker 3: when people talk about how this is a credit facility

194
00:10:04,520 --> 00:10:07,360
Speaker 3: backed by GPUs, it's not really backed by GPUs. It's

195
00:10:07,400 --> 00:10:12,040
Speaker 3: backed by you know, commercial contracts with large international enterprises

196
00:10:12,080 --> 00:10:14,480
Speaker 3: that may have triple a credit, right, So you know

197
00:10:14,520 --> 00:10:15,600
Speaker 3: it's it's the framing of the.

198
00:10:15,679 --> 00:10:17,240
Speaker 1: Aid receivables finance.

199
00:10:17,280 --> 00:10:20,400
Speaker 3: Basically it's closer to trade receivables financing than it is Hey,

200
00:10:20,400 --> 00:10:22,200
Speaker 3: we're going to go leverage up a bunch of GPUs

201
00:10:22,240 --> 00:10:22,959
Speaker 3: and see what happens.

202
00:10:23,080 --> 00:10:26,480
Speaker 1: Huh, okay, well walk us through the I guess like

203
00:10:26,760 --> 00:10:31,440
Speaker 1: the sequence in some of these financing agreements. So you know,

204
00:10:31,480 --> 00:10:33,959
Speaker 1: if a customer comes to you and they say, we

205
00:10:34,040 --> 00:10:37,280
Speaker 1: want a certain amount of compute, can you do this

206
00:10:37,400 --> 00:10:40,800
Speaker 1: for us? And you start going down the process of like, okay,

207
00:10:40,840 --> 00:10:43,320
Speaker 1: what do we need to make this happen? What do

208
00:10:43,440 --> 00:10:47,440
Speaker 1: those like financial agreements actually look like. And who's bearing

209
00:10:47,480 --> 00:10:50,200
Speaker 1: the initial risk? Is it the customer? Is it you?

210
00:10:51,360 --> 00:10:51,880
Speaker 2: Good question?

211
00:10:52,000 --> 00:10:54,080
Speaker 3: So when we're approached by a customer, right, you know,

212
00:10:54,160 --> 00:10:57,360
Speaker 3: the ask is typically going to be pretty pretty general,

213
00:10:57,840 --> 00:11:00,360
Speaker 3: and they're going to say, hey, we're looking for facity

214
00:11:00,400 --> 00:11:02,080
Speaker 3: in Q one of next year. What's the largest thing

215
00:11:02,120 --> 00:11:06,200
Speaker 3: you can do? And you know, we take that effectively

216
00:11:06,240 --> 00:11:08,960
Speaker 3: as a mandate of okay, hey, you know this customer.

217
00:11:08,679 --> 00:11:09,199
Speaker 2: We're not business.

218
00:11:09,240 --> 00:11:11,080
Speaker 3: But before you know, we're really comfortable with them, we

219
00:11:11,120 --> 00:11:12,840
Speaker 3: know that we're going to get a contract done. We'll

220
00:11:12,840 --> 00:11:14,719
Speaker 3: go out and we'll try to secure an asset to

221
00:11:15,080 --> 00:11:16,240
Speaker 3: you know, to go build it. And we may have

222
00:11:16,280 --> 00:11:18,160
Speaker 3: it in our portfolio already. We maybe it may have

223
00:11:18,200 --> 00:11:20,720
Speaker 3: been a strategic investment that we made. But once we

224
00:11:20,760 --> 00:11:22,360
Speaker 3: find the data center asset, that's when we go back

225
00:11:22,360 --> 00:11:24,160
Speaker 3: to the customer and say, okay, like we can commit

226
00:11:24,200 --> 00:11:26,680
Speaker 3: to doing this. This is the timeline. We'll structure a

227
00:11:26,679 --> 00:11:29,600
Speaker 3: contract around it. Depending upon who the customer is. There

228
00:11:29,640 --> 00:11:31,640
Speaker 3: may or may not be some credit support associated with

229
00:11:31,679 --> 00:11:34,280
Speaker 3: it around the scaling of the you know, that asset,

230
00:11:35,080 --> 00:11:38,200
Speaker 3: and then we'll get a commercial contract in place, and

231
00:11:38,760 --> 00:11:42,600
Speaker 3: we will initially fund a large portion of that project

232
00:11:42,720 --> 00:11:44,800
Speaker 3: off of our own balance sheet. Right. It's why you

233
00:11:44,840 --> 00:11:47,160
Speaker 3: also see us raising equity, right, is we have to

234
00:11:47,320 --> 00:11:49,800
Speaker 3: have the capital to accelerate the business. And then once

235
00:11:49,840 --> 00:11:52,439
Speaker 3: we have that and we're making progress, you know, think

236
00:11:52,440 --> 00:11:54,360
Speaker 3: about it as you're building real estate. Right, you have

237
00:11:54,400 --> 00:11:56,520
Speaker 3: a construction loan and then you have a stabilized asset loan,

238
00:11:57,040 --> 00:11:59,439
Speaker 3: and we basically fund the construction loan piece off of

239
00:11:59,440 --> 00:12:01,880
Speaker 3: our balance sheet. When we get to a more stabilized asset,

240
00:12:01,880 --> 00:12:03,679
Speaker 3: that's when we go out and kind of do that

241
00:12:03,800 --> 00:12:07,720
Speaker 3: trade financing or trade receivables financing our with our partner lenders.

242
00:12:08,160 --> 00:12:09,760
Speaker 3: You know, they worked with us before, they know that

243
00:12:09,800 --> 00:12:11,040
Speaker 3: these things are going to stand up, They know how

244
00:12:11,040 --> 00:12:13,480
Speaker 3: they perform, and at that point in time, it's it's

245
00:12:13,480 --> 00:12:21,760
Speaker 3: pretty easy for them to underwrite that risk.

246
00:12:31,160 --> 00:12:35,479
Speaker 2: It's funny. Tracy and I had coffee with someone yesterday

247
00:12:36,160 --> 00:12:39,000
Speaker 2: who is sort of in the space I want docs here,

248
00:12:39,080 --> 00:12:40,560
Speaker 2: And I was like, what should we ask Brian? And

249
00:12:40,600 --> 00:12:42,880
Speaker 2: he's like, ask him why he won't let my company,

250
00:12:43,440 --> 00:12:45,240
Speaker 2: why I'm still on the waiting list or something, or

251
00:12:45,280 --> 00:12:48,480
Speaker 2: why he hasn't approved my company to use core weave.

252
00:12:48,679 --> 00:12:51,679
Speaker 2: But what are some of the bars or the threshold?

253
00:12:51,760 --> 00:12:54,280
Speaker 2: So you know, I apparently there's a lot of demand

254
00:12:54,440 --> 00:12:57,360
Speaker 2: for compute these days. What does it take to get

255
00:12:57,360 --> 00:12:59,880
Speaker 2: in the door and get access to some of your

256
00:13:00,200 --> 00:13:01,319
Speaker 2: chips and electricity?

257
00:13:01,840 --> 00:13:05,000
Speaker 3: So it's it's a great question. It's a question that

258
00:13:05,040 --> 00:13:07,240
Speaker 3: we get all the time from our sales teams, right

259
00:13:07,400 --> 00:13:09,760
Speaker 3: is you know, we're faced a lot with a sales

260
00:13:09,840 --> 00:13:13,400
Speaker 3: team that is incredible at delivering product to customer and

261
00:13:13,440 --> 00:13:15,880
Speaker 3: we don't have anything to sell. And it's kind of

262
00:13:15,880 --> 00:13:19,760
Speaker 3: my job. As the strategy organization at Core, We've were

263
00:13:19,800 --> 00:13:24,360
Speaker 3: responsible for two things. It's product and infrastructure. Capacity, and

264
00:13:24,559 --> 00:13:26,280
Speaker 3: you know, I spend most of my time going out

265
00:13:26,280 --> 00:13:28,160
Speaker 3: and finding those data centers and being able to support

266
00:13:28,200 --> 00:13:31,040
Speaker 3: those deals and the growth that we had over the

267
00:13:31,040 --> 00:13:34,600
Speaker 3: past twelve months. The company was pretty flat out right

268
00:13:34,679 --> 00:13:37,760
Speaker 3: in building and delivering this infrastructure. You know, publicly on

269
00:13:37,760 --> 00:13:40,400
Speaker 3: our documentation page it says that we have three regions.

270
00:13:40,600 --> 00:13:42,360
Speaker 3: We'll have twenty eight regions online by the end of

271
00:13:42,400 --> 00:13:44,439
Speaker 3: the year. I think we delivered eleven of them in

272
00:13:44,559 --> 00:13:47,320
Speaker 3: Q one alone, Right, So we're building at a scale,

273
00:13:48,080 --> 00:13:50,160
Speaker 3: you know, i'd say that almost larger than some of

274
00:13:50,200 --> 00:13:54,120
Speaker 3: the three big hyperscalers. But in terms of how do

275
00:13:54,200 --> 00:13:56,959
Speaker 3: you become a customer of Core, it's really relationship driven,

276
00:13:57,160 --> 00:13:59,040
Speaker 3: right is. We want to make sure that we're going

277
00:13:59,120 --> 00:14:00,880
Speaker 3: to be able to be successfu with our customers and

278
00:14:00,920 --> 00:14:03,679
Speaker 3: have an engineering relationship and we're aligned on what they

279
00:14:03,720 --> 00:14:04,079
Speaker 3: need and.

280
00:14:04,040 --> 00:14:05,120
Speaker 2: We can deliver what they need.

281
00:14:05,800 --> 00:14:07,240
Speaker 3: The last thing that we want is for somebody to

282
00:14:07,280 --> 00:14:09,480
Speaker 3: walk in the door and say, hey, I need this

283
00:14:09,520 --> 00:14:12,080
Speaker 3: for three weeks and two weeks into it, they're unhappy

284
00:14:12,640 --> 00:14:14,720
Speaker 3: and we can't give them what they need to be successful.

285
00:14:14,800 --> 00:14:17,520
Speaker 3: Right is, you know, our customers are making such large

286
00:14:17,600 --> 00:14:20,760
Speaker 3: investments in this infrastructure, that we have to have, you know,

287
00:14:20,960 --> 00:14:23,720
Speaker 3: a lot of conviction that we will be successful with

288
00:14:23,760 --> 00:14:26,880
Speaker 3: them and provide a good experience. So it's not that

289
00:14:26,960 --> 00:14:29,240
Speaker 3: we're trying to keep people out, it's we're trying to

290
00:14:29,320 --> 00:14:31,960
Speaker 3: ensure positive experiences for people that we do bring on board.

291
00:14:32,320 --> 00:14:37,000
Speaker 2: Do you build complete housed facilities or is it all

292
00:14:37,480 --> 00:14:40,720
Speaker 2: you're going to bring your chips and expertise into an

293
00:14:40,760 --> 00:14:44,200
Speaker 2: existing Tier one data center and essentially rent floor space

294
00:14:44,240 --> 00:14:44,640
Speaker 2: from them.

295
00:14:44,760 --> 00:14:47,640
Speaker 3: Yeah, so a year ago it was we were effectively

296
00:14:47,640 --> 00:14:51,000
Speaker 3: just a co location tenant, and now we've gone a

297
00:14:51,040 --> 00:14:54,760
Speaker 3: lot more vertical for some strategic builds where we're either

298
00:14:54,760 --> 00:14:56,640
Speaker 3: a partner in the project where we own equity and

299
00:14:56,680 --> 00:15:00,120
Speaker 3: the development company, or we're building the project ourselves. We've

300
00:15:00,120 --> 00:15:02,040
Speaker 3: been scaling that team up over the past six months,

301
00:15:02,480 --> 00:15:04,800
Speaker 3: and we had to at our scale to be able

302
00:15:04,800 --> 00:15:07,400
Speaker 3: to guarantee outcomes. Right, is, we were in a position

303
00:15:07,440 --> 00:15:09,600
Speaker 3: where we had data centers getting delayed with things that

304
00:15:09,640 --> 00:15:12,560
Speaker 3: weren't communicated to us, and you know, we had to

305
00:15:12,560 --> 00:15:15,920
Speaker 3: go build the capability to handle that situation and you know,

306
00:15:15,960 --> 00:15:17,640
Speaker 3: make sure we can still deliver for our customers.

307
00:15:17,880 --> 00:15:20,400
Speaker 1: One of the differentiators that you and some of your

308
00:15:20,400 --> 00:15:23,800
Speaker 1: colleagues have emphasized previously, is this idea that you're designing

309
00:15:24,320 --> 00:15:27,680
Speaker 1: the server clusters kind of from the ground up, whereas

310
00:15:27,760 --> 00:15:31,880
Speaker 1: like other hyperscalers maybe are doing it on a sort

311
00:15:31,880 --> 00:15:34,800
Speaker 1: of different mass scale. But can you walk us through

312
00:15:34,840 --> 00:15:38,760
Speaker 1: like what is the benefit of doing it that way?

313
00:15:38,840 --> 00:15:42,640
Speaker 1: And then secondly, does that end up being an impediment

314
00:15:43,040 --> 00:15:47,640
Speaker 1: to I guess efficiencies or economics of scale and how

315
00:15:47,680 --> 00:15:49,560
Speaker 1: customized Like do you really get here?

316
00:15:49,960 --> 00:15:54,080
Speaker 3: So from a customization perspective, it's aggressive, right, And I

317
00:15:54,200 --> 00:15:56,960
Speaker 3: say that because you know, our customers are involved in

318
00:15:57,040 --> 00:15:59,200
Speaker 3: the design of you know, our network topology of the

319
00:15:59,200 --> 00:16:02,400
Speaker 3: East West fabric for the GPU to GPU communication, for

320
00:16:02,520 --> 00:16:04,920
Speaker 3: things like cooling. You know, I have customers that toward

321
00:16:04,920 --> 00:16:07,680
Speaker 3: the data centers under construction process with me like once

322
00:16:07,720 --> 00:16:13,640
Speaker 3: a week, and it's to the point that they're impacting

323
00:16:14,120 --> 00:16:17,320
Speaker 3: how we build the base level networking products to ensure

324
00:16:17,400 --> 00:16:20,080
Speaker 3: they have enough throughput to you know, meet their use

325
00:16:20,120 --> 00:16:23,240
Speaker 3: case needs. Whereas in you know, what I what we

326
00:16:23,280 --> 00:16:27,960
Speaker 3: call the legacy hyperscaler installations, It maybe they have a

327
00:16:27,960 --> 00:16:30,080
Speaker 3: couple thousand GPUs that are in a data center that

328
00:16:30,120 --> 00:16:34,160
Speaker 3: was really built for CPU computation or to provide services

329
00:16:34,160 --> 00:16:37,280
Speaker 3: to ten thousand customers that is really with a much

330
00:16:37,360 --> 00:16:40,080
Speaker 3: lower base expectation of what they're going to be doing. Right,

331
00:16:40,160 --> 00:16:43,720
Speaker 3: So it's things around connectivity for storage, it's things around

332
00:16:43,760 --> 00:16:47,480
Speaker 3: power and cooling, It's things around how they want to

333
00:16:47,520 --> 00:16:51,360
Speaker 3: be able to optimize their workloads inside of the GPU

334
00:16:51,400 --> 00:16:53,680
Speaker 3: to GPU communication. You know, we have some customers that

335
00:16:53,720 --> 00:16:57,200
Speaker 3: even customize their infiniban fabrics and the size of those

336
00:16:57,240 --> 00:16:59,200
Speaker 3: fabrics and how they connect together. So you know, we

337
00:16:59,280 --> 00:17:01,240
Speaker 3: work with them to really understand what their use case is,

338
00:17:01,280 --> 00:17:03,440
Speaker 3: where they're worried currently and in the future, and then

339
00:17:03,480 --> 00:17:07,080
Speaker 3: design around that. So it's a pretty comprehensive program when

340
00:17:07,080 --> 00:17:09,080
Speaker 3: we're building something from the ground up.

341
00:17:09,200 --> 00:17:12,240
Speaker 1: And how much complexity does that introduce into the business

342
00:17:12,320 --> 00:17:15,320
Speaker 1: and does it end up being a limiting factor on

343
00:17:15,359 --> 00:17:18,199
Speaker 1: your growth or is demand just so strong at the

344
00:17:18,240 --> 00:17:20,280
Speaker 1: moment that it's not really an issue.

345
00:17:20,440 --> 00:17:22,480
Speaker 3: The customization that we do is typically going to be

346
00:17:22,520 --> 00:17:26,040
Speaker 3: above what our base level offering is, meaning the environment

347
00:17:26,080 --> 00:17:29,040
Speaker 3: will be more performant because the customer required it. So

348
00:17:29,200 --> 00:17:31,640
Speaker 3: it's typically not going to be limiting to us from

349
00:17:31,720 --> 00:17:34,840
Speaker 3: a future you know, revenue or resale perspective. It's going

350
00:17:34,880 --> 00:17:37,560
Speaker 3: to make the asset more valuable. But you know, we're

351
00:17:37,840 --> 00:17:40,440
Speaker 3: we're designing our reference builds for ninety nine percent of

352
00:17:40,520 --> 00:17:43,359
Speaker 3: use cases, and we're trying to price it efficiently, and

353
00:17:43,400 --> 00:17:45,879
Speaker 3: then when customer wants something above and beyond, you know,

354
00:17:45,920 --> 00:17:49,360
Speaker 3: it impacts price. But for these installations it's probably deminimus, right,

355
00:17:49,440 --> 00:17:51,800
Speaker 3: So you know, it doesn't really add a lot of

356
00:17:51,800 --> 00:17:54,919
Speaker 3: complexity for us from a business perspective, so we're happy

357
00:17:54,920 --> 00:17:55,239
Speaker 3: to do it.

358
00:17:55,600 --> 00:17:59,560
Speaker 2: You mentioned that some of the hyperscalers, yes they have GPUs,

359
00:17:59,600 --> 00:18:05,960
Speaker 2: but they like built in an environment for like legacy CPUs.

360
00:18:06,200 --> 00:18:08,960
Speaker 2: Can you talk a little bit about a just the

361
00:18:09,080 --> 00:18:12,879
Speaker 2: difference between the legacy architectures and the new one and

362
00:18:12,920 --> 00:18:15,520
Speaker 2: then in the design, like what kind of bottlenecks you

363
00:18:15,560 --> 00:18:18,320
Speaker 2: run into? Is there issues with labor like the types

364
00:18:18,359 --> 00:18:20,840
Speaker 2: of people who know how to string these things together well,

365
00:18:21,000 --> 00:18:24,960
Speaker 2: or other different cooling requirements for this type of compute

366
00:18:25,080 --> 00:18:28,439
Speaker 2: environment that did not exist, Like what are what are

367
00:18:28,480 --> 00:18:32,040
Speaker 2: the challenges in building out these sort of like fundamentally

368
00:18:32,720 --> 00:18:33,600
Speaker 2: different environments.

369
00:18:33,680 --> 00:18:36,000
Speaker 3: Yeah, so that that's changed also in the last twelve

370
00:18:36,040 --> 00:18:38,640
Speaker 3: months in that you used to be able to take

371
00:18:38,680 --> 00:18:41,560
Speaker 3: what was an enterprise data center and you know, creatively

372
00:18:41,600 --> 00:18:45,080
Speaker 3: retrofit it to be capable of supporting the AI workloads

373
00:18:45,440 --> 00:18:48,000
Speaker 3: to a certain density level. Okay, right, Like instead of

374
00:18:48,040 --> 00:18:49,800
Speaker 3: filling up a cabinet, you could put two servers in

375
00:18:49,840 --> 00:18:52,439
Speaker 3: a cabinet and you could meet the power and cooling

376
00:18:52,440 --> 00:18:55,840
Speaker 3: requirements of the installation. It you use a lot more

377
00:18:55,920 --> 00:18:59,360
Speaker 3: floor space, but it was doable. One of the incredible

378
00:18:59,359 --> 00:19:01,879
Speaker 3: things about is that they're always pushing the boundary on

379
00:19:01,880 --> 00:19:04,919
Speaker 3: the engineering side, and their next generation of chips is

380
00:19:05,040 --> 00:19:08,400
Speaker 3: largely dependent upon much more aggressive heat transfer, and they've

381
00:19:08,440 --> 00:19:11,600
Speaker 3: introduced liquid cooling to the reference architectures. So as liquid

382
00:19:11,600 --> 00:19:14,800
Speaker 3: cooling comes in, it changes what type of data center

383
00:19:14,840 --> 00:19:18,239
Speaker 3: is capable of doing this, and it truly requires that

384
00:19:18,359 --> 00:19:23,200
Speaker 3: ground up redesign and almost greenfield only build to support it.

385
00:19:23,240 --> 00:19:25,640
Speaker 3: Is you've gone from an environment where you could take

386
00:19:25,680 --> 00:19:28,520
Speaker 3: an enterprise data center and deploy less servers per cabinet

387
00:19:28,560 --> 00:19:31,560
Speaker 3: and get away with it to hey, nobody's ever built

388
00:19:31,560 --> 00:19:33,960
Speaker 3: this before. It's at an incredible scale and it has

389
00:19:34,000 --> 00:19:37,399
Speaker 3: to happen on a yearly cadence now, so the data

390
00:19:37,400 --> 00:19:40,280
Speaker 3: center industry is in't a full sprint to figure out, Okay,

391
00:19:40,280 --> 00:19:42,000
Speaker 3: how do we do this? How do we do it quickly?

392
00:19:42,240 --> 00:19:44,840
Speaker 3: How do we operationalize it right? And you know that's

393
00:19:44,920 --> 00:19:46,680
Speaker 3: kind of where I've been spending all of my time

394
00:19:46,920 --> 00:19:48,000
Speaker 3: over the past six months.

395
00:19:48,200 --> 00:19:51,200
Speaker 1: Can I ask a really basic question, and we've done

396
00:19:51,280 --> 00:19:53,959
Speaker 1: episodes on this, but I would be very interested in

397
00:19:54,000 --> 00:19:58,760
Speaker 1: your opinion, But why does it feel like customers and

398
00:19:58,880 --> 00:20:02,719
Speaker 1: AI customers in particular, are so I don't know if

399
00:20:02,720 --> 00:20:05,960
Speaker 1: addicted is the right word, but like so devoted to

400
00:20:06,359 --> 00:20:10,160
Speaker 1: in Nvidia chips, Like what is it about them specifically

401
00:20:10,800 --> 00:20:13,920
Speaker 1: that is so attractive? How much of it is due

402
00:20:13,960 --> 00:20:17,879
Speaker 1: to like the technology versus say, maybe the interoperability.

403
00:20:18,280 --> 00:20:20,800
Speaker 3: So you have to understand that when you're an AI

404
00:20:20,880 --> 00:20:25,239
Speaker 3: lab that has just started and it is a it's

405
00:20:25,280 --> 00:20:27,439
Speaker 3: an arms race in the industry to deliver product and

406
00:20:27,440 --> 00:20:30,880
Speaker 3: models as fast as possible, that it's an existential risk

407
00:20:30,960 --> 00:20:36,280
Speaker 3: to you that you don't have your infrastructure be like

408
00:20:36,480 --> 00:20:40,639
Speaker 3: your Achilles heel. Right, And and Vidia has proven to

409
00:20:40,720 --> 00:20:44,119
Speaker 3: be a number of things. One is they're the engineers

410
00:20:44,160 --> 00:20:49,879
Speaker 3: of the best products, right. They are an engineering organization first,

411
00:20:49,880 --> 00:20:52,640
Speaker 3: and that they identify and solve problems. They push the limits.

412
00:20:52,960 --> 00:20:54,960
Speaker 3: You know, they're willing to listen to customers and help

413
00:20:55,000 --> 00:20:58,080
Speaker 3: you solve problems and design things around new use cases.

414
00:20:58,800 --> 00:21:02,200
Speaker 3: But it's not just creating good hardware. It's creating good

415
00:21:02,240 --> 00:21:05,080
Speaker 3: hardware that's scales and they can support at scale. And

416
00:21:05,119 --> 00:21:07,720
Speaker 3: when you're building these installations that are hundreds of thousands

417
00:21:07,720 --> 00:21:11,280
Speaker 3: of components on the accelerator side and the infinband link side,

418
00:21:11,440 --> 00:21:14,000
Speaker 3: it all has to work together well. And when you

419
00:21:14,040 --> 00:21:16,240
Speaker 3: go to somebody like in Video that has done this

420
00:21:16,440 --> 00:21:19,680
Speaker 3: for so long at scale, with such engineering expertise, they

421
00:21:19,720 --> 00:21:22,760
Speaker 3: eliminate so much of that existential risk for these startups. Right.

422
00:21:22,760 --> 00:21:24,240
Speaker 3: So when I look at it and I see some

423
00:21:24,280 --> 00:21:26,520
Speaker 3: of these smaller startups saying we're going to go a

424
00:21:26,560 --> 00:21:29,640
Speaker 3: different route, I'm like, what are you doing? Right? You're

425
00:21:29,720 --> 00:21:32,640
Speaker 3: taking so much risk for no reason here? Right, this

426
00:21:32,680 --> 00:21:35,240
Speaker 3: is a proven solution, it's the best solution, and it

427
00:21:35,280 --> 00:21:38,399
Speaker 3: has the most community support, right, Like go the easy

428
00:21:38,400 --> 00:21:41,280
Speaker 3: path because the venture you're embarking on is hard enough.

429
00:21:41,600 --> 00:21:44,040
Speaker 1: Is it like the old what was that old adage?

430
00:21:44,080 --> 00:21:46,879
Speaker 1: Like no one ever got fired for buying Microsoft? Is

431
00:21:46,920 --> 00:21:49,960
Speaker 1: it like no, yeah, or IBM something like that.

432
00:21:50,040 --> 00:21:53,840
Speaker 3: But the thing here is that it's not even nobody's

433
00:21:53,880 --> 00:21:56,840
Speaker 3: getting fired for buying the tried and true and slower

434
00:21:56,880 --> 00:21:59,840
Speaker 3: moving thing. It's nobody's getting fired for buying the tried,

435
00:22:00,160 --> 00:22:03,080
Speaker 3: true and best performing and you know bleeding edge thing.

436
00:22:03,600 --> 00:22:03,800
Speaker 2: Right.

437
00:22:03,880 --> 00:22:06,560
Speaker 3: So I look at the folks that are buying other

438
00:22:06,640 --> 00:22:10,000
Speaker 3: products and investing and other products almost as like they're trying.

439
00:22:10,200 --> 00:22:11,600
Speaker 3: They almost have a chip on their shoulder and they're

440
00:22:11,600 --> 00:22:13,520
Speaker 3: going against the mold just to do it.

441
00:22:14,160 --> 00:22:17,880
Speaker 2: There are competitors to in video that they claim cheaper

442
00:22:18,600 --> 00:22:23,600
Speaker 2: or more application specific chips. I think Intel came out

443
00:22:23,680 --> 00:22:26,600
Speaker 2: with something like that. First of all, from the core

444
00:22:26,680 --> 00:22:30,640
Speaker 2: weave perspective, are you all in on in video hardware?

445
00:22:31,119 --> 00:22:31,439
Speaker 3: We are?

446
00:22:32,119 --> 00:22:33,000
Speaker 2: Could that change?

447
00:22:33,320 --> 00:22:35,480
Speaker 3: The party line is that we're always going to be

448
00:22:35,520 --> 00:22:38,200
Speaker 3: driven by customers, right, and we're going to be driven

449
00:22:38,240 --> 00:22:43,240
Speaker 3: by customers to the chip that is most performant, provides

450
00:22:43,280 --> 00:22:47,479
Speaker 3: the best TCO, is best supported and right now and

451
00:22:47,520 --> 00:22:50,399
Speaker 3: in what I think is the foreseeable future, like I

452
00:22:50,440 --> 00:22:51,800
Speaker 3: believe that is strongly in video.

453
00:22:52,560 --> 00:22:54,720
Speaker 2: Think about okay, maybe one day you guys IPO And

454
00:22:54,760 --> 00:22:56,760
Speaker 2: I'm looking through the risk factors, and one of the

455
00:22:56,840 --> 00:22:59,919
Speaker 2: risk factors, right, we have a heavy reliance on in

456
00:23:00,080 --> 00:23:02,320
Speaker 2: video chips. There is a risk that a competitor thing,

457
00:23:02,600 --> 00:23:05,919
Speaker 2: what would it take for one of these competitors that

458
00:23:05,960 --> 00:23:10,600
Speaker 2: does ostensibly over cheaper or hardware or perhaps lower electricity

459
00:23:10,720 --> 00:23:14,240
Speaker 2: consumption in your view, To make one of those risk

460
00:23:14,320 --> 00:23:15,120
Speaker 2: factors real.

461
00:23:15,720 --> 00:23:17,960
Speaker 3: I think that they'd have to be willing to quote

462
00:23:18,000 --> 00:23:20,840
Speaker 3: unquote buy the market. And when I say that, I

463
00:23:20,880 --> 00:23:24,199
Speaker 3: mean they'd have to subsidize their hardware to get a

464
00:23:24,480 --> 00:23:27,439
Speaker 3: material market share. And from what I've seen, there's no

465
00:23:27,480 --> 00:23:29,640
Speaker 3: one else that's really been willing to do that so far.

466
00:23:30,280 --> 00:23:34,040
Speaker 2: And what about Meta with Piedtorch and all their chips.

467
00:23:33,960 --> 00:23:36,479
Speaker 3: So they're in house chips. I think they have those

468
00:23:36,560 --> 00:23:41,240
Speaker 3: for very very specific production applications, but they're not really

469
00:23:41,280 --> 00:23:44,000
Speaker 3: general purpose chips, okay, right, And I think that when

470
00:23:44,040 --> 00:23:46,000
Speaker 3: you're building something for general purpose and there has to

471
00:23:46,000 --> 00:23:49,159
Speaker 3: be flexibility in the use case. While you can go

472
00:23:49,200 --> 00:23:52,399
Speaker 3: build a custom AASIC to solve very specific problems, I

473
00:23:52,440 --> 00:23:54,679
Speaker 3: don't think it makes sense to invest in those to

474
00:23:54,760 --> 00:23:56,920
Speaker 3: go to be a five year ass set if you

475
00:23:56,960 --> 00:23:58,320
Speaker 3: don't necessarily know what you're going to do with it.

476
00:23:58,760 --> 00:24:03,439
Speaker 1: So you talked about the advantages of Nvidia hardware like

477
00:24:03,480 --> 00:24:05,919
Speaker 1: the chips themselves, but one of the things you sometimes

478
00:24:06,000 --> 00:24:09,480
Speaker 1: hear is that those same chips might perform differently in

479
00:24:09,560 --> 00:24:13,200
Speaker 1: different clouds. So what is it that you can do

480
00:24:13,440 --> 00:24:16,200
Speaker 1: to sort of boost the performance of the same chip

481
00:24:16,760 --> 00:24:21,800
Speaker 1: in your structure or ecosystem versus say an AWS or

482
00:24:21,800 --> 00:24:22,439
Speaker 1: someone like that.

483
00:24:22,680 --> 00:24:24,440
Speaker 3: Sure, a great question. We do a lot of work

484
00:24:24,440 --> 00:24:26,959
Speaker 3: around this internally and it's a big part of our

485
00:24:26,960 --> 00:24:30,919
Speaker 3: technical differentiation. And what we call it internally is mission control.

486
00:24:31,400 --> 00:24:34,480
Speaker 3: And mission control is effectively a portfolio of different services

487
00:24:34,480 --> 00:24:37,399
Speaker 3: that we run on our infrastructure to make sure that

488
00:24:37,440 --> 00:24:42,400
Speaker 3: these incredibly complex supercomputers are healthy and performant and are optimized,

489
00:24:43,440 --> 00:24:46,000
Speaker 3: you know, where we take a lot of that responsibility

490
00:24:46,040 --> 00:24:49,760
Speaker 3: off of our customer engineering teams, right, And it sounds

491
00:24:49,760 --> 00:24:51,520
Speaker 3: like that might be an easy lift, but when you're

492
00:24:52,040 --> 00:24:54,679
Speaker 3: running supercomputer scale, you know you need a team of

493
00:24:54,680 --> 00:24:56,680
Speaker 3: fifty to do that, right, So we provide a ton

494
00:24:56,720 --> 00:24:59,680
Speaker 3: of software automation around that, providing that health checking and

495
00:24:59,680 --> 00:25:03,680
Speaker 3: observed ability to our customers. But it's also the engineering engagement, right,

496
00:25:03,840 --> 00:25:06,040
Speaker 3: is you know, working with our customers to understand, Okay,

497
00:25:06,080 --> 00:25:08,359
Speaker 3: what are you doing, what's the best way to optimize this,

498
00:25:08,560 --> 00:25:10,720
Speaker 3: how do we you know, how did we design the

499
00:25:10,760 --> 00:25:12,919
Speaker 3: data center to be more performant, to make sure your

500
00:25:12,920 --> 00:25:16,000
Speaker 3: storage solution was correct, Your networking solution was correct. So

501
00:25:16,400 --> 00:25:20,120
Speaker 3: it's not just a hey core we've provides like this

502
00:25:20,160 --> 00:25:23,200
Speaker 3: one little thing that makes it better. It's the comprehensive solutions,

503
00:25:23,240 --> 00:25:26,320
Speaker 3: starting from the data center design, through the software automation

504
00:25:26,400 --> 00:25:29,080
Speaker 3: and health checking and monitoring, via mission control, via the

505
00:25:29,119 --> 00:25:31,480
Speaker 3: engineering relationships that really add that value.

506
00:25:31,880 --> 00:25:35,000
Speaker 2: Let's talk about electricity, because this has become this huge

507
00:25:35,040 --> 00:25:37,560
Speaker 2: talking point that this is the major constraint and now

508
00:25:37,600 --> 00:25:40,240
Speaker 2: that you're becoming more vertically integrated and having to stand

509
00:25:40,320 --> 00:25:43,560
Speaker 2: up more of your operations. We talked to one guy

510
00:25:43,720 --> 00:25:46,199
Speaker 2: formerly at Microsoft who said, you know, one of the

511
00:25:46,240 --> 00:25:48,760
Speaker 2: issues that there may be a backlash in some communities

512
00:25:48,800 --> 00:25:52,320
Speaker 2: who don't want, you know, their scarce electricity to go

513
00:25:52,359 --> 00:25:55,520
Speaker 2: to data centers when they could go to household air conditioning.

514
00:25:55,760 --> 00:25:57,840
Speaker 2: What are you running into right now or what are

515
00:25:57,880 --> 00:25:58,240
Speaker 2: you seeing?

516
00:25:58,560 --> 00:26:01,560
Speaker 3: So we've been very very selective on where we put

517
00:26:01,640 --> 00:26:04,600
Speaker 3: data centers. We don't have anything in Ashburn, Virginia, right

518
00:26:04,640 --> 00:26:07,320
Speaker 3: and the Northern Virginia market, I think is incredibly saturated.

519
00:26:07,680 --> 00:26:10,280
Speaker 3: There's a lot of growing backlash in that market around

520
00:26:10,320 --> 00:26:13,240
Speaker 3: power usage and you know, just thinking about how do

521
00:26:13,280 --> 00:26:15,440
Speaker 3: you get enough diesel trucks in there to refill generators

522
00:26:15,480 --> 00:26:17,080
Speaker 3: that they have a prolonged outage.

523
00:26:17,200 --> 00:26:17,360
Speaker 1: Right.

524
00:26:17,880 --> 00:26:20,159
Speaker 3: So I think that there's some markets where it's just

525
00:26:20,200 --> 00:26:23,160
Speaker 3: like okay, like to stay away from that, and when

526
00:26:23,200 --> 00:26:26,159
Speaker 3: the grids have issues and that market hasn't really had

527
00:26:26,160 --> 00:26:29,080
Speaker 3: an issue yet, it becomes an acute problem immediately. Like

528
00:26:29,200 --> 00:26:32,159
Speaker 3: just think about the Texas power market crisis back in

529
00:26:32,320 --> 00:26:34,800
Speaker 3: I think it's twenty twenty one, twenty twenty, where the

530
00:26:34,840 --> 00:26:36,840
Speaker 3: grid wasn't really set up to be able to handle

531
00:26:37,320 --> 00:26:40,600
Speaker 3: the frigid temperatures and they had natural gas valves that

532
00:26:40,600 --> 00:26:43,560
Speaker 3: were freezing off at the natural gas generation plants that

533
00:26:43,840 --> 00:26:46,480
Speaker 3: didn't allow them to actually come online and produce electricity

534
00:26:46,640 --> 00:26:49,000
Speaker 3: no matter how high the price was. Right. So there's

535
00:26:49,080 --> 00:26:51,239
Speaker 3: there's going to be these acute issues that you know,

536
00:26:51,520 --> 00:26:53,439
Speaker 3: people are going to learn from and the regulators are

537
00:26:53,440 --> 00:26:55,119
Speaker 3: going to learn from to make sure they don't happen again.

538
00:26:55,600 --> 00:26:58,520
Speaker 3: And we're kind of citing our our plants and markets

539
00:26:58,520 --> 00:27:01,000
Speaker 3: where our data centers and markets where we think the

540
00:27:01,040 --> 00:27:04,160
Speaker 3: grid infrastructure is capable of handling it right, And it's

541
00:27:04,200 --> 00:27:07,040
Speaker 3: not just is there enough power, it's also on things.

542
00:27:07,080 --> 00:27:10,160
Speaker 3: You know, AI workloads are pretty volatile in how much

543
00:27:10,200 --> 00:27:12,760
Speaker 3: power they use, and they're volatile because you know, every

544
00:27:12,840 --> 00:27:15,800
Speaker 3: fifteen minutes or every thirty minutes, you effectively stop the

545
00:27:15,920 --> 00:27:19,560
Speaker 3: job to save the progress you've made, right, and it's

546
00:27:19,600 --> 00:27:21,600
Speaker 3: so expensive to run these clusters that you don't want

547
00:27:21,640 --> 00:27:24,359
Speaker 3: to lose hundreds of thousands of dollars of progress, So

548
00:27:24,400 --> 00:27:26,680
Speaker 3: they take a minute, they do what's called checkpointing, where

549
00:27:26,680 --> 00:27:29,560
Speaker 3: they write the current state of the job back to storage,

550
00:27:29,960 --> 00:27:33,000
Speaker 3: and that checkpointing time, your power usage basically goes from

551
00:27:33,000 --> 00:27:35,560
Speaker 3: one hundred percent to like ten percent, and then it

552
00:27:35,600 --> 00:27:37,439
Speaker 3: goes right back up again when it's done saving it.

553
00:27:37,760 --> 00:27:41,560
Speaker 3: So that load volatility on a local market will create

554
00:27:41,600 --> 00:27:44,960
Speaker 3: either voltage spikes or voltage SAgs, and a voltage sag

555
00:27:45,000 --> 00:27:47,120
Speaker 3: is what you see is what causes a brown out

556
00:27:47,359 --> 00:27:48,560
Speaker 3: that we used to see a lot of times when

557
00:27:48,560 --> 00:27:51,359
Speaker 3: people turn their cognitioners on and it's thinking through, Okay,

558
00:27:51,359 --> 00:27:55,440
Speaker 3: how do I ensure that, you know, my AI installation

559
00:27:55,600 --> 00:27:57,760
Speaker 3: doesn't cause a brown out when people are turning their

560
00:27:58,080 --> 00:28:00,239
Speaker 3: you know, during checkpointing, when people are turning the air

561
00:28:00,240 --> 00:28:02,320
Speaker 3: conditioners on. Like that's the type of stuff that we're

562
00:28:02,320 --> 00:28:03,920
Speaker 3: thoughtful around, like how do we make sure we don't

563
00:28:03,960 --> 00:28:07,680
Speaker 3: do this right. And you know, talking to engineerings and

564
00:28:07,680 --> 00:28:10,800
Speaker 3: in Video's engineering expertise, like they're working on this problem

565
00:28:10,800 --> 00:28:13,400
Speaker 3: as well, and there they've solved this for the next generation.

566
00:28:14,600 --> 00:28:17,160
Speaker 3: So it's everything from is there enough power there? What's

567
00:28:17,200 --> 00:28:19,359
Speaker 3: the source of that power? You know, how clean is it?

568
00:28:19,440 --> 00:28:21,280
Speaker 3: How do we make sure that we're investing in solar

569
00:28:21,320 --> 00:28:23,080
Speaker 3: and stuff in the area to make sure that we're

570
00:28:23,520 --> 00:28:26,120
Speaker 3: not just taking power from the grid. To also when

571
00:28:26,119 --> 00:28:27,800
Speaker 3: we're using that power, how is it going to impact

572
00:28:27,840 --> 00:28:28,920
Speaker 3: the consumers around us?

573
00:28:29,119 --> 00:28:31,240
Speaker 1: I want to ask you more about what in Nvidia

574
00:28:31,480 --> 00:28:34,160
Speaker 1: is doing, but just on that note, what's the most

575
00:28:34,240 --> 00:28:39,960
Speaker 1: important metric for evaluating a data center's quality or performance?

576
00:28:40,080 --> 00:28:44,600
Speaker 1: Is it like days without brownouts or an interrupted power supply,

577
00:28:44,720 --> 00:28:48,200
Speaker 1: or is it measures of efficiency like power usage effectiveness

578
00:28:48,240 --> 00:28:50,400
Speaker 1: or something like that. If I'm serving a bunch of

579
00:28:50,520 --> 00:28:52,400
Speaker 1: data centers, I want to pick a good one. What

580
00:28:52,440 --> 00:28:53,440
Speaker 1: should I be looking for?

581
00:28:53,840 --> 00:28:56,920
Speaker 3: So right now, the market's pretty thin, So right now.

582
00:28:58,040 --> 00:29:02,040
Speaker 1: Options Okay, I imagine I'm like the biggest customer on

583
00:29:02,160 --> 00:29:04,960
Speaker 1: earth and I can get in anywhere. What should I

584
00:29:04,960 --> 00:29:05,560
Speaker 1: be looking for?

585
00:29:06,440 --> 00:29:10,400
Speaker 3: So it's the first thing goes back to the electricity piece, right,

586
00:29:10,560 --> 00:29:13,719
Speaker 3: is the grid stable? Is there enough power supply? You know,

587
00:29:13,880 --> 00:29:16,680
Speaker 3: is there excess renewable generation in the area that doesn't

588
00:29:16,720 --> 00:29:19,200
Speaker 3: have the ability to make it too downstream consumers? Right?

589
00:29:19,240 --> 00:29:20,880
Speaker 3: A lot of the renewables that we have in the

590
00:29:20,960 --> 00:29:24,200
Speaker 3: US are built in places that don't necessarily have the consumers.

591
00:29:24,840 --> 00:29:27,920
Speaker 3: So you're citing these data centers in places where you

592
00:29:28,000 --> 00:29:30,800
Speaker 3: have this excess supply, So that that's the first piece, right,

593
00:29:30,880 --> 00:29:34,280
Speaker 3: is how good is the electricity supply? And how angry

594
00:29:34,280 --> 00:29:35,520
Speaker 3: are the people around me going to be if I

595
00:29:35,520 --> 00:29:39,200
Speaker 3: take it? Now? You go from there into everything else

596
00:29:39,240 --> 00:29:41,560
Speaker 3: is kind of solvable, right, And the way that you

597
00:29:41,640 --> 00:29:44,520
Speaker 3: design it, and if you're building a green field, it's okay.

598
00:29:44,560 --> 00:29:46,719
Speaker 3: You know what type of ups systems am I putting in?

599
00:29:46,800 --> 00:29:49,160
Speaker 3: Are they capable of handling that load volatility?

600
00:29:50,040 --> 00:29:50,280
Speaker 2: You know?

601
00:29:50,640 --> 00:29:54,440
Speaker 3: How am I thinking about my cooling solutions? There's been

602
00:29:54,480 --> 00:29:58,600
Speaker 3: a big shift to liquid cooling, right, and liquid cooling

603
00:29:58,640 --> 00:30:02,600
Speaker 3: from a PE perspective, isn't a thirty to forty percent

604
00:30:03,200 --> 00:30:06,480
Speaker 3: decrease in electricity utilization like people think? It's more like

605
00:30:06,520 --> 00:30:09,840
Speaker 3: sixty to seventy percent, right, And the reason for that

606
00:30:09,920 --> 00:30:13,760
Speaker 3: is it's not just the efficiency of the data center plant.

607
00:30:14,040 --> 00:30:16,400
Speaker 3: It's also that now if you're not cooling things with air,

608
00:30:16,440 --> 00:30:18,480
Speaker 3: you don't have to run the fans inside the servers

609
00:30:18,480 --> 00:30:21,960
Speaker 3: as well. And for these AI installations, because they're so dense,

610
00:30:22,240 --> 00:30:25,280
Speaker 3: the fans consume a lot of energy. Right. So everything

611
00:30:25,280 --> 00:30:27,360
Speaker 3: that we're building now is a combination of liquid and

612
00:30:27,400 --> 00:30:30,760
Speaker 3: air cooling, right. And the liquid cooling piece has solved

613
00:30:30,800 --> 00:30:34,280
Speaker 3: the PUE issue, right, And we're everything we're doing is

614
00:30:34,320 --> 00:30:37,640
Speaker 3: trying to say, Okay, how much power can we use

615
00:30:37,840 --> 00:30:42,680
Speaker 3: only for running our critical IT operations versus cooling the

616
00:30:42,760 --> 00:30:46,200
Speaker 3: environment making sure the environment's running correctly from a resiliency perspective,

617
00:30:47,000 --> 00:30:48,880
Speaker 3: And there's been big strides made there over the last

618
00:30:48,880 --> 00:31:01,560
Speaker 3: whole months.

619
00:31:06,120 --> 00:31:11,400
Speaker 1: Does colocation trump grid reliability? Like if I'm Elon Musk

620
00:31:11,880 --> 00:31:14,960
Speaker 1: building some sort of new AI thing as I think

621
00:31:15,000 --> 00:31:18,880
Speaker 1: he's doing in Texas, say like, am I just going

622
00:31:18,960 --> 00:31:21,200
Speaker 1: to have to find a data center in Texas? Or

623
00:31:21,240 --> 00:31:25,640
Speaker 1: how much flexibility do I have to use one further away?

624
00:31:25,880 --> 00:31:30,360
Speaker 3: So great question, it's it's a different answer for different

625
00:31:30,440 --> 00:31:33,960
Speaker 3: use cases at different times. And right now, you know,

626
00:31:34,000 --> 00:31:36,400
Speaker 3: we were in the middle of this rush to train

627
00:31:37,200 --> 00:31:40,400
Speaker 3: whether they're open source or proprietary foundation models at the largest,

628
00:31:40,440 --> 00:31:43,600
Speaker 3: most valuable companies in the world, and they're mostly worried

629
00:31:43,600 --> 00:31:47,400
Speaker 3: about access to contiguous compute capacity. Right, how much compute

630
00:31:47,400 --> 00:31:49,959
Speaker 3: can I get in one location, all connected together so

631
00:31:50,040 --> 00:31:52,320
Speaker 3: I can go faster than the next guy. But when

632
00:31:52,720 --> 00:31:56,080
Speaker 3: the models are trained, they want that compute to then

633
00:31:56,120 --> 00:31:58,440
Speaker 3: be local to their customer base, right, is how do

634
00:31:58,520 --> 00:32:00,560
Speaker 3: they take it from the middle of nowhere and then

635
00:32:00,600 --> 00:32:03,200
Speaker 3: go serve it in the metropolitan markets. And as the

636
00:32:03,320 --> 00:32:06,000
Speaker 3: use cases are more distilled and they get more real time,

637
00:32:07,080 --> 00:32:09,800
Speaker 3: think like the type ahead suggestions that you get in

638
00:32:09,840 --> 00:32:12,120
Speaker 3: your Gmail account right as you're typing something, and it's

639
00:32:12,120 --> 00:32:15,120
Speaker 3: getting better and better. It's you know, that's an AI

640
00:32:15,200 --> 00:32:18,200
Speaker 3: model somewhere like predicting what you would want to say next,

641
00:32:18,640 --> 00:32:21,560
Speaker 3: And they want to make sure that's delivered at human speed.

642
00:32:21,920 --> 00:32:26,040
Speaker 3: So that human speed is a latency consideration. Right as

643
00:32:26,080 --> 00:32:28,480
Speaker 3: you're citing those GPUs and you're citing that compute to

644
00:32:28,520 --> 00:32:30,240
Speaker 3: be locals to the people that are using it. So

645
00:32:32,000 --> 00:32:36,120
Speaker 3: that move has started probably four months ago where we

646
00:32:36,200 --> 00:32:39,600
Speaker 3: saw customers finally becoming concern around latency for their serving

647
00:32:39,720 --> 00:32:42,720
Speaker 3: use cases. So initially training people don't really care where

648
00:32:42,760 --> 00:32:46,000
Speaker 3: it is cheap power, reliable grid. They just need it

649
00:32:46,040 --> 00:32:48,440
Speaker 3: all contiguous and they need it fast. And then down

650
00:32:48,440 --> 00:32:51,080
Speaker 3: the road as their applications find success, they're more worried

651
00:32:51,080 --> 00:32:52,640
Speaker 3: about where the compute is for their customers.

652
00:32:53,040 --> 00:32:54,720
Speaker 2: What are some of the areas that are going to

653
00:32:54,720 --> 00:32:57,880
Speaker 2: be the next Northern Virginia when it comes to data

654
00:32:57,920 --> 00:32:58,880
Speaker 2: center clusters.

655
00:32:59,680 --> 00:33:02,960
Speaker 3: So I think we're seeing it in Atlanta already, where

656
00:33:03,120 --> 00:33:06,920
Speaker 3: Georgia has paused or has attempted to pause some of

657
00:33:06,920 --> 00:33:08,840
Speaker 3: their tax incentives around it because they want to make

658
00:33:08,880 --> 00:33:12,280
Speaker 3: sure they do grid studies. I think that we're we're

659
00:33:12,320 --> 00:33:14,560
Speaker 3: probably going to see it in some of the other hotspots.

660
00:33:14,760 --> 00:33:14,960
Speaker 2: You know.

661
00:33:16,280 --> 00:33:18,560
Speaker 3: You know, you see aws up in Oregon who is

662
00:33:18,560 --> 00:33:22,120
Speaker 3: trying to find creative alternative ways to power their data

663
00:33:22,160 --> 00:33:25,560
Speaker 3: centers from non grid generation to alleviate some concerns there.

664
00:33:26,960 --> 00:33:29,960
Speaker 3: But you know, I think that the market has to

665
00:33:30,040 --> 00:33:33,040
Speaker 3: solve this problem. And you know, you're starting to see

666
00:33:33,040 --> 00:33:36,640
Speaker 3: some of the startups around nuclear generation in you know,

667
00:33:36,720 --> 00:33:39,960
Speaker 3: the small reactors at the data center level. As people

668
00:33:39,960 --> 00:33:41,920
Speaker 3: are you know, being thoughtful for five to ten years

669
00:33:41,920 --> 00:33:42,760
Speaker 3: from now, do.

670
00:33:42,720 --> 00:33:46,320
Speaker 1: You have any influence on the type of power being

671
00:33:46,360 --> 00:33:49,200
Speaker 1: built in certain areas? You know, could you say to

672
00:33:49,400 --> 00:33:53,240
Speaker 1: a utility company of some sort, we're here, we need

673
00:33:53,280 --> 00:33:56,200
Speaker 1: access to energy, but we want it to come in

674
00:33:56,240 --> 00:33:57,640
Speaker 1: a particular form.

675
00:33:57,640 --> 00:34:00,040
Speaker 3: So you can. But you have to understand that the

676
00:34:00,080 --> 00:34:02,600
Speaker 3: investment cycles and the physical build cycles for those are

677
00:34:02,640 --> 00:34:06,040
Speaker 3: so much longer than you know how quickly our customers

678
00:34:06,120 --> 00:34:08,399
Speaker 3: need infrastructure, right. So you may go to a market

679
00:34:08,480 --> 00:34:09,960
Speaker 3: and say, hey, we're going to be here over the

680
00:34:09,960 --> 00:34:11,880
Speaker 3: next ten years, we'd like you to install X y Z,

681
00:34:12,200 --> 00:34:15,560
Speaker 3: you know, renewable, and they're happy to do it. It's

682
00:34:15,600 --> 00:34:17,240
Speaker 3: just that you have to find a medium term solution

683
00:34:17,719 --> 00:34:18,800
Speaker 3: while that's being built.

684
00:34:19,000 --> 00:34:20,880
Speaker 2: I'm going to ask a question. So there was a

685
00:34:20,920 --> 00:34:23,800
Speaker 2: news story, and maybe you won't comment on the news story,

686
00:34:23,880 --> 00:34:27,160
Speaker 2: specifically about core Weave having made a one billion dollar

687
00:34:27,239 --> 00:34:33,320
Speaker 2: offer for a bitcoin miner called core Scientific, apparently was rejected.

688
00:34:33,360 --> 00:34:36,759
Speaker 2: According to things I've read in the news. Setting aside

689
00:34:37,000 --> 00:34:39,480
Speaker 2: this deal, there's you know, there used to be a

690
00:34:39,480 --> 00:34:43,439
Speaker 2: lot of crypto mining and then ethereum went from proof

691
00:34:43,480 --> 00:34:45,440
Speaker 2: of work to proof of steak and that all basically

692
00:34:45,480 --> 00:34:49,160
Speaker 2: disappeared overnight. There are still bitcoin miners. I never get

693
00:34:49,160 --> 00:34:51,920
Speaker 2: the impression it's like that great of business. But whatever

694
00:34:52,320 --> 00:34:55,960
Speaker 2: are there bitcoin miners that have latent value in the

695
00:34:56,000 --> 00:34:58,719
Speaker 2: fact that they I mean, I know those chips don't

696
00:34:58,800 --> 00:35:01,680
Speaker 2: the bitcoin mining chip, the actual acis don't work for

697
00:35:01,800 --> 00:35:04,960
Speaker 2: AI because all they are is bitcoin mining chips. But

698
00:35:05,080 --> 00:35:08,560
Speaker 2: are there by dint of their access to electricity, space,

699
00:35:08,600 --> 00:35:12,600
Speaker 2: et cetera, is there a fair amount of latent value

700
00:35:12,800 --> 00:35:16,520
Speaker 2: in the general physical structures that they've built for the mining.

701
00:35:16,960 --> 00:35:18,960
Speaker 3: So I'm just not going to answer your question at all.

702
00:35:19,000 --> 00:35:19,960
Speaker 3: I'm gonna go on a tangent.

703
00:35:20,120 --> 00:35:21,280
Speaker 2: Okay, that's fine.

704
00:35:21,360 --> 00:35:24,520
Speaker 3: So I think that when I think about core Weave

705
00:35:25,000 --> 00:35:29,440
Speaker 3: and what our mission is, it's to find creative solutions

706
00:35:29,440 --> 00:35:32,919
Speaker 3: to problems in in you know, various markets, and those

707
00:35:33,000 --> 00:35:36,120
Speaker 3: various markets can be blocking for us and our customers to.

708
00:35:36,080 --> 00:35:36,840
Speaker 2: Achieve our goals.

709
00:35:37,400 --> 00:35:40,920
Speaker 3: So if power is a concern for us, and power

710
00:35:40,960 --> 00:35:43,720
Speaker 3: availability and substations and substation.

711
00:35:43,360 --> 00:35:45,440
Speaker 2: Transform, coin miners definitely have access to power.

712
00:35:46,080 --> 00:35:46,959
Speaker 3: That that is true.

713
00:35:47,680 --> 00:35:50,040
Speaker 2: I'm just stating fact you could keep doing it.

714
00:35:50,719 --> 00:35:53,040
Speaker 3: So you know, as we go and we try to

715
00:35:53,080 --> 00:35:56,200
Speaker 3: solve these problems, you know, we're going to go to

716
00:35:56,440 --> 00:35:59,799
Speaker 3: places that others may not have thought of, and we're

717
00:35:59,800 --> 00:36:02,120
Speaker 3: going to go do due diligence and I'm going to

718
00:36:02,200 --> 00:36:04,239
Speaker 3: personally go and walk the sites and I'm going to

719
00:36:04,440 --> 00:36:07,239
Speaker 3: you know, look through and see, okay, can we.

720
00:36:07,239 --> 00:36:07,920
Speaker 2: Pull this off?

721
00:36:08,360 --> 00:36:10,680
Speaker 3: And we're going to get our engineering partners in to

722
00:36:11,040 --> 00:36:14,000
Speaker 3: help us design retrofits. And you know, we're going to

723
00:36:14,040 --> 00:36:16,560
Speaker 3: do deals with the companies that we believe have the

724
00:36:16,600 --> 00:36:17,760
Speaker 3: ability to provide us value.

725
00:36:19,000 --> 00:36:22,680
Speaker 1: Since we're doing stuff in the news. This has been

726
00:36:22,680 --> 00:36:24,440
Speaker 1: in the news for a while, so it doesn't really count.

727
00:36:24,440 --> 00:36:29,600
Speaker 1: But the new Nvidia chips, the GB two hundreds, what

728
00:36:29,680 --> 00:36:32,320
Speaker 1: will those do for core weave and when would you

729
00:36:32,360 --> 00:36:33,360
Speaker 1: expect to get them?

730
00:36:33,680 --> 00:36:35,520
Speaker 3: What will they do for us? It's more about what

731
00:36:35,560 --> 00:36:38,680
Speaker 3: they're going to do for our customers, right, and I think.

732
00:36:38,440 --> 00:36:39,799
Speaker 2: That they are.

733
00:36:41,960 --> 00:36:45,720
Speaker 3: This is a great question. They are going to open

734
00:36:45,840 --> 00:36:49,400
Speaker 3: up a lot of both training and inference use cases

735
00:36:50,000 --> 00:36:53,520
Speaker 3: in the AI side that I think our customers have

736
00:36:53,560 --> 00:36:58,080
Speaker 3: been blocked by UH with the existing generation in that

737
00:36:58,960 --> 00:37:02,080
Speaker 3: you're now able to think seventy two of these GPUs

738
00:37:02,120 --> 00:37:05,640
Speaker 3: together to work almost as one unit, and previously that

739
00:37:05,680 --> 00:37:08,799
Speaker 3: was limited to eight. They have a much larger what's

740
00:37:08,800 --> 00:37:10,719
Speaker 3: called the frame buffer, which is how much memory that's

741
00:37:10,760 --> 00:37:14,080
Speaker 3: usable for their matrix operations. So you know, I think

742
00:37:14,080 --> 00:37:15,839
Speaker 3: that we're going to see a lot of new use

743
00:37:15,880 --> 00:37:18,560
Speaker 3: cases show up for this stuff, but I think it

744
00:37:18,760 --> 00:37:22,520
Speaker 3: extends well beyond AI as well, and it's going to

745
00:37:22,560 --> 00:37:25,000
Speaker 3: be a lot more useful for things like scientific computing.

746
00:37:26,000 --> 00:37:28,200
Speaker 3: One of the things that has me really excited is

747
00:37:28,320 --> 00:37:32,760
Speaker 3: the computational fluidynamics and I'm specifically thinking about the uses

748
00:37:32,840 --> 00:37:35,359
Speaker 3: for that in F one under the new regulation in

749
00:37:35,400 --> 00:37:39,279
Speaker 3: twenty twenty six. I'm excited for the new platform. I

750
00:37:39,320 --> 00:37:40,840
Speaker 3: think in a year and a half people are going

751
00:37:40,920 --> 00:37:42,680
Speaker 3: to be using it for things that are different than

752
00:37:42,880 --> 00:37:47,560
Speaker 3: anybody expects today. And that's to me. The pace at

753
00:37:47,640 --> 00:37:50,239
Speaker 3: which this is changing is the piece that's really cool.

754
00:37:50,360 --> 00:37:51,960
Speaker 1: Wait, I'm sorry, I hate sports.

755
00:37:52,000 --> 00:37:55,920
Speaker 2: What's the six? Explain how the invidio is.

756
00:37:56,000 --> 00:37:59,440
Speaker 3: Yeah, So the F one platform, they have very tight

757
00:37:59,480 --> 00:38:01,960
Speaker 3: restrictions around what type of compute and how much compute

758
00:38:01,960 --> 00:38:04,680
Speaker 3: you can use to do aerodynamic testing in your cars,

759
00:38:05,200 --> 00:38:06,840
Speaker 3: and you can either do real life testing in a

760
00:38:06,840 --> 00:38:09,880
Speaker 3: wind tunnel or you can do it through CFD analysis.

761
00:38:10,080 --> 00:38:13,960
Speaker 3: And what are the great uses for the you know,

762
00:38:13,960 --> 00:38:17,080
Speaker 3: the Grace Blackwell and the Grace Hopper architectures. Impairing that

763
00:38:17,200 --> 00:38:20,719
Speaker 3: Grace super chip with the GPU is they're great for

764
00:38:20,760 --> 00:38:23,319
Speaker 3: CFD workloads, right, and the.

765
00:38:23,920 --> 00:38:27,239
Speaker 2: DAFD stands for computational fluid dynamics yep, yep.

766
00:38:27,400 --> 00:38:30,520
Speaker 3: And the regulations around the existing program in F one

767
00:38:30,680 --> 00:38:33,359
Speaker 3: are they're only able to use CPUs. They have very

768
00:38:33,600 --> 00:38:35,880
Speaker 3: like specific limitations around it. But there's been a lot

769
00:38:35,920 --> 00:38:39,080
Speaker 3: of talk of that changing for twenty twenty six car models,

770
00:38:39,520 --> 00:38:43,239
Speaker 3: and for me, like, that's pretty cool and I'm gung

771
00:38:43,320 --> 00:38:46,200
Speaker 3: ho excited about possibly supporting that.

772
00:38:46,200 --> 00:38:48,920
Speaker 2: That does sound very fun. I want to get back

773
00:38:48,960 --> 00:38:52,480
Speaker 2: to actually the financing a little bit because I guess

774
00:38:52,560 --> 00:38:57,120
Speaker 2: two questions. So the logic of why you would borrow

775
00:38:57,239 --> 00:39:00,800
Speaker 2: money both I guess for the equal position of chips,

776
00:39:00,800 --> 00:39:03,320
Speaker 2: and the chips are sort of collateral, but I understand

777
00:39:03,360 --> 00:39:08,080
Speaker 2: they're not really chip back loans per se. A. Do

778
00:39:08,160 --> 00:39:12,799
Speaker 2: you see your clients getting more into debt financing rather

779
00:39:12,840 --> 00:39:15,759
Speaker 2: than equity financing. I mean, there's a whole generation of

780
00:39:16,280 --> 00:39:19,160
Speaker 2: software companies from the Zerp era that was just you know,

781
00:39:20,160 --> 00:39:22,640
Speaker 2: all equity and never had any debt at all, and

782
00:39:22,680 --> 00:39:25,560
Speaker 2: they never really had to think about like their compute costs,

783
00:39:25,680 --> 00:39:29,000
Speaker 2: or they did, but not as much. Do you think

784
00:39:29,040 --> 00:39:31,920
Speaker 2: that will rise their own use of debt instead of

785
00:39:32,000 --> 00:39:35,600
Speaker 2: equity in terms of their own financing. And another topic

786
00:39:35,600 --> 00:39:37,480
Speaker 2: we talk about a lot on the show private credit,

787
00:39:37,600 --> 00:39:41,000
Speaker 2: like there is there an emergence of an ecosystem of

788
00:39:41,280 --> 00:39:44,799
Speaker 2: lenders for whom this is going to become a specialty

789
00:39:44,840 --> 00:39:45,359
Speaker 2: of some sort.

790
00:39:46,080 --> 00:39:48,400
Speaker 3: So the first piece of the question, I don't believe

791
00:39:48,480 --> 00:39:51,239
Speaker 3: that the venture backed kind of AI lab startups will

792
00:39:51,239 --> 00:39:55,200
Speaker 3: ever take on debt in this type of environment, largely

793
00:39:55,200 --> 00:39:57,520
Speaker 3: because they don't have the collateral to back it. If

794
00:39:57,560 --> 00:40:01,200
Speaker 3: they're buying cloud services to run their infrastructure. And you

795
00:40:01,239 --> 00:40:04,080
Speaker 3: may see some that start to buy their own infrastructure

796
00:40:04,080 --> 00:40:06,560
Speaker 3: and to do that themselves, but it is a herculean

797
00:40:06,600 --> 00:40:08,719
Speaker 3: task to do this at scale. Right, There's a reason

798
00:40:08,719 --> 00:40:11,040
Speaker 3: why clouds exist is that there's a lot of complexity

799
00:40:11,080 --> 00:40:14,160
Speaker 3: that they abstract away. On the second question around are

800
00:40:14,239 --> 00:40:16,279
Speaker 3: is there a private credit sector that's going to be

801
00:40:16,320 --> 00:40:19,160
Speaker 3: built to do this? I think that it's more you're

802
00:40:19,200 --> 00:40:22,480
Speaker 3: seeing public lenders that are extending into the private credit

803
00:40:22,520 --> 00:40:25,600
Speaker 3: space because the opportunities are there. And I'm going to

804
00:40:25,760 --> 00:40:29,479
Speaker 3: give you the party line answer that my CEO gives

805
00:40:29,480 --> 00:40:31,839
Speaker 3: all the time is that you know, as we're thinking

806
00:40:31,840 --> 00:40:34,880
Speaker 3: about financing our business, the biggest thing for us is

807
00:40:34,880 --> 00:40:37,000
Speaker 3: our cost to capital, and we're always going to do

808
00:40:37,040 --> 00:40:38,960
Speaker 3: the things that provide us the lowest cost of capital.

809
00:40:39,000 --> 00:40:42,759
Speaker 3: And you know the lenders that we work with, including Blackstone,

810
00:40:43,200 --> 00:40:45,399
Speaker 3: that have been so wonderful for us, you know, them

811
00:40:45,440 --> 00:40:48,000
Speaker 3: extending on the private credit side as we go to

812
00:40:48,000 --> 00:40:50,520
Speaker 3: the public markets because we're dragged there by cost of

813
00:40:50,560 --> 00:40:54,480
Speaker 3: capital concerns, I would expect them to be involved as well, right, So,

814
00:40:54,680 --> 00:40:56,680
Speaker 3: I think it's a continuation of the business they've been

815
00:40:56,680 --> 00:40:58,520
Speaker 3: doing in the public markets, just kind of extending into

816
00:40:58,520 --> 00:40:59,960
Speaker 3: this capital intensive business.

817
00:41:00,560 --> 00:41:04,440
Speaker 1: Wait, what was I guess you can't get into specific details,

818
00:41:04,480 --> 00:41:07,680
Speaker 1: but my impression was for these types of loans that

819
00:41:07,719 --> 00:41:10,960
Speaker 1: the interest rate is usually higher than like a basic

820
00:41:11,320 --> 00:41:14,560
Speaker 1: bank loan or say issuing a corporate bond.

821
00:41:15,200 --> 00:41:17,239
Speaker 3: I would definitely say our cost of capital is lower

822
00:41:17,280 --> 00:41:19,320
Speaker 3: than some of the corporate issuance is out there, Okay,

823
00:41:20,560 --> 00:41:24,319
Speaker 3: but you know it's definitely higher than if our cost

824
00:41:24,360 --> 00:41:25,960
Speaker 3: of capital today is definitely higher than if we were

825
00:41:25,960 --> 00:41:27,239
Speaker 3: republican public entity.

826
00:41:27,400 --> 00:41:30,840
Speaker 1: But specifically on the GPU backed loans, and I know

827
00:41:30,920 --> 00:41:33,040
Speaker 1: you keep saying it's not really a GPU back loan,

828
00:41:33,080 --> 00:41:36,080
Speaker 1: but that's sort of an uphill battle to call it

829
00:41:36,160 --> 00:41:39,359
Speaker 1: trade receivables financing instead. It sounds so much better that way,

830
00:41:39,520 --> 00:41:43,120
Speaker 1: I know, I know, but like on that in particular, Okay,

831
00:41:43,160 --> 00:41:46,640
Speaker 1: there's collateral, so maybe that brings the overall like borrowing

832
00:41:46,719 --> 00:41:48,640
Speaker 1: rate down. But on the other hand, it's kind of

833
00:41:48,640 --> 00:41:51,960
Speaker 1: a new thing, new structure. How does that compare with

834
00:41:52,040 --> 00:41:53,640
Speaker 1: more traditional types of finance.

835
00:41:53,920 --> 00:41:57,239
Speaker 3: Yeah, so you know that every credit facility that we do,

836
00:41:57,600 --> 00:42:01,280
Speaker 3: the cost of capital declines, and it's declining because it's

837
00:42:01,560 --> 00:42:04,839
Speaker 3: the execution risk and the ongoing concern risk are reduced. Right.

838
00:42:04,960 --> 00:42:07,200
Speaker 3: And you know, when we first did this, people like

839
00:42:07,239 --> 00:42:09,200
Speaker 3: you guys are crazy. You have no history of execution.

840
00:42:09,880 --> 00:42:12,120
Speaker 3: And as we've gone through and we've done it, like

841
00:42:12,200 --> 00:42:14,879
Speaker 3: now there's a path that everybody that's underwriting these loans

842
00:42:14,880 --> 00:42:16,719
Speaker 3: now understands. Okay, this is what happens, this is how

843
00:42:16,719 --> 00:42:19,240
Speaker 3: it reforms, This is what we should expect from the customers.

844
00:42:19,239 --> 00:42:21,239
Speaker 3: This is what we should expect from receivables. They get

845
00:42:21,280 --> 00:42:23,960
Speaker 3: more comfortable, they're willing to do it at more aggressive rates, right,

846
00:42:24,000 --> 00:42:25,960
Speaker 3: so that the risk premium associated with it has just

847
00:42:26,080 --> 00:42:26,959
Speaker 3: decreased over time.

848
00:42:27,080 --> 00:42:27,480
Speaker 1: Got it.

849
00:42:27,640 --> 00:42:30,040
Speaker 2: I just have one last question I sort of touched

850
00:42:30,080 --> 00:42:33,560
Speaker 2: on it earlier. But Okay, we know that power is scarce.

851
00:42:34,040 --> 00:42:37,200
Speaker 2: We know that, you know, there's not an infinite number

852
00:42:37,320 --> 00:42:41,360
Speaker 2: of Nvidia chips et cetera. Like those are quite scarce

853
00:42:41,760 --> 00:42:43,840
Speaker 2: for the other stuff. You know, we've done episodes in

854
00:42:43,880 --> 00:42:47,760
Speaker 2: the past like talking about like just generic electrical gear components,

855
00:42:47,760 --> 00:42:50,239
Speaker 2: and we've certainly done a lot on like labor shortages.

856
00:42:50,480 --> 00:42:52,600
Speaker 2: What are you seeing on that front sort of like

857
00:42:53,000 --> 00:42:56,239
Speaker 2: simple gear and the sort of basic building blocks of

858
00:42:56,239 --> 00:42:59,640
Speaker 2: a new construction and how difficult that is to acquire.

859
00:42:59,760 --> 00:43:01,839
Speaker 2: Verse to say, if you were doing this, you know

860
00:43:02,000 --> 00:43:03,920
Speaker 2: you started in twenty seventeen, I imagine a lot of

861
00:43:03,920 --> 00:43:05,279
Speaker 2: the things were more plentiful back then.

862
00:43:05,520 --> 00:43:09,280
Speaker 3: Yeah, so it's not even that they're less plentiful today

863
00:43:09,320 --> 00:43:11,600
Speaker 3: than they were. You know, the lead times were always

864
00:43:11,640 --> 00:43:14,560
Speaker 3: the lead times for this electrical gear. It's that there

865
00:43:14,640 --> 00:43:18,160
Speaker 3: was capacity to go buy off the shelf, right there

866
00:43:18,160 --> 00:43:20,360
Speaker 3: was inventory in the data center market. And the inventory

867
00:43:20,440 --> 00:43:23,520
Speaker 3: is basically gone. And you know, I see deals today

868
00:43:23,960 --> 00:43:25,960
Speaker 3: that get brought to me and there's seven people bidding

869
00:43:25,960 --> 00:43:27,919
Speaker 3: on the same deal and they're all trying to sell

870
00:43:27,920 --> 00:43:31,000
Speaker 3: it to like similar customers. So the market has gotten

871
00:43:31,040 --> 00:43:33,200
Speaker 3: pretty thin. So now you're looking at it, going Okay,

872
00:43:33,480 --> 00:43:36,600
Speaker 3: my only option here is for new built, and you're

873
00:43:36,680 --> 00:43:39,240
Speaker 3: looking at lead times that haven't really shifted that much

874
00:43:39,360 --> 00:43:43,080
Speaker 3: on things inside of the data center. The substation transformers

875
00:43:43,120 --> 00:43:46,840
Speaker 3: are multiple years out, and part of that reason is

876
00:43:46,880 --> 00:43:49,160
Speaker 3: that it takes a year for them to cure after

877
00:43:49,200 --> 00:43:51,440
Speaker 3: they're manufactured. Like, there's no getting around that, there's no

878
00:43:51,480 --> 00:43:52,279
Speaker 3: speeding that piece up.

879
00:43:52,600 --> 00:43:53,600
Speaker 2: I mean, it takes a year.

880
00:43:53,800 --> 00:43:56,840
Speaker 3: You when the transformer is built, that's taking on so

881
00:43:56,920 --> 00:44:00,560
Speaker 3: much power that whatever the process is, it has to

882
00:44:00,600 --> 00:44:03,360
Speaker 3: sit for a year and harden before it's able to

883
00:44:03,360 --> 00:44:05,560
Speaker 3: take on that electrical load. So even if you went

884
00:44:05,600 --> 00:44:06,800
Speaker 3: and said, hey, I'm going to build ten more of

885
00:44:06,840 --> 00:44:08,840
Speaker 3: these this year, it's still a year away before you

886
00:44:08,840 --> 00:44:09,399
Speaker 3: can use them.

887
00:44:09,520 --> 00:44:10,399
Speaker 2: Huh right.

888
00:44:10,440 --> 00:44:12,799
Speaker 3: And those are the types of things from a manufacturing

889
00:44:12,800 --> 00:44:15,200
Speaker 3: perspective you just can't get around, and it takes time

890
00:44:15,239 --> 00:44:17,400
Speaker 3: for the supply chain to catch up. But you know,

891
00:44:17,719 --> 00:44:19,480
Speaker 3: the problems that I'm solving on a day to day

892
00:44:19,480 --> 00:44:22,920
Speaker 3: basis in these builds isn't even around the substation transformers.

893
00:44:22,960 --> 00:44:25,799
Speaker 3: It's around like small components that somebody missed it when

894
00:44:25,840 --> 00:44:28,480
Speaker 3: they ordered the gear sixteen weeks ago. And now you

895
00:44:28,560 --> 00:44:30,799
Speaker 3: have to go scramble and call in favors across the

896
00:44:30,800 --> 00:44:32,600
Speaker 3: country of Hey, who has this part? I need it

897
00:44:32,600 --> 00:44:35,400
Speaker 3: by tomorrow because I have fifty thousand GPUs that are

898
00:44:35,400 --> 00:44:38,319
Speaker 3: blocked by this one little thing, right, So it's a

899
00:44:38,320 --> 00:44:40,880
Speaker 3: lot of it is logistical and human coordination and solving

900
00:44:40,960 --> 00:44:42,239
Speaker 3: dumb problems in real time.

901
00:44:42,640 --> 00:44:45,560
Speaker 2: Ryan Venturro, thank you so much for coming on odd Laws.

902
00:44:45,600 --> 00:45:01,280
Speaker 2: That was fantastic. Thanks for having me, Tracy. I'm really

903
00:45:01,360 --> 00:45:04,240
Speaker 2: glad we did that conversation because there are a number

904
00:45:04,280 --> 00:45:07,160
Speaker 2: of these sort of like big picture ideas in there

905
00:45:07,160 --> 00:45:09,439
Speaker 2: that we've sort of hit on of course, about data

906
00:45:09,480 --> 00:45:12,359
Speaker 2: centers and AI and electricity consumption, and it was really

907
00:45:12,400 --> 00:45:15,600
Speaker 2: interesting to hear some of them. So, like, for example,

908
00:45:16,440 --> 00:45:19,200
Speaker 2: just this idea of like northern Virginia is out and

909
00:45:19,239 --> 00:45:22,960
Speaker 2: like needing this sort of hunt to find these spots

910
00:45:23,440 --> 00:45:27,840
Speaker 2: in the country where there is ample electricity and basically

911
00:45:28,080 --> 00:45:30,120
Speaker 2: nobody local is going to get upset at you for

912
00:45:30,239 --> 00:45:30,600
Speaker 2: using it.

913
00:45:31,320 --> 00:45:33,359
Speaker 1: Yeah, no one will come out with pitchforks. The thing

914
00:45:33,400 --> 00:45:35,839
Speaker 1: that stood out to me from a bunch of these

915
00:45:35,880 --> 00:45:39,319
Speaker 1: conversations at this point is the arms race aspect of it,

916
00:45:39,400 --> 00:45:43,239
Speaker 1: and how urgent building out AI is for a lot

917
00:45:43,239 --> 00:45:46,480
Speaker 1: of these companies, and then there seems to be this

918
00:45:46,640 --> 00:45:52,399
Speaker 1: mismatch between the immediate need for scale and compute and

919
00:45:52,960 --> 00:45:58,960
Speaker 1: energy now versus these really long timelines of actually building

920
00:45:58,960 --> 00:46:04,239
Speaker 1: the stuff out and Brian mentioning the substation transformers taking

921
00:46:04,280 --> 00:46:05,000
Speaker 1: a care of cure.

922
00:46:05,080 --> 00:46:06,040
Speaker 2: I had no idea about that.

923
00:46:06,080 --> 00:46:08,240
Speaker 1: I didn't know that either. But that's a really good example.

924
00:46:08,640 --> 00:46:11,480
Speaker 2: That's super interesting, and of course now we have to

925
00:46:11,520 --> 00:46:14,440
Speaker 2: do a how do you build a substation transform.

926
00:46:14,080 --> 00:46:16,120
Speaker 1: How do you cure a substation transformer?

927
00:46:16,200 --> 00:46:18,600
Speaker 2: Totally? I mean maybe this is probably something that electrical

928
00:46:18,640 --> 00:46:21,000
Speaker 2: engineer is not interesting to them at all, But for me,

929
00:46:21,120 --> 00:46:23,759
Speaker 2: I did not realize that there was this one year long,

930
00:46:25,440 --> 00:46:28,400
Speaker 2: one year long curing process. You know, I think there

931
00:46:28,400 --> 00:46:31,520
Speaker 2: are like a couple other things that now I want

932
00:46:31,560 --> 00:46:35,200
Speaker 2: to talk more about, so I'm interested. I mean, like

933
00:46:35,280 --> 00:46:38,399
Speaker 2: Coreweave is an in video company. It's not owned by Video,

934
00:46:38,440 --> 00:46:41,120
Speaker 2: but you know it's joined at the hip in many respects.

935
00:46:41,280 --> 00:46:44,680
Speaker 2: So how difficult is it going to be either for

936
00:46:45,280 --> 00:46:49,040
Speaker 2: some other maker of chips, whether it's an Intel or

937
00:46:49,120 --> 00:46:54,160
Speaker 2: some other maker of software environments, whether it's Meta and

938
00:46:54,719 --> 00:46:58,160
Speaker 2: PyTorch going against Kuda or whatever, like that's a really

939
00:46:58,440 --> 00:47:02,080
Speaker 2: interesting question to me, Like, you know, we have to

940
00:47:02,120 --> 00:47:05,000
Speaker 2: do more essentially on like how much of a lock

941
00:47:05,080 --> 00:47:06,640
Speaker 2: and video really has on this industry.

942
00:47:06,719 --> 00:47:09,440
Speaker 1: Yeah, this seems to be the really big question. And

943
00:47:09,480 --> 00:47:12,160
Speaker 1: then the other thing I was thinking about, and I

944
00:47:12,160 --> 00:47:15,600
Speaker 1: know Brian emphasized this and other Core Weave executives have

945
00:47:15,719 --> 00:47:20,200
Speaker 1: emphasized this before, but this idea that hyperscalers maybe are

946
00:47:20,239 --> 00:47:23,799
Speaker 1: starting from a point of being disadvantaged because they have

947
00:47:23,880 --> 00:47:28,840
Speaker 1: to retrofit all this old infrastructure for this new AI

948
00:47:29,040 --> 00:47:32,880
Speaker 1: technology totally, and like I can see that. But on

949
00:47:32,920 --> 00:47:37,439
Speaker 1: the other hand, these are insanely impressive companies. You are

950
00:47:37,520 --> 00:47:41,200
Speaker 1: explicitly trying to compete against Core Weave in this business,

951
00:47:41,560 --> 00:47:43,799
Speaker 1: and they're not going to stand still. And so I

952
00:47:43,800 --> 00:47:46,279
Speaker 1: guess there's an open question over how much progress they're

953
00:47:46,320 --> 00:47:49,400
Speaker 1: making or how fast that progress is actually happening.

954
00:47:49,600 --> 00:47:53,880
Speaker 2: Right, Large companies always are going to have some challenges

955
00:47:53,920 --> 00:47:57,000
Speaker 2: when there's like a new model or something. But these

956
00:47:57,000 --> 00:47:59,640
Speaker 2: companies have all the money in the entire world, right,

957
00:48:00,080 --> 00:48:01,520
Speaker 2: and they also have all you know, one of the

958
00:48:01,520 --> 00:48:03,680
Speaker 2: things that Brian said is like they if they were

959
00:48:03,719 --> 00:48:05,279
Speaker 2: if one of them are going to do it, they

960
00:48:05,280 --> 00:48:06,719
Speaker 2: would have to go out and to buy a big

961
00:48:06,800 --> 00:48:09,160
Speaker 2: chunk of the market, which again they have all the

962
00:48:09,160 --> 00:48:12,799
Speaker 2: money in the entire world. So theoretically, whether it's the

963
00:48:12,800 --> 00:48:16,240
Speaker 2: big companies and retrofitting the clouds or building new clouds,

964
00:48:16,719 --> 00:48:18,600
Speaker 2: or you know a lot of them like a Google,

965
00:48:18,680 --> 00:48:22,759
Speaker 2: even if they're for now using their TPUs internally primarily like,

966
00:48:23,440 --> 00:48:26,480
Speaker 2: it does seem like in theory the opportunities out there,

967
00:48:26,680 --> 00:48:31,239
Speaker 2: particularly with the the sky high amount you know, valuation

968
00:48:31,360 --> 00:48:33,840
Speaker 2: that a company like in video is getting.

969
00:48:34,000 --> 00:48:36,279
Speaker 1: Oh yeah, you mentioned the sky high valuation. That was

970
00:48:36,280 --> 00:48:38,719
Speaker 1: something that also stood out to me, just on the

971
00:48:38,760 --> 00:48:42,240
Speaker 1: financing side. So this idea of you know, the debt

972
00:48:42,239 --> 00:48:45,920
Speaker 1: financing deal that they did, and I'm not going to

973
00:48:45,960 --> 00:48:47,839
Speaker 1: call it trade receivables because.

974
00:48:47,680 --> 00:48:49,239
Speaker 2: No one GPU backed loan.

975
00:48:49,360 --> 00:48:51,520
Speaker 1: Yeah, no one will be interested when we start talking

976
00:48:51,520 --> 00:48:55,320
Speaker 1: about trade receivables. But the GPU back loan. This idea

977
00:48:55,360 --> 00:48:58,080
Speaker 1: that like, okay, it's a new structure, but the more

978
00:48:58,120 --> 00:49:01,120
Speaker 1: you do it, the more the cost of particular capital

979
00:49:01,239 --> 00:49:03,680
Speaker 1: starts to fall, the more the market gets comfortable with it.

980
00:49:03,840 --> 00:49:05,839
Speaker 1: I mean, we can talk about whether or not it's

981
00:49:05,920 --> 00:49:10,239
Speaker 1: priced correctly for a new type of unfamiliar risk, but

982
00:49:11,040 --> 00:49:13,480
Speaker 1: it does seem like that might be a new avenue

983
00:49:13,680 --> 00:49:16,520
Speaker 1: for the vast amounts of capital that are needed for

984
00:49:16,600 --> 00:49:17,160
Speaker 1: this business.

985
00:49:17,360 --> 00:49:20,799
Speaker 2: So one, it's interesting to think about the idea that, like,

986
00:49:21,560 --> 00:49:24,160
Speaker 2: you know, I don't think it's like totally true. You

987
00:49:24,160 --> 00:49:26,960
Speaker 2: know that if you need compute at scale for AI,

988
00:49:27,520 --> 00:49:29,640
Speaker 2: that you don't just get to call up core weave

989
00:49:29,800 --> 00:49:31,600
Speaker 2: and get it, and you actually have to prove that

990
00:49:31,680 --> 00:49:34,279
Speaker 2: you're going to be a good customer and so like

991
00:49:34,560 --> 00:49:37,560
Speaker 2: have something that is probably going to be sustainable, have

992
00:49:37,680 --> 00:49:40,839
Speaker 2: the balance sheet capacity. So this even if the sort

993
00:49:40,880 --> 00:49:45,520
Speaker 2: of software the end users aren't themselves raising debt, it

994
00:49:45,560 --> 00:49:47,239
Speaker 2: does sound like they have to have a lot of

995
00:49:47,320 --> 00:49:53,319
Speaker 2: equity upfront just so that they're perceived as a sustainable,

996
00:49:53,480 --> 00:49:56,960
Speaker 2: viable customer for a company like corewev. I also thought

997
00:49:57,040 --> 00:49:59,759
Speaker 2: on the electricity front, like obviously we talk all the

998
00:49:59,760 --> 00:50:03,160
Speaker 2: time about just sort of the raw demand for electricity.

999
00:50:03,520 --> 00:50:05,680
Speaker 2: But this idea what he said, and I hadn't heard

1000
00:50:05,719 --> 00:50:09,200
Speaker 2: anyone say it that the runs the modeling runs stop

1001
00:50:09,239 --> 00:50:11,640
Speaker 2: everyone do you say thirty minutes and have to be saved.

1002
00:50:11,680 --> 00:50:14,600
Speaker 2: Oh yeah, And so you have this big variability at times,

1003
00:50:14,600 --> 00:50:17,880
Speaker 2: and that creates its own specific issue because it's not

1004
00:50:18,000 --> 00:50:21,600
Speaker 2: just steady state flow of electricity and solving for that.

1005
00:50:21,600 --> 00:50:26,640
Speaker 2: That's probably another area in which the legacy data centers

1006
00:50:26,760 --> 00:50:30,439
Speaker 2: or cloud companies. Perhaps my guess would be that they're

1007
00:50:30,480 --> 00:50:33,719
Speaker 2: just sort of the demand is more constant and therefore

1008
00:50:34,120 --> 00:50:36,200
Speaker 2: something that would be a novelty for them.

1009
00:50:36,239 --> 00:50:38,279
Speaker 1: Just thinking about the financing more, I do kind of

1010
00:50:38,280 --> 00:50:41,440
Speaker 1: wonder how much of this is like AI built on

1011
00:50:41,520 --> 00:50:44,960
Speaker 1: top of AI on top of AI. Like, yeah, to

1012
00:50:45,000 --> 00:50:49,200
Speaker 1: the point where if if the bubble were to burst,

1013
00:50:49,320 --> 00:50:52,000
Speaker 1: or if funding was suddenly pulled from a bunch of

1014
00:50:52,000 --> 00:50:56,960
Speaker 1: these startups, like what would that mean for core weaves financing?

1015
00:50:57,480 --> 00:51:00,200
Speaker 1: And what would that mean for black Rock, which lent

1016
00:51:00,320 --> 00:51:03,480
Speaker 1: money based on the GPUs that the clients are taking on,

1017
00:51:03,640 --> 00:51:05,480
Speaker 1: who might not be there anymore. I don't know.

1018
00:51:05,760 --> 00:51:07,359
Speaker 2: By the way, have you ever looked at a chart

1019
00:51:07,440 --> 00:51:09,000
Speaker 2: of riot lockschain?

1020
00:51:09,200 --> 00:51:12,200
Speaker 1: Oh no, not for a while?

1021
00:51:12,280 --> 00:51:14,120
Speaker 2: Yeah, well, I mean they're still there as a minor,

1022
00:51:14,200 --> 00:51:16,080
Speaker 2: but like here we are in the midst of this

1023
00:51:16,320 --> 00:51:18,399
Speaker 2: pretty big crypto bal run. I mean, I guess it's

1024
00:51:18,400 --> 00:51:20,600
Speaker 2: cooled a little bit, but and that stock is done

1025
00:51:20,719 --> 00:51:24,560
Speaker 2: terribly so it's interesting to wonder, and apparently it doesn't

1026
00:51:24,560 --> 00:51:26,440
Speaker 2: seem like anyone's made a bid for them. But it

1027
00:51:26,600 --> 00:51:30,800
Speaker 2: is interesting to wonder, like, Okay, those chips are useless

1028
00:51:31,000 --> 00:51:36,759
Speaker 2: for AI because they don't work for that, but you know,

1029
00:51:36,840 --> 00:51:41,560
Speaker 2: they do have capacity and they do have electricity agreements

1030
00:51:41,640 --> 00:51:44,360
Speaker 2: already in place. So it does make you wonder whether,

1031
00:51:44,480 --> 00:51:46,879
Speaker 2: like some of the bitcoin mining companies which aren't really

1032
00:51:46,960 --> 00:51:50,760
Speaker 2: getting a very the market is not excited about them, clearly,

1033
00:51:50,840 --> 00:51:53,320
Speaker 2: even in the midst of this crypto bal run.

1034
00:51:53,239 --> 00:51:56,680
Speaker 1: Maybe they should go back to being a diagnostics company.

1035
00:51:56,800 --> 00:51:59,120
Speaker 1: That's what they were before, is it. I think so.

1036
00:51:59,760 --> 00:52:01,520
Speaker 1: I think they're one of the ones that changed their

1037
00:52:01,600 --> 00:52:05,000
Speaker 1: name and then like there something including blockchain, and then

1038
00:52:05,040 --> 00:52:07,520
Speaker 1: their shares went up enormously and now they're back down.

1039
00:52:07,640 --> 00:52:11,760
Speaker 2: Well they have been. Riot Platforms has been around, Okay,

1040
00:52:11,840 --> 00:52:16,120
Speaker 2: now I'm curious. Yeah, so it's a bitcoin mining company,

1041
00:52:16,160 --> 00:52:17,960
Speaker 2: but it's been the stock has been around since two

1042
00:52:17,960 --> 00:52:22,759
Speaker 2: thousand and three. So pretty clearly, uh, pretty clearly they

1043
00:52:22,760 --> 00:52:24,360
Speaker 2: were in some other business. I don't know what.

1044
00:52:24,520 --> 00:52:26,680
Speaker 1: Yeah, I'm looking on the terminal, it says Riot Blockchain,

1045
00:52:26,880 --> 00:52:31,880
Speaker 1: formerly Bioptics, has ditched the drug diagnostic machinery business for

1046
00:52:31,960 --> 00:52:34,040
Speaker 1: the digital currency trade.

1047
00:52:34,800 --> 00:52:37,000
Speaker 2: Well, there you go. So if you have some sort

1048
00:52:37,000 --> 00:52:38,799
Speaker 2: of computing power or something. I don't know what they

1049
00:52:38,800 --> 00:52:41,520
Speaker 2: were doing before, but maybe it is interesting to think about.

1050
00:52:41,719 --> 00:52:44,000
Speaker 2: Maybe some of the option value for some of these

1051
00:52:44,040 --> 00:52:47,960
Speaker 2: miners isn't there. Non is in all the infrastructure other

1052
00:52:48,040 --> 00:52:49,800
Speaker 2: than the bitcoin mining operation.

1053
00:52:50,080 --> 00:52:51,360
Speaker 1: Maybe we should put in a bid.

1054
00:52:51,560 --> 00:52:52,040
Speaker 2: Let's do it.

1055
00:52:52,120 --> 00:52:56,239
Speaker 1: We can crowdfund and start our own business. Okay, maybe

1056
00:52:56,239 --> 00:52:56,960
Speaker 1: we should leave it there.

1057
00:52:57,040 --> 00:52:57,799
Speaker 2: Let's leave it there.

1058
00:52:57,960 --> 00:53:00,719
Speaker 1: This has been another episode of the All Thought podcast.

1059
00:53:00,840 --> 00:53:03,960
Speaker 1: I'm Tracy Alloway. You can follow me at Tracy Alloway and.

1060
00:53:03,920 --> 00:53:06,560
Speaker 2: I'm Joe Wisenthal. You can follow me at the Stalwart.

1061
00:53:06,840 --> 00:53:10,440
Speaker 2: Follow our guest Brian Venturo. He's at Brian Venturo. Follow

1062
00:53:10,480 --> 00:53:13,880
Speaker 2: our producers Carmen Rodriguez at Carman Erman dash Ol Bennett

1063
00:53:13,920 --> 00:53:17,040
Speaker 2: at Dashbot, and Kilbrooks at Kilbrooks. Thank you to our

1064
00:53:17,080 --> 00:53:20,040
Speaker 2: producer Moses Ondam. For more odd Lots content, go to

1065
00:53:20,080 --> 00:53:22,719
Speaker 2: Bloomberg dot com slash odd Lots, where we have transcripts,

1066
00:53:22,719 --> 00:53:25,680
Speaker 2: a blog, and a newsletter and you can chat about

1067
00:53:25,680 --> 00:53:30,280
Speaker 2: all of these topics, including AI, including semiconductors, including energy

1068
00:53:30,400 --> 00:53:33,759
Speaker 2: in our discord discord gg slash.

1069
00:53:33,400 --> 00:53:36,239
Speaker 1: Hot Lots and if you enjoy all thoughts, if you

1070
00:53:36,440 --> 00:53:38,960
Speaker 1: like it when we talk about AI and chips and

1071
00:53:39,160 --> 00:53:41,600
Speaker 1: energy and all that stuff, then please leave us a

1072
00:53:41,680 --> 00:53:45,560
Speaker 1: positive review on your favorite podcast platform. And remember, if

1073
00:53:45,600 --> 00:53:48,399
Speaker 1: you are a Bloomberg subscriber, you can listen to all

1074
00:53:48,440 --> 00:53:51,600
Speaker 1: of our episodes absolutely ad free. All you need to

1075
00:53:51,600 --> 00:53:55,000
Speaker 1: do is connect your Bloomberg account with Apple Podcasts. In

1076
00:53:55,120 --> 00:53:57,560
Speaker 1: order to do that, just find the Bloomberg channel on

1077
00:53:57,640 --> 00:54:01,120
Speaker 1: Apple Podcasts and follow the instructions there. Thanks for listening.