1
00:00:10,160 --> 00:00:14,400
Speaker 1: Hello, and welcome to another episode of the Odd Lots Podcast.

2
00:00:14,480 --> 00:00:16,680
Speaker 2: I'm Joe Wisenthal and I'm Tracy Alloway.

3
00:00:16,800 --> 00:00:19,360
Speaker 1: Tracy, have you looked at in video stock chart lately?

4
00:00:19,440 --> 00:00:21,080
Speaker 1: And by lately, I don't mean like over the last

5
00:00:21,120 --> 00:00:22,680
Speaker 1: two years. I mean like just like over the last

6
00:00:22,720 --> 00:00:24,520
Speaker 1: like two weeks or two months.

7
00:00:24,560 --> 00:00:26,400
Speaker 2: I don't need to look at it because everyone keeps

8
00:00:26,440 --> 00:00:28,760
Speaker 2: talking about it. So I know, I know what's happening.

9
00:00:28,960 --> 00:00:30,880
Speaker 1: You know what, I'm pretty happy about. Could I just say,

10
00:00:31,160 --> 00:00:33,960
Speaker 1: you know, we did that episode like two months ago, yes,

11
00:00:34,200 --> 00:00:37,519
Speaker 1: with Stacy Raskin, and we were like, what's what's up

12
00:00:37,520 --> 00:00:39,239
Speaker 1: a good video, like well, you know, I know it's

13
00:00:39,240 --> 00:00:42,920
Speaker 1: at the center of the AI chips boom and whatever.

14
00:00:43,479 --> 00:00:45,199
Speaker 1: And then like we did that episode and it came

15
00:00:45,200 --> 00:00:46,680
Speaker 1: out and then a week later like they just like

16
00:00:46,800 --> 00:00:47,600
Speaker 1: knocked it out of the park.

17
00:00:48,479 --> 00:00:49,639
Speaker 3: Yeah, so you.

18
00:00:49,640 --> 00:00:51,000
Speaker 2: Know, we were early.

19
00:00:51,880 --> 00:00:53,640
Speaker 3: We were at least like, you know, a good like

20
00:00:53,680 --> 00:00:54,400
Speaker 3: two weeks earlier.

21
00:00:55,160 --> 00:00:56,840
Speaker 2: Hey, hey, two weeks I'll take it.

22
00:00:57,440 --> 00:00:58,360
Speaker 3: I'll take it.

23
00:00:58,440 --> 00:01:01,800
Speaker 1: So clearly something that you know, we and we talked

24
00:01:01,800 --> 00:01:03,760
Speaker 1: about this with Stacy, like you know, something that in

25
00:01:03,840 --> 00:01:06,840
Speaker 1: Nvidia has is like everyone's trying to buy it. Everyone's

26
00:01:06,840 --> 00:01:08,839
Speaker 1: trying to get it, But then it raises the next

27
00:01:08,880 --> 00:01:11,240
Speaker 1: question of like, okay, but what is that market? Like

28
00:01:11,520 --> 00:01:12,560
Speaker 1: how do you buy a chip?

29
00:01:13,000 --> 00:01:14,959
Speaker 2: Yeah? How do you buy a chip? And then I

30
00:01:14,959 --> 00:01:17,080
Speaker 2: guess what do you actually do with it once you

31
00:01:17,200 --> 00:01:19,280
Speaker 2: have it? Because my impression is that for a lot

32
00:01:19,319 --> 00:01:23,279
Speaker 2: of these AI applications, the way you use the chips,

33
00:01:23,319 --> 00:01:26,119
Speaker 2: the way you set up the data centers is very,

34
00:01:26,240 --> 00:01:29,399
Speaker 2: very different to what we've seen in the past. And

35
00:01:29,440 --> 00:01:32,280
Speaker 2: I think also what in Vidia is doing now is

36
00:01:32,360 --> 00:01:34,959
Speaker 2: kind of different. But maybe we can get into this

37
00:01:35,040 --> 00:01:36,920
Speaker 2: with our guests. My impression is they're trying to create

38
00:01:37,280 --> 00:01:40,960
Speaker 2: a sort of like holistic approach for customers where they

39
00:01:41,000 --> 00:01:44,840
Speaker 2: provide not just the hardware, but also some services to

40
00:01:44,880 --> 00:01:45,680
Speaker 2: go along with it.

41
00:01:46,000 --> 00:01:48,720
Speaker 1: Yes, right, and like all the software and Stacy talked

42
00:01:48,760 --> 00:01:50,880
Speaker 1: about that with the Kuda ecosystemica that.

43
00:01:50,840 --> 00:01:53,160
Speaker 3: Was it, how dominant that is? But right, like what

44
00:01:53,200 --> 00:01:55,360
Speaker 3: do you do with it? Like how do you get one? If?

45
00:01:55,400 --> 00:01:56,800
Speaker 3: Like what you know, what would we do.

46
00:01:56,760 --> 00:02:00,760
Speaker 1: Tracy if a big palette of in Vidia chips wound

47
00:02:00,840 --> 00:02:01,840
Speaker 1: up here?

48
00:02:01,480 --> 00:02:03,920
Speaker 2: Do you want to know a secret? Yeah, my basement

49
00:02:03,960 --> 00:02:06,600
Speaker 2: is filled with h one hundred chips. Just got a

50
00:02:06,640 --> 00:02:08,080
Speaker 2: pile of them. It came with the house.

51
00:02:08,440 --> 00:02:10,880
Speaker 1: It was on that ship that was stuck off the Chesapeake,

52
00:02:10,960 --> 00:02:12,960
Speaker 1: and instead of getting your cowards, you got it.

53
00:02:13,080 --> 00:02:15,720
Speaker 2: I just caught a palette of age four hundreds.

54
00:02:15,440 --> 00:02:19,680
Speaker 1: That that well, we're manifesting that into reality. So anyway,

55
00:02:19,760 --> 00:02:22,600
Speaker 1: I like how this world works so essentially, like the

56
00:02:22,800 --> 00:02:27,160
Speaker 1: trading and dealing of these, like the hottest commodity in

57
00:02:27,200 --> 00:02:30,920
Speaker 1: the world right which is these these advanced chips from AI,

58
00:02:31,000 --> 00:02:32,920
Speaker 1: and how that works and who can get one? I

59
00:02:32,960 --> 00:02:35,520
Speaker 1: still think is like a sort of mystery that we

60
00:02:35,639 --> 00:02:38,160
Speaker 1: need to delve further into this question.

61
00:02:38,400 --> 00:02:40,359
Speaker 2: I agree, And there is also there's a lot of

62
00:02:40,440 --> 00:02:43,600
Speaker 2: excitement around it right now for the obvious reasons of

63
00:02:43,680 --> 00:02:48,239
Speaker 2: everyone's really into generative AI and in video stock is exploding,

64
00:02:48,240 --> 00:02:50,640
Speaker 2: as we already talked about, but we're also seeing a

65
00:02:50,680 --> 00:02:55,560
Speaker 2: lot of previous I guess consumers of chips, like the

66
00:02:55,600 --> 00:02:59,079
Speaker 2: crypto miners start to pivot into the space, and I'd

67
00:02:59,080 --> 00:03:01,720
Speaker 2: be curious to see what they're doing in it as well,

68
00:03:01,800 --> 00:03:04,960
Speaker 2: and how much of that is just you know, desperation

69
00:03:05,520 --> 00:03:08,959
Speaker 2: versus versus a real business opportunity.

70
00:03:08,440 --> 00:03:09,440
Speaker 3: In the video game market.

71
00:03:09,520 --> 00:03:11,079
Speaker 2: Yeah, oh totally, I forgot about.

72
00:03:11,160 --> 00:03:13,320
Speaker 1: Which was like the other thing. It's like, for years

73
00:03:13,360 --> 00:03:15,519
Speaker 1: I thought of Nvidia is the video game company. Yeah,

74
00:03:15,560 --> 00:03:17,639
Speaker 1: because they had their logo on xboxes.

75
00:03:17,720 --> 00:03:21,680
Speaker 2: And how realistic is that pivot? What proportion of those

76
00:03:21,720 --> 00:03:23,840
Speaker 2: types of chips can be used for AI?

77
00:03:24,000 --> 00:03:27,080
Speaker 1: Now, well, I'm very excited. We do have I believe

78
00:03:27,240 --> 00:03:29,400
Speaker 1: the perfect guest. We are going to be speaking with

79
00:03:29,440 --> 00:03:32,919
Speaker 1: Brandon McBee. He is the chief strategy officer and co

80
00:03:33,000 --> 00:03:36,720
Speaker 1: founder of core Weave, which is a specialized cloud services

81
00:03:36,800 --> 00:03:40,680
Speaker 1: provider that's basically providing this sort of like high volume

82
00:03:40,800 --> 00:03:45,160
Speaker 1: compute to AI type companies. They recently raised over four

83
00:03:45,240 --> 00:03:47,560
Speaker 1: hundred million dollars. Have been in this space for a

84
00:03:47,600 --> 00:03:50,280
Speaker 1: little while. So Brandon, thank you so much for coming

85
00:03:50,320 --> 00:03:51,160
Speaker 1: on odd lots.

86
00:03:51,520 --> 00:03:53,640
Speaker 4: Thanks for the opportunity. Guys, really excited to chat with

87
00:03:53,640 --> 00:03:54,160
Speaker 4: you all today.

88
00:03:54,720 --> 00:03:57,240
Speaker 1: So let's just let me sorry if Tracy and I, like,

89
00:03:57,360 --> 00:03:58,880
Speaker 1: I don't know why they would do this, but if

90
00:03:58,920 --> 00:04:00,880
Speaker 1: like some VC was like, you know, we want you

91
00:04:00,960 --> 00:04:03,680
Speaker 1: to do on launch GPT. We want you to like

92
00:04:03,800 --> 00:04:07,880
Speaker 1: do a pore base large language model off of all

93
00:04:07,920 --> 00:04:09,640
Speaker 1: the work you've done. We want you to compete with

94
00:04:09,720 --> 00:04:12,240
Speaker 1: open AI. And they gave us like I don't know

95
00:04:12,440 --> 00:04:15,400
Speaker 1: some like, you know, one hundred million dollar rays, they said,

96
00:04:15,440 --> 00:04:18,640
Speaker 1: go start, do your startup? Could I call in video

97
00:04:19,320 --> 00:04:21,560
Speaker 1: and buy chips? Would I be able to get in

98
00:04:21,600 --> 00:04:22,120
Speaker 1: the door there?

99
00:04:22,600 --> 00:04:25,440
Speaker 4: Gosh? I mean you're I think you and everyone else

100
00:04:25,800 --> 00:04:27,960
Speaker 4: is asking that question, and you're going to have a

101
00:04:28,240 --> 00:04:31,039
Speaker 4: huge problem doing that. Right now, it's mostly just around

102
00:04:31,720 --> 00:04:35,520
Speaker 4: how much in demand this infrastructure became, right I mean,

103
00:04:35,600 --> 00:04:38,599
Speaker 4: you could argue it's one of the most critical pieces

104
00:04:38,600 --> 00:04:41,840
Speaker 4: of information technology resources on the planet right now and

105
00:04:42,400 --> 00:04:45,279
Speaker 4: suddenly everyone needs it, and you know, I like to

106
00:04:45,279 --> 00:04:49,800
Speaker 4: contextualize it in that, you know, the piece of software

107
00:04:49,800 --> 00:04:53,080
Speaker 4: adoption for AIS like one of the fastest adoption curves

108
00:04:53,080 --> 00:04:57,440
Speaker 4: we've ever seen, right Like, you're you're hitting these milestones

109
00:04:57,480 --> 00:05:01,280
Speaker 4: faster than any other software platform previously, and now all

110
00:05:01,320 --> 00:05:04,560
Speaker 4: of a sudden, you're asking infrastructure build to keep up

111
00:05:04,600 --> 00:05:07,360
Speaker 4: with that, right a space that traditionally takes more time,

112
00:05:07,480 --> 00:05:12,120
Speaker 4: and it's created this massive supply demand and balance just

113
00:05:12,200 --> 00:05:16,280
Speaker 4: on in place infrastructure today and not only infratructure is

114
00:05:16,279 --> 00:05:20,560
Speaker 4: available to purchase, and it's an issue that is going

115
00:05:20,600 --> 00:05:22,480
Speaker 4: to be ongoing for a bit as well, we think.

116
00:05:23,360 --> 00:05:26,479
Speaker 2: So can I ask the basic question, which is core weave.

117
00:05:27,400 --> 00:05:30,880
Speaker 2: What do you do exactly? Joe mentioned the capital raise,

118
00:05:30,920 --> 00:05:33,320
Speaker 2: which I think has you valued at something like two

119
00:05:33,320 --> 00:05:37,719
Speaker 2: billion dollars, So congrats, but what exactly are you doing here?

120
00:05:38,160 --> 00:05:41,400
Speaker 4: Yeah? Thank you. So Corey is a specialized cloud service

121
00:05:41,400 --> 00:05:45,920
Speaker 4: provider that is focused on highly parallelizable workloads. So we

122
00:05:46,320 --> 00:05:50,440
Speaker 4: build and operate the world's most performant GPU infrastructure at

123
00:05:50,480 --> 00:05:54,760
Speaker 4: scale and predominantly serve three sectors. That's the artificial intelligence sector,

124
00:05:54,920 --> 00:05:58,520
Speaker 4: the media and entertainment sector, and the computational chemistry sector.

125
00:05:58,640 --> 00:06:04,320
Speaker 4: So we build specialize in building this infrastructure at supercompute scale.

126
00:06:04,480 --> 00:06:07,560
Speaker 4: It's like quite literally, you know, it's sixteen thousand GPU

127
00:06:07,760 --> 00:06:09,640
Speaker 4: fabric and we can get into all the details and

128
00:06:09,680 --> 00:06:12,200
Speaker 4: how complex that is. But we build that so that

129
00:06:12,320 --> 00:06:16,280
Speaker 4: entities can come in and train these next generation foundation

130
00:06:16,440 --> 00:06:19,040
Speaker 4: machine learning models on. And you know, we found ourselves

131
00:06:19,080 --> 00:06:20,760
Speaker 4: in a spot where we can do that better than

132
00:06:20,880 --> 00:06:23,279
Speaker 4: literally anyone else in the market and do it on

133
00:06:23,320 --> 00:06:27,159
Speaker 4: a timeline that's faster or I think the only entity

134
00:06:27,200 --> 00:06:32,080
Speaker 4: with H one hundred available to clients at scale globally today.

135
00:06:32,800 --> 00:06:35,120
Speaker 2: So you have an actual basement full of H one

136
00:06:35,200 --> 00:06:38,960
Speaker 2: hundred chips. Well, can you talk to us. You know,

137
00:06:39,040 --> 00:06:42,560
Speaker 2: when you say infrastructure, we help clients build out the infrastructure,

138
00:06:42,640 --> 00:06:47,599
Speaker 2: help us conceptualize this. What does what does the infrastructure

139
00:06:48,080 --> 00:06:51,880
Speaker 2: for this type of AI actually look like? And how

140
00:06:51,880 --> 00:06:55,720
Speaker 2: does it differ to infrastructure for other types of large

141
00:06:55,720 --> 00:06:57,440
Speaker 2: scale technology projects.

142
00:06:57,920 --> 00:07:01,000
Speaker 4: Yeah, totally, so, you know, I I think during the

143
00:07:01,080 --> 00:07:04,200
Speaker 4: last in video quarterly earnings called Jensen put this a

144
00:07:04,200 --> 00:07:06,839
Speaker 4: really great way in the Q and A section, he

145
00:07:06,960 --> 00:07:10,000
Speaker 4: said that we are at the first year of a

146
00:07:10,120 --> 00:07:13,760
Speaker 4: decade long modernization of the data center, or like making

147
00:07:13,760 --> 00:07:16,640
Speaker 4: the data center intelligent. Right, you can kind of you

148
00:07:16,640 --> 00:07:19,840
Speaker 4: could suggest that the last generation or the twenty tens

149
00:07:20,040 --> 00:07:23,280
Speaker 4: data center was comprised of CPU, compute, storage and these

150
00:07:23,320 --> 00:07:26,840
Speaker 4: things that didn't really work together that intelligently. And the

151
00:07:26,880 --> 00:07:29,880
Speaker 4: way that in Nvidia has positioned itself is to make

152
00:07:29,880 --> 00:07:32,320
Speaker 4: it a smart data center that's like smart routing of

153
00:07:32,480 --> 00:07:36,160
Speaker 4: data packets of different pieces of infrastructure in there. That's

154
00:07:36,200 --> 00:07:39,680
Speaker 4: all focused on how do you expand the throughput in

155
00:07:39,720 --> 00:07:46,440
Speaker 4: communicability of and between pieces of infrastructure. Right, It's just

156
00:07:46,840 --> 00:07:50,840
Speaker 4: an amazingly different approach to data center deployments. And so

157
00:07:51,680 --> 00:07:54,160
Speaker 4: the way that we're building it and we're working with

158
00:07:54,480 --> 00:07:58,360
Speaker 4: Nvidia infrastructure. We design everything to a DGX reference back

159
00:07:58,400 --> 00:08:01,160
Speaker 4: in dgx's in videos like how do you draw the

160
00:08:01,200 --> 00:08:04,720
Speaker 4: most performance out of Nvidia infrastructure is possible with all

161
00:08:04,720 --> 00:08:08,000
Speaker 4: the anciliary components associated with it. So all this stuff

162
00:08:08,040 --> 00:08:10,960
Speaker 4: is going into what's qualified as a Tier three or

163
00:08:11,000 --> 00:08:14,440
Speaker 4: a Tier four data center. We collocate with within these things,

164
00:08:14,440 --> 00:08:17,520
Speaker 4: so we're not quite building in a basement, even though

165
00:08:17,800 --> 00:08:21,400
Speaker 4: like in our past history we certainly you know, had

166
00:08:21,840 --> 00:08:24,640
Speaker 4: time doing that, but this is within you know, just

167
00:08:25,120 --> 00:08:28,640
Speaker 4: amazing collocation sites that are operated by our partners such

168
00:08:28,640 --> 00:08:30,800
Speaker 4: as switch right. So a Tier three a Tier four

169
00:08:30,920 --> 00:08:35,680
Speaker 4: site is something that's qualified based on its ability to

170
00:08:35,720 --> 00:08:39,480
Speaker 4: serve workloads with an extremely high uptime. So we're talking

171
00:08:39,520 --> 00:08:43,840
Speaker 4: like ninety nine point nine nine percent uptime rate, and

172
00:08:43,880 --> 00:08:49,760
Speaker 4: that's guaranteed by its power redundancy, it's Internet redundancy, and

173
00:08:49,800 --> 00:08:53,480
Speaker 4: its security and then ultimately like it's connectivity to the

174
00:08:53,559 --> 00:08:56,959
Speaker 4: Internet backbone. Right, So as it's like, as a first step,

175
00:08:57,360 --> 00:09:02,520
Speaker 4: you're housed within these data centers that are just critical

176
00:09:02,600 --> 00:09:06,880
Speaker 4: parts of the Internet infrastructure, and then from there you

177
00:09:06,920 --> 00:09:09,079
Speaker 4: start building out the servers within there. And I can

178
00:09:09,240 --> 00:09:10,160
Speaker 4: go into that detail.

179
00:09:10,600 --> 00:09:13,320
Speaker 1: So you mentioned actually I want to just get sort

180
00:09:13,320 --> 00:09:15,480
Speaker 1: of defined some terms. Can you just real quickly before

181
00:09:15,520 --> 00:09:17,240
Speaker 1: we move on Tier three tier four?

182
00:09:17,240 --> 00:09:18,080
Speaker 3: What do you mean by this?

183
00:09:18,880 --> 00:09:21,880
Speaker 4: Yeah? So tier three, tier four. This all goes back

184
00:09:21,920 --> 00:09:24,360
Speaker 4: to like the quality of the data center that you're

185
00:09:24,360 --> 00:09:26,760
Speaker 4: in it. It's all about the reliability and up time

186
00:09:26,880 --> 00:09:28,640
Speaker 4: that you should be able to achieve out of that

187
00:09:28,760 --> 00:09:32,600
Speaker 4: data center. It's another way to qualify the services around it.

188
00:09:32,600 --> 00:09:35,960
Speaker 4: It's like power. You get redundant power, right like multiple

189
00:09:35,960 --> 00:09:39,880
Speaker 4: power services in case one goes offline, there's another one

190
00:09:40,080 --> 00:09:44,320
Speaker 4: you get, you know, redundant cooling, you get redundant Internet connectivity.

191
00:09:44,360 --> 00:09:47,480
Speaker 4: It's all these services that like have extra fiil safs

192
00:09:47,840 --> 00:09:50,760
Speaker 4: that allow for you to operate at the highest up

193
00:09:50,800 --> 00:09:52,400
Speaker 4: time and security level possible.

194
00:09:52,640 --> 00:09:54,760
Speaker 1: Is higher tier better? Like tier three four? Is that

195
00:09:54,800 --> 00:09:56,120
Speaker 1: better than Tier one and tier two?

196
00:09:57,040 --> 00:09:57,760
Speaker 4: That's correct?

197
00:09:57,960 --> 00:10:01,400
Speaker 1: Okay, so quick follow up question. Then you know we're

198
00:10:01,400 --> 00:10:03,280
Speaker 1: interested in, like, okay, where the rubber hits the road.

199
00:10:03,280 --> 00:10:08,040
Speaker 1: The scarcity is here. Let's say Tracy miraculously opens her

200
00:10:08,120 --> 00:10:10,400
Speaker 1: basement and there really is like you know, all these

201
00:10:10,440 --> 00:10:14,480
Speaker 1: palettes of these video chips, there is there capacity at

202
00:10:14,520 --> 00:10:16,440
Speaker 1: the data centers right now, She's like, you know, what

203
00:10:16,440 --> 00:10:19,120
Speaker 1: we want to co locate with you. You guys have great power,

204
00:10:19,679 --> 00:10:22,080
Speaker 1: pretty well connected to the internet. You have like good

205
00:10:22,120 --> 00:10:24,920
Speaker 1: security guards. So there's operated twenty four to seven. We

206
00:10:24,920 --> 00:10:27,000
Speaker 1: want to set something up, like is there space there?

207
00:10:28,240 --> 00:10:30,839
Speaker 4: Yeah, it's a fantastic question. It's a it's an issue

208
00:10:30,840 --> 00:10:33,480
Speaker 4: that didn't really pop up until really in the last

209
00:10:33,559 --> 00:10:34,560
Speaker 4: eight weeks or so.

210
00:10:34,800 --> 00:10:37,760
Speaker 3: Oh, it's really happening that fast.

211
00:10:38,400 --> 00:10:39,760
Speaker 4: It's happening that fast, Joe.

212
00:10:39,880 --> 00:10:40,880
Speaker 3: And it's okay, So.

213
00:10:40,960 --> 00:10:43,440
Speaker 2: That we said the two week lead time on in

214
00:10:43,520 --> 00:10:45,080
Speaker 2: video was very important, Joe.

215
00:10:45,360 --> 00:10:48,400
Speaker 3: Yeah, you're right, you're right. Is wow? Wait what happened?

216
00:10:48,600 --> 00:10:48,800
Speaker 4: Wait?

217
00:10:49,240 --> 00:10:52,600
Speaker 1: What happened sixteen described? Sixteen weeks ago?

218
00:10:52,720 --> 00:10:53,760
Speaker 3: Verus eight weeks ago?

219
00:10:54,520 --> 00:10:57,720
Speaker 4: Sure, it even last year? Right, So this is a space,

220
00:10:57,760 --> 00:11:02,240
Speaker 4: the data centers space, collocation space that's been fairly chronically

221
00:11:02,360 --> 00:11:05,440
Speaker 4: underinvested in because the Hyperscale has just built out their

222
00:11:05,480 --> 00:11:09,680
Speaker 4: own data centers, right instead. But what's happened is the

223
00:11:09,760 --> 00:11:13,600
Speaker 4: infrastructure changed. The type of compute that we're putting in

224
00:11:13,640 --> 00:11:17,240
Speaker 4: these data centers, it's different than the last generation, right,

225
00:11:17,280 --> 00:11:20,800
Speaker 4: so we're predominantly focused on GPU compute instead of CPU

226
00:11:20,840 --> 00:11:25,439
Speaker 4: compute and GPU compute. It's about four times more power

227
00:11:25,559 --> 00:11:29,839
Speaker 4: dance than CPU compute, and that throws the data center

228
00:11:29,960 --> 00:11:33,440
Speaker 4: planning into chaos, right because ultimately, let's say you have

229
00:11:33,480 --> 00:11:36,319
Speaker 4: a ten thousand square foot room in the data center, right,

230
00:11:36,360 --> 00:11:38,160
Speaker 4: and you have a certain amount of power it's called

231
00:11:38,160 --> 00:11:39,960
Speaker 4: one hundred units of power that go into that ten

232
00:11:40,000 --> 00:11:44,079
Speaker 4: thousand square feet. Well, because I'm four times more power DNS,

233
00:11:44,760 --> 00:11:47,320
Speaker 4: it means that now I take those hundred units of power,

234
00:11:47,400 --> 00:11:50,560
Speaker 4: but I only require about twenty five percent of that

235
00:11:50,640 --> 00:11:53,439
Speaker 4: data center footprint or in other words, twenty five hundred

236
00:11:53,440 --> 00:11:56,400
Speaker 4: square feet within that ten thousand square foot footprint. So

237
00:11:56,800 --> 00:12:00,480
Speaker 4: that then leads to like, not only is the space

238
00:12:00,760 --> 00:12:03,960
Speaker 4: in the data center being used inefficiently now because you

239
00:12:04,240 --> 00:12:06,720
Speaker 4: theoretically have to run more power into the data center

240
00:12:06,800 --> 00:12:08,800
Speaker 4: to use that full ten thousand square feet due to

241
00:12:08,800 --> 00:12:12,360
Speaker 4: the Poara density delta, but now you have cooling issues,

242
00:12:12,960 --> 00:12:15,720
Speaker 4: right because you designed that footprint to be able to

243
00:12:15,720 --> 00:12:19,920
Speaker 4: cool ten thousand square feet spread out across that entire area.

244
00:12:20,000 --> 00:12:21,040
Speaker 4: But now you're dropping storry.

245
00:12:21,559 --> 00:12:23,640
Speaker 1: Sorry, I just want to back up because this is

246
00:12:23,880 --> 00:12:26,040
Speaker 1: extremely interesting, so I don't want to I just want

247
00:12:26,080 --> 00:12:27,079
Speaker 1: to get this detail right.

248
00:12:27,640 --> 00:12:30,840
Speaker 3: Just sorry, just to and then move on.

249
00:12:30,920 --> 00:12:34,439
Speaker 1: But the let's given an x amount of power at

250
00:12:34,480 --> 00:12:37,720
Speaker 1: one hundred units of power. What you're saying is that

251
00:12:37,800 --> 00:12:41,880
Speaker 1: with this next generation of compute, it now only gets

252
00:12:42,120 --> 00:12:42,720
Speaker 1: that's now.

253
00:12:42,600 --> 00:12:44,720
Speaker 3: Only sufficient for a quarter of the data center.

254
00:12:44,760 --> 00:12:48,040
Speaker 1: In other words, that to power that whole that space,

255
00:12:48,520 --> 00:12:50,800
Speaker 1: and that to then power the whole space, you really

256
00:12:50,840 --> 00:12:52,360
Speaker 1: would need like four x the power.

257
00:12:53,320 --> 00:12:57,160
Speaker 4: That's accurate. Okay. The complication really arises out of the

258
00:12:57,160 --> 00:13:00,400
Speaker 4: cooling that that's required from that, right, So if you

259
00:13:00,400 --> 00:13:03,160
Speaker 4: imagine you can cool a ten thousand square foot space

260
00:13:03,160 --> 00:13:05,400
Speaker 4: and you designed for that, that's one thing. But now

261
00:13:05,440 --> 00:13:08,080
Speaker 4: if you have to cool in a much more dense area,

262
00:13:08,559 --> 00:13:12,400
Speaker 4: that's a different type of cooling requirement. And so that's

263
00:13:12,480 --> 00:13:15,719
Speaker 4: led to this issue where there's only a certain subset

264
00:13:15,960 --> 00:13:18,640
Speaker 4: of Tier three and four data centers across the US

265
00:13:19,120 --> 00:13:23,440
Speaker 4: that can are currently designed for or can quickly be

266
00:13:23,559 --> 00:13:27,880
Speaker 4: designed and changed to be able to accommodate this new

267
00:13:27,960 --> 00:13:31,760
Speaker 4: power density issue. So now not only like if you

268
00:13:31,840 --> 00:13:34,440
Speaker 4: had all those eighth one hundreds in your basement, you

269
00:13:34,520 --> 00:13:36,480
Speaker 4: might not have a place to plug them into. And

270
00:13:37,320 --> 00:13:40,160
Speaker 4: that's become a pretty big problem for the industry very quickly,

271
00:13:40,160 --> 00:13:42,520
Speaker 4: and truly has only arisen in the last eight weeks

272
00:13:42,640 --> 00:13:45,120
Speaker 4: or so, and it's going to persist for a few quarters.

273
00:13:46,040 --> 00:13:50,040
Speaker 2: So you were describing the difference between CPU and GPU.

274
00:13:50,600 --> 00:13:54,600
Speaker 2: How do you actually connect these newer types or these

275
00:13:54,640 --> 00:13:58,600
Speaker 2: different types of chips together, because I imagine, you know,

276
00:13:58,760 --> 00:14:01,160
Speaker 2: old data centers you just have a bunch of like

277
00:14:01,240 --> 00:14:04,520
Speaker 2: Ethernet cables or something like that. But for this type

278
00:14:04,559 --> 00:14:06,640
Speaker 2: of processing power, do you need something different?

279
00:14:08,000 --> 00:14:12,160
Speaker 4: That's exactly correct, Chracy. So what we so the legacy

280
00:14:12,280 --> 00:14:15,240
Speaker 4: the generalized compute data centers are really what the hyperscalers

281
00:14:15,280 --> 00:14:19,560
Speaker 4: look like. You know, Amazon, Google, Microsoft, Oracle. They predominantly

282
00:14:19,640 --> 00:14:23,240
Speaker 4: use something that's called Ethernet to connect all the service together.

283
00:14:23,280 --> 00:14:25,600
Speaker 4: And the reason you use that was, you know, you

284
00:14:25,600 --> 00:14:29,120
Speaker 4: don't really need to have high data throughput to connect

285
00:14:29,240 --> 00:14:31,040
Speaker 4: all these servers together, right, They just need to be

286
00:14:31,080 --> 00:14:33,040
Speaker 4: able to send some messages back and forth. They talk

287
00:14:33,080 --> 00:14:35,960
Speaker 4: to each other about what they're working on, but they're not,

288
00:14:36,480 --> 00:14:41,000
Speaker 4: you know, necessarily doing highly collaborative tasks that require moving

289
00:14:41,080 --> 00:14:44,720
Speaker 4: lots of data in between each other. That's changed. So

290
00:14:45,080 --> 00:14:48,160
Speaker 4: so today what people are focused on and need to

291
00:14:48,200 --> 00:14:52,400
Speaker 4: build are these effectively supercomputers. Right, and so we refer

292
00:14:52,520 --> 00:14:56,000
Speaker 4: to the connectivity between them, the network between them as

293
00:14:56,120 --> 00:14:59,560
Speaker 4: a fabric, right, it's called a network fabric. So if

294
00:14:59,600 --> 00:15:02,480
Speaker 4: we're build holding something to help train like the next

295
00:15:02,520 --> 00:15:07,600
Speaker 4: generation GPT model, typically clients are coming to us saying, hey,

296
00:15:07,640 --> 00:15:11,400
Speaker 4: I need a sixteen thousand GPU fabric of H one hundred.

297
00:15:11,960 --> 00:15:15,800
Speaker 4: So that's there's about eight GPUs that go into each server,

298
00:15:16,120 --> 00:15:18,960
Speaker 4: and then you have to run this connectivity between each

299
00:15:19,080 --> 00:15:21,240
Speaker 4: one of those servers. But it's now done in a

300
00:15:21,280 --> 00:15:25,480
Speaker 4: different way to your point, So we're using a in

301
00:15:25,520 --> 00:15:30,560
Speaker 4: Nvidio technology called InfiniBand which has the highest data throughput

302
00:15:30,680 --> 00:15:34,000
Speaker 4: to connect each of these devices together. And you know,

303
00:15:34,080 --> 00:15:38,560
Speaker 4: taking this sixteen thousand GPU cluster as an example, there's

304
00:15:38,600 --> 00:15:41,680
Speaker 4: two crazy numbers in here. One is that there are

305
00:15:41,960 --> 00:15:47,200
Speaker 4: forty eight thousand discrete connections that need to be made,

306
00:15:47,400 --> 00:15:50,280
Speaker 4: right like plugging one thing in from one computer to

307
00:15:50,320 --> 00:15:54,680
Speaker 4: another computer. But there's lots of switches and routers that

308
00:15:54,720 --> 00:15:57,200
Speaker 4: are between there. But you need to that forty eight

309
00:15:57,240 --> 00:16:03,120
Speaker 4: thousand times, and it takes over five hundred miles of

310
00:16:03,200 --> 00:16:07,200
Speaker 4: fiber optic cabling to do that successfully across the sixteen

311
00:16:07,240 --> 00:16:09,840
Speaker 4: thousand GPU cluster. And now again you're doing that within

312
00:16:09,880 --> 00:16:11,800
Speaker 4: a small space with a ton of power density, with

313
00:16:11,840 --> 00:16:14,720
Speaker 4: a ton of cooling, and it's just a completely different

314
00:16:14,760 --> 00:16:17,960
Speaker 4: way to build this infrastructure. It's just because the requirements

315
00:16:17,960 --> 00:16:20,560
Speaker 4: have changed, right, Like we've moved into this, like this

316
00:16:20,720 --> 00:16:25,040
Speaker 4: area where we are designing next generation AI models and

317
00:16:25,120 --> 00:16:27,840
Speaker 4: it requires a completely different type of compute, and it's

318
00:16:27,920 --> 00:16:31,320
Speaker 4: just it's caught the whole sector by surprise so much

319
00:16:31,360 --> 00:16:34,880
Speaker 4: so that you know, it's really challenging to go procure

320
00:16:34,880 --> 00:16:38,240
Speaker 4: it at the hyperscalers today because they didn't specialize in

321
00:16:38,280 --> 00:16:40,560
Speaker 4: building it. And that's you know where where core we've

322
00:16:40,560 --> 00:16:43,440
Speaker 4: comes in is we only focus on building this type

323
00:16:43,480 --> 00:16:46,240
Speaker 4: of compute for clients. It's our specialty. We hire all

324
00:16:46,240 --> 00:16:48,400
Speaker 4: of our engineering around it, all of our research goes

325
00:16:48,440 --> 00:16:51,040
Speaker 4: into it, and it's you know, it's been a fantastic

326
00:16:51,040 --> 00:16:53,640
Speaker 4: spot to be but our goal at the end of

327
00:16:53,640 --> 00:16:54,760
Speaker 4: the day is just to be able to get this

328
00:16:54,840 --> 00:16:57,480
Speaker 4: infrastructure into the hands of end consumers so that they

329
00:16:57,480 --> 00:17:00,640
Speaker 4: can build the amazing AI companies that have ones looking

330
00:17:00,680 --> 00:17:16,439
Speaker 4: forward to using and incorporating to enterprises and software companies.

331
00:17:21,520 --> 00:17:25,560
Speaker 2: You know, you mentioned these special or purpose built connections

332
00:17:25,680 --> 00:17:28,480
Speaker 2: that Nvidia is making, and this kind of leads nicely

333
00:17:28,640 --> 00:17:31,760
Speaker 2: into my next question, which is what exactly is your

334
00:17:31,880 --> 00:17:37,879
Speaker 2: relationship with Nvidia and in order to provide this type

335
00:17:37,960 --> 00:17:42,280
Speaker 2: of service, you know, vast amounts of processing power that

336
00:17:42,440 --> 00:17:46,080
Speaker 2: is well suited to a particular type of technology in

337
00:17:46,119 --> 00:17:49,160
Speaker 2: this case AI, do you have to have a really

338
00:17:49,200 --> 00:17:51,919
Speaker 2: good relationship with Nvidia to make that work? Like do

339
00:17:52,000 --> 00:17:55,120
Speaker 2: you have to have special access to H one, hundreds

340
00:17:55,160 --> 00:17:56,840
Speaker 2: and other chips.

341
00:17:57,840 --> 00:18:00,119
Speaker 4: It's a great question, and I'll try to offer or

342
00:18:00,680 --> 00:18:03,159
Speaker 4: from Nvidia's perspective, and it goes a little bit back

343
00:18:03,200 --> 00:18:05,160
Speaker 4: to the answer I just provided as well in that

344
00:18:06,200 --> 00:18:09,560
Speaker 4: I would think from in Nvidia's seat, what's most important

345
00:18:09,720 --> 00:18:13,640
Speaker 4: is empowering end users of their compute to be able

346
00:18:13,640 --> 00:18:18,200
Speaker 4: to access their compute and the most performant variant possible

347
00:18:19,160 --> 00:18:21,440
Speaker 4: at scale, and to be able to access it quickly, right,

348
00:18:21,480 --> 00:18:23,000
Speaker 4: Like a new generation comes out, they want to be

349
00:18:23,000 --> 00:18:25,240
Speaker 4: able to get their hands on it, right. And we've

350
00:18:25,320 --> 00:18:29,200
Speaker 4: built Core. We've around hitting every single one of those checkboxes. Right.

351
00:18:29,200 --> 00:18:31,440
Speaker 4: We build it at DGX reference back, we build it

352
00:18:31,560 --> 00:18:34,160
Speaker 4: at scale, and we bring it online on a timeline

353
00:18:34,200 --> 00:18:37,720
Speaker 4: that's you know, within months of a next generation chipset launch,

354
00:18:37,840 --> 00:18:41,680
Speaker 4: as opposed to you know, the more traditional legacy hyperscalers

355
00:18:41,680 --> 00:18:45,679
Speaker 4: that take quarters at a times, so US being in

356
00:18:45,720 --> 00:18:49,399
Speaker 4: a position to do that has has enabled us fantastic

357
00:18:49,640 --> 00:18:53,600
Speaker 4: access within Nvidia, and we have a history of consistently

358
00:18:53,640 --> 00:18:56,960
Speaker 4: executing on exactly what we've what we say we'll do right,

359
00:18:57,000 --> 00:19:02,120
Speaker 4: we under promise and over deliver as a business, and

360
00:19:02,520 --> 00:19:04,480
Speaker 4: I think that's just put us in this place where

361
00:19:04,800 --> 00:19:08,720
Speaker 4: Nvidia has the confidence in allocating infrastructure to us because

362
00:19:08,760 --> 00:19:10,680
Speaker 4: they know it's going to come online, they know it's

363
00:19:10,680 --> 00:19:14,240
Speaker 4: going to get to consumers faster than anyone else in

364
00:19:14,280 --> 00:19:15,760
Speaker 4: the market, and they know it's going to be delivered

365
00:19:15,760 --> 00:19:18,520
Speaker 4: in its most performance configuration that exists.

366
00:19:19,760 --> 00:19:22,359
Speaker 1: You know, I was thinking as I listened to some

367
00:19:22,400 --> 00:19:25,240
Speaker 1: of these answers, I keep having like these like imagines,

368
00:19:25,280 --> 00:19:29,520
Speaker 1: like you know, there's probably like some random industrial company

369
00:19:29,600 --> 00:19:32,159
Speaker 1: that's like traded like you know on the like S

370
00:19:32,200 --> 00:19:36,320
Speaker 1: and P four hundred that makes some cooling fluid whose

371
00:19:36,359 --> 00:19:38,080
Speaker 1: like sales are going to be up ten x. So

372
00:19:38,119 --> 00:19:40,040
Speaker 1: I'm like googling while we're talking, like what is a

373
00:19:40,040 --> 00:19:42,000
Speaker 1: company that makes kool aid fluid? Or like who is

374
00:19:42,040 --> 00:19:44,000
Speaker 1: some company that's like really good at making these like

375
00:19:44,040 --> 00:19:48,360
Speaker 1: infinite bands, Because it just likes.

376
00:19:46,840 --> 00:19:48,479
Speaker 3: Right, yeah, like what are the anyway?

377
00:19:48,560 --> 00:19:48,760
Speaker 4: Right?

378
00:19:48,920 --> 00:19:50,840
Speaker 1: But like right, like you know there's going to be

379
00:19:50,880 --> 00:19:54,760
Speaker 1: some yeah urchiery plate that are like thirty x up.

380
00:19:54,960 --> 00:19:56,760
Speaker 1: But you know, I want to get a sense from

381
00:19:56,800 --> 00:20:00,879
Speaker 1: you of so it's really changed a lot, and I

382
00:20:01,000 --> 00:20:02,600
Speaker 1: kind of you know, in the last several months.

383
00:20:02,600 --> 00:20:04,080
Speaker 3: Could we see it from in video results?

384
00:20:04,119 --> 00:20:08,040
Speaker 1: What you're describing, like how big is the market getting

385
00:20:08,119 --> 00:20:09,520
Speaker 1: and the way I think you know, I know, like

386
00:20:09,520 --> 00:20:13,040
Speaker 1: with AI, there's training and they sort of build the

387
00:20:13,080 --> 00:20:15,199
Speaker 1: model and then there's inference, and the inference is how

388
00:20:15,240 --> 00:20:17,600
Speaker 1: they spit out the results. Can you talk a little

389
00:20:17,600 --> 00:20:20,920
Speaker 1: bit about what you're seeing in terms of the growth

390
00:20:21,400 --> 00:20:24,760
Speaker 1: of both of those aspects of AI, which is bigger

391
00:20:25,000 --> 00:20:27,240
Speaker 1: and which is growing faster? And how do they compare

392
00:20:27,280 --> 00:20:29,720
Speaker 1: to like the size of the installed compute base that

393
00:20:29,760 --> 00:20:30,440
Speaker 1: already exists.

394
00:20:31,240 --> 00:20:34,760
Speaker 4: Oh. Absolutely, So this is one of my favorite topics

395
00:20:34,800 --> 00:20:37,280
Speaker 4: because it's just mind blowing the scale that's going to

396
00:20:37,280 --> 00:20:41,719
Speaker 4: be needed to support AI and scale this infrastructure. So okay,

397
00:20:41,800 --> 00:20:45,600
Speaker 4: so today most of the funding that's going into the

398
00:20:45,600 --> 00:20:49,800
Speaker 4: AI space is too for funding to train next generation

399
00:20:50,600 --> 00:20:53,400
Speaker 4: foundation models. Right, So when a company's raising a bunch

400
00:20:53,440 --> 00:20:55,119
Speaker 4: of money at the end of the day, most of

401
00:20:55,119 --> 00:20:57,680
Speaker 4: that money is going into cloud compute to go train

402
00:20:57,760 --> 00:21:01,200
Speaker 4: this next generation found model to build that intellectual property.

403
00:21:01,280 --> 00:21:03,360
Speaker 4: So they have this model and they can go bring

404
00:21:03,359 --> 00:21:06,800
Speaker 4: it into the inference market. And what I would say is,

405
00:21:07,280 --> 00:21:12,160
Speaker 4: we're having a supply demand issue like a chip access

406
00:21:12,200 --> 00:21:17,280
Speaker 4: crunch in the training phase, where in reality, the scale

407
00:21:17,440 --> 00:21:20,800
Speaker 4: of the inference market is where we're all the demand

408
00:21:21,000 --> 00:21:24,560
Speaker 4: truly is going to sit. So what I'd offer to

409
00:21:24,600 --> 00:21:28,720
Speaker 4: help contextualize that is, let's take you know, there's some

410
00:21:28,760 --> 00:21:30,880
Speaker 4: well known models in the market today. Let's let's say

411
00:21:30,920 --> 00:21:35,080
Speaker 4: there's a preach an in market trained model and it

412
00:21:35,160 --> 00:21:37,679
Speaker 4: took about let's say ten thousand. A one hundred or

413
00:21:37,680 --> 00:21:39,479
Speaker 4: so to train A one hundred is the last generation

414
00:21:39,560 --> 00:21:41,400
Speaker 4: in chip, but you know it still applies in terms

415
00:21:41,440 --> 00:21:45,520
Speaker 4: of relative scale here. So that company that used ten

416
00:21:46,520 --> 00:21:50,840
Speaker 4: train their model, our understanding is they're going to need

417
00:21:50,880 --> 00:21:55,760
Speaker 4: about a million GPUs within one to two years of

418
00:21:55,840 --> 00:21:59,760
Speaker 4: launch to support the entire inference demand.

419
00:22:00,720 --> 00:22:04,440
Speaker 1: So ten you could train the model on ten thousand

420
00:22:04,440 --> 00:22:06,920
Speaker 1: of these chips, ten thousand of these sisters, whatever they're

421
00:22:07,119 --> 00:22:09,080
Speaker 1: and then if they're actually going to be in the

422
00:22:09,119 --> 00:22:13,160
Speaker 1: market and sell something or provide some service to make

423
00:22:13,160 --> 00:22:13,720
Speaker 1: it worthwhile.

424
00:22:13,720 --> 00:22:15,040
Speaker 3: They're going to need a million.

425
00:22:15,680 --> 00:22:17,960
Speaker 4: A million, and I think that's just within first two

426
00:22:18,000 --> 00:22:21,440
Speaker 4: years of launch. Show like we're we're talking about something

427
00:22:21,440 --> 00:22:25,720
Speaker 4: that's going to continue growing afterwards. And so what does

428
00:22:25,720 --> 00:22:28,439
Speaker 4: a million GPUs mean? Obviously? Right, so you know a

429
00:22:28,440 --> 00:22:31,520
Speaker 4: couple I think it was like into last year, all

430
00:22:31,560 --> 00:22:35,600
Speaker 4: the hyperscalers combined, right, Amazon, Google, Microsoft, Oracle. You can

431
00:22:35,600 --> 00:22:37,479
Speaker 4: throw a COORWY from there, it was about, you know,

432
00:22:37,640 --> 00:22:42,400
Speaker 4: five hundred thousand GPUs globally right available across those platforms.

433
00:22:42,720 --> 00:22:44,760
Speaker 4: I'd see it end of this year, it'll be closer

434
00:22:44,800 --> 00:22:48,720
Speaker 4: to a million or so. But that's suggesting then that

435
00:22:48,920 --> 00:22:53,440
Speaker 4: one AI company with one model could consume the entire

436
00:22:53,520 --> 00:22:58,360
Speaker 4: global footprint of GPUs. And and now you start to think, wait,

437
00:22:58,440 --> 00:23:01,040
Speaker 4: aren't there a bunch of other companies training these models

438
00:23:01,040 --> 00:23:03,760
Speaker 4: in market right now? And I leo'ld say, yes, there are.

439
00:23:03,920 --> 00:23:08,240
Speaker 4: So it can imply that there are in the short term,

440
00:23:08,280 --> 00:23:12,480
Speaker 4: the demand of several million GPUs just to support the

441
00:23:12,680 --> 00:23:17,000
Speaker 4: inference market. And there's just there's just nowhere near enough

442
00:23:17,320 --> 00:23:20,639
Speaker 4: globally of this infrastructure, and it's it's going to be

443
00:23:20,680 --> 00:23:24,119
Speaker 4: a big challenge for the market as we exit this

444
00:23:24,200 --> 00:23:26,800
Speaker 4: training phase and move into the productization or really just

445
00:23:26,840 --> 00:23:29,080
Speaker 4: the commercialization of these models, like how do you generate

446
00:23:29,119 --> 00:23:32,640
Speaker 4: revenue off them? And it's it's something that I don't

447
00:23:32,680 --> 00:23:36,440
Speaker 4: think many people truly understand just the amount of scale

448
00:23:36,520 --> 00:23:39,800
Speaker 4: and construction that needs to take place. And now you

449
00:23:39,840 --> 00:23:42,119
Speaker 4: put that in the same framework of the data centers

450
00:23:42,119 --> 00:23:43,920
Speaker 4: that we were talking about, right, So there's this lack

451
00:23:43,960 --> 00:23:46,280
Speaker 4: of data center space, there's lack of chipset supply, like

452
00:23:46,359 --> 00:23:49,920
Speaker 4: it's it's going to be an issue for years that

453
00:23:49,960 --> 00:23:50,560
Speaker 4: we see.

454
00:23:50,840 --> 00:23:53,800
Speaker 2: So when it comes to scale, you know, you keep

455
00:23:53,840 --> 00:23:57,679
Speaker 2: mentioning the hyper scalers, which is a great term, but

456
00:23:57,960 --> 00:24:02,560
Speaker 2: people like Amazon, Google, I guess, Microsoft, IBM, et cetera.

457
00:24:03,359 --> 00:24:07,560
Speaker 2: How quickly or what is your impression of how quickly

458
00:24:07,880 --> 00:24:11,159
Speaker 2: they are able to ramp up in this space? Like

459
00:24:11,320 --> 00:24:15,200
Speaker 2: how fast could they react to some of the trends

460
00:24:15,240 --> 00:24:16,280
Speaker 2: that you've been outlining.

461
00:24:17,520 --> 00:24:20,600
Speaker 4: Yeah, so I can offer what I'm seeing today. You know,

462
00:24:20,760 --> 00:24:25,159
Speaker 4: the h one hundreds started to be distributed globally to

463
00:24:26,359 --> 00:24:28,720
Speaker 4: all of us, right, like all the entities that have

464
00:24:28,800 --> 00:24:31,439
Speaker 4: these you know, kind of upper tier relationships with Nvidia

465
00:24:31,720 --> 00:24:35,080
Speaker 4: back in March, right, so we started getting them the

466
00:24:35,080 --> 00:24:37,959
Speaker 4: s infrastructure online in April, really scaling in May, and

467
00:24:38,040 --> 00:24:40,160
Speaker 4: you know, we have builds going on at ten data

468
00:24:40,160 --> 00:24:42,719
Speaker 4: centers across the US right now, and we're delivering it

469
00:24:42,720 --> 00:24:46,680
Speaker 4: to clients. The guidance that we're seeing from the hyperscalers

470
00:24:47,240 --> 00:24:50,760
Speaker 4: is that they're not going to begin delivering scale access

471
00:24:51,080 --> 00:24:54,720
Speaker 4: to the H one hundred chipset until late Q three,

472
00:24:55,880 --> 00:24:57,920
Speaker 4: maybe mid Q four, and some of them are even

473
00:24:57,960 --> 00:25:01,159
Speaker 4: beginning to guide into Q one. And it's all driven

474
00:25:01,400 --> 00:25:04,040
Speaker 4: by the fact that this is just a different type

475
00:25:04,040 --> 00:25:07,919
Speaker 4: of compute that they're building relative to last generation. Right

476
00:25:07,920 --> 00:25:10,720
Speaker 4: You're no longer just running Ethernet to your point, between

477
00:25:10,760 --> 00:25:13,360
Speaker 4: all these devices, you're not just plugging in CPU blades.

478
00:25:13,440 --> 00:25:15,720
Speaker 4: You're having to deal with like totally different data center

479
00:25:15,760 --> 00:25:20,240
Speaker 4: power density and cooling requirements. You're having to build supercomputers instead,

480
00:25:20,359 --> 00:25:24,119
Speaker 4: with five hundred miles of fiber and all these connections.

481
00:25:24,119 --> 00:25:26,480
Speaker 4: It's just it's a completely different way to build the cloud,

482
00:25:26,520 --> 00:25:29,760
Speaker 4: and it's taking them some time to catch up because

483
00:25:29,760 --> 00:25:32,159
Speaker 4: you have to retrain entire organizations to do this. So

484
00:25:32,960 --> 00:25:35,399
Speaker 4: you know, as of now, i'd say the direct answer

485
00:25:35,480 --> 00:25:38,920
Speaker 4: is three quarters after a chip set launch, But it's

486
00:25:38,920 --> 00:25:41,760
Speaker 4: seeming it might take longer, And I think that's all

487
00:25:41,760 --> 00:25:45,720
Speaker 4: going to contribute to this just kind of slower ability

488
00:25:45,800 --> 00:25:50,760
Speaker 4: to scale infrastructure than what's being dictated by the adoption

489
00:25:50,920 --> 00:25:53,480
Speaker 4: rate of AI software, And it's going to lead to

490
00:25:53,480 --> 00:25:56,600
Speaker 4: this supply demand imbalance that will just last for a while.

491
00:25:57,640 --> 00:26:00,479
Speaker 2: You know, you keep mentioning or we both keep mentioning

492
00:26:00,520 --> 00:26:03,800
Speaker 2: the H one hundred for obvious reasons, But do you

493
00:26:03,840 --> 00:26:08,199
Speaker 2: look at other chips or what would happen to you know,

494
00:26:08,280 --> 00:26:11,439
Speaker 2: your own business if, for instance, a new chip was

495
00:26:11,480 --> 00:26:14,920
Speaker 2: developed that could do the same thing or better than

496
00:26:15,080 --> 00:26:17,640
Speaker 2: an Nvidia H one hundred. Like, for instance, I hear

497
00:26:17,640 --> 00:26:19,919
Speaker 2: a lot of excitement about some of the stuff that

498
00:26:19,960 --> 00:26:25,040
Speaker 2: AMD is developing. And I'm not a chips expert, except

499
00:26:25,119 --> 00:26:29,159
Speaker 2: maybe when it comes to freedo's or layers. But like,

500
00:26:30,080 --> 00:26:33,840
Speaker 2: how how big a difference would that make to you

501
00:26:34,200 --> 00:26:39,199
Speaker 2: if we suddenly got a different chip manufacturer gain prominence

502
00:26:39,320 --> 00:26:39,720
Speaker 2: in AI.

503
00:26:40,800 --> 00:26:45,520
Speaker 4: Sure? So I'd offer kind of two broad responses. One, typically,

504
00:26:45,720 --> 00:26:50,080
Speaker 4: when you train a model you're going to use, you're

505
00:26:50,119 --> 00:26:52,719
Speaker 4: going to use the same chips for inference on that

506
00:26:52,840 --> 00:26:55,520
Speaker 4: model as well. Right, So two PT four, For example,

507
00:26:55,600 --> 00:26:58,399
Speaker 4: I was trained on a one hundreds, they're predominantly going

508
00:26:58,480 --> 00:27:00,440
Speaker 4: to use a one hundreds going for or you might

509
00:27:00,440 --> 00:27:03,400
Speaker 4: fit in some kind of newer generation hyper efficient chips

510
00:27:03,400 --> 00:27:06,199
Speaker 4: into there, but it's not like you need a quote

511
00:27:07,000 --> 00:27:10,200
Speaker 4: a GP with more vram on it, right, Like you're

512
00:27:10,240 --> 00:27:12,439
Speaker 4: in a need your forty gig or your your eighty

513
00:27:12,480 --> 00:27:15,879
Speaker 4: grid gig ram chip because that's the size of the

514
00:27:15,880 --> 00:27:18,200
Speaker 4: model that that you trained, Right, You're not gonna need

515
00:27:18,280 --> 00:27:20,680
Speaker 4: like next multiple generations. You're not going to like really

516
00:27:20,680 --> 00:27:23,479
Speaker 4: be able to adopt them to change the efficiency of

517
00:27:23,760 --> 00:27:26,480
Speaker 4: serving that model. So what we view is that a

518
00:27:26,560 --> 00:27:30,560
Speaker 4: chip's lifespan is like this first two to three years

519
00:27:30,800 --> 00:27:34,240
Speaker 4: is spent training models, and then it's next four to

520
00:27:34,359 --> 00:27:38,480
Speaker 4: five years is spent doing inference for those models that

521
00:27:38,560 --> 00:27:42,159
Speaker 4: it trained. And then within there as well, you do

522
00:27:42,200 --> 00:27:44,879
Speaker 4: this thing called fine tuning, which is updating the model

523
00:27:44,960 --> 00:27:47,199
Speaker 4: with new information, right, Like how do you keep a

524
00:27:47,200 --> 00:27:49,560
Speaker 4: model like up to date with what's happened on a

525
00:27:49,600 --> 00:27:53,399
Speaker 4: Twitter or what's happened on in the media. Right, you

526
00:27:53,480 --> 00:27:55,680
Speaker 4: have to keep retraining it, right, and you'll use those

527
00:27:55,680 --> 00:27:58,080
Speaker 4: same chips to do that. But so your question on

528
00:27:58,240 --> 00:28:01,000
Speaker 4: other chip sets, and this is something that we have

529
00:28:01,040 --> 00:28:05,000
Speaker 4: a particularly interesting view into because we have like you know,

530
00:28:05,080 --> 00:28:07,720
Speaker 4: call it six hundred and fifty AI clients right, and

531
00:28:07,760 --> 00:28:10,960
Speaker 4: we're having conversations with them daily to ensure that we're

532
00:28:11,000 --> 00:28:13,960
Speaker 4: meeting their scaling demands. So it gives us a look

533
00:28:13,960 --> 00:28:17,480
Speaker 4: into six to twelve months into the future what type

534
00:28:17,480 --> 00:28:21,480
Speaker 4: of infrastructure they expect to need, and it's it's overwhelmingly

535
00:28:21,840 --> 00:28:26,479
Speaker 4: people still want access to Nvidia chips and the reason

536
00:28:26,560 --> 00:28:29,320
Speaker 4: for this is something that dates back, I think it's

537
00:28:29,359 --> 00:28:34,000
Speaker 4: nearly fifteen years when Nvidia and Jensen made the decision

538
00:28:34,040 --> 00:28:38,400
Speaker 4: to open source Kuda and to make this software set

539
00:28:38,600 --> 00:28:42,840
Speaker 4: accessible to the machine learning community. And you know today

540
00:28:42,840 --> 00:28:44,640
Speaker 4: if you go to GitHub and you search a machine

541
00:28:44,680 --> 00:28:49,200
Speaker 4: learning project, they's all reference Kuda drivers. And he's established

542
00:28:49,200 --> 00:28:55,320
Speaker 4: this utter dominance of ecosystem around his compute within the

543
00:28:55,440 --> 00:28:58,640
Speaker 4: mL space really similar to like the x eighty six

544
00:28:58,840 --> 00:29:02,360
Speaker 4: instruction set for CP you versus ARM, right, like X

545
00:29:02,400 --> 00:29:04,880
Speaker 4: eighty six is is used prodominantly. ARM has been trying

546
00:29:04,880 --> 00:29:07,400
Speaker 4: to find its way into the space for a while

547
00:29:07,480 --> 00:29:10,360
Speaker 4: now and it's just really struggled because all the engineers

548
00:29:10,360 --> 00:29:13,680
Speaker 4: and developers are used to x eighty six. Similar to

549
00:29:13,800 --> 00:29:16,520
Speaker 4: how all the engineers and developers in the AI space

550
00:29:16,600 --> 00:29:20,280
Speaker 4: are used to using KUDA, so it's something that like

551
00:29:20,360 --> 00:29:25,320
Speaker 4: obviously AMD is highly incentivized to find a way into

552
00:29:25,360 --> 00:29:28,480
Speaker 4: the sector, but they just don't have the ecosystem and

553
00:29:28,520 --> 00:29:30,760
Speaker 4: it's a huge moat to deal with. Then, you know,

554
00:29:30,880 --> 00:29:35,280
Speaker 4: kudos to Nvidia for establishing themselves and having the patients

555
00:29:35,320 --> 00:29:37,720
Speaker 4: to stick with it and to continue to support that

556
00:29:37,760 --> 00:29:40,120
Speaker 4: community over the last fifteen years, and it's it's really

557
00:29:40,120 --> 00:29:43,040
Speaker 4: paying off for them in spades today. You know, if

558
00:29:43,080 --> 00:29:46,560
Speaker 4: the demand comes for that infrastructure at some point, it's

559
00:29:46,680 --> 00:29:49,960
Speaker 4: you know, we can run other pieces of infrastructure within

560
00:29:50,000 --> 00:29:54,000
Speaker 4: our data center. But I also find that Nvidia has

561
00:29:54,080 --> 00:29:57,440
Speaker 4: such an advantage on the competition with not only its GPUs,

562
00:29:57,480 --> 00:30:00,600
Speaker 4: but all its components that support the gp USE, like

563
00:30:00,760 --> 00:30:04,400
Speaker 4: the infinite band fabric, that it's it's gonna be a

564
00:30:04,440 --> 00:30:08,320
Speaker 4: really difficult company to displace from the market in terms

565
00:30:08,360 --> 00:30:12,280
Speaker 4: of the best standard for AI infrastructure.

566
00:30:12,840 --> 00:30:14,680
Speaker 1: Can I ask a question, and I'm gonna I want

567
00:30:14,680 --> 00:30:17,000
Speaker 1: to ask this politely because it's not intended to be

568
00:30:17,120 --> 00:30:20,560
Speaker 1: accusatory or anything like that, So I don't want you

569
00:30:20,560 --> 00:30:22,840
Speaker 1: to you know here it is like but like when

570
00:30:22,840 --> 00:30:29,200
Speaker 1: you're like talking about like hyperscalers and you're like you know, Amazon, Google, Microsoft,

571
00:30:29,200 --> 00:30:31,000
Speaker 1: and you know kind of core Weave, and it's like, okay,

572
00:30:31,000 --> 00:30:33,640
Speaker 1: those are trillion dollar companies and you're a two billion

573
00:30:33,680 --> 00:30:36,760
Speaker 1: dollar company. Like why like I don't still don't think

574
00:30:36,800 --> 00:30:38,560
Speaker 1: I like wrap my head around, like and I know,

575
00:30:38,640 --> 00:30:41,000
Speaker 1: like they're all like in they're all talking about.

576
00:30:40,760 --> 00:30:41,520
Speaker 3: AI et cetera.

577
00:30:41,600 --> 00:30:43,360
Speaker 1: Like can you still just like explain to me a

578
00:30:43,400 --> 00:30:46,080
Speaker 1: little bit, like why aren't they just gonna frankly like

579
00:30:46,200 --> 00:30:49,080
Speaker 1: steamroll you or be able to let's put it this way,

580
00:30:49,360 --> 00:30:52,160
Speaker 1: be able to Okay, maybe it'll take a few quarters

581
00:30:52,200 --> 00:30:55,840
Speaker 1: to re evaluate things, but like you know, eventually this

582
00:30:56,040 --> 00:30:59,360
Speaker 1: just becomes this sort of de facto offering from these

583
00:30:59,400 --> 00:31:02,960
Speaker 1: big is that have these huge cloud budgets that must

584
00:31:02,960 --> 00:31:05,680
Speaker 1: be orders of magnitude large larger than yours.

585
00:31:06,320 --> 00:31:08,120
Speaker 4: Yeah. Yeah, I would really love to be able to

586
00:31:08,160 --> 00:31:11,040
Speaker 4: have access to their their cost of capital, that's pretty sure.

587
00:31:12,120 --> 00:31:15,480
Speaker 4: So the way, look, it's the way I had to talk

588
00:31:15,480 --> 00:31:18,960
Speaker 4: about this is we don't have a silver bullet necessarily, right.

589
00:31:18,960 --> 00:31:21,600
Speaker 4: I can't point to like a super secret piece of

590
00:31:21,640 --> 00:31:24,400
Speaker 4: technology that we put inside of our servers or anything

591
00:31:24,400 --> 00:31:27,680
Speaker 4: along those lines. But the way I like to broadly contextualize,

592
00:31:27,680 --> 00:31:33,680
Speaker 4: it is referencing another sector, and it's that like Ford

593
00:31:34,080 --> 00:31:37,000
Speaker 4: should be able to produce a model, why right, Like

594
00:31:37,240 --> 00:31:39,280
Speaker 4: they have the budget, they have the people, they have

595
00:31:39,400 --> 00:31:43,920
Speaker 4: the decades of expertise. But in order to ask them

596
00:31:43,920 --> 00:31:46,320
Speaker 4: to produce a model, why you would have to ask

597
00:31:46,360 --> 00:31:50,560
Speaker 4: them to foundationally change the way that they produce a vehicle,

598
00:31:50,880 --> 00:31:55,280
Speaker 4: all the way from research to servicing and that entire mechanism.

599
00:31:55,680 --> 00:31:58,120
Speaker 4: Like it's a giant organization. Now you have to go

600
00:31:58,200 --> 00:32:01,240
Speaker 4: ask that huge organization and people to change the way

601
00:32:01,240 --> 00:32:03,960
Speaker 4: that they go about producing things.

602
00:32:04,000 --> 00:32:05,560
Speaker 1: And I get that, but just to push back a

603
00:32:05,600 --> 00:32:07,280
Speaker 1: little bit, and I get And this is like a

604
00:32:07,280 --> 00:32:10,000
Speaker 1: theme that we that comes up in various flavors on

605
00:32:10,120 --> 00:32:12,800
Speaker 1: odd lots a lot, which is that like companies have

606
00:32:13,000 --> 00:32:16,600
Speaker 1: internal it's really hard to replicate sort of like tacit

607
00:32:16,720 --> 00:32:20,280
Speaker 1: knowledge within a corporation. And we see that with companies

608
00:32:20,280 --> 00:32:22,680
Speaker 1: that make semiconductor equipment. We see that with companies that

609
00:32:22,720 --> 00:32:25,840
Speaker 1: make airplanes. We see that with real estate developers that

610
00:32:26,040 --> 00:32:28,479
Speaker 1: know how to turn an office building into a condo.

611
00:32:28,520 --> 00:32:30,479
Speaker 1: And so I think this is like a deep point,

612
00:32:31,200 --> 00:32:34,000
Speaker 1: but you know they are offering AI stuff like I

613
00:32:34,000 --> 00:32:36,800
Speaker 1: can look at Google right now, like there's Cloud AI

614
00:32:37,160 --> 00:32:39,600
Speaker 1: like and there's Asia AI, and they all have their announcements.

615
00:32:39,680 --> 00:32:41,480
Speaker 1: So I'm still trying to understand, like what is it

616
00:32:42,080 --> 00:32:45,120
Speaker 1: that you're offering that all the hyper scalers they all have,

617
00:32:45,240 --> 00:32:47,680
Speaker 1: they all say they have AI offering, So what is

618
00:32:47,720 --> 00:32:49,920
Speaker 1: the difference between sort of like what you have and

619
00:32:50,000 --> 00:32:52,920
Speaker 1: what they say is like their you know, AI compute platforms.

620
00:32:54,000 --> 00:32:56,320
Speaker 4: Absolutely, and this will really depend on how much technical

621
00:32:56,360 --> 00:32:59,360
Speaker 4: detail you like me to get into. But broadly, through

622
00:33:00,120 --> 00:33:03,960
Speaker 4: structure differentiation like literally using different components to build our cloud,

623
00:33:04,360 --> 00:33:07,520
Speaker 4: and through software differentiation we use different pieces of software

624
00:33:07,560 --> 00:33:11,400
Speaker 4: to operate and optimize our cloud, we're able to deliver

625
00:33:11,480 --> 00:33:15,240
Speaker 4: a product that's about forty to sixty percent more efficient

626
00:33:15,600 --> 00:33:18,000
Speaker 4: on a workload adjusted basis than what you find across

627
00:33:18,040 --> 00:33:21,440
Speaker 4: any of the hyperscalers. So, in other words, if you

628
00:33:21,440 --> 00:33:23,560
Speaker 4: were to take the same workload or like go do

629
00:33:23,680 --> 00:33:27,880
Speaker 4: the same process at a hyperscaler on the exact same

630
00:33:28,240 --> 00:33:32,360
Speaker 4: GPU compute versus core weave, we're going to be forty

631
00:33:32,360 --> 00:33:36,040
Speaker 4: to sixty percent more efficient at doing that because of

632
00:33:36,040 --> 00:33:39,520
Speaker 4: the way that we've configured everything relative to the hyperscalers

633
00:33:39,520 --> 00:33:43,040
Speaker 4: and it comes back to this analogy between like why

634
00:33:43,080 --> 00:33:45,960
Speaker 4: Forward can't produce the model? Why again like they can.

635
00:33:46,120 --> 00:33:49,320
Speaker 4: These are trillion dollar companies we're talking about. To your point,

636
00:33:49,320 --> 00:33:51,160
Speaker 4: they have the budget, they have the personnel, and they

637
00:33:51,200 --> 00:33:54,200
Speaker 4: certainly have the motivation to do so. But you know,

638
00:33:54,280 --> 00:33:57,280
Speaker 4: it's not just one singular thing they have to change.

639
00:33:57,280 --> 00:34:01,040
Speaker 4: It's a completely different way to building their their business

640
00:34:01,200 --> 00:34:04,440
Speaker 4: that they would have to orchestrate. And it's what's the analogies.

641
00:34:04,760 --> 00:34:07,520
Speaker 4: However many miles it takes to turn an aircraft carrier, right,

642
00:34:07,560 --> 00:34:09,279
Speaker 4: like it's it's going to take them a while to

643
00:34:09,320 --> 00:34:12,160
Speaker 4: do that. And I think if they do get there

644
00:34:12,400 --> 00:34:14,799
Speaker 4: at some point, which you know, I don't disagree with you,

645
00:34:14,920 --> 00:34:18,200
Speaker 4: they're certainly motivated to, it's it's going to have taken

646
00:34:18,239 --> 00:34:20,960
Speaker 4: them some time, literally years to get there, and they're

647
00:34:21,000 --> 00:34:23,839
Speaker 4: going to look really similar to us. And meanwhile, I've

648
00:34:24,280 --> 00:34:27,640
Speaker 4: dominated mortgage share and I've really established my product and

649
00:34:27,719 --> 00:34:30,440
Speaker 4: market and I continue when I'll continue to differentiate myself

650
00:34:30,480 --> 00:34:32,160
Speaker 4: on the software asign business as well.

651
00:34:49,120 --> 00:34:52,760
Speaker 2: Since we're on the topic of adaptation, can I ask about,

652
00:34:53,239 --> 00:34:56,319
Speaker 2: you know, your own evolution as a company, because I

653
00:34:56,360 --> 00:35:00,160
Speaker 2: think I've read that you started out in ethereum mining,

654
00:35:00,320 --> 00:35:03,719
Speaker 2: and at one point I'm pretty sure crypto mining was

655
00:35:03,800 --> 00:35:07,520
Speaker 2: a substantial, if not the biggest, portion of your business.

656
00:35:08,000 --> 00:35:14,000
Speaker 2: But you have clearly adapted or pivoted into this AI space.

657
00:35:14,080 --> 00:35:16,200
Speaker 2: So what has that been like and can you maybe

658
00:35:16,239 --> 00:35:20,040
Speaker 2: describe some of the trends that you've seen over your history.

659
00:35:21,160 --> 00:35:24,319
Speaker 4: Yes, absolutely, and you're right. We did start within the

660
00:35:24,320 --> 00:35:28,080
Speaker 4: cryptocurrency space back in twenty seventeen or so, and that

661
00:35:28,280 --> 00:35:32,279
Speaker 4: was spawned out of just frankly curiosity from a group

662
00:35:32,320 --> 00:35:36,640
Speaker 4: of former commodity traders. So myself, my two co founders,

663
00:35:36,840 --> 00:35:39,920
Speaker 4: we ran hedge funds, we ran family offices, so we

664
00:35:40,000 --> 00:35:42,480
Speaker 4: traded in these energy markets. So we were always attracted

665
00:35:42,520 --> 00:35:45,520
Speaker 4: to supply demand mechanics. But what attracted us but then

666
00:35:45,560 --> 00:35:49,800
Speaker 4: cryptocurrency was there's this arbitrage opportunity that was a permissionless

667
00:35:49,920 --> 00:35:52,400
Speaker 4: revenue stream, right, Like I knew the cost of power,

668
00:35:52,800 --> 00:35:55,239
Speaker 4: I knew what the hardware could generate in terms of

669
00:35:55,280 --> 00:35:58,719
Speaker 4: revenue with using a power input. Thus it's effectively an

670
00:35:58,800 --> 00:36:02,880
Speaker 4: arbitrage right, So we explored that we had some of

671
00:36:02,880 --> 00:36:06,359
Speaker 4: the infrastructure operating literally in our basements. As you said,

672
00:36:09,200 --> 00:36:12,720
Speaker 4: then that like quickly turned into scaling across warehouses, and

673
00:36:13,160 --> 00:36:16,520
Speaker 4: at some point in twenty I think it was twenty eighteen,

674
00:36:16,640 --> 00:36:21,560
Speaker 4: maybe late twenty eighteen, we were the largest ethereum miner

675
00:36:21,880 --> 00:36:26,160
Speaker 4: in North America. We were operating over fifty thousand GPUs,

676
00:36:26,200 --> 00:36:29,880
Speaker 4: we represented over one percent of the ethereum network. But

677
00:36:31,040 --> 00:36:33,400
Speaker 4: during that whole time, we just kept coming back to

678
00:36:33,440 --> 00:36:37,400
Speaker 4: the idea that there's no moat, there's no advantage that

679
00:36:37,840 --> 00:36:41,680
Speaker 4: we could create for ourselves relative to our competitors, right, Like, sure,

680
00:36:41,680 --> 00:36:44,160
Speaker 4: you could maybe focus on power price and just kind

681
00:36:44,200 --> 00:36:46,480
Speaker 4: of chase the cheapest power, but that just felt like

682
00:36:46,560 --> 00:36:49,080
Speaker 4: chasing to the bottom of the bucket, right. You know,

683
00:36:49,560 --> 00:36:51,200
Speaker 4: I think an area we could have gone into is

684
00:36:51,200 --> 00:36:53,640
Speaker 4: producing your own chips, right because if you produce the

685
00:36:53,640 --> 00:36:56,359
Speaker 4: own chips and you run the mining equipment before anyone

686
00:36:56,400 --> 00:36:58,920
Speaker 4: else has access to it, then you have an advantage

687
00:36:58,920 --> 00:37:00,719
Speaker 4: for that period. But you know, we weren't going to

688
00:37:00,719 --> 00:37:04,200
Speaker 4: go design and fabric own chips. So what we kept

689
00:37:04,239 --> 00:37:08,440
Speaker 4: coming back to was this GPU compute, Man, what if

690
00:37:08,480 --> 00:37:10,520
Speaker 4: we could do other things, right, like what if they

691
00:37:10,560 --> 00:37:15,240
Speaker 4: were what if we could develop uncorrelated optionality into multiple

692
00:37:15,320 --> 00:37:19,200
Speaker 4: high growth markets right in those markets or where we

693
00:37:19,440 --> 00:37:22,760
Speaker 4: predominantly sit today with an artificial intelligence, media and entertainment,

694
00:37:22,800 --> 00:37:27,040
Speaker 4: and computational chemistry. And the original thesis was, well, whenever

695
00:37:27,040 --> 00:37:31,360
Speaker 4: our compute isn't being allocated into those sectors, we'll just

696
00:37:31,400 --> 00:37:34,280
Speaker 4: have it mining cryptocurrency and we'll build out this fantastic

697
00:37:34,320 --> 00:37:37,279
Speaker 4: company that has one hundred percent utilization rate across the

698
00:37:37,280 --> 00:37:40,920
Speaker 4: infrastructure because it could switch immediately from being released from

699
00:37:40,960 --> 00:37:43,600
Speaker 4: an AI workload into going back into the Ethereum network.

700
00:37:44,360 --> 00:37:47,319
Speaker 4: And we did get a brief glimpse of being able

701
00:37:47,360 --> 00:37:51,440
Speaker 4: to operate that way in twenty twenty one, as we

702
00:37:51,560 --> 00:37:55,320
Speaker 4: had our cloud live and we had AI clients in place,

703
00:37:55,640 --> 00:37:59,880
Speaker 4: but Ethereum mining effectively ended during the merge in Q

704
00:38:00,200 --> 00:38:04,480
Speaker 4: three of twenty twenty two. But I'd say the other

705
00:38:04,560 --> 00:38:10,239
Speaker 4: thing that we never appreciated was the utter complexity of

706
00:38:10,320 --> 00:38:14,760
Speaker 4: running a CSP forgetting about the software side of the business,

707
00:38:14,760 --> 00:38:16,680
Speaker 4: which in and of itself, you know, we spent about

708
00:38:16,719 --> 00:38:20,320
Speaker 4: four years developing the software to build a modern cloud

709
00:38:20,360 --> 00:38:23,600
Speaker 4: to do infrastructure orchestration and actually be a cloud service provider.

710
00:38:24,200 --> 00:38:29,400
Speaker 4: The components themselves that the sector broadly used for crypto

711
00:38:29,480 --> 00:38:33,920
Speaker 4: mining were these retail grade GPUs, right, the kind of

712
00:38:33,960 --> 00:38:37,359
Speaker 4: things that you plug in your desktop to go play video.

713
00:38:37,600 --> 00:38:41,040
Speaker 3: They were like selling them on stock X. Yes, yes

714
00:38:41,480 --> 00:38:41,719
Speaker 3: it was.

715
00:38:41,760 --> 00:38:44,080
Speaker 4: It was crazy during that period to get your hands

716
00:38:44,080 --> 00:38:45,839
Speaker 4: on that infrastructure for crypto.

717
00:38:45,560 --> 00:38:48,560
Speaker 1: Mining, and all the video gamers hated the crypto people,

718
00:38:48,640 --> 00:38:50,319
Speaker 1: right because they're like, I want to like play this

719
00:38:50,400 --> 00:38:52,320
Speaker 1: game and they would like line up what is it?

720
00:38:52,920 --> 00:38:55,239
Speaker 1: Game stop and like the geek wire shop and all

721
00:38:55,280 --> 00:38:57,560
Speaker 1: that or whatever it is, and they like couldn't get

722
00:38:57,560 --> 00:38:58,759
Speaker 1: it because you got it, not you.

723
00:38:58,920 --> 00:39:01,359
Speaker 3: But yeah, they're orbating.

724
00:39:01,040 --> 00:39:03,319
Speaker 1: Access to the chips first and getting more value out

725
00:39:03,360 --> 00:39:04,520
Speaker 1: of them so that you could bit them up.

726
00:39:05,440 --> 00:39:08,920
Speaker 4: We were certainly part of the problem, and that's absolutely correct.

727
00:39:09,600 --> 00:39:12,640
Speaker 4: But you know what we found ultimately is like those chips,

728
00:39:13,560 --> 00:39:16,479
Speaker 4: that's not what you run enterprise grade workloads on. That's

729
00:39:16,520 --> 00:39:19,319
Speaker 4: not what's supporting you know, the largest AI companies in

730
00:39:19,360 --> 00:39:23,240
Speaker 4: the world, And starting in twenty nineteen, we stopped buying

731
00:39:23,400 --> 00:39:27,160
Speaker 4: any of those chips and only focused on purchasing enterprise

732
00:39:27,200 --> 00:39:30,960
Speaker 4: grade GPU chip sets that you know, Nvidia has a

733
00:39:31,000 --> 00:39:34,200
Speaker 4: probably about twelve different SKUs that they offer, including a

734
00:39:34,360 --> 00:39:38,440
Speaker 4: one hundred and h one hundred chips, and really oriented

735
00:39:38,440 --> 00:39:42,200
Speaker 4: our business around it. So it's a. I don't expect

736
00:39:42,320 --> 00:39:46,760
Speaker 4: to see much repurposing of this kind of older retail

737
00:39:46,880 --> 00:39:49,640
Speaker 4: grade GPU equipment that was used for crypto mining, because

738
00:39:50,360 --> 00:39:52,200
Speaker 4: in crypto mining, you want to buy the cheapest chip

739
00:39:52,239 --> 00:39:54,480
Speaker 4: that can do the thing for it, right, that can

740
00:39:54,520 --> 00:39:57,960
Speaker 4: participate in crypto mining. But there's a huge difference in

741
00:39:58,040 --> 00:40:00,880
Speaker 4: price between a retail plugin into your computer so you

742
00:40:00,880 --> 00:40:03,840
Speaker 4: can play video games chip and an enterprise grade you

743
00:40:03,880 --> 00:40:05,680
Speaker 4: can run it twenty four to seven. There's not going

744
00:40:05,719 --> 00:40:07,680
Speaker 4: to be downtime, You're going to have a low failure rate.

745
00:40:08,040 --> 00:40:10,640
Speaker 4: Like there's a large technology difference and there's a large

746
00:40:10,640 --> 00:40:13,840
Speaker 4: pricing difference between those and the crypto miners. You only

747
00:40:13,960 --> 00:40:16,360
Speaker 4: needed the retail grade ship because you know, if it

748
00:40:16,400 --> 00:40:19,239
Speaker 4: went down for two percent five percent of the time

749
00:40:19,280 --> 00:40:21,759
Speaker 4: for a failure rate, that's not a big deal. But

750
00:40:22,080 --> 00:40:27,200
Speaker 4: the tolerance, the uptime tolerance for these enterprise grade workloads

751
00:40:27,320 --> 00:40:30,719
Speaker 4: is measured on the thousandths of a percent, and it's

752
00:40:30,760 --> 00:40:33,960
Speaker 4: a different type of infrastructure, so we don't expect to

753
00:40:34,000 --> 00:40:38,040
Speaker 4: see the components really being reused, if at all. And

754
00:40:38,080 --> 00:40:40,640
Speaker 4: then the other variable, going back to the very beginning

755
00:40:40,640 --> 00:40:43,120
Speaker 4: of our conversation are the data centers in which these

756
00:40:43,120 --> 00:40:47,319
Speaker 4: are housed. So Joe, to your point earlier, you know,

757
00:40:47,400 --> 00:40:49,840
Speaker 4: we sit within tier three tier four data centers, and

758
00:40:49,880 --> 00:40:53,080
Speaker 4: that's the basically the broad industry standard for being able

759
00:40:53,080 --> 00:40:57,560
Speaker 4: to serve these kind of workloads. The crypto miners sat

760
00:40:57,600 --> 00:41:01,399
Speaker 4: within tier zero tier one data centers, and these things

761
00:41:01,440 --> 00:41:05,200
Speaker 4: are like highly interruptible. They do like really interesting things

762
00:41:05,280 --> 00:41:09,040
Speaker 4: like helping load balance the power markets in places like

763
00:41:09,080 --> 00:41:11,520
Speaker 4: ir Cot right, Like they'll shut down when power prices

764
00:41:11,560 --> 00:41:13,840
Speaker 4: go too high and it load balances the grid. But

765
00:41:14,960 --> 00:41:19,000
Speaker 4: enterprise AI workloads don't have a tolerance for that. Their

766
00:41:19,080 --> 00:41:22,200
Speaker 4: tolerance again is measured on the thousands of a percentage

767
00:41:22,239 --> 00:41:26,280
Speaker 4: in terms of uptime. So not only does the infrastructure

768
00:41:26,680 --> 00:41:30,839
Speaker 4: not work from cryptomning, but the data centers that they

769
00:41:31,040 --> 00:41:35,960
Speaker 4: built within don't work either the way that they're currently configured. Now,

770
00:41:36,360 --> 00:41:39,680
Speaker 4: they could potentially convert their sites into tier three and

771
00:41:39,719 --> 00:41:42,279
Speaker 4: tier four data centers. I'll tell you that in and

772
00:41:42,320 --> 00:41:46,200
Speaker 4: of itself, that is an extremely challenging task and it

773
00:41:46,200 --> 00:41:49,840
Speaker 4: takes a lot of proprietary knowledge and industry expertise to

774
00:41:49,880 --> 00:41:51,719
Speaker 4: do so. It's not just throwing a few fans in

775
00:41:51,760 --> 00:41:54,279
Speaker 4: a room and a few air conditioning units. It's a

776
00:41:55,320 --> 00:41:57,840
Speaker 4: it's it. Honestly, it feels like walking to a spaceship.

777
00:41:57,960 --> 00:41:59,960
Speaker 3: Tracy, this is this is an episode.

778
00:42:00,360 --> 00:42:02,560
Speaker 1: I don't know about you, Tracy, there's like six another

779
00:42:02,840 --> 00:42:06,279
Speaker 1: like six follow on episodes. No, seriously, like the whole

780
00:42:06,280 --> 00:42:10,400
Speaker 1: like data center market and the coolant and all, you know,

781
00:42:10,440 --> 00:42:13,600
Speaker 1: the electricity, Like there's so many different rabbit holes you

782
00:42:13,640 --> 00:42:15,960
Speaker 1: could go down, just like with the infrastructure you're.

783
00:42:15,800 --> 00:42:18,920
Speaker 2: Talking about, for sure, And I think the estimates that

784
00:42:19,000 --> 00:42:23,839
Speaker 2: I've seen on repurposing crypto GPUs, I think I've seen

785
00:42:23,880 --> 00:42:28,400
Speaker 2: like five to fifteen percent, so to Brannon's point, but

786
00:42:28,680 --> 00:42:31,000
Speaker 2: I'm sure, I'm sure there will be people out there

787
00:42:31,040 --> 00:42:31,600
Speaker 2: who try.

788
00:42:32,680 --> 00:42:35,560
Speaker 4: You got to try, right, because what if it works right,

789
00:42:35,960 --> 00:42:38,919
Speaker 4: If you can make that work, that's amazing. But we're

790
00:42:39,080 --> 00:42:41,920
Speaker 4: just you know, coming as an entity that was an

791
00:42:41,920 --> 00:42:45,759
Speaker 4: extremely large operator of that infrastructure and has built, you know,

792
00:42:45,880 --> 00:42:48,920
Speaker 4: one of the largest cloud service providers for AI workloads.

793
00:42:49,719 --> 00:42:51,959
Speaker 4: I can tell you it's it's gonna be really really

794
00:42:52,000 --> 00:42:54,239
Speaker 4: hard to do it because we've had exposure in both

795
00:42:54,280 --> 00:42:56,160
Speaker 4: those places and at the end of the day, they're

796
00:42:56,200 --> 00:42:59,719
Speaker 4: just very very different businesses, both from the type of

797
00:43:00,080 --> 00:43:02,400
Speaker 4: nearing and developers that you employed to the infrastructure to

798
00:43:02,440 --> 00:43:03,719
Speaker 4: the data centers that you sit within.

799
00:43:04,920 --> 00:43:07,160
Speaker 1: So can I just go back you know, yeah, just

800
00:43:07,239 --> 00:43:09,680
Speaker 1: sort of like big picture, and I guess it sort

801
00:43:09,680 --> 00:43:11,920
Speaker 1: of goes back to like who gets access to what?

802
00:43:12,000 --> 00:43:14,040
Speaker 3: Who gets access to chips?

803
00:43:14,080 --> 00:43:17,279
Speaker 1: And I imagine that you know, not only do you

804
00:43:17,320 --> 00:43:19,440
Speaker 1: need a lot of money to like build a relationship

805
00:43:19,520 --> 00:43:22,680
Speaker 1: with like Nvidio, you also probably need like a you know,

806
00:43:23,040 --> 00:43:25,399
Speaker 1: expectation you're going to be back the next year, back

807
00:43:25,440 --> 00:43:26,920
Speaker 1: the next year, back the next year, and that you

808
00:43:26,960 --> 00:43:30,040
Speaker 1: actually like have relationship and so forth. But I have

809
00:43:30,080 --> 00:43:33,200
Speaker 1: to imagine like planning is really tough, and when like

810
00:43:33,280 --> 00:43:35,880
Speaker 1: you know, you have this sort of like AI Machine

811
00:43:35,920 --> 00:43:39,960
Speaker 1: language whatever like industry, and then something like chet GPT

812
00:43:40,200 --> 00:43:43,040
Speaker 1: comes out and like suddenly everyone like, oh I need

813
00:43:43,080 --> 00:43:45,600
Speaker 1: to like have AI access. Talk to us about like

814
00:43:45,640 --> 00:43:47,759
Speaker 1: the sort of like challenge of just sort of like

815
00:43:48,120 --> 00:43:51,879
Speaker 1: planning to build when it can move that fast, and

816
00:43:51,960 --> 00:43:54,560
Speaker 1: like everyone is just sort of guessing how big this

817
00:43:54,640 --> 00:43:56,359
Speaker 1: market is going to be in two to three years.

818
00:43:57,080 --> 00:44:00,960
Speaker 4: Oh my gosh, it's it's it's been utterly insane right

819
00:44:01,040 --> 00:44:04,080
Speaker 4: like the it you know, back to last year, you know,

820
00:44:04,120 --> 00:44:06,200
Speaker 4: the supply chain and the ability to get your hands

821
00:44:06,200 --> 00:44:08,640
Speaker 4: on components. You know, you would call your your OEM.

822
00:44:08,960 --> 00:44:12,160
Speaker 4: The OEM is the original equipment manufacturer, Like those are

823
00:44:12,160 --> 00:44:14,760
Speaker 4: the super micros, the gigabytes of the world, who actually

824
00:44:14,840 --> 00:44:17,200
Speaker 4: you know, build the nodes, build the servers, and you're

825
00:44:17,360 --> 00:44:20,040
Speaker 4: you're buying through them, and then they buy the GPUs

826
00:44:20,080 --> 00:44:23,239
Speaker 4: from Nvidia and build all the components together. Right, So

827
00:44:23,480 --> 00:44:25,160
Speaker 4: if you called them and said, hey, I need this

828
00:44:25,239 --> 00:44:29,880
Speaker 4: many nodes to be delivered, they'll say, great, we'll start assembling.

829
00:44:30,040 --> 00:44:32,360
Speaker 4: Takes us, you know, a week to two weeks to

830
00:44:32,360 --> 00:44:34,399
Speaker 4: get the parts in assembling, and then it's another week

831
00:44:34,440 --> 00:44:36,160
Speaker 4: for them to ship them to you, and then it

832
00:44:36,160 --> 00:44:37,880
Speaker 4: takes us two to three weeks to plug them in

833
00:44:37,880 --> 00:44:40,799
Speaker 4: and put them online get them going. Right now, that's

834
00:44:40,920 --> 00:44:45,040
Speaker 4: completely changed, as you know, like all the supply chain

835
00:44:45,040 --> 00:44:48,000
Speaker 4: has gotten thrown off so much so that you know,

836
00:44:48,080 --> 00:44:52,480
Speaker 4: in Vidia is fully allocated, like they've fully sold out

837
00:44:52,480 --> 00:44:56,160
Speaker 4: their infrastructure through the end of the year. Right, you

838
00:44:56,200 --> 00:44:58,319
Speaker 4: can't call them, You can't call the OEM and just

839
00:44:58,320 --> 00:45:01,000
Speaker 4: say you need more compute chips like that, that's not possible.

840
00:45:01,160 --> 00:45:03,560
Speaker 4: So much so that you know, when clients are coming

841
00:45:03,640 --> 00:45:06,279
Speaker 4: to us today and they're asking for like a four

842
00:45:06,320 --> 00:45:09,560
Speaker 4: thousand GPU cluster to be built for them. We're telling

843
00:45:09,560 --> 00:45:12,879
Speaker 4: them Q one, and increasingly it's moving towards Q two

844
00:45:12,880 --> 00:45:14,760
Speaker 4: at this point because Q one is starting to get

845
00:45:15,160 --> 00:45:18,160
Speaker 4: booked up right now, So it's something that a lot

846
00:45:18,200 --> 00:45:20,879
Speaker 4: of time has been added to it. And then there's

847
00:45:20,920 --> 00:45:25,080
Speaker 4: other supply chain variables within there as well. You know,

848
00:45:25,120 --> 00:45:28,520
Speaker 4: we had a client earlier this year that we were

849
00:45:28,520 --> 00:45:30,560
Speaker 4: in negotiations with them on the contract and you know,

850
00:45:30,680 --> 00:45:33,600
Speaker 4: we really wanted to perform well on timing for it.

851
00:45:34,200 --> 00:45:38,080
Speaker 4: So we knew because of our orientation within the supply

852
00:45:38,160 --> 00:45:41,400
Speaker 4: chain that there were some critical components that needed to

853
00:45:41,400 --> 00:45:44,800
Speaker 4: be ordered ahead of time so that it would reduce

854
00:45:45,120 --> 00:45:47,680
Speaker 4: our time to bring in the infrastructure online. And at

855
00:45:47,680 --> 00:45:50,120
Speaker 4: that point it was the power supply units and the

856
00:45:50,200 --> 00:45:53,840
Speaker 4: fans for the nodes that the Oliams were putting together,

857
00:45:54,480 --> 00:45:56,719
Speaker 4: and if we hadn't have done that, it would have

858
00:45:56,719 --> 00:45:59,600
Speaker 4: been another i think eight weeks on top of the

859
00:45:59,640 --> 00:46:03,279
Speaker 4: build process, just because not all the components would have

860
00:46:03,280 --> 00:46:06,359
Speaker 4: been there at the same time. So you're navigating this,

861
00:46:06,960 --> 00:46:10,960
Speaker 4: you know, within other kind of global supply chain disruptions

862
00:46:10,960 --> 00:46:13,160
Speaker 4: and inflation and all these other things that are going

863
00:46:13,160 --> 00:46:16,960
Speaker 4: on right now and it's just an insanely complex task

864
00:46:17,080 --> 00:46:21,640
Speaker 4: that I think, you know, the generation of software developers

865
00:46:21,880 --> 00:46:25,759
Speaker 4: and founders that we're working with today were used to

866
00:46:25,840 --> 00:46:28,480
Speaker 4: being able to go to a cloud service provider and

867
00:46:28,560 --> 00:46:31,560
Speaker 4: just getting whatever infrastructure they needed. Right. You go to

868
00:46:31,600 --> 00:46:33,960
Speaker 4: your hyperscalers and say all right, any of this and

869
00:46:34,000 --> 00:46:37,680
Speaker 4: it was just there and available. And that just doesn't

870
00:46:37,719 --> 00:46:42,560
Speaker 4: exist today because the pace of demand growth that we've

871
00:46:42,600 --> 00:46:45,439
Speaker 4: been on and just the lack of this infrastructure's availability,

872
00:46:45,680 --> 00:46:49,080
Speaker 4: and it's just caught everyone by surprise. Again. You're you're

873
00:46:49,120 --> 00:46:54,120
Speaker 4: asking infrastructure to keep pace with the fastest adoption of

874
00:46:55,080 --> 00:46:57,760
Speaker 4: a new piece of software that's ever occurred.

875
00:46:58,480 --> 00:47:01,440
Speaker 1: Brandon McBee, core Leave, Thank you so much. That was

876
00:47:01,440 --> 00:47:03,960
Speaker 1: a great conversation. Like I said, I always sort of

877
00:47:04,040 --> 00:47:06,520
Speaker 1: measure the quality of a conversation of like do I

878
00:47:06,520 --> 00:47:07,200
Speaker 1: get seven.

879
00:47:07,160 --> 00:47:09,279
Speaker 2: How many additional episode like that is.

880
00:47:09,200 --> 00:47:11,520
Speaker 1: A pretty good proxy for a good conversation. Do you

881
00:47:11,520 --> 00:47:13,480
Speaker 1: get like eight ideas for future episodes? We got a

882
00:47:13,480 --> 00:47:15,920
Speaker 1: bunch there, So thank you so much for coming.

883
00:47:15,719 --> 00:47:16,360
Speaker 3: On the podcast.

884
00:47:16,640 --> 00:47:18,319
Speaker 4: Always happy to chat with you guys, and thank you

885
00:47:18,360 --> 00:47:19,080
Speaker 4: for the invitation.

886
00:47:32,880 --> 00:47:33,280
Speaker 3: Tracy.

887
00:47:33,320 --> 00:47:35,920
Speaker 1: I want to find that company that makes the coolant

888
00:47:36,280 --> 00:47:39,200
Speaker 1: for the data No, seriously, for the data centers. That

889
00:47:39,320 --> 00:47:43,040
Speaker 1: allows them to pack more compute and more energy into

890
00:47:43,080 --> 00:47:45,200
Speaker 1: this space, because it's like it feels like they're probably

891
00:47:45,200 --> 00:47:46,239
Speaker 1: going to make a fortune of the.

892
00:47:46,160 --> 00:47:47,759
Speaker 2: Next Joe, I think you just want to talk to

893
00:47:47,800 --> 00:47:51,640
Speaker 2: an HVACT contractor that's like installing out.

894
00:47:51,880 --> 00:47:52,520
Speaker 3: Can we talk to it?

895
00:47:52,920 --> 00:47:56,000
Speaker 1: Just some random like I love the Maybe it was

896
00:47:56,040 --> 00:47:59,200
Speaker 1: such a funny thought like these like really advanced data centers,

897
00:47:59,280 --> 00:48:00,960
Speaker 1: like oh, do we have like a local air conditioning

898
00:48:00,960 --> 00:48:01,759
Speaker 1: guy who can like.

899
00:48:01,880 --> 00:48:03,520
Speaker 2: But I imagine actually that would have been a good

900
00:48:03,600 --> 00:48:07,120
Speaker 2: question for Brandon, wouldn't it. Like the labor constraints in

901
00:48:07,600 --> 00:48:10,799
Speaker 2: building and adapting those data centers. But there was so

902
00:48:10,920 --> 00:48:12,960
Speaker 2: much in there. One of the things, one of the

903
00:48:12,960 --> 00:48:15,680
Speaker 2: things that I was thinking about was the point about how, well, okay,

904
00:48:15,960 --> 00:48:18,799
Speaker 2: if you train a model on one type of chip,

905
00:48:18,840 --> 00:48:21,600
Speaker 2: you're going to keep using that type of chip. And

906
00:48:21,640 --> 00:48:24,000
Speaker 2: I guess, I guess it's kind of obvious, but it

907
00:48:24,080 --> 00:48:28,560
Speaker 2: does suggest that there's some stickiness there, Like if you

908
00:48:28,640 --> 00:48:32,239
Speaker 2: start out using an Nvidia each one hundred, you're going

909
00:48:32,280 --> 00:48:33,880
Speaker 2: to keep using them, and in fact, you're going to

910
00:48:33,920 --> 00:48:37,800
Speaker 2: consume even more because the processing power required the compute

911
00:48:37,840 --> 00:48:40,880
Speaker 2: required for the inference is higher than for the actual

912
00:48:40,920 --> 00:48:42,080
Speaker 2: initial training.

913
00:48:42,320 --> 00:48:44,760
Speaker 1: Which I knew that that was the case because Stacy

914
00:48:44,800 --> 00:48:47,200
Speaker 1: said so as well, but I did not realize quite

915
00:48:47,200 --> 00:48:51,600
Speaker 1: the scale of like how much more Like okay, like

916
00:48:51,680 --> 00:48:53,400
Speaker 1: if you train a model and then we try to

917
00:48:53,440 --> 00:48:54,840
Speaker 1: take it to market product tize it.

918
00:48:55,000 --> 00:48:56,080
Speaker 3: As a business person.

919
00:48:55,920 --> 00:48:58,319
Speaker 1: I'd say, if we try to productize, like how much

920
00:48:58,480 --> 00:49:02,680
Speaker 1: more computing power or we would need for the inferant aspect?

921
00:49:02,880 --> 00:49:04,520
Speaker 1: And meanwhile we have to keep training it all the

922
00:49:04,560 --> 00:49:05,960
Speaker 1: time to keep it up with fresh data and.

923
00:49:05,920 --> 00:49:06,480
Speaker 3: Stuff like that.

924
00:49:06,600 --> 00:49:10,320
Speaker 2: Yeah, totally. And the other thing that I was thinking about,

925
00:49:10,400 --> 00:49:13,839
Speaker 2: and again Stacy mentioned this in our discussion with him

926
00:49:13,840 --> 00:49:17,480
Speaker 2: as well, but this idea of Nvidia building a kind

927
00:49:17,520 --> 00:49:21,279
Speaker 2: of large ecosystem around the hardware. So you have the

928
00:49:21,360 --> 00:49:24,799
Speaker 2: open source software Kudo, which we talked about a little bit,

929
00:49:25,120 --> 00:49:28,640
Speaker 2: and then you have these sort of high touch partnerships

930
00:49:28,719 --> 00:49:32,520
Speaker 2: with companies like core Weave where they're trying to make

931
00:49:32,560 --> 00:49:35,160
Speaker 2: it as easy as possible for you to use their

932
00:49:35,239 --> 00:49:38,440
Speaker 2: chips and set them up in a way that works

933
00:49:38,480 --> 00:49:42,880
Speaker 2: for you. It feels like maybe it feels almost like

934
00:49:42,920 --> 00:49:44,120
Speaker 2: what bitmin used to do.

935
00:49:44,120 --> 00:49:46,960
Speaker 3: Do you remember that, uh, no.

936
00:49:46,920 --> 00:49:49,680
Speaker 2: Maybe they're still doing it anyway, but it does feel

937
00:49:49,719 --> 00:49:52,920
Speaker 2: like they're trying to build this like ecosystem mote around

938
00:49:53,080 --> 00:49:54,160
Speaker 2: the chip technology.

939
00:49:54,440 --> 00:49:54,680
Speaker 4: Yeah.

940
00:49:54,760 --> 00:49:56,319
Speaker 3: No, absolutely true.

941
00:49:56,360 --> 00:49:58,320
Speaker 1: And you know, I really do take that point that

942
00:49:58,440 --> 00:50:02,040
Speaker 1: Brandon made about like every company has a sort of

943
00:50:02,080 --> 00:50:04,759
Speaker 1: like knowledge that cannot be written down on a piece

944
00:50:04,800 --> 00:50:06,640
Speaker 1: of paper. Yeah, which is a Dan Wong point that

945
00:50:06,680 --> 00:50:08,759
Speaker 1: we've been talking about for years. And so it's like

946
00:50:09,120 --> 00:50:10,840
Speaker 1: to your point, you know, like you have to like

947
00:50:11,080 --> 00:50:13,720
Speaker 1: use different types of connectors and different types of power

948
00:50:13,760 --> 00:50:16,040
Speaker 1: and all these stuff like the ease with which any

949
00:50:16,040 --> 00:50:21,680
Speaker 1: sort of traditional cloud provider or data center provider can

950
00:50:21,920 --> 00:50:23,880
Speaker 1: you know, sort of switch to it's like a you know,

951
00:50:24,000 --> 00:50:25,840
Speaker 1: it's not trivial even with lots of no.

952
00:50:26,400 --> 00:50:29,760
Speaker 2: But coming away, I'm coming away from that conversation thinking,

953
00:50:29,840 --> 00:50:32,799
Speaker 2: like the big question here is how quickly can those

954
00:50:32,840 --> 00:50:37,120
Speaker 2: other hyperscalers adapt and like how big a moat can

955
00:50:37,239 --> 00:50:38,960
Speaker 2: Nvidia build around this business?

956
00:50:39,000 --> 00:50:41,160
Speaker 1: And then I mean the other question I have is

957
00:50:41,280 --> 00:50:44,040
Speaker 1: like what if none of these companies make any money

958
00:50:44,400 --> 00:50:46,600
Speaker 1: building AI models, Like I still don't think like that's

959
00:50:46,640 --> 00:50:48,800
Speaker 1: been proven and so you can have this like huge

960
00:50:48,840 --> 00:50:50,520
Speaker 1: boom and like, hey, we got to build any by

961
00:50:50,640 --> 00:50:52,920
Speaker 1: a model is what we're going to build, like you know,

962
00:50:53,120 --> 00:50:56,440
Speaker 1: outlaws GPT for like data stuff and whatever. But it

963
00:50:56,640 --> 00:50:59,759
Speaker 1: all is somewhat predicated on these companies being successful and

964
00:50:59,760 --> 00:51:02,040
Speaker 1: making a lot of money. And if they're not, and

965
00:51:02,040 --> 00:51:04,720
Speaker 1: if it turns out that like the monetization of AI

966
00:51:04,840 --> 00:51:08,320
Speaker 1: products is trickier than expected, then that also raises this

967
00:51:08,440 --> 00:51:09,640
Speaker 1: question about like how long.

968
00:51:09,480 --> 00:51:12,480
Speaker 2: This Like I'm sorry, Joe, so you're saying that tech

969
00:51:12,520 --> 00:51:16,319
Speaker 2: companies should make money? Is that it? Are you sure?

970
00:51:17,120 --> 00:51:20,640
Speaker 3: That's right? That's it's real post zerp thinking of it?

971
00:51:20,960 --> 00:51:23,000
Speaker 2: I know, all right? Shall we leave it there?

972
00:51:23,040 --> 00:51:23,719
Speaker 3: Let's leave it there.

973
00:51:23,800 --> 00:51:26,640
Speaker 2: This has been another episode of the All Thoughts podcast.

974
00:51:26,680 --> 00:51:29,279
Speaker 2: I'm Tracy Alloway. You can follow me on Twitter at

975
00:51:29,320 --> 00:51:30,520
Speaker 2: Tracy Alloway.

976
00:51:30,160 --> 00:51:32,480
Speaker 1: And I'm Joe Wisenthal. You can follow me on Twitter

977
00:51:32,560 --> 00:51:35,160
Speaker 1: at the Stalwart. Follow our guest Brannon McBee.

978
00:51:35,200 --> 00:51:36,680
Speaker 3: He's at Brannon McBee.

979
00:51:36,840 --> 00:51:40,279
Speaker 1: Follow our producers Carmen Rodriguez at Carmen Arman and dash

980
00:51:40,280 --> 00:51:42,600
Speaker 1: Ol Bennett at dashbot. And check out all of the

981
00:51:42,600 --> 00:51:46,200
Speaker 1: Bloomberg podcasts under the handle at podcasts, and for more

982
00:51:46,239 --> 00:51:49,359
Speaker 1: Odd Lots content, go to Bloomberg dot com slash odd

983
00:51:49,360 --> 00:51:52,560
Speaker 1: lots where we have transcripts, a blog, and a newsletter

984
00:51:52,600 --> 00:51:55,760
Speaker 1: that comes out each Friday. And check out our Discord.

985
00:51:55,920 --> 00:51:58,920
Speaker 1: We have an AI channel and a semiconductor channel in

986
00:51:58,960 --> 00:52:01,280
Speaker 1: there so people talk about these topics twenty four to seven.

987
00:52:01,480 --> 00:52:02,120
Speaker 3: Maybe they'll be.

988
00:52:02,040 --> 00:52:04,799
Speaker 1: Talking about them in both of those rooms when this

989
00:52:04,880 --> 00:52:08,080
Speaker 1: comes out. Discord dot gg, slash.

990
00:52:07,719 --> 00:52:11,520
Speaker 2: Outline and if you enjoy all thoughts, if you appreciate

991
00:52:11,520 --> 00:52:14,080
Speaker 2: conversations like the one we just had with Brandon McBee,

992
00:52:14,160 --> 00:52:17,960
Speaker 2: then please leave us a positive review on your favorite

993
00:52:18,080 --> 00:52:20,080
Speaker 2: podcast platform. Thanks for listening.