1 00:00:02,720 --> 00:00:15,880 Speaker 1: Bloomberg Audio Studios, Podcasts, Radio News. 2 00:00:18,640 --> 00:00:21,480 Speaker 2: Hello and welcome to another episode of the Out Thoughts podcast. 3 00:00:21,560 --> 00:00:22,840 Speaker 2: I'm Tracy Alloway. 4 00:00:22,640 --> 00:00:23,200 Speaker 3: And I'm Jill. 5 00:00:23,400 --> 00:00:27,480 Speaker 2: Wasn't thal so Joe. We're still continuing our series. Recorded 6 00:00:27,520 --> 00:00:29,440 Speaker 2: from the live show in New York. We had a 7 00:00:29,440 --> 00:00:32,320 Speaker 2: bunch of great conversations. A couple of them were building 8 00:00:32,360 --> 00:00:35,479 Speaker 2: off of discussions that we had had previously, and one 9 00:00:35,520 --> 00:00:39,160 Speaker 2: of those discussions was in Chicago at another live show 10 00:00:39,280 --> 00:00:42,279 Speaker 2: about six or seven months ago. Back in October, we 11 00:00:42,320 --> 00:00:46,280 Speaker 2: spoke with Don Wilson of DRW about the trading environment, 12 00:00:46,400 --> 00:00:48,479 Speaker 2: but also about his new venture. 13 00:00:48,920 --> 00:00:51,839 Speaker 3: Right and so his new venture is one that actually 14 00:00:51,880 --> 00:00:54,440 Speaker 3: there's quite a bit of competition in and quite of 15 00:00:54,560 --> 00:00:59,080 Speaker 3: excitement in, and it's essentially like okay, GPUs. We know 16 00:00:59,080 --> 00:01:02,560 Speaker 3: they're very important for the AI boom, et cetera. The 17 00:01:02,640 --> 00:01:07,240 Speaker 3: question is can GPU capacity, which is scarce, can it 18 00:01:07,319 --> 00:01:11,920 Speaker 3: become a tradable commodity such that I can buy futures 19 00:01:11,959 --> 00:01:15,479 Speaker 3: to lock in my price of access to compute power. 20 00:01:15,959 --> 00:01:20,000 Speaker 3: Could I resell those futures? Will there be speculators speculating 21 00:01:20,000 --> 00:01:21,759 Speaker 3: on the upper down price of like an H one 22 00:01:21,840 --> 00:01:24,120 Speaker 3: hundred running an H one hundred in video chip for 23 00:01:24,160 --> 00:01:26,200 Speaker 3: an hour. This is a big question. We know there's 24 00:01:26,200 --> 00:01:28,960 Speaker 3: a lot of interest in the actual compute, but whether 25 00:01:29,000 --> 00:01:32,800 Speaker 3: there's interested in compute futures. It's tradable instruments is very TVD. 26 00:01:33,000 --> 00:01:36,000 Speaker 2: Yeah, And the analogy that everyone always uses is compute 27 00:01:36,080 --> 00:01:38,240 Speaker 2: is the new oil, right, so why can't it have 28 00:01:38,640 --> 00:01:41,120 Speaker 2: you know, a market structure that looks somewhat like the 29 00:01:41,120 --> 00:01:45,520 Speaker 2: oil market. And there are challenges. Fungibility is a big one, 30 00:01:45,560 --> 00:01:48,800 Speaker 2: like one chip might not necessarily be equal to another 31 00:01:48,880 --> 00:01:49,880 Speaker 2: chip or one ship. 32 00:01:50,040 --> 00:01:52,880 Speaker 3: The same chip at one data center might equal to 33 00:01:52,880 --> 00:01:55,160 Speaker 3: the same chip at a different data center exactly. 34 00:01:55,200 --> 00:01:58,760 Speaker 2: And so even if you're not interested in AI, what 35 00:01:58,840 --> 00:02:02,400 Speaker 2: I say here is like the market structure questions and 36 00:02:02,480 --> 00:02:06,360 Speaker 2: the idea of building an entirely new market is really 37 00:02:06,400 --> 00:02:08,680 Speaker 2: fascinating to me, and I think others will find it 38 00:02:08,720 --> 00:02:11,320 Speaker 2: interesting too. And we really do have the perfect guest. 39 00:02:11,360 --> 00:02:14,000 Speaker 2: We're speaking with Carmen Lee. She is the CEO of 40 00:02:14,080 --> 00:02:17,360 Speaker 2: Compute Exchange and Silicon Data. These are the two companies 41 00:02:17,600 --> 00:02:21,400 Speaker 2: that Wilson is invested in, and they've already announced that 42 00:02:21,440 --> 00:02:25,080 Speaker 2: they're doing futures with the CME. So really the perfect 43 00:02:25,160 --> 00:02:28,760 Speaker 2: person to speak to So take a listen. Last October 44 00:02:28,840 --> 00:02:32,440 Speaker 2: we spoke with Don Wilson of DRW fame and he 45 00:02:32,639 --> 00:02:35,680 Speaker 2: was talking to us about his new project, which was 46 00:02:35,720 --> 00:02:39,720 Speaker 2: basically building out this compute exchange. Now we're here with 47 00:02:39,760 --> 00:02:42,720 Speaker 2: you six months later. You're actually the one leading it. 48 00:02:43,480 --> 00:02:46,320 Speaker 2: How far are you in this endeavor? And remind us 49 00:02:46,520 --> 00:02:48,160 Speaker 2: what exactly are you trying to do here? 50 00:02:48,440 --> 00:02:52,480 Speaker 4: Yeah, so thank you for the the quizing one great audience. 51 00:02:52,919 --> 00:02:54,919 Speaker 4: Before I do that, I actually going to call back 52 00:02:55,120 --> 00:02:58,840 Speaker 4: to six months ago in the DOM podcast you did. 53 00:02:59,120 --> 00:03:02,920 Speaker 4: You asked them question, what if compute prices keep going 54 00:03:03,000 --> 00:03:07,519 Speaker 4: to go up? At the time September October, compute prices 55 00:03:07,560 --> 00:03:11,040 Speaker 4: were going down across our chips. Now see what happened. 56 00:03:11,240 --> 00:03:13,239 Speaker 4: I think you called it. I think you called the 57 00:03:13,480 --> 00:03:16,040 Speaker 4: market called it. So I'm the fund of CEO for 58 00:03:16,120 --> 00:03:19,679 Speaker 4: Cilicond Data, so that's the index provider for GPU indcees. 59 00:03:20,360 --> 00:03:23,800 Speaker 4: We recently announced partnership with CME, so we all been 60 00:03:23,840 --> 00:03:28,760 Speaker 4: launching gp futurning options ME in a couple of months, 61 00:03:29,240 --> 00:03:33,280 Speaker 4: pending the FDIC approval. Obviously that's quite exciting. We've been 62 00:03:33,320 --> 00:03:36,520 Speaker 4: working on GPU indices for past two and a half years, 63 00:03:37,160 --> 00:03:40,000 Speaker 4: starting twenty twenty four April, so it's been a while, 64 00:03:40,480 --> 00:03:44,320 Speaker 4: and we launched for world's first GPU inducees at Bloomberg 65 00:03:44,360 --> 00:03:48,200 Speaker 4: Terminal in twenty twenty five. A year later we launched 66 00:03:48,200 --> 00:03:51,600 Speaker 4: the partnership with CMME, so it's quite exciting. Separately, I 67 00:03:51,640 --> 00:03:53,720 Speaker 4: heard you mentioned Compute Change before, so thank you for 68 00:03:53,760 --> 00:03:56,840 Speaker 4: doing AGIL. I'm the CEO for compute Change, which is 69 00:03:56,880 --> 00:04:01,200 Speaker 4: sport marketplace for GPO procurement. So we do reserve contracts 70 00:04:01,640 --> 00:04:04,480 Speaker 4: for contracts as well as ree for pitch contracts. 71 00:04:04,720 --> 00:04:08,200 Speaker 3: Let's talk about the variety of options that we have 72 00:04:08,320 --> 00:04:12,280 Speaker 3: to financialize compute and so forth. So this, I mean, 73 00:04:12,280 --> 00:04:14,520 Speaker 3: this came up in our conversation. The first conversation we 74 00:04:14,600 --> 00:04:17,120 Speaker 3: had with we had with Ian Dunning. Who is the 75 00:04:17,160 --> 00:04:21,159 Speaker 3: type of buyer who would want to buy compute on 76 00:04:21,200 --> 00:04:24,680 Speaker 3: a spot market because right you talk about typically we 77 00:04:24,720 --> 00:04:27,839 Speaker 3: think it's like these multi year contracts where some entity 78 00:04:28,200 --> 00:04:30,320 Speaker 3: enters into a contract with a data center or a 79 00:04:30,360 --> 00:04:33,200 Speaker 3: new cloud whatever, and this they have this for a while. 80 00:04:33,400 --> 00:04:36,080 Speaker 3: So who is the market, who is the buyer or 81 00:04:36,120 --> 00:04:39,279 Speaker 3: the user of these instruments that might want to buy 82 00:04:39,600 --> 00:04:43,680 Speaker 3: spot compute or very short term short dated compute futures. 83 00:04:43,760 --> 00:04:46,640 Speaker 4: It's a great question. So the compute market right now, 84 00:04:46,760 --> 00:04:50,880 Speaker 4: for comput change, we have all our provider mostly are 85 00:04:50,920 --> 00:04:54,479 Speaker 4: neo clouds around the world. It's one side. Another site 86 00:04:54,600 --> 00:04:58,479 Speaker 4: is a big variety AI start up. So even though 87 00:04:58,480 --> 00:05:02,440 Speaker 4: they are start up, millions of dollars on GPUs already. 88 00:05:02,880 --> 00:05:06,640 Speaker 4: There are enterprises who are traditional businesses, but they are 89 00:05:06,760 --> 00:05:09,240 Speaker 4: needing a note to notes, a few service here and 90 00:05:09,279 --> 00:05:13,200 Speaker 4: there for their inferencing or I don't know, other deployment needs. 91 00:05:13,560 --> 00:05:17,680 Speaker 4: They are providers. They are influencing providers right they don't 92 00:05:17,680 --> 00:05:21,960 Speaker 4: own GPUs, but they provide open source, open weights model 93 00:05:22,000 --> 00:05:25,600 Speaker 4: support for other use cases. So what see big variety. 94 00:05:26,160 --> 00:05:31,200 Speaker 4: Most North American firms they do a variety of combination contracts. 95 00:05:31,839 --> 00:05:35,640 Speaker 4: Obviously on demand give you the most flexibility. You don't 96 00:05:35,680 --> 00:05:38,360 Speaker 4: pay and you don't use it. However, you're also at 97 00:05:38,360 --> 00:05:40,960 Speaker 4: a mercy of defense. Apply curve at a given time, 98 00:05:41,279 --> 00:05:44,039 Speaker 4: so translate to your price can go from three dollars 99 00:05:44,040 --> 00:05:47,000 Speaker 4: to six to nine depends on demand supply curve shifting, 100 00:05:47,440 --> 00:05:51,880 Speaker 4: so that doesn't help when you can have a predictable margin. 101 00:05:52,560 --> 00:05:56,360 Speaker 4: And also in terms of scarce, you're not guaranteed to 102 00:05:56,400 --> 00:06:00,360 Speaker 4: feel GPU resources for next hour or next month. So 103 00:06:00,440 --> 00:06:02,680 Speaker 4: you see a lot of people shifting from on demand 104 00:06:02,720 --> 00:06:06,320 Speaker 4: to reserve even four contracts right, so full contracts you 105 00:06:06,360 --> 00:06:10,600 Speaker 4: back the lock in a deliverables for next whatever month, right, 106 00:06:10,640 --> 00:06:13,080 Speaker 4: starting in September maybe or starting no member of forbare 107 00:06:13,120 --> 00:06:16,760 Speaker 4: weapons right. So this all comes because of market condition. 108 00:06:17,120 --> 00:06:21,599 Speaker 4: So computing cover that the physical GPO procurement also token, 109 00:06:21,680 --> 00:06:23,680 Speaker 4: so we love to talk about token as well. On 110 00:06:23,760 --> 00:06:27,760 Speaker 4: flip side, who's going to use the future options can 111 00:06:27,839 --> 00:06:29,080 Speaker 4: be a similar set of people. 112 00:06:29,320 --> 00:06:29,520 Speaker 1: Right. 113 00:06:29,920 --> 00:06:32,799 Speaker 4: You look at oil market, which we all love double 114 00:06:32,880 --> 00:06:36,159 Speaker 4: TM brand. Right, the people use double TM brand. A 115 00:06:36,240 --> 00:06:39,880 Speaker 4: lot of them are naturally long oil. So the shells 116 00:06:40,080 --> 00:06:44,120 Speaker 4: ran the producers. They need to hatch your revenue volatility 117 00:06:44,640 --> 00:06:48,360 Speaker 4: by shorting futures or port options if you're naturally short 118 00:06:48,400 --> 00:06:52,680 Speaker 4: oil americandline right, they want to control their cost volatility. 119 00:06:53,080 --> 00:06:55,960 Speaker 4: They want to obviously use future options as well. 120 00:06:56,680 --> 00:06:57,120 Speaker 3: Simple to the. 121 00:06:57,160 --> 00:07:01,520 Speaker 4: Compute your new cloud or you're at anyone have the servers. 122 00:07:01,800 --> 00:07:04,560 Speaker 4: Ideally you want to have predictable revenue streams. 123 00:07:04,240 --> 00:07:07,040 Speaker 3: So the neo cloud would be the shell on this example. 124 00:07:06,760 --> 00:07:11,720 Speaker 4: Exactly, you have GPS, right, all the banks were GPUs 125 00:07:11,760 --> 00:07:15,720 Speaker 4: on your balance sheet right your long GPUs, then naturally 126 00:07:15,760 --> 00:07:18,800 Speaker 4: you want to make sure revenue right is stable to 127 00:07:18,840 --> 00:07:21,040 Speaker 4: certain degree, and then you want to use future to 128 00:07:21,040 --> 00:07:23,800 Speaker 4: do so. If you are naturally short GPU, which is 129 00:07:23,840 --> 00:07:25,880 Speaker 4: everybody in this room, amas you tell me you have 130 00:07:25,920 --> 00:07:30,120 Speaker 4: GPUs right, then you depends how much you use. If 131 00:07:30,120 --> 00:07:32,480 Speaker 4: you want to control your cost volatility, you want to 132 00:07:32,520 --> 00:07:33,680 Speaker 4: use future to hatch as well. 133 00:07:34,080 --> 00:07:38,280 Speaker 2: Just on the compute exchange side of things, if someone 134 00:07:38,360 --> 00:07:41,120 Speaker 2: is buying like off the spot market, how do you 135 00:07:41,160 --> 00:07:44,440 Speaker 2: guarantee I'm not sure quality is the right word for this, 136 00:07:44,560 --> 00:07:47,840 Speaker 2: but how do you guarantee they're getting what they expect? 137 00:07:47,960 --> 00:07:50,640 Speaker 4: This is a great question. So I canna flip to 138 00:07:50,760 --> 00:07:55,200 Speaker 4: slide if you don't mind. So I usually don't let 139 00:07:55,240 --> 00:07:57,960 Speaker 4: use slides, but this time because you mentioned really good questions. 140 00:07:58,000 --> 00:08:01,600 Speaker 4: So we actually call it GPU lottery. So we published 141 00:08:01,600 --> 00:08:06,320 Speaker 4: a paper earlier this year at GPGPO conference with Jefferson 142 00:08:06,400 --> 00:08:09,880 Speaker 4: Lab on GPO performances. Well actually, so well we can 143 00:08:09,960 --> 00:08:12,360 Speaker 4: have your quit link, you know, to the audience later on. 144 00:08:12,600 --> 00:08:14,240 Speaker 4: We actually this is a one hundred by the way, 145 00:08:14,360 --> 00:08:16,280 Speaker 4: I know we didn't put tang on the day. Is 146 00:08:16,440 --> 00:08:20,080 Speaker 4: a one hundred forty gigabytes memory bandwidth. We prove there's 147 00:08:20,120 --> 00:08:24,160 Speaker 4: thirty eight percent performance variants for the same trip, and 148 00:08:24,320 --> 00:08:27,880 Speaker 4: we decompose it into the chip self, into provider and 149 00:08:28,040 --> 00:08:31,720 Speaker 4: inter provider. And there's many reasons for that, right and 150 00:08:31,840 --> 00:08:34,719 Speaker 4: to your point, you can you don't know until you 151 00:08:34,800 --> 00:08:39,280 Speaker 4: get your GPS. We have a PLAT for GPU. CARFAX 152 00:08:39,360 --> 00:08:41,600 Speaker 4: for GPU depends how to look at it, so we 153 00:08:41,800 --> 00:08:45,720 Speaker 4: in compute change. You actually verify the GPU before delivered 154 00:08:45,720 --> 00:08:48,320 Speaker 4: to you. So basically you are you are. You CANFQ 155 00:08:48,720 --> 00:08:50,720 Speaker 4: for say, hey, I one A two hundred BI two 156 00:08:50,800 --> 00:08:54,160 Speaker 4: hundred notes. Obviously it will give you specs back and 157 00:08:54,200 --> 00:08:58,360 Speaker 4: the commercial back same time, independently verify the performances on flops, 158 00:08:58,559 --> 00:09:04,079 Speaker 4: memory bandwidth totals another information as ACE and other things. 159 00:09:04,320 --> 00:09:07,079 Speaker 4: And as a user you can decide is price your 160 00:09:07,080 --> 00:09:10,400 Speaker 4: most important criteria? Maybe it is, or maybe you're willing 161 00:09:10,440 --> 00:09:13,319 Speaker 4: to pay a premium for dual location or the performances 162 00:09:13,400 --> 00:09:17,040 Speaker 4: that you care more about on latency. Right, we believe 163 00:09:17,080 --> 00:09:18,959 Speaker 4: gave people the option and transparency it is the most 164 00:09:19,000 --> 00:09:19,559 Speaker 4: important thing. 165 00:09:19,840 --> 00:09:22,600 Speaker 3: Let's stick with the oil analogy for a second. You know, 166 00:09:22,679 --> 00:09:25,480 Speaker 3: there's a few benchmarks that we all know about. There's brand. 167 00:09:25,520 --> 00:09:28,360 Speaker 3: There's WTI, there's others, but those are the two that 168 00:09:28,440 --> 00:09:31,720 Speaker 3: we talk about. If we transpose this to chips for 169 00:09:31,800 --> 00:09:34,520 Speaker 3: a second, okay, we say you have an H one 170 00:09:34,600 --> 00:09:38,520 Speaker 3: hundred index. We did an episode of the podcast last week, 171 00:09:38,600 --> 00:09:41,600 Speaker 3: I think with the CEO of Servius, which is another 172 00:09:41,800 --> 00:09:45,360 Speaker 3: amazing company. Yep, yeah, but there are different Another type 173 00:09:45,360 --> 00:09:50,880 Speaker 3: of chip for inference is your assumption that these indices 174 00:09:50,920 --> 00:09:54,880 Speaker 3: are going to be close enough to the cost such 175 00:09:55,040 --> 00:09:58,680 Speaker 3: that if you're okay, I'm running inference maybe on some 176 00:09:58,840 --> 00:10:03,079 Speaker 3: service or TPUs or training whatever, some of these others, 177 00:10:03,280 --> 00:10:06,040 Speaker 3: that an H one hundred index will be good enough 178 00:10:06,160 --> 00:10:07,240 Speaker 3: as a hedging instrument. 179 00:10:07,440 --> 00:10:10,480 Speaker 4: This is the whole goal for me sitting here. Actually, right, 180 00:10:11,080 --> 00:10:14,880 Speaker 4: there's a meaningful every financial products in the functional reason 181 00:10:15,400 --> 00:10:19,960 Speaker 4: for commodity it is for hadging. Right, This speculation is great, 182 00:10:20,080 --> 00:10:23,480 Speaker 4: but really for people to hatch their relativity, to do 183 00:10:23,600 --> 00:10:27,920 Speaker 4: risk allocation, to do risk transfer, and then asset capitala location. 184 00:10:28,360 --> 00:10:31,160 Speaker 4: If we can't do what you said, then we fail 185 00:10:31,200 --> 00:10:34,040 Speaker 4: at our job. Right, So that's why we when all 186 00:10:34,080 --> 00:10:36,600 Speaker 4: the way back the way we cut we develop our 187 00:10:36,800 --> 00:10:39,840 Speaker 4: index model is not a simple math. It's not Hey, 188 00:10:39,880 --> 00:10:43,040 Speaker 4: you have two h one hundred to simple average, right, 189 00:10:43,080 --> 00:10:45,400 Speaker 4: because then then you compare Apple to oranges, the two 190 00:10:45,559 --> 00:10:48,000 Speaker 4: h one hundred can have different CPU different RAM defend 191 00:10:48,040 --> 00:10:51,840 Speaker 4: this differential location depend memory, bandwidth. You cannot do simple math. 192 00:10:52,200 --> 00:10:54,840 Speaker 4: What we do is we usually collect six months of 193 00:10:54,920 --> 00:10:58,880 Speaker 4: historical training data from over one hundred data sources, and 194 00:10:58,920 --> 00:11:03,559 Speaker 4: we see which factor to have the price differentiation. So 195 00:11:03,600 --> 00:11:07,240 Speaker 4: every day over one hundred and fifty thousand traded prices 196 00:11:07,400 --> 00:11:11,280 Speaker 4: in just our platform, and we normalize the traded prices 197 00:11:11,600 --> 00:11:15,560 Speaker 4: based on the different characteristics of the model itself and 198 00:11:15,720 --> 00:11:18,280 Speaker 4: normalized to a base case. And then we do the 199 00:11:18,320 --> 00:11:22,680 Speaker 4: math of settlement price calculation. Right, So then this price 200 00:11:22,880 --> 00:11:27,000 Speaker 4: will be highly correlated, ideally as much as it can 201 00:11:27,080 --> 00:11:30,760 Speaker 4: to the price you pay at a new cloud for example. However, 202 00:11:30,760 --> 00:11:33,040 Speaker 4: it won't be the same just like basis trading, right, 203 00:11:33,080 --> 00:11:35,760 Speaker 4: like every other commodity is a basis risk. We're helping 204 00:11:35,760 --> 00:11:38,199 Speaker 4: client colpeling the basis risk. So you know, hey, you 205 00:11:38,440 --> 00:11:42,079 Speaker 4: us east, you may be BIPs higher or two than 206 00:11:42,160 --> 00:11:47,920 Speaker 4: there's expectation a manageable correlation understanding of the indicies. 207 00:11:47,600 --> 00:11:50,280 Speaker 2: You mentioned volatility just then, I mean the reason people 208 00:11:50,320 --> 00:11:54,520 Speaker 2: need to hedg just because of volatility. Are you seeing 209 00:11:54,600 --> 00:11:57,440 Speaker 2: enough of that in GPU prices that like, this model 210 00:11:57,640 --> 00:12:00,520 Speaker 2: makes sense because if it's just a steady line up 211 00:12:00,800 --> 00:12:03,319 Speaker 2: or steady line down, like it's gonna be a kind 212 00:12:03,320 --> 00:12:04,080 Speaker 2: of boring market. 213 00:12:04,480 --> 00:12:07,520 Speaker 4: So it's interesting. So last year, when GP prices are 214 00:12:07,600 --> 00:12:09,640 Speaker 4: going down, the big conversation is why do you need 215 00:12:09,679 --> 00:12:13,040 Speaker 4: indicies for something price will always go down? And this 216 00:12:13,160 --> 00:12:15,040 Speaker 4: year is why do you want to inducease when price 217 00:12:15,040 --> 00:12:17,199 Speaker 4: always go up? Literally is all the questions. 218 00:12:17,240 --> 00:12:17,599 Speaker 1: I get it. 219 00:12:17,720 --> 00:12:20,199 Speaker 4: It's pretty fascinating. So when we will look at volatility, 220 00:12:20,240 --> 00:12:23,079 Speaker 4: we look at daily bomb volatility movement, not the price 221 00:12:23,160 --> 00:12:26,000 Speaker 4: up and down. Right, the daily volatility for eight one 222 00:12:26,120 --> 00:12:29,000 Speaker 4: hundred h one hundred is around twenty to thirty. So 223 00:12:29,000 --> 00:12:33,880 Speaker 4: it's a very healthy commodity volatility range. So I don't 224 00:12:33,960 --> 00:12:36,400 Speaker 4: manage volatility. It just happened to be the volatility that 225 00:12:36,520 --> 00:12:39,520 Speaker 4: can change. It's all because we normalize it. If you 226 00:12:39,559 --> 00:12:44,199 Speaker 4: look at each individual chip configuration at differential location, the 227 00:12:44,320 --> 00:12:48,119 Speaker 4: voltity are different. There's some chips with eight percent volatility, 228 00:12:48,280 --> 00:12:52,360 Speaker 4: some chips with over one hundred. Because normalization of indices, 229 00:12:52,600 --> 00:12:55,160 Speaker 4: you actually get very healthy twenty to thirty daily bomb. 230 00:13:11,000 --> 00:13:14,480 Speaker 3: I'm always fascinated by like, you know, we look at 231 00:13:14,520 --> 00:13:17,200 Speaker 3: the Bloomberg terminal, for example, and there's a price on 232 00:13:17,240 --> 00:13:20,240 Speaker 3: the screen and it's just there, and we started taking 233 00:13:20,280 --> 00:13:22,559 Speaker 3: for granted that like it had to come from somewhere, 234 00:13:22,960 --> 00:13:26,520 Speaker 3: and maybe some commodities have like a you know, there's 235 00:13:26,559 --> 00:13:29,480 Speaker 3: an existing exchange and a public price, and then there's 236 00:13:29,520 --> 00:13:32,560 Speaker 3: also a lot of commodities just bilateral traits. What is 237 00:13:32,600 --> 00:13:37,520 Speaker 3: the actual process by which you collect the most recent data? 238 00:13:37,520 --> 00:13:40,719 Speaker 3: So if you say, okay, an hour of h one 239 00:13:40,800 --> 00:13:44,400 Speaker 3: hundred usage costs x right, whatever it is right now, 240 00:13:44,800 --> 00:13:47,800 Speaker 3: how did you assemble that number? How did you gather 241 00:13:47,920 --> 00:13:50,439 Speaker 3: that information from, say the inference providers? 242 00:13:50,520 --> 00:13:53,600 Speaker 4: So it is a very can be lengthy, depends on 243 00:13:53,640 --> 00:13:58,040 Speaker 4: what data sources, the nature of GPU spawn markets conpictures 244 00:13:58,080 --> 00:13:59,439 Speaker 4: just one of them, and then many of my in 245 00:13:59,480 --> 00:14:05,640 Speaker 4: neoclaushyperscular marketplaces all have very different contras, size, durations, backs, 246 00:14:05,760 --> 00:14:09,480 Speaker 4: and their way to manage their data right, So it's 247 00:14:09,640 --> 00:14:14,160 Speaker 4: a lot of licensing, conversation, negotiation well and also context 248 00:14:14,240 --> 00:14:16,880 Speaker 4: love myself, I don't know, I was used for boomboard data, 249 00:14:17,440 --> 00:14:20,320 Speaker 4: so I was in data basiness for period of time. 250 00:14:20,560 --> 00:14:24,640 Speaker 4: So everything is pretty intuitive to me. It's very important 251 00:14:24,680 --> 00:14:27,520 Speaker 4: to get a variety of data sources, especially for computing. 252 00:14:27,600 --> 00:14:29,520 Speaker 3: Like do you call them up? Like so it's like, okay, 253 00:14:29,600 --> 00:14:32,120 Speaker 3: the price is different on the sum them up. 254 00:14:32,960 --> 00:14:36,280 Speaker 4: Well, you first of conversations say hey, I love what 255 00:14:36,320 --> 00:14:39,840 Speaker 4: you do. You're bring your cloud? Can I license your data? 256 00:14:39,920 --> 00:14:42,440 Speaker 4: And usually your feedback is what is in for me? 257 00:14:42,920 --> 00:14:43,080 Speaker 1: Right? 258 00:14:43,120 --> 00:14:45,480 Speaker 4: And I will tell our commercials And then your concern 259 00:14:45,560 --> 00:14:47,720 Speaker 4: could be, hey, you know, if I give you all 260 00:14:47,720 --> 00:14:50,360 Speaker 4: my data, I give my way all my secrets and 261 00:14:50,360 --> 00:14:53,560 Speaker 4: I will go through traditional licensing agreement. What can I disclose? 262 00:14:53,640 --> 00:14:55,280 Speaker 4: What I want from you? What I do not want 263 00:14:55,320 --> 00:14:57,960 Speaker 4: from you? What's the pipeline look like? Are you right? 264 00:14:58,040 --> 00:15:00,440 Speaker 4: You use a street market job? You are running my 265 00:15:00,600 --> 00:15:03,440 Speaker 4: API to yours? Are you running into mind? It's a 266 00:15:03,480 --> 00:15:07,240 Speaker 4: lot of conversations. It's actually pretty standard conversation. And right 267 00:15:07,320 --> 00:15:10,440 Speaker 4: now with eight million pricing points globally around two hundred 268 00:15:10,520 --> 00:15:14,200 Speaker 4: data sources, it's pretty much bau A lot. People will say, hey, 269 00:15:14,480 --> 00:15:17,520 Speaker 4: always bring up you can I have your data? It's 270 00:15:17,560 --> 00:15:18,320 Speaker 4: always my ending. 271 00:15:18,520 --> 00:15:22,920 Speaker 2: You know, we were talking about GPU indices and you're 272 00:15:22,960 --> 00:15:25,600 Speaker 2: not the only one doing GPU price indices for. 273 00:15:25,640 --> 00:15:26,560 Speaker 4: Sure, not anymore. 274 00:15:26,680 --> 00:15:28,600 Speaker 2: Yeah, not anymore. But when you look at some of 275 00:15:28,600 --> 00:15:32,600 Speaker 2: the other ones, like sometimes they show different numbers or 276 00:15:32,640 --> 00:15:37,400 Speaker 2: even different longer term trends, what accounts for the discrepancy there? 277 00:15:37,640 --> 00:15:40,200 Speaker 2: What are you doing differently or what are they doing differently? 278 00:15:40,240 --> 00:15:40,720 Speaker 2: I guess. 279 00:15:40,920 --> 00:15:43,480 Speaker 4: So I can't come on other people's mythology because I 280 00:15:43,480 --> 00:15:48,360 Speaker 4: actually don't know different raw data. Different mythology will eventually 281 00:15:48,400 --> 00:15:52,240 Speaker 4: draft different prices. So the way I would look at 282 00:15:52,280 --> 00:15:56,400 Speaker 4: this is, you know, it's always smart for anyone to 283 00:15:56,400 --> 00:15:59,640 Speaker 4: look at multiple data sources and then figure out what 284 00:15:59,800 --> 00:16:04,320 Speaker 4: is the actual decision you have to make, which datasps 285 00:16:04,320 --> 00:16:08,320 Speaker 4: do you trust? The market always volte whilst us things 286 00:16:08,320 --> 00:16:11,760 Speaker 4: start trading. The market will always gravitates what things actually 287 00:16:11,760 --> 00:16:15,480 Speaker 4: help them hatch? Right, if you easily manupulatable, if you 288 00:16:15,720 --> 00:16:18,760 Speaker 4: are not data source people acting actions, do you hatch? 289 00:16:19,200 --> 00:16:22,600 Speaker 4: There was the point a sideh one speculation, right, So 290 00:16:22,920 --> 00:16:24,920 Speaker 4: you know I'd love to say, I mean, I also 291 00:16:24,920 --> 00:16:27,360 Speaker 4: strongly believe what the best but again I will let 292 00:16:27,360 --> 00:16:29,280 Speaker 4: the market decide, which will happen very soon. 293 00:16:29,640 --> 00:16:33,040 Speaker 3: So of course, like yes, there's the economic rationale for 294 00:16:33,080 --> 00:16:36,440 Speaker 3: the existence of a hedging instrument, and we can understand 295 00:16:36,520 --> 00:16:39,440 Speaker 3: that someone who is an entity that from time that 296 00:16:39,520 --> 00:16:43,400 Speaker 3: needs compute their short implicitly short GPUs, they want to hedge, 297 00:16:43,680 --> 00:16:46,600 Speaker 3: et cetera. But the liquid markets also really do need 298 00:16:46,640 --> 00:16:49,680 Speaker 3: speculators and they need people betting on price. What are 299 00:16:49,720 --> 00:16:54,200 Speaker 3: you seeing right now in terms of traders or institutions, 300 00:16:54,200 --> 00:16:57,280 Speaker 3: et cetera, who economically can take both sides of the 301 00:16:57,320 --> 00:17:00,720 Speaker 3: trade and how active is this getting where it's just 302 00:17:00,760 --> 00:17:05,439 Speaker 3: a compute trading desk that is separate from their economic needs. 303 00:17:05,640 --> 00:17:08,040 Speaker 4: The conversation has been going on for a very long 304 00:17:08,080 --> 00:17:12,919 Speaker 4: time with various banks, various multi participants, speculators. They are 305 00:17:13,040 --> 00:17:16,200 Speaker 4: very excited. So some banks obviously have those both sides 306 00:17:16,200 --> 00:17:18,239 Speaker 4: of the trade right, so they can cross off some 307 00:17:18,280 --> 00:17:21,880 Speaker 4: positions internally. That's great. Always some they have to use 308 00:17:22,040 --> 00:17:25,399 Speaker 4: leverage external products. So that's where we come in. The 309 00:17:25,440 --> 00:17:28,919 Speaker 4: way I encourage them to do is I selfishly, I 310 00:17:28,960 --> 00:17:31,680 Speaker 4: want them to start trading desk and compute. The more 311 00:17:31,720 --> 00:17:34,840 Speaker 4: people trade, the better for me, right selfishly. The same time, 312 00:17:35,000 --> 00:17:38,560 Speaker 4: it is important for people understand GPU trading. It's not 313 00:17:38,760 --> 00:17:41,919 Speaker 4: like you can just move someone from oil electricity with 314 00:17:42,160 --> 00:17:46,000 Speaker 4: no background context jobing to GPU compute futures. There's a 315 00:17:46,000 --> 00:17:48,159 Speaker 4: lot of context where number one GPU it is not 316 00:17:48,200 --> 00:17:51,560 Speaker 4: homogenous product number two. You have to understand the use 317 00:17:51,600 --> 00:17:53,760 Speaker 4: cases of eight one hundred h one hundred right now, 318 00:17:53,800 --> 00:17:56,520 Speaker 4: they are not that correlated. Is that right? Maybe that's 319 00:17:56,520 --> 00:17:59,040 Speaker 4: not right. I don't know. There are use cases which 320 00:17:59,080 --> 00:18:01,800 Speaker 4: they're pretty separated, but maybe their use cases they can 321 00:18:01,880 --> 00:18:05,040 Speaker 4: be transferred and also their software layer to this. Right, 322 00:18:05,240 --> 00:18:07,560 Speaker 4: so right now you can art give sooner use cases 323 00:18:07,600 --> 00:18:11,080 Speaker 4: some large amount of models cannot be deployed and the 324 00:18:11,200 --> 00:18:14,560 Speaker 4: legacy chips, but doesn't mean six months later you cannot 325 00:18:14,560 --> 00:18:18,800 Speaker 4: do so. As the software layer compression model compression gets better, 326 00:18:18,840 --> 00:18:22,560 Speaker 4: optimization gets better, things can change. So really understand not 327 00:18:22,640 --> 00:18:26,560 Speaker 4: just the hardware configuration, this local supply demand curve for 328 00:18:26,640 --> 00:18:31,080 Speaker 4: the service self also software layer that's kind of critical, right, 329 00:18:31,160 --> 00:18:33,320 Speaker 4: that's really changed the supply demand curve and all the 330 00:18:33,359 --> 00:18:36,280 Speaker 4: way to the user behavior. So it's all it's going 331 00:18:36,320 --> 00:18:38,919 Speaker 4: to take some times as we have engaged with a 332 00:18:38,920 --> 00:18:41,520 Speaker 4: lot of participants make sure they have the right set up. 333 00:18:42,119 --> 00:18:44,280 Speaker 2: I have what is possibly a dumb question, but the 334 00:18:44,560 --> 00:18:48,320 Speaker 2: compute futures, how are those actually settled? Because I have 335 00:18:48,359 --> 00:18:51,400 Speaker 2: like images in my mind of taking physical delivery of 336 00:18:51,440 --> 00:18:53,280 Speaker 2: like maybe one of those big. 337 00:18:53,240 --> 00:18:59,240 Speaker 4: Okay, that'll be fun. So for the CME, futures will 338 00:18:59,280 --> 00:19:03,520 Speaker 4: be financially settled just like the traditional oil settlement price. 339 00:19:03,560 --> 00:19:06,560 Speaker 4: That priceis goes four contracts. Well obviously we do four 340 00:19:06,640 --> 00:19:09,119 Speaker 4: right now, I can fut change, but we always open 341 00:19:09,200 --> 00:19:11,960 Speaker 4: to do you know, physically deliver futures, especially given we 342 00:19:12,119 --> 00:19:15,800 Speaker 4: do have silica mark which is GPO benchmarking. So imagine 343 00:19:15,800 --> 00:19:18,200 Speaker 4: the future. You can do, Hey, I want to twenty 344 00:19:18,640 --> 00:19:22,879 Speaker 4: grade A B two hundred, this configuration, this shape of 345 00:19:23,280 --> 00:19:26,520 Speaker 4: servers in US East and then at the end we'll 346 00:19:26,560 --> 00:19:29,480 Speaker 4: get that. Well, usually API costs so you don't get 347 00:19:29,480 --> 00:19:31,560 Speaker 4: physical and it's not as coo as physically give you 348 00:19:31,600 --> 00:19:33,399 Speaker 4: a way way for but you get APIA costs. 349 00:19:33,400 --> 00:19:34,159 Speaker 2: One can dream. 350 00:19:34,680 --> 00:19:37,480 Speaker 3: How do you literally trade it is in like let's 351 00:19:37,480 --> 00:19:40,080 Speaker 3: say there's probably some very bright people in the room 352 00:19:40,520 --> 00:19:43,600 Speaker 3: now with an institution. When it's all listed in everything, 353 00:19:44,119 --> 00:19:46,120 Speaker 3: does it need to go through like a future's broker? 354 00:19:46,880 --> 00:19:48,760 Speaker 3: Is it like a could it be like a prediction 355 00:19:48,800 --> 00:19:50,359 Speaker 3: market if you just go to a website? Like what 356 00:19:50,480 --> 00:19:54,760 Speaker 3: is the actual How does someone actually get in this setting? 357 00:19:54,800 --> 00:19:57,040 Speaker 3: Aside whether they're sophisticated enough of whether they know what 358 00:19:57,040 --> 00:19:58,840 Speaker 3: they're doing a lot of people trade to have no 359 00:19:58,920 --> 00:20:01,800 Speaker 3: idea what they're doing. Yeah, setting all this side. Yes, 360 00:20:01,840 --> 00:20:04,199 Speaker 3: you know, only trade what you know. But like, what 361 00:20:04,600 --> 00:20:06,800 Speaker 3: is it through a prime broker? Like how will people 362 00:20:06,880 --> 00:20:08,600 Speaker 3: actually be able to participate. 363 00:20:08,119 --> 00:20:08,720 Speaker 2: In this market? 364 00:20:08,760 --> 00:20:11,359 Speaker 4: The beauty of CME is you can do the same 365 00:20:11,480 --> 00:20:15,120 Speaker 4: thing you're doing now treating semi products. Okay, the same 366 00:20:15,160 --> 00:20:19,880 Speaker 4: process and processing margin. That's why you get great margin optimization. Right, 367 00:20:20,440 --> 00:20:25,080 Speaker 4: everything is BAU it's no different. We don't have anything 368 00:20:25,160 --> 00:20:25,480 Speaker 4: right now. 369 00:20:25,680 --> 00:20:29,040 Speaker 3: So any commodities broker that someone has, they will be 370 00:20:29,119 --> 00:20:32,359 Speaker 3: able to on that platform. They will have access to these. 371 00:20:32,240 --> 00:20:35,240 Speaker 4: Instruments exactly right. Yeah, we make it easy for people. 372 00:20:35,600 --> 00:20:38,280 Speaker 2: Would you be upset if a prediction market set up 373 00:20:38,359 --> 00:20:41,280 Speaker 2: a GPU price contract of some sort with that into 374 00:20:41,320 --> 00:20:42,640 Speaker 2: your business? 375 00:20:42,680 --> 00:20:46,160 Speaker 4: Not at all. So we actually work with poly market. 376 00:20:46,840 --> 00:20:49,679 Speaker 4: Last year someone actually listed my product at polymarket with 377 00:20:50,200 --> 00:20:53,439 Speaker 4: my consent. It's pretty it's always start like that. And 378 00:20:53,480 --> 00:20:55,960 Speaker 4: then someone told me that, and we try to Polymarket 379 00:20:55,960 --> 00:20:57,399 Speaker 4: and say, hey, do you want to do something you 380 00:20:57,440 --> 00:21:02,520 Speaker 4: know more real? So we did. Fabruary settled and April 381 00:21:02,560 --> 00:21:06,400 Speaker 4: settled a few contracts on polymarket. I just with test 382 00:21:06,440 --> 00:21:10,760 Speaker 4: the water right obviously we're exclusively with CME right now. 383 00:21:11,280 --> 00:21:13,080 Speaker 4: But yeah, so I think obviously you have to do 384 00:21:13,080 --> 00:21:16,440 Speaker 4: it right, licensing, nonminal, pillared, it all the right things. Yeah, 385 00:21:17,000 --> 00:21:19,440 Speaker 4: you know, I don't. Mark can do whatever they want, 386 00:21:19,560 --> 00:21:21,920 Speaker 4: and then people will choose the best product for them 387 00:21:21,960 --> 00:21:22,320 Speaker 4: to use. 388 00:21:22,880 --> 00:21:27,560 Speaker 3: Sitting aside the financial instruments for the moment, would people 389 00:21:27,640 --> 00:21:30,919 Speaker 3: think about AI and they think about the use of GPUs, 390 00:21:31,480 --> 00:21:34,359 Speaker 3: they mostly still probably in their mind think of like 391 00:21:34,680 --> 00:21:39,480 Speaker 3: open AI and Thropic and Google basically, and that's kind 392 00:21:39,520 --> 00:21:43,680 Speaker 3: of it. But obviously, as you've stated, like the world 393 00:21:43,840 --> 00:21:47,480 Speaker 3: of entities that serve inference in some form or another 394 00:21:47,960 --> 00:21:51,280 Speaker 3: is much greater than these three companies that we talk about. 395 00:21:51,560 --> 00:21:53,480 Speaker 3: Talk to us a little bit more about what the 396 00:21:53,560 --> 00:21:58,240 Speaker 3: actual world of inference provision looks like outside of the 397 00:21:58,280 --> 00:21:59,880 Speaker 3: big household AI name. 398 00:22:00,160 --> 00:22:03,280 Speaker 4: So the ones you mentioned, they mostly are closed source 399 00:22:03,320 --> 00:22:05,479 Speaker 4: models as we call it, right, but they do have 400 00:22:05,520 --> 00:22:09,040 Speaker 4: some open source versions, but they're famous for their closed 401 00:22:09,040 --> 00:22:12,520 Speaker 4: source models. So we actually track three hundred open source 402 00:22:13,359 --> 00:22:17,200 Speaker 4: open weights closed source models globally. Upon pricing and consumption 403 00:22:17,440 --> 00:22:21,040 Speaker 4: point of view, it is really interesting if we have actually, 404 00:22:21,520 --> 00:22:24,840 Speaker 4: you know, we haven't really formally launched token in disease. 405 00:22:25,200 --> 00:22:27,800 Speaker 4: You can currently look at Bloomberg and it's on Bloomberg. 406 00:22:28,240 --> 00:22:33,680 Speaker 4: What's interesting is people are depends. It's all based on 407 00:22:33,720 --> 00:22:38,160 Speaker 4: your choices. Right now. The price actually doubled file indusees 408 00:22:38,160 --> 00:22:41,160 Speaker 4: from now from December first last year. It's like two 409 00:22:41,200 --> 00:22:44,520 Speaker 4: dollars twenty one dollar per million token. It's a mixture 410 00:22:44,560 --> 00:22:47,680 Speaker 4: of input up and token prices reach weighted by consumption 411 00:22:47,800 --> 00:22:52,400 Speaker 4: by buscut models. It's not here. This is a GPU unfortunately, which. 412 00:22:52,160 --> 00:22:55,320 Speaker 3: Are just well, since we have this specific chart up 413 00:22:55,440 --> 00:22:58,920 Speaker 3: right now, what is the y access in this church? 414 00:22:59,240 --> 00:23:03,680 Speaker 4: So you look yet that the per GPO power rental 415 00:23:03,760 --> 00:23:07,280 Speaker 4: rate on demand for three chips. The top one the 416 00:23:07,359 --> 00:23:10,080 Speaker 4: yellow line is B two hundred new cloud on demand 417 00:23:10,080 --> 00:23:13,440 Speaker 4: per flower. So it's a mouthful. The line the air 418 00:23:13,480 --> 00:23:16,080 Speaker 4: line is interesting, right. So every new chip we came 419 00:23:16,119 --> 00:23:19,760 Speaker 4: out based on historical data one hundred usually came out 420 00:23:19,760 --> 00:23:23,440 Speaker 4: to be high and then comes down as more supply 421 00:23:24,080 --> 00:23:27,240 Speaker 4: you know, came life, and then price came down and 422 00:23:27,240 --> 00:23:30,040 Speaker 4: then stabilizes. So that's the trend we have absorbed for 423 00:23:30,080 --> 00:23:32,680 Speaker 4: A one hundred and for H one hundred. So when 424 00:23:32,680 --> 00:23:34,640 Speaker 4: B two country came out, we pollished the data last 425 00:23:34,680 --> 00:23:38,240 Speaker 4: year at Bloomberg this early this year, the price was 426 00:23:38,320 --> 00:23:40,359 Speaker 4: high and it came down, which is kind of what 427 00:23:40,440 --> 00:23:43,920 Speaker 4: I expected. But the slope was less steep than I expected. 428 00:23:44,359 --> 00:23:47,639 Speaker 4: I was like, hm, that's interesting, the slope wasn't as steep, 429 00:23:48,000 --> 00:23:51,560 Speaker 4: and I'm quickly observed the price just came up and 430 00:23:51,600 --> 00:23:54,320 Speaker 4: now it's higher than the initial opened whatever you call that, 431 00:23:54,440 --> 00:23:57,119 Speaker 4: right launch prices. That shows your demands apply craft in 432 00:23:57,160 --> 00:24:01,239 Speaker 4: a different stage than whatever stage we had before. So 433 00:24:01,280 --> 00:24:04,080 Speaker 4: the A one the red line is H one hundred 434 00:24:04,600 --> 00:24:07,760 Speaker 4: neil cloud on demand per jubile power rate. So you 435 00:24:07,800 --> 00:24:09,800 Speaker 4: can see the price came down last year a little 436 00:24:09,800 --> 00:24:11,800 Speaker 4: bit sort about the skill so you don't see much, 437 00:24:11,840 --> 00:24:15,120 Speaker 4: but came down and came back up quite a bit. 438 00:24:15,520 --> 00:24:17,480 Speaker 4: I think the last three months came up to like 439 00:24:18,040 --> 00:24:21,720 Speaker 4: eight percent for AGE one hundred. The one hundred is older, 440 00:24:21,840 --> 00:24:25,800 Speaker 4: oldest chips among the three, right, they're pretty you know, 441 00:24:26,440 --> 00:24:28,760 Speaker 4: put don't you calm audit? At this point the price 442 00:24:28,840 --> 00:24:31,560 Speaker 4: came down, they stabilized, but the price came up about 443 00:24:31,600 --> 00:24:34,840 Speaker 4: ten to fifteen percent for the past three months. Remember 444 00:24:34,880 --> 00:24:37,040 Speaker 4: the A one hundred right there, they're not the latest 445 00:24:37,080 --> 00:24:40,200 Speaker 4: and greatest at all. So this also tells you to 446 00:24:40,200 --> 00:24:41,399 Speaker 4: supply deman croft shifting. 447 00:24:41,480 --> 00:24:43,280 Speaker 2: Oh yeah, actually that reminds me, could you talk to 448 00:24:43,400 --> 00:24:46,840 Speaker 2: us because you're doing refurbishment of chips as well, right, like, 449 00:24:46,880 --> 00:24:51,399 Speaker 2: which seems like challenging in many ways and kind of 450 00:24:51,440 --> 00:24:54,040 Speaker 2: reminds me a lot about like the sort of carbona 451 00:24:54,200 --> 00:24:59,080 Speaker 2: model of compute or something like that. How are you 452 00:24:59,160 --> 00:25:01,320 Speaker 2: actually doing this? Like how does that business work? 453 00:25:01,720 --> 00:25:05,000 Speaker 4: So this is cool in two different things. One is 454 00:25:06,280 --> 00:25:08,600 Speaker 4: for people come to conpect change saying that, hey, I 455 00:25:08,640 --> 00:25:11,440 Speaker 4: want to you know, it's Union Cloud provider. Right, if 456 00:25:11,480 --> 00:25:13,199 Speaker 4: you get a piece of land, you again a g 457 00:25:13,359 --> 00:25:17,000 Speaker 4: CO location, Great, congratulations? Then your option is number one. 458 00:25:17,040 --> 00:25:19,760 Speaker 4: Should I get the latest and greatest, the B three hundred, 459 00:25:19,840 --> 00:25:22,320 Speaker 4: the gbs the roubmen with a few months or do 460 00:25:22,359 --> 00:25:26,160 Speaker 4: you want to get refrabit trips and turn out maybe sooner? Right? Then, 461 00:25:26,200 --> 00:25:29,359 Speaker 4: to you, it's become our I calculation for the most part. Right, 462 00:25:29,680 --> 00:25:34,240 Speaker 4: what's you expected, you know, future revenue generation? What's your 463 00:25:34,240 --> 00:25:37,880 Speaker 4: residual VALUD calculation? How much you can purchase by right, 464 00:25:37,960 --> 00:25:42,560 Speaker 4: it's actually pretty simple cash flow based in our I calculation. 465 00:25:43,320 --> 00:25:47,040 Speaker 4: So the way we approach residual value and refurb trans 466 00:25:47,240 --> 00:25:50,199 Speaker 4: transaction is you know, based on our I you know 467 00:25:50,280 --> 00:25:53,360 Speaker 4: this is your potential break even look at h one hundred, right, 468 00:25:53,400 --> 00:25:55,560 Speaker 4: obviously you're not going to charge its high speed two hundred, 469 00:25:55,880 --> 00:25:58,760 Speaker 4: but your cost base is also lower, so you can 470 00:25:58,840 --> 00:26:01,600 Speaker 4: do the future. You're assume a few years of four 471 00:26:01,680 --> 00:26:04,800 Speaker 4: country's signing three years this kind of cash flowback. That's 472 00:26:04,880 --> 00:26:08,000 Speaker 4: your misgivell now, right, So we do that calculation with 473 00:26:08,040 --> 00:26:10,800 Speaker 4: people so they understand, Hey, what was the value supposed 474 00:26:10,800 --> 00:26:14,000 Speaker 4: to generate? And it was the treating the market prices 475 00:26:14,040 --> 00:26:17,280 Speaker 4: for refurbished or used GPUA and you have to test 476 00:26:17,320 --> 00:26:19,240 Speaker 4: you to make sure things works, and there's out of 477 00:26:19,280 --> 00:26:21,639 Speaker 4: nuances to that. But we helped go to the understanding 478 00:26:21,680 --> 00:26:24,080 Speaker 4: of the whole research of value and that's why the 479 00:26:24,080 --> 00:26:26,159 Speaker 4: whole bubble thinking about but go ahead, what. 480 00:26:26,160 --> 00:26:30,119 Speaker 3: Month was it last year when like everyone got really 481 00:26:30,160 --> 00:26:33,879 Speaker 3: obsessed with like the life span of chips. Remember, like 482 00:26:33,920 --> 00:26:37,720 Speaker 3: tweeted something about He's like, oh the lifespan, They're like right, 483 00:26:37,960 --> 00:26:40,159 Speaker 3: and everyone was free, spent like three weeks free and 484 00:26:40,200 --> 00:26:43,159 Speaker 3: then moved on from that conversation, right, like that was 485 00:26:43,680 --> 00:26:46,240 Speaker 3: what do we know about chip life spans? Are there 486 00:26:46,240 --> 00:26:49,959 Speaker 3: misconceptions out there about the how long yes, these can 487 00:26:49,960 --> 00:26:51,119 Speaker 3: get be productive. 488 00:26:51,320 --> 00:26:53,440 Speaker 4: I'm going in to view a few times, but I don't. 489 00:26:53,720 --> 00:26:56,040 Speaker 4: I mean, I'm not important. Still, I'm more important, but 490 00:26:56,040 --> 00:26:59,159 Speaker 4: back then that even less relevant. I was telling reporters. 491 00:26:59,280 --> 00:27:01,359 Speaker 4: I was like, look, I don't know what did that 492 00:27:01,400 --> 00:27:04,000 Speaker 4: you're looking at based on my I actually have blocks. 493 00:27:04,000 --> 00:27:06,000 Speaker 4: So my website, which is completely can to search for 494 00:27:06,040 --> 00:27:09,920 Speaker 4: it last year because of that conversation, I want you 495 00:27:10,200 --> 00:27:13,080 Speaker 4: currect later on the second year h one hundred residual 496 00:27:13,160 --> 00:27:16,200 Speaker 4: value of reseal value for refribit chips about eighty five 497 00:27:16,280 --> 00:27:18,640 Speaker 4: cents one dollar, So a year later you can sell 498 00:27:18,680 --> 00:27:21,280 Speaker 4: eighty five to one dollar. That's pretty good. I would 499 00:27:21,280 --> 00:27:23,760 Speaker 4: say the thirty is eighty four cents one dollar, and 500 00:27:24,040 --> 00:27:26,200 Speaker 4: I think my cards appreciate win more than that, right, 501 00:27:26,440 --> 00:27:28,800 Speaker 4: And I enjoined my car forboard every ten ten years, 502 00:27:29,160 --> 00:27:31,680 Speaker 4: So it's I had a data. But again I'm not 503 00:27:31,720 --> 00:27:34,000 Speaker 4: going to argue it against narrative, which is so. 504 00:27:34,080 --> 00:27:36,960 Speaker 3: But there was there's a fairly sived a decent drop 505 00:27:37,000 --> 00:27:39,440 Speaker 3: from year one to year two. That's what after that. 506 00:27:39,560 --> 00:27:40,680 Speaker 3: You see your general level. 507 00:27:40,840 --> 00:27:43,280 Speaker 4: That's November December analysis. Right now, it's a little different. 508 00:27:43,320 --> 00:27:45,480 Speaker 4: I haven't refreshed to study but our code is there. 509 00:27:45,480 --> 00:27:47,240 Speaker 4: If you're my data client, you can around my cold 510 00:27:47,320 --> 00:27:50,040 Speaker 4: you get a number right away. Another thing I want 511 00:27:50,080 --> 00:27:53,720 Speaker 4: to point out is l forties. They're like you know 512 00:27:53,720 --> 00:27:57,720 Speaker 4: the ogs right at that time, the people still use them. 513 00:27:57,960 --> 00:28:00,520 Speaker 4: They charge you hypercure charge your forties as p JB 514 00:28:00,680 --> 00:28:03,880 Speaker 4: per hour. So you know, I don't know about two 515 00:28:03,960 --> 00:28:06,520 Speaker 4: years where's the net number coming from? But I will 516 00:28:06,600 --> 00:28:08,200 Speaker 4: I will do that trade every single day. You sell 517 00:28:08,280 --> 00:28:10,119 Speaker 4: you're two years old at ten cents a dollars, I 518 00:28:10,160 --> 00:28:10,639 Speaker 4: will buy it. 519 00:28:11,080 --> 00:28:13,960 Speaker 2: There's a sort of big question looming in the background 520 00:28:13,960 --> 00:28:17,280 Speaker 2: of a lot of these discussions, which is the B question. 521 00:28:17,320 --> 00:28:19,879 Speaker 2: I guess whether or not we're in an AI bubble right, 522 00:28:19,960 --> 00:28:22,280 Speaker 2: And you sort of touched on it earlier. You have 523 00:28:22,400 --> 00:28:26,960 Speaker 2: all this granular data on how people are actually using compute, 524 00:28:26,960 --> 00:28:31,560 Speaker 2: GPU prices, all of that. What what's your take on 525 00:28:31,640 --> 00:28:32,359 Speaker 2: the big question? 526 00:28:32,840 --> 00:28:35,360 Speaker 4: So as an index provider, I cannot give any full 527 00:28:35,359 --> 00:28:40,320 Speaker 4: of guidance, no blamer, nor do I know, right being fairness, 528 00:28:40,320 --> 00:28:41,920 Speaker 4: I know what do I know? So the way I 529 00:28:41,960 --> 00:28:44,800 Speaker 4: look at is we had defined a bubble right, so 530 00:28:44,840 --> 00:28:47,080 Speaker 4: you look at style bubble right and the nest that 531 00:28:47,200 --> 00:28:49,040 Speaker 4: showed up to two hundred percent and came back down 532 00:28:49,160 --> 00:28:53,080 Speaker 4: eighty four percent whatever back. Then that's it's a bubble. Right. 533 00:28:53,480 --> 00:28:56,280 Speaker 4: The way I look at bubble is can is your valiation? 534 00:28:57,600 --> 00:29:03,040 Speaker 4: Can your future cash flows support today's valuation of yours? Right? 535 00:29:03,600 --> 00:29:06,200 Speaker 4: So then I'm not talking about like opening and everyone 536 00:29:06,200 --> 00:29:08,680 Speaker 4: else valuation. I'm not ob busy. I don't. I don't. 537 00:29:08,720 --> 00:29:12,000 Speaker 4: I don't understand that process. The way I look at 538 00:29:12,080 --> 00:29:16,120 Speaker 4: GPUs the machines, it's actually pretty simple. Look at a 539 00:29:16,160 --> 00:29:19,320 Speaker 4: future cash flow of your forward contracts and then you 540 00:29:19,440 --> 00:29:22,880 Speaker 4: discounted back. Can you get money back for the price 541 00:29:22,960 --> 00:29:28,080 Speaker 4: you pay? Right, It's actually pretty straightforward for the machine level. Right, 542 00:29:28,560 --> 00:29:30,600 Speaker 4: But to your point, right, you can say, hey, what 543 00:29:30,720 --> 00:29:35,240 Speaker 4: happened if demand jobbed? No one gonna use your whatever 544 00:29:35,440 --> 00:29:38,480 Speaker 4: things you have. But remember the four contract is a 545 00:29:38,560 --> 00:29:42,400 Speaker 4: signed contract. If you have that, you can't know obviously 546 00:29:42,440 --> 00:29:45,960 Speaker 4: if you have things, the biggest concern is people have 547 00:29:46,400 --> 00:29:50,960 Speaker 4: over built. You've overbuilt, then by theory, then all your 548 00:29:51,040 --> 00:29:54,760 Speaker 4: prices will calmed down because it over supply of the market. Right, 549 00:29:54,800 --> 00:29:58,120 Speaker 4: So then you talk about supply demain equilibrium. How do 550 00:29:58,160 --> 00:30:02,360 Speaker 4: we know about future demand of GPUs? Right? I don't 551 00:30:02,400 --> 00:30:05,640 Speaker 4: know that everyone's guess is better than mine. Probably the 552 00:30:05,680 --> 00:30:08,200 Speaker 4: way I look at it is not. It's not that 553 00:30:08,280 --> 00:30:11,800 Speaker 4: easy to bring any GP online, right you can use 554 00:30:11,920 --> 00:30:13,920 Speaker 4: you hear all those side big deals in the bill, 555 00:30:14,080 --> 00:30:17,520 Speaker 4: twenty five million dollars invested, but they don't translate to 556 00:30:17,600 --> 00:30:22,280 Speaker 4: immediate GPU availability right in the servers which you have 557 00:30:22,320 --> 00:30:25,720 Speaker 4: to be waylisted if you buy brand new stuff co location, 558 00:30:26,080 --> 00:30:30,160 Speaker 4: you need optic fiber, so there's a lot of unfortunately 559 00:30:30,800 --> 00:30:31,920 Speaker 4: STAR has to be aligned. 560 00:30:32,080 --> 00:30:35,280 Speaker 2: Wait, but people can default on contracts, right, so even 561 00:30:35,320 --> 00:30:37,560 Speaker 2: if you have a long term contract signed like that 562 00:30:37,960 --> 00:30:41,800 Speaker 2: could not work out. Could you envision like credit default 563 00:30:41,920 --> 00:30:45,520 Speaker 2: swaps or something in the compute market like so. 564 00:30:45,400 --> 00:30:48,200 Speaker 4: That happens in every other markets, right, aman market if 565 00:30:48,200 --> 00:30:50,840 Speaker 4: you do OTC trade, you have the rice some monual 566 00:30:50,880 --> 00:30:53,800 Speaker 4: defon you. Doesn't matter who they are, right, So there's 567 00:30:53,960 --> 00:30:57,240 Speaker 4: vary There's a lot of mechanism to hatch that. The 568 00:30:57,320 --> 00:31:00,960 Speaker 4: things you cannot hatch is GPU cost write the price 569 00:31:01,000 --> 00:31:04,040 Speaker 4: you entered. So that's something exactly will seemmy futures for 570 00:31:04,640 --> 00:31:08,480 Speaker 4: you can have a transparency the liquidity and then the 571 00:31:08,600 --> 00:31:11,320 Speaker 4: easiness of treating it out and had your position. 572 00:31:11,720 --> 00:31:14,560 Speaker 3: Carmen Lee, thank you so much for joining us. 573 00:31:29,680 --> 00:31:33,040 Speaker 2: That was our conversation with Carmen Lee of Compute Exchange 574 00:31:33,080 --> 00:31:35,800 Speaker 2: and Silicon Data, recorded live at our New York show. 575 00:31:36,080 --> 00:31:39,200 Speaker 2: I'm Tracy Alloway. You can follow me at Tracy Alloway. 576 00:31:38,880 --> 00:31:41,720 Speaker 3: And I'm Joe Wisenthal. You can follow me at the Stalwart. 577 00:31:42,160 --> 00:31:45,000 Speaker 3: Follow our guest Carmen Lee at Carmen Lee. Follow our 578 00:31:45,000 --> 00:31:48,960 Speaker 3: producers Carmen Rodriguez at Carmen armand dash Ol Bennett at Dashbot, 579 00:31:49,320 --> 00:31:52,320 Speaker 3: kil Brooks at Kile Brooks and Kevin Lozano at Kevin 580 00:31:52,360 --> 00:31:54,920 Speaker 3: Lloyd Lozano. And from our odd Laws content, go to 581 00:31:54,920 --> 00:31:57,880 Speaker 3: Bloomberg dot com slash odd Lads, where the daily newsletter 582 00:31:57,880 --> 00:32:00,000 Speaker 3: and all of our episodes, and you could shout up 583 00:32:00,280 --> 00:32:03,560 Speaker 3: all these topics twenty four seven in our discord Discord 584 00:32:03,640 --> 00:32:05,520 Speaker 3: dot gg slash odd Lots And. 585 00:32:05,480 --> 00:32:07,400 Speaker 2: If you enjoy odd Lots, if you like it when 586 00:32:07,440 --> 00:32:09,640 Speaker 2: we do these live shows and talk about the birth 587 00:32:09,680 --> 00:32:11,960 Speaker 2: of a new market, then please leave us a positive 588 00:32:12,000 --> 00:32:15,000 Speaker 2: review on your favorite podcast platform. And remember, if you 589 00:32:15,040 --> 00:32:17,440 Speaker 2: are a Bloomberg subscriber, you can listen to all of 590 00:32:17,440 --> 00:32:20,200 Speaker 2: our episodes absolutely id free. All you need to do 591 00:32:20,280 --> 00:32:22,880 Speaker 2: is find the Bloomberg channel on Apple Podcasts and follow 592 00:32:22,920 --> 00:32:25,040 Speaker 2: the instructions there. Thanks for listening.