WEBVTT - Carmen Li's Plan to Build a Futures Market for Compute 0:00:02.720 --> 0:00:15.880 Bloomberg Audio Studios, Podcasts, Radio News. 0:00:18.640 --> 0:00:21.480 Hello and welcome to another episode of the Out Thoughts podcast. 0:00:21.560 --> 0:00:22.840 I'm Tracy Alloway. 0:00:22.640 --> 0:00:23.200 And I'm Jill. 0:00:23.400 --> 0:00:27.480 Wasn't thal so Joe. We're still continuing our series. Recorded 0:00:27.520 --> 0:00:29.440 from the live show in New York. We had a 0:00:29.440 --> 0:00:32.320 bunch of great conversations. A couple of them were building 0:00:32.360 --> 0:00:35.479 off of discussions that we had had previously, and one 0:00:35.520 --> 0:00:39.160 of those discussions was in Chicago at another live show 0:00:39.280 --> 0:00:42.279 about six or seven months ago. Back in October, we 0:00:42.320 --> 0:00:46.280 spoke with Don Wilson of DRW about the trading environment, 0:00:46.400 --> 0:00:48.479 but also about his new venture. 0:00:48.920 --> 0:00:51.839 Right and so his new venture is one that actually 0:00:51.880 --> 0:00:54.440 there's quite a bit of competition in and quite of 0:00:54.560 --> 0:00:59.080 excitement in, and it's essentially like okay, GPUs. We know 0:00:59.080 --> 0:01:02.560 they're very important for the AI boom, et cetera. The 0:01:02.640 --> 0:01:07.240 question is can GPU capacity, which is scarce, can it 0:01:07.319 --> 0:01:11.920 become a tradable commodity such that I can buy futures 0:01:11.959 --> 0:01:15.479 to lock in my price of access to compute power. 0:01:15.959 --> 0:01:20.000 Could I resell those futures? Will there be speculators speculating 0:01:20.000 --> 0:01:21.759 on the upper down price of like an H one 0:01:21.840 --> 0:01:24.120 hundred running an H one hundred in video chip for 0:01:24.160 --> 0:01:26.200 an hour. This is a big question. We know there's 0:01:26.200 --> 0:01:28.960 a lot of interest in the actual compute, but whether 0:01:29.000 --> 0:01:32.800 there's interested in compute futures. It's tradable instruments is very TVD. 0:01:33.000 --> 0:01:36.000 Yeah, And the analogy that everyone always uses is compute 0:01:36.080 --> 0:01:38.240 is the new oil, right, so why can't it have 0:01:38.640 --> 0:01:41.120 you know, a market structure that looks somewhat like the 0:01:41.120 --> 0:01:45.520 oil market. And there are challenges. Fungibility is a big one, 0:01:45.560 --> 0:01:48.800 like one chip might not necessarily be equal to another 0:01:48.880 --> 0:01:49.880 chip or one ship. 0:01:50.040 --> 0:01:52.880 The same chip at one data center might equal to 0:01:52.880 --> 0:01:55.160 the same chip at a different data center exactly. 0:01:55.200 --> 0:01:58.760 And so even if you're not interested in AI, what 0:01:58.840 --> 0:02:02.400 I say here is like the market structure questions and 0:02:02.480 --> 0:02:06.360 the idea of building an entirely new market is really 0:02:06.400 --> 0:02:08.680 fascinating to me, and I think others will find it 0:02:08.720 --> 0:02:11.320 interesting too. And we really do have the perfect guest. 0:02:11.360 --> 0:02:14.000 We're speaking with Carmen Lee. She is the CEO of 0:02:14.080 --> 0:02:17.360 Compute Exchange and Silicon Data. These are the two companies 0:02:17.600 --> 0:02:21.400 that Wilson is invested in, and they've already announced that 0:02:21.440 --> 0:02:25.080 they're doing futures with the CME. So really the perfect 0:02:25.160 --> 0:02:28.760 person to speak to So take a listen. Last October 0:02:28.840 --> 0:02:32.440 we spoke with Don Wilson of DRW fame and he 0:02:32.639 --> 0:02:35.680 was talking to us about his new project, which was 0:02:35.720 --> 0:02:39.720 basically building out this compute exchange. Now we're here with 0:02:39.760 --> 0:02:42.720 you six months later. You're actually the one leading it. 0:02:43.480 --> 0:02:46.320 How far are you in this endeavor? And remind us 0:02:46.520 --> 0:02:48.160 what exactly are you trying to do here? 0:02:48.440 --> 0:02:52.480 Yeah, so thank you for the the quizing one great audience. 0:02:52.919 --> 0:02:54.919 Before I do that, I actually going to call back 0:02:55.120 --> 0:02:58.840 to six months ago in the DOM podcast you did. 0:02:59.120 --> 0:03:02.920 You asked them question, what if compute prices keep going 0:03:03.000 --> 0:03:07.519 to go up? At the time September October, compute prices 0:03:07.560 --> 0:03:11.040 were going down across our chips. Now see what happened. 0:03:11.240 --> 0:03:13.239 I think you called it. I think you called the 0:03:13.480 --> 0:03:16.040 market called it. So I'm the fund of CEO for 0:03:16.120 --> 0:03:19.679 Cilicond Data, so that's the index provider for GPU indcees. 0:03:20.360 --> 0:03:23.800 We recently announced partnership with CME, so we all been 0:03:23.840 --> 0:03:28.760 launching gp futurning options ME in a couple of months, 0:03:29.240 --> 0:03:33.280 pending the FDIC approval. Obviously that's quite exciting. We've been 0:03:33.320 --> 0:03:36.520 working on GPU indices for past two and a half years, 0:03:37.160 --> 0:03:40.000 starting twenty twenty four April, so it's been a while, 0:03:40.480 --> 0:03:44.320 and we launched for world's first GPU inducees at Bloomberg 0:03:44.360 --> 0:03:48.200 Terminal in twenty twenty five. A year later we launched 0:03:48.200 --> 0:03:51.600 the partnership with CMME, so it's quite exciting. Separately, I 0:03:51.640 --> 0:03:53.720 heard you mentioned Compute Change before, so thank you for 0:03:53.760 --> 0:03:56.840 doing AGIL. I'm the CEO for compute Change, which is 0:03:56.880 --> 0:04:01.200 sport marketplace for GPO procurement. So we do reserve contracts 0:04:01.640 --> 0:04:04.480 for contracts as well as ree for pitch contracts. 0:04:04.720 --> 0:04:08.200 Let's talk about the variety of options that we have 0:04:08.320 --> 0:04:12.280 to financialize compute and so forth. So this, I mean, 0:04:12.280 --> 0:04:14.520 this came up in our conversation. The first conversation we 0:04:14.600 --> 0:04:17.120 had with we had with Ian Dunning. Who is the 0:04:17.160 --> 0:04:21.159 type of buyer who would want to buy compute on 0:04:21.200 --> 0:04:24.680 a spot market because right you talk about typically we 0:04:24.720 --> 0:04:27.839 think it's like these multi year contracts where some entity 0:04:28.200 --> 0:04:30.320 enters into a contract with a data center or a 0:04:30.360 --> 0:04:33.200 new cloud whatever, and this they have this for a while. 0:04:33.400 --> 0:04:36.080 So who is the market, who is the buyer or 0:04:36.120 --> 0:04:39.279 the user of these instruments that might want to buy 0:04:39.600 --> 0:04:43.680 spot compute or very short term short dated compute futures. 0:04:43.760 --> 0:04:46.640 It's a great question. So the compute market right now, 0:04:46.760 --> 0:04:50.880 for comput change, we have all our provider mostly are 0:04:50.920 --> 0:04:54.479 neo clouds around the world. It's one side. Another site 0:04:54.600 --> 0:04:58.479 is a big variety AI start up. So even though 0:04:58.480 --> 0:05:02.440 they are start up, millions of dollars on GPUs already. 0:05:02.880 --> 0:05:06.640 There are enterprises who are traditional businesses, but they are 0:05:06.760 --> 0:05:09.240 needing a note to notes, a few service here and 0:05:09.279 --> 0:05:13.200 there for their inferencing or I don't know, other deployment needs. 0:05:13.560 --> 0:05:17.680 They are providers. They are influencing providers right they don't 0:05:17.680 --> 0:05:21.960 own GPUs, but they provide open source, open weights model 0:05:22.000 --> 0:05:25.600 support for other use cases. So what see big variety. 0:05:26.160 --> 0:05:31.200 Most North American firms they do a variety of combination contracts. 0:05:31.839 --> 0:05:35.640 Obviously on demand give you the most flexibility. You don't 0:05:35.680 --> 0:05:38.360 pay and you don't use it. However, you're also at 0:05:38.360 --> 0:05:40.960 a mercy of defense. Apply curve at a given time, 0:05:41.279 --> 0:05:44.039 so translate to your price can go from three dollars 0:05:44.040 --> 0:05:47.000 to six to nine depends on demand supply curve shifting, 0:05:47.440 --> 0:05:51.880 so that doesn't help when you can have a predictable margin. 0:05:52.560 --> 0:05:56.360 And also in terms of scarce, you're not guaranteed to 0:05:56.400 --> 0:06:00.360 feel GPU resources for next hour or next month. So 0:06:00.440 --> 0:06:02.680 you see a lot of people shifting from on demand 0:06:02.720 --> 0:06:06.320 to reserve even four contracts right, so full contracts you 0:06:06.360 --> 0:06:10.600 back the lock in a deliverables for next whatever month, right, 0:06:10.640 --> 0:06:13.080 starting in September maybe or starting no member of forbare 0:06:13.120 --> 0:06:16.760 weapons right. So this all comes because of market condition. 0:06:17.120 --> 0:06:21.599 So computing cover that the physical GPO procurement also token, 0:06:21.680 --> 0:06:23.680 so we love to talk about token as well. On 0:06:23.760 --> 0:06:27.760 flip side, who's going to use the future options can 0:06:27.839 --> 0:06:29.080 be a similar set of people. 0:06:29.320 --> 0:06:29.520 Right. 0:06:29.920 --> 0:06:32.799 You look at oil market, which we all love double 0:06:32.880 --> 0:06:36.159 TM brand. Right, the people use double TM brand. A 0:06:36.240 --> 0:06:39.880 lot of them are naturally long oil. So the shells 0:06:40.080 --> 0:06:44.120 ran the producers. They need to hatch your revenue volatility 0:06:44.640 --> 0:06:48.360 by shorting futures or port options if you're naturally short 0:06:48.400 --> 0:06:52.680 oil americandline right, they want to control their cost volatility. 0:06:53.080 --> 0:06:55.960 They want to obviously use future options as well. 0:06:56.680 --> 0:06:57.120 Simple to the. 0:06:57.160 --> 0:07:01.520 Compute your new cloud or you're at anyone have the servers. 0:07:01.800 --> 0:07:04.560 Ideally you want to have predictable revenue streams. 0:07:04.240 --> 0:07:07.040 So the neo cloud would be the shell on this example. 0:07:06.760 --> 0:07:11.720 Exactly, you have GPS, right, all the banks were GPUs 0:07:11.760 --> 0:07:15.720 on your balance sheet right your long GPUs, then naturally 0:07:15.760 --> 0:07:18.800 you want to make sure revenue right is stable to 0:07:18.840 --> 0:07:21.040 certain degree, and then you want to use future to 0:07:21.040 --> 0:07:23.800 do so. If you are naturally short GPU, which is 0:07:23.840 --> 0:07:25.880 everybody in this room, amas you tell me you have 0:07:25.920 --> 0:07:30.120 GPUs right, then you depends how much you use. If 0:07:30.120 --> 0:07:32.480 you want to control your cost volatility, you want to 0:07:32.520 --> 0:07:33.680 use future to hatch as well. 0:07:34.080 --> 0:07:38.280 Just on the compute exchange side of things, if someone 0:07:38.360 --> 0:07:41.120 is buying like off the spot market, how do you 0:07:41.160 --> 0:07:44.440 guarantee I'm not sure quality is the right word for this, 0:07:44.560 --> 0:07:47.840 but how do you guarantee they're getting what they expect? 0:07:47.960 --> 0:07:50.640 This is a great question. So I canna flip to 0:07:50.760 --> 0:07:55.200 slide if you don't mind. So I usually don't let 0:07:55.240 --> 0:07:57.960 use slides, but this time because you mentioned really good questions. 0:07:58.000 --> 0:08:01.600 So we actually call it GPU lottery. So we published 0:08:01.600 --> 0:08:06.320 a paper earlier this year at GPGPO conference with Jefferson 0:08:06.400 --> 0:08:09.880 Lab on GPO performances. Well actually, so well we can 0:08:09.960 --> 0:08:12.360 have your quit link, you know, to the audience later on. 0:08:12.600 --> 0:08:14.240 We actually this is a one hundred by the way, 0:08:14.360 --> 0:08:16.280 I know we didn't put tang on the day. Is 0:08:16.440 --> 0:08:20.080 a one hundred forty gigabytes memory bandwidth. We prove there's 0:08:20.120 --> 0:08:24.160 thirty eight percent performance variants for the same trip, and 0:08:24.320 --> 0:08:27.880 we decompose it into the chip self, into provider and 0:08:28.040 --> 0:08:31.720 inter provider. And there's many reasons for that, right and 0:08:31.840 --> 0:08:34.719 to your point, you can you don't know until you 0:08:34.800 --> 0:08:39.280 get your GPS. We have a PLAT for GPU. CARFAX 0:08:39.360 --> 0:08:41.600 for GPU depends how to look at it, so we 0:08:41.800 --> 0:08:45.720 in compute change. You actually verify the GPU before delivered 0:08:45.720 --> 0:08:48.320 to you. So basically you are you are. You CANFQ 0:08:48.720 --> 0:08:50.720 for say, hey, I one A two hundred BI two 0:08:50.800 --> 0:08:54.160 hundred notes. Obviously it will give you specs back and 0:08:54.200 --> 0:08:58.360 the commercial back same time, independently verify the performances on flops, 0:08:58.559 --> 0:09:04.079 memory bandwidth totals another information as ACE and other things. 0:09:04.320 --> 0:09:07.079 And as a user you can decide is price your 0:09:07.080 --> 0:09:10.400 most important criteria? Maybe it is, or maybe you're willing 0:09:10.440 --> 0:09:13.319 to pay a premium for dual location or the performances 0:09:13.400 --> 0:09:17.040 that you care more about on latency. Right, we believe 0:09:17.080 --> 0:09:18.959 gave people the option and transparency it is the most 0:09:19.000 --> 0:09:19.559 important thing. 0:09:19.840 --> 0:09:22.600 Let's stick with the oil analogy for a second. You know, 0:09:22.679 --> 0:09:25.480 there's a few benchmarks that we all know about. There's brand. 0:09:25.520 --> 0:09:28.360 There's WTI, there's others, but those are the two that 0:09:28.440 --> 0:09:31.720 we talk about. If we transpose this to chips for 0:09:31.800 --> 0:09:34.520 a second, okay, we say you have an H one 0:09:34.600 --> 0:09:38.520 hundred index. We did an episode of the podcast last week, 0:09:38.600 --> 0:09:41.600 I think with the CEO of Servius, which is another 0:09:41.800 --> 0:09:45.360 amazing company. Yep, yeah, but there are different Another type 0:09:45.360 --> 0:09:50.880 of chip for inference is your assumption that these indices 0:09:50.920 --> 0:09:54.880 are going to be close enough to the cost such 0:09:55.040 --> 0:09:58.680 that if you're okay, I'm running inference maybe on some 0:09:58.840 --> 0:10:03.079 service or TPUs or training whatever, some of these others, 0:10:03.280 --> 0:10:06.040 that an H one hundred index will be good enough 0:10:06.160 --> 0:10:07.240 as a hedging instrument. 0:10:07.440 --> 0:10:10.480 This is the whole goal for me sitting here. Actually, right, 0:10:11.080 --> 0:10:14.880 there's a meaningful every financial products in the functional reason 0:10:15.400 --> 0:10:19.960 for commodity it is for hadging. Right, This speculation is great, 0:10:20.080 --> 0:10:23.480 but really for people to hatch their relativity, to do 0:10:23.600 --> 0:10:27.920 risk allocation, to do risk transfer, and then asset capitala location. 0:10:28.360 --> 0:10:31.160 If we can't do what you said, then we fail 0:10:31.200 --> 0:10:34.040 at our job. Right, So that's why we when all 0:10:34.080 --> 0:10:36.600 the way back the way we cut we develop our 0:10:36.800 --> 0:10:39.840 index model is not a simple math. It's not Hey, 0:10:39.880 --> 0:10:43.040 you have two h one hundred to simple average, right, 0:10:43.080 --> 0:10:45.400 because then then you compare Apple to oranges, the two 0:10:45.559 --> 0:10:48.000 h one hundred can have different CPU different RAM defend 0:10:48.040 --> 0:10:51.840 this differential location depend memory, bandwidth. You cannot do simple math. 0:10:52.200 --> 0:10:54.840 What we do is we usually collect six months of 0:10:54.920 --> 0:10:58.880 historical training data from over one hundred data sources, and 0:10:58.920 --> 0:11:03.559 we see which factor to have the price differentiation. So 0:11:03.600 --> 0:11:07.240 every day over one hundred and fifty thousand traded prices 0:11:07.400 --> 0:11:11.280 in just our platform, and we normalize the traded prices 0:11:11.600 --> 0:11:15.560 based on the different characteristics of the model itself and 0:11:15.720 --> 0:11:18.280 normalized to a base case. And then we do the 0:11:18.320 --> 0:11:22.680 math of settlement price calculation. Right, So then this price 0:11:22.880 --> 0:11:27.000 will be highly correlated, ideally as much as it can 0:11:27.080 --> 0:11:30.760 to the price you pay at a new cloud for example. However, 0:11:30.760 --> 0:11:33.040 it won't be the same just like basis trading, right, 0:11:33.080 --> 0:11:35.760 like every other commodity is a basis risk. We're helping 0:11:35.760 --> 0:11:38.199 client colpeling the basis risk. So you know, hey, you 0:11:38.440 --> 0:11:42.079 us east, you may be BIPs higher or two than 0:11:42.160 --> 0:11:47.920 there's expectation a manageable correlation understanding of the indicies. 0:11:47.600 --> 0:11:50.280 You mentioned volatility just then, I mean the reason people 0:11:50.320 --> 0:11:54.520 need to hedg just because of volatility. Are you seeing 0:11:54.600 --> 0:11:57.440 enough of that in GPU prices that like, this model 0:11:57.640 --> 0:12:00.520 makes sense because if it's just a steady line up 0:12:00.800 --> 0:12:03.319 or steady line down, like it's gonna be a kind 0:12:03.320 --> 0:12:04.080 of boring market. 0:12:04.480 --> 0:12:07.520 So it's interesting. So last year, when GP prices are 0:12:07.600 --> 0:12:09.640 going down, the big conversation is why do you need 0:12:09.679 --> 0:12:13.040 indicies for something price will always go down? And this 0:12:13.160 --> 0:12:15.040 year is why do you want to inducease when price 0:12:15.040 --> 0:12:17.199 always go up? Literally is all the questions. 0:12:17.240 --> 0:12:17.599 I get it. 0:12:17.720 --> 0:12:20.199 It's pretty fascinating. So when we will look at volatility, 0:12:20.240 --> 0:12:23.079 we look at daily bomb volatility movement, not the price 0:12:23.160 --> 0:12:26.000 up and down. Right, the daily volatility for eight one 0:12:26.120 --> 0:12:29.000 hundred h one hundred is around twenty to thirty. So 0:12:29.000 --> 0:12:33.880 it's a very healthy commodity volatility range. So I don't 0:12:33.960 --> 0:12:36.400 manage volatility. It just happened to be the volatility that 0:12:36.520 --> 0:12:39.520 can change. It's all because we normalize it. If you 0:12:39.559 --> 0:12:44.199 look at each individual chip configuration at differential location, the 0:12:44.320 --> 0:12:48.119 voltity are different. There's some chips with eight percent volatility, 0:12:48.280 --> 0:12:52.360 some chips with over one hundred. Because normalization of indices, 0:12:52.600 --> 0:12:55.160 you actually get very healthy twenty to thirty daily bomb. 0:13:11.000 --> 0:13:14.480 I'm always fascinated by like, you know, we look at 0:13:14.520 --> 0:13:17.200 the Bloomberg terminal, for example, and there's a price on 0:13:17.240 --> 0:13:20.240 the screen and it's just there, and we started taking 0:13:20.280 --> 0:13:22.559 for granted that like it had to come from somewhere, 0:13:22.960 --> 0:13:26.520 and maybe some commodities have like a you know, there's 0:13:26.559 --> 0:13:29.480 an existing exchange and a public price, and then there's 0:13:29.520 --> 0:13:32.560 also a lot of commodities just bilateral traits. What is 0:13:32.600 --> 0:13:37.520 the actual process by which you collect the most recent data? 0:13:37.520 --> 0:13:40.719 So if you say, okay, an hour of h one 0:13:40.800 --> 0:13:44.400 hundred usage costs x right, whatever it is right now, 0:13:44.800 --> 0:13:47.800 how did you assemble that number? How did you gather 0:13:47.920 --> 0:13:50.439 that information from, say the inference providers? 0:13:50.520 --> 0:13:53.600 So it is a very can be lengthy, depends on 0:13:53.640 --> 0:13:58.040 what data sources, the nature of GPU spawn markets conpictures 0:13:58.080 --> 0:13:59.439 just one of them, and then many of my in 0:13:59.480 --> 0:14:05.640 neoclaushyperscular marketplaces all have very different contras, size, durations, backs, 0:14:05.760 --> 0:14:09.480 and their way to manage their data right, So it's 0:14:09.640 --> 0:14:14.160 a lot of licensing, conversation, negotiation well and also context 0:14:14.240 --> 0:14:16.880 love myself, I don't know, I was used for boomboard data, 0:14:17.440 --> 0:14:20.320 so I was in data basiness for period of time. 0:14:20.560 --> 0:14:24.640 So everything is pretty intuitive to me. It's very important 0:14:24.680 --> 0:14:27.520 to get a variety of data sources, especially for computing. 0:14:27.600 --> 0:14:29.520 Like do you call them up? Like so it's like, okay, 0:14:29.600 --> 0:14:32.120 the price is different on the sum them up. 0:14:32.960 --> 0:14:36.280 Well, you first of conversations say hey, I love what 0:14:36.320 --> 0:14:39.840 you do. You're bring your cloud? Can I license your data? 0:14:39.920 --> 0:14:42.440 And usually your feedback is what is in for me? 0:14:42.920 --> 0:14:43.080 Right? 0:14:43.120 --> 0:14:45.480 And I will tell our commercials And then your concern 0:14:45.560 --> 0:14:47.720 could be, hey, you know, if I give you all 0:14:47.720 --> 0:14:50.360 my data, I give my way all my secrets and 0:14:50.360 --> 0:14:53.560 I will go through traditional licensing agreement. What can I disclose? 0:14:53.640 --> 0:14:55.280 What I want from you? What I do not want 0:14:55.320 --> 0:14:57.960 from you? What's the pipeline look like? Are you right? 0:14:58.040 --> 0:15:00.440 You use a street market job? You are running my 0:15:00.600 --> 0:15:03.440 API to yours? Are you running into mind? It's a 0:15:03.480 --> 0:15:07.240 lot of conversations. It's actually pretty standard conversation. And right 0:15:07.320 --> 0:15:10.440 now with eight million pricing points globally around two hundred 0:15:10.520 --> 0:15:14.200 data sources, it's pretty much bau A lot. People will say, hey, 0:15:14.480 --> 0:15:17.520 always bring up you can I have your data? It's 0:15:17.560 --> 0:15:18.320 always my ending. 0:15:18.520 --> 0:15:22.920 You know, we were talking about GPU indices and you're 0:15:22.960 --> 0:15:25.600 not the only one doing GPU price indices for. 0:15:25.640 --> 0:15:26.560 Sure, not anymore. 0:15:26.680 --> 0:15:28.600 Yeah, not anymore. But when you look at some of 0:15:28.600 --> 0:15:32.600 the other ones, like sometimes they show different numbers or 0:15:32.640 --> 0:15:37.400 even different longer term trends, what accounts for the discrepancy there? 0:15:37.640 --> 0:15:40.200 What are you doing differently or what are they doing differently? 0:15:40.240 --> 0:15:40.720 I guess. 0:15:40.920 --> 0:15:43.480 So I can't come on other people's mythology because I 0:15:43.480 --> 0:15:48.360 actually don't know different raw data. Different mythology will eventually 0:15:48.400 --> 0:15:52.240 draft different prices. So the way I would look at 0:15:52.280 --> 0:15:56.400 this is, you know, it's always smart for anyone to 0:15:56.400 --> 0:15:59.640 look at multiple data sources and then figure out what 0:15:59.800 --> 0:16:04.320 is the actual decision you have to make, which datasps 0:16:04.320 --> 0:16:08.320 do you trust? The market always volte whilst us things 0:16:08.320 --> 0:16:11.760 start trading. The market will always gravitates what things actually 0:16:11.760 --> 0:16:15.480 help them hatch? Right, if you easily manupulatable, if you 0:16:15.720 --> 0:16:18.760 are not data source people acting actions, do you hatch? 0:16:19.200 --> 0:16:22.600 There was the point a sideh one speculation, right, So 0:16:22.920 --> 0:16:24.920 you know I'd love to say, I mean, I also 0:16:24.920 --> 0:16:27.360 strongly believe what the best but again I will let 0:16:27.360 --> 0:16:29.280 the market decide, which will happen very soon. 0:16:29.640 --> 0:16:33.040 So of course, like yes, there's the economic rationale for 0:16:33.080 --> 0:16:36.440 the existence of a hedging instrument, and we can understand 0:16:36.520 --> 0:16:39.440 that someone who is an entity that from time that 0:16:39.520 --> 0:16:43.400 needs compute their short implicitly short GPUs, they want to hedge, 0:16:43.680 --> 0:16:46.600 et cetera. But the liquid markets also really do need 0:16:46.640 --> 0:16:49.680 speculators and they need people betting on price. What are 0:16:49.720 --> 0:16:54.200 you seeing right now in terms of traders or institutions, 0:16:54.200 --> 0:16:57.280 et cetera, who economically can take both sides of the 0:16:57.320 --> 0:17:00.720 trade and how active is this getting where it's just 0:17:00.760 --> 0:17:05.439 a compute trading desk that is separate from their economic needs. 0:17:05.640 --> 0:17:08.040 The conversation has been going on for a very long 0:17:08.080 --> 0:17:12.919 time with various banks, various multi participants, speculators. They are 0:17:13.040 --> 0:17:16.200 very excited. So some banks obviously have those both sides 0:17:16.200 --> 0:17:18.239 of the trade right, so they can cross off some 0:17:18.280 --> 0:17:21.880 positions internally. That's great. Always some they have to use 0:17:22.040 --> 0:17:25.399 leverage external products. So that's where we come in. The 0:17:25.440 --> 0:17:28.919 way I encourage them to do is I selfishly, I 0:17:28.960 --> 0:17:31.680 want them to start trading desk and compute. The more 0:17:31.720 --> 0:17:34.840 people trade, the better for me, right selfishly. The same time, 0:17:35.000 --> 0:17:38.560 it is important for people understand GPU trading. It's not 0:17:38.760 --> 0:17:41.919 like you can just move someone from oil electricity with 0:17:42.160 --> 0:17:46.000 no background context jobing to GPU compute futures. There's a 0:17:46.000 --> 0:17:48.159 lot of context where number one GPU it is not 0:17:48.200 --> 0:17:51.560 homogenous product number two. You have to understand the use 0:17:51.600 --> 0:17:53.760 cases of eight one hundred h one hundred right now, 0:17:53.800 --> 0:17:56.520 they are not that correlated. Is that right? Maybe that's 0:17:56.520 --> 0:17:59.040 not right. I don't know. There are use cases which 0:17:59.080 --> 0:18:01.800 they're pretty separated, but maybe their use cases they can 0:18:01.880 --> 0:18:05.040 be transferred and also their software layer to this. Right, 0:18:05.240 --> 0:18:07.560 so right now you can art give sooner use cases 0:18:07.600 --> 0:18:11.080 some large amount of models cannot be deployed and the 0:18:11.200 --> 0:18:14.560 legacy chips, but doesn't mean six months later you cannot 0:18:14.560 --> 0:18:18.800 do so. As the software layer compression model compression gets better, 0:18:18.840 --> 0:18:22.560 optimization gets better, things can change. So really understand not 0:18:22.640 --> 0:18:26.560 just the hardware configuration, this local supply demand curve for 0:18:26.640 --> 0:18:31.080 the service self also software layer that's kind of critical, right, 0:18:31.160 --> 0:18:33.320 that's really changed the supply demand curve and all the 0:18:33.359 --> 0:18:36.280 way to the user behavior. So it's all it's going 0:18:36.320 --> 0:18:38.919 to take some times as we have engaged with a 0:18:38.920 --> 0:18:41.520 lot of participants make sure they have the right set up. 0:18:42.119 --> 0:18:44.280 I have what is possibly a dumb question, but the 0:18:44.560 --> 0:18:48.320 compute futures, how are those actually settled? Because I have 0:18:48.359 --> 0:18:51.400 like images in my mind of taking physical delivery of 0:18:51.440 --> 0:18:53.280 like maybe one of those big. 0:18:53.240 --> 0:18:59.240 Okay, that'll be fun. So for the CME, futures will 0:18:59.280 --> 0:19:03.520 be financially settled just like the traditional oil settlement price. 0:19:03.560 --> 0:19:06.560 That priceis goes four contracts. Well obviously we do four 0:19:06.640 --> 0:19:09.119 right now, I can fut change, but we always open 0:19:09.200 --> 0:19:11.960 to do you know, physically deliver futures, especially given we 0:19:12.119 --> 0:19:15.800 do have silica mark which is GPO benchmarking. So imagine 0:19:15.800 --> 0:19:18.200 the future. You can do, Hey, I want to twenty 0:19:18.640 --> 0:19:22.879 grade A B two hundred, this configuration, this shape of 0:19:23.280 --> 0:19:26.520 servers in US East and then at the end we'll 0:19:26.560 --> 0:19:29.480 get that. Well, usually API costs so you don't get 0:19:29.480 --> 0:19:31.560 physical and it's not as coo as physically give you 0:19:31.600 --> 0:19:33.399 a way way for but you get APIA costs. 0:19:33.400 --> 0:19:34.159 One can dream. 0:19:34.680 --> 0:19:37.480 How do you literally trade it is in like let's 0:19:37.480 --> 0:19:40.080 say there's probably some very bright people in the room 0:19:40.520 --> 0:19:43.600 now with an institution. When it's all listed in everything, 0:19:44.119 --> 0:19:46.120 does it need to go through like a future's broker? 0:19:46.880 --> 0:19:48.760 Is it like a could it be like a prediction 0:19:48.800 --> 0:19:50.359 market if you just go to a website? Like what 0:19:50.480 --> 0:19:54.760 is the actual How does someone actually get in this setting? 0:19:54.800 --> 0:19:57.040 Aside whether they're sophisticated enough of whether they know what 0:19:57.040 --> 0:19:58.840 they're doing a lot of people trade to have no 0:19:58.920 --> 0:20:01.800 idea what they're doing. Yeah, setting all this side. Yes, 0:20:01.840 --> 0:20:04.199 you know, only trade what you know. But like, what 0:20:04.600 --> 0:20:06.800 is it through a prime broker? Like how will people 0:20:06.880 --> 0:20:08.600 actually be able to participate. 0:20:08.119 --> 0:20:08.720 In this market? 0:20:08.760 --> 0:20:11.359 The beauty of CME is you can do the same 0:20:11.480 --> 0:20:15.120 thing you're doing now treating semi products. Okay, the same 0:20:15.160 --> 0:20:19.880 process and processing margin. That's why you get great margin optimization. Right, 0:20:20.440 --> 0:20:25.080 everything is BAU it's no different. We don't have anything 0:20:25.160 --> 0:20:25.480 right now. 0:20:25.680 --> 0:20:29.040 So any commodities broker that someone has, they will be 0:20:29.119 --> 0:20:32.359 able to on that platform. They will have access to these. 0:20:32.240 --> 0:20:35.240 Instruments exactly right. Yeah, we make it easy for people. 0:20:35.600 --> 0:20:38.280 Would you be upset if a prediction market set up 0:20:38.359 --> 0:20:41.280 a GPU price contract of some sort with that into 0:20:41.320 --> 0:20:42.640 your business? 0:20:42.680 --> 0:20:46.160 Not at all. So we actually work with poly market. 0:20:46.840 --> 0:20:49.679 Last year someone actually listed my product at polymarket with 0:20:50.200 --> 0:20:53.439 my consent. It's pretty it's always start like that. And 0:20:53.480 --> 0:20:55.960 then someone told me that, and we try to Polymarket 0:20:55.960 --> 0:20:57.399 and say, hey, do you want to do something you 0:20:57.440 --> 0:21:02.520 know more real? So we did. Fabruary settled and April 0:21:02.560 --> 0:21:06.400 settled a few contracts on polymarket. I just with test 0:21:06.440 --> 0:21:10.760 the water right obviously we're exclusively with CME right now. 0:21:11.280 --> 0:21:13.080 But yeah, so I think obviously you have to do 0:21:13.080 --> 0:21:16.440 it right, licensing, nonminal, pillared, it all the right things. Yeah, 0:21:17.000 --> 0:21:19.440 you know, I don't. Mark can do whatever they want, 0:21:19.560 --> 0:21:21.920 and then people will choose the best product for them 0:21:21.960 --> 0:21:22.320 to use. 0:21:22.880 --> 0:21:27.560 Sitting aside the financial instruments for the moment, would people 0:21:27.640 --> 0:21:30.919 think about AI and they think about the use of GPUs, 0:21:31.480 --> 0:21:34.359 they mostly still probably in their mind think of like 0:21:34.680 --> 0:21:39.480 open AI and Thropic and Google basically, and that's kind 0:21:39.520 --> 0:21:43.680 of it. But obviously, as you've stated, like the world 0:21:43.840 --> 0:21:47.480 of entities that serve inference in some form or another 0:21:47.960 --> 0:21:51.280 is much greater than these three companies that we talk about. 0:21:51.560 --> 0:21:53.480 Talk to us a little bit more about what the 0:21:53.560 --> 0:21:58.240 actual world of inference provision looks like outside of the 0:21:58.280 --> 0:21:59.880 big household AI name. 0:22:00.160 --> 0:22:03.280 So the ones you mentioned, they mostly are closed source 0:22:03.320 --> 0:22:05.479 models as we call it, right, but they do have 0:22:05.520 --> 0:22:09.040 some open source versions, but they're famous for their closed 0:22:09.040 --> 0:22:12.520