WEBVTT - Coreweave's CSO on the Business of Building AI Datacenters 0:00:03.120 --> 0:00:12.680 Bloomberg Audio Studios, Podcasts, radio News. 0:00:20.040 --> 0:00:24.040 Hello and welcome to another episode of the Odd Lots podcast. 0:00:24.120 --> 0:00:25.880 I'm Joe Wisenthal. 0:00:25.400 --> 0:00:26.520 And I'm Tracy Alloway. 0:00:26.920 --> 0:00:29.640 Tracy, you know, we've done tons of course on like 0:00:29.800 --> 0:00:34.559 electricity and AI and data centers and all that stuff, 0:00:34.880 --> 0:00:38.320 but we've never actually done like a well, we've never 0:00:38.440 --> 0:00:41.560 talked to someone who is building data centers. 0:00:42.120 --> 0:00:43.479 Putting it all together, you mean. 0:00:43.440 --> 0:00:45.800 Yeah, putting it all together like what you know, just 0:00:45.800 --> 0:00:47.880 a bunch of you know, I've had consultants, so we 0:00:47.960 --> 0:00:50.920 talked to energy people, but like, how does this business 0:00:51.560 --> 0:00:54.920 of essentially, I guess, building a building, putting a bunch 0:00:54.960 --> 0:00:57.880 of chips in there, getting the electricity, and then in theory, 0:00:58.440 --> 0:01:00.440 selling all of that at a markup? Like, how does 0:01:00.480 --> 0:01:01.240 it actually work? 0:01:01.520 --> 0:01:01.720 You know? 0:01:01.760 --> 0:01:04.120 What I was reading recently This is kind of a tangent, 0:01:04.240 --> 0:01:07.399 but not really because we're talking about the physical and 0:01:07.760 --> 0:01:12.920 financial process of building these things. But I saw this 0:01:12.959 --> 0:01:17.240 is online. There's a guide to the like physical Planning 0:01:18.080 --> 0:01:23.200 around an IBM system three sixty from like nineteen sixty 0:01:23.280 --> 0:01:27.399 three or something, and it's two hundred and thirteen pages long. 0:01:27.560 --> 0:01:28.480 Have you read it yet? 0:01:28.840 --> 0:01:32.160 I did flip through it, there's like there's guidance on 0:01:32.240 --> 0:01:36.760 minimizing vibrations obviously, like temperature and humidity and stuff like that. 0:01:36.800 --> 0:01:39.000 I did not read the full two hundred pages, but 0:01:39.040 --> 0:01:42.560 I'm kind of thinking like if this is what if 0:01:42.560 --> 0:01:45.640 this is all the thinking that had to go into 0:01:45.720 --> 0:01:49.800 like one computer, albeit a supercomputer in the nineteen sixties, 0:01:49.840 --> 0:01:52.120 but like a pretty basic machine. When we look back 0:01:52.160 --> 0:01:54.960 on it now, how much planning and thinking has to 0:01:55.000 --> 0:01:58.720 go into building like these huge cloud servers and all 0:01:58.760 --> 0:02:02.760 their associated infrastry structure, both physical and software as well. 0:02:03.080 --> 0:02:05.600 No, totally, and you know, we you know, one of 0:02:05.640 --> 0:02:07.880 the ways that we've touched on this subject a little 0:02:07.880 --> 0:02:11.200 bit is in our conversations with Steve Eisman, who's been 0:02:11.360 --> 0:02:13.600 investing at least as far as we know, in a 0:02:13.600 --> 0:02:19.079 lot of these like industrial HVAC companies and electricity gear 0:02:19.240 --> 0:02:22.280 companies and stuff like that. So like companies that have 0:02:22.320 --> 0:02:25.280 actually been around for a really long time, sort of 0:02:25.280 --> 0:02:29.000 standard cyclical businesses, and then they've like caught the secular 0:02:29.160 --> 0:02:32.960 tailwind because with this boom in AI data center construction, 0:02:33.560 --> 0:02:36.520 suddenly there's this sort of continuous bid for all their 0:02:36.520 --> 0:02:37.640 gear and services. 0:02:37.760 --> 0:02:41.600 I'm going to start an anti vibration floor maker or something. 0:02:41.600 --> 0:02:43.560 Do you think that's a viable business? Does anyone care 0:02:43.560 --> 0:02:44.840 about vibrations anymore? 0:02:44.880 --> 0:02:48.600 I am certain that in various high tech environments you 0:02:48.680 --> 0:02:50.960 do not want to have vibrations. You know, you have 0:02:51.040 --> 0:02:53.240 like a valuable chips, you don't want them to be 0:02:53.280 --> 0:02:54.639 like degrading. 0:02:54.280 --> 0:02:56.080 Because people are walking around. 0:02:55.919 --> 0:02:58.280 Yeah, or just you know what, all the machine and 0:02:58.320 --> 0:03:01.560 all your air conditioners and equipment and all that stuff, 0:03:01.600 --> 0:03:03.080 you can't be having that stuff degrade. 0:03:03.200 --> 0:03:06.160 Well, the other interesting thing that's happening in the space now, 0:03:06.240 --> 0:03:10.200 So in addition to the physical challenge of building a 0:03:10.200 --> 0:03:13.600 bunch of this stuff, there's also the financial aspect of it. 0:03:13.720 --> 0:03:17.400 And I guess as AI becomes more and more of 0:03:17.440 --> 0:03:19.600 a thing, and clearly, as you laid out, there's a 0:03:19.639 --> 0:03:22.400 lot of enthusiasm around the space. At the moment, you 0:03:22.480 --> 0:03:26.520 are seeing a bunch of financial entities get interested as well. 0:03:26.600 --> 0:03:29.799 So obviously venture capital has been pouring money into the space, 0:03:29.840 --> 0:03:33.440 but we're starting to see some new types of financial 0:03:33.520 --> 0:03:37.560 investments in AI. And I'm thinking about one thing in particular, 0:03:37.960 --> 0:03:42.360 and it is the recent GPU or chip backed loan 0:03:42.840 --> 0:03:46.320 that was reported by the Wall Street Journal and I 0:03:46.360 --> 0:03:48.080 think we should talk about that aspect. 0:03:47.760 --> 0:03:50.080 Of it too, totally, because one of the things that's 0:03:50.200 --> 0:03:53.960 happening in tech is this big sort of shift from like, okay, 0:03:54.000 --> 0:03:56.200 we're all of your costs in the past, where a 0:03:56.200 --> 0:03:58.720 lot of them were sort of op x, the cost 0:03:58.720 --> 0:04:02.200 of engineers, et cetera. And now suddenly tech companies have 0:04:02.240 --> 0:04:04.680 to think about CAPEX for the first time, these big 0:04:04.800 --> 0:04:06.720 upfront costs that are in theory going to pay off 0:04:06.760 --> 0:04:09.480 for a long time, which in theory then changes how 0:04:09.520 --> 0:04:11.080 you should think about the financing model. 0:04:11.160 --> 0:04:12.880 Absolutely well, I am. 0:04:12.840 --> 0:04:16.640 Excited to say because we literally do have the perfect 0:04:16.720 --> 0:04:20.000 guest we're going to be speaking with, Brian Venturo. He 0:04:20.080 --> 0:04:24.120 is the chief strategy officer at core Weave. Corewave. For 0:04:24.160 --> 0:04:27.480 those who don't know, it's probably the company right now 0:04:27.760 --> 0:04:31.120 that people most associate with being at the heart of 0:04:31.200 --> 0:04:35.359 the AI data center boom. They have a bunch of 0:04:35.360 --> 0:04:38.440 in video chips, they have investments from in Nvidio right 0:04:38.560 --> 0:04:41.279 here in the sweet spot. As you mentioned, one of 0:04:41.320 --> 0:04:44.200 the interesting things that's going on is they not long 0:04:44.240 --> 0:04:49.440 ago announced a debt financing facility sit back basically by 0:04:49.480 --> 0:04:53.679 the GPUs that they would acquire, so literally the perfect 0:04:53.680 --> 0:04:58.920 person to understand like the business of these AI cloud 0:04:59.360 --> 0:05:02.080 data center So Brian, thank you so much for coming in. 0:05:02.360 --> 0:05:04.080 Thanks for having me. It's the second time I've been 0:05:04.080 --> 0:05:04.760 on the podcast. 0:05:05.040 --> 0:05:08.200 That's right. We talked to Brian years ago. It's interesting 0:05:08.440 --> 0:05:10.920 to think about at that time because I think that 0:05:11.040 --> 0:05:13.320 may have been like twenty twenty or twenty one, and 0:05:13.360 --> 0:05:16.000 the excitement then was that these chips could be used 0:05:16.200 --> 0:05:19.520 for crypto mining and other things like sort of distributed 0:05:19.600 --> 0:05:23.520 video editing and stuff like that, and then Ethereum stopped 0:05:23.600 --> 0:05:26.320 using mining. But it was sort of fortuitous timing because 0:05:26.400 --> 0:05:30.040 right around then AI went crazy and that's probably I 0:05:30.080 --> 0:05:31.960 don't know, in my view, maybe a higher use of 0:05:32.000 --> 0:05:34.479 these chips before we get to that. Do you worry 0:05:34.520 --> 0:05:37.920 about vibration in your data center? 0:05:38.120 --> 0:05:42.760 So everywhere that's close to a fault line is designed 0:05:42.760 --> 0:05:45.440 around that and is part of code. So you know, 0:05:45.440 --> 0:05:47.800 the engineering firms that help us build these data centers 0:05:47.800 --> 0:05:49.680 have taken all of that into account, and all of 0:05:49.720 --> 0:05:54.600 our racks are you know, seismically tuned to make sure 0:05:54.600 --> 0:05:57.239 that we can withstand the normal vibration from the Earth. 0:05:57.640 --> 0:06:00.360 So yeah, it's been something that's been in those annuals 0:06:00.360 --> 0:06:03.360 for a long time. Some of our hardwer manufacturers actually 0:06:03.360 --> 0:06:06.479 have vibration testing labs where they put the racks on 0:06:06.520 --> 0:06:08.839 top of a big kind of platform that shakes, and 0:06:08.880 --> 0:06:12.240 it's pretty dangerous and uncontrollable and hard to watch. But 0:06:12.720 --> 0:06:14.239 you know, there's people out there that have been solving 0:06:14.240 --> 0:06:15.080 this problem for decades. 0:06:15.120 --> 0:06:17.800 Now I missed the boat on that business choir. It 0:06:17.920 --> 0:06:22.400 sounds like it's been dealt with decades ago. Okay, well, actually, 0:06:22.760 --> 0:06:25.760 why didn't I start with a very simple question, which 0:06:25.839 --> 0:06:29.400 is when when you're looking at the business of core Weave, 0:06:29.880 --> 0:06:34.560 so a specialized cloud service provider, let's put it that way, 0:06:34.960 --> 0:06:38.440 what are the different components that you have to think about? 0:06:38.480 --> 0:06:41.040 You know, Joe kind of alluded to all these different 0:06:41.240 --> 0:06:44.200 ingredients that go into the business, but walk us through 0:06:44.360 --> 0:06:45.680 what those actually are. 0:06:46.120 --> 0:06:49.120 Sure, so, there's there's three pieces that as a management team, 0:06:49.120 --> 0:06:52.680 we think are incredibly critical to the business. The first is, 0:06:53.240 --> 0:06:55.479 you know, our technology services that we provide on top 0:06:55.520 --> 0:06:57.880 of the hardware, right and this is everything from the 0:06:57.920 --> 0:07:00.480 software layer through the support organization to you know, how 0:07:00.480 --> 0:07:03.000 we work with our customers. This isn't the type of 0:07:03.040 --> 0:07:04.960 thing that you just go plug in and it works. 0:07:05.600 --> 0:07:08.279 In these large supercomputer clusters, there may be two hundred 0:07:08.360 --> 0:07:12.080 thousand infinibank connections that connect all the GPUs together, and 0:07:12.120 --> 0:07:15.120 if one of those connections fails for whatever reason, the 0:07:15.280 --> 0:07:18.040 job will completely stop and have to restart from its 0:07:18.080 --> 0:07:20.720 previous checkpoints. So, you know, everything that we do on 0:07:20.760 --> 0:07:22.800 the software side and engineering side is to make sure 0:07:22.800 --> 0:07:25.800 these clusters are as resilient and performant as they possibly 0:07:25.840 --> 0:07:28.680 can be to ensure you know, our customers can run 0:07:28.720 --> 0:07:33.240 their jobs, you know, increase efficiency and get all of 0:07:33.320 --> 0:07:36.000 the kind of monetary value they can out of the chips. 0:07:36.920 --> 0:07:39.520 So technology piece is really hard. It's something that I 0:07:39.560 --> 0:07:42.440 think is very overlooked by the market, but it's just 0:07:42.480 --> 0:07:44.520 as hard as the two other kind of pieces that 0:07:44.520 --> 0:07:47.560 this business stands on. The second is, you know, the 0:07:47.600 --> 0:07:49.680 physical nature of the business in that you have to 0:07:49.720 --> 0:07:52.680 actually build and run these data centers and those hundreds 0:07:52.680 --> 0:07:56.160 of thousands connections inside the supercomputers. Like somebody has to 0:07:56.160 --> 0:07:58.640 go put those together and make sure they're clean and 0:07:58.680 --> 0:08:01.200 make sure they're labeled correctly to be able to remediate failures. 0:08:01.720 --> 0:08:05.600 And when you're building a thirty two thousand GPU supercomputer 0:08:05.880 --> 0:08:08.800 that is one of the fastest three computers in the planet. 0:08:09.800 --> 0:08:13.360 You know, you're running thousands of miles of cable inside 0:08:13.400 --> 0:08:16.520 a very dense space, right. These data centers are built 0:08:16.640 --> 0:08:19.480 very tiny to make sure that you can connect everything together, 0:08:20.160 --> 0:08:22.880 and that becomes a huge logistical challenge. So, you know, 0:08:22.920 --> 0:08:25.120 the data centerpiece, which we're going to talk more about today, 0:08:25.880 --> 0:08:28.440 is very challenging to design for the use case. And 0:08:28.520 --> 0:08:30.080 then the third piece is how the hell do you 0:08:30.200 --> 0:08:32.959 finance the whole thing? Right, And you know, we've been 0:08:33.480 --> 0:08:36.800 very successful in the financing aspect of this, but you know, 0:08:36.840 --> 0:08:39.880 whether you're financing technology operations or the physical build of 0:08:39.920 --> 0:08:43.880 these things, it is an incredibly capital intensive business and 0:08:44.240 --> 0:08:48.320 constructing those financial instruments to back our business is very hard, 0:08:48.360 --> 0:08:51.120 and we have to be very very thoughtful around who 0:08:51.120 --> 0:08:53.360 the counterparties are, how do we think about credit risk, 0:08:53.679 --> 0:08:56.160 how do our investors think about that credit risk, How 0:08:56.160 --> 0:08:59.319 do we deal with contingencies inside the contracts to make 0:08:59.360 --> 0:09:01.160 sure that they are financeable on the scale that we've 0:09:01.160 --> 0:09:02.400 done over the last eighteen months. 0:09:02.640 --> 0:09:04.880 Talk to us a little bit more. We could probably 0:09:05.120 --> 0:09:09.319 talk about data center financing credit and have have that 0:09:09.440 --> 0:09:12.520 be a whole episode, but when you think about you 0:09:12.640 --> 0:09:16.079 have to think about your counter party's credit risk. Talk 0:09:16.080 --> 0:09:18.160 to us a little bit about what you're who those are, 0:09:18.320 --> 0:09:20.080 what the type of entity is. 0:09:20.400 --> 0:09:22.800 Sure, so I'll get myself in trouble if I just 0:09:22.840 --> 0:09:25.400 start naming them off. Yeah, some of them are more 0:09:25.400 --> 0:09:28.559 public than others. You know, I'm going to refer to 0:09:28.600 --> 0:09:33.040 them as you know, hyperscale customers. We have AI lab customers, 0:09:33.679 --> 0:09:37.800 we have large enterprise customers. We've really constructed our portfolio 0:09:37.800 --> 0:09:40.600 of business around the idea that you know, if we're 0:09:40.600 --> 0:09:42.960 going to build ten billion dollars of infrastructure for somebody, 0:09:43.000 --> 0:09:44.640 we have to know there's a balance sheet we can 0:09:44.720 --> 0:09:48.800 lean into behind it, right, and we're the pace at 0:09:48.840 --> 0:09:53.320 which we've grown. You know, our customers are demanding scale 0:09:53.440 --> 0:09:57.320 so quickly that the credit of the counterparty is incredibly 0:09:57.360 --> 0:09:59.400 important to find the low cost of capital we have 0:09:59.480 --> 0:10:02.120 with these ADIT facilities we've announced, right, So you know, 0:10:02.160 --> 0:10:04.520 when people talk about how this is a credit facility 0:10:04.520 --> 0:10:07.360 backed by GPUs, it's not really backed by GPUs. It's 0:10:07.400 --> 0:10:12.040 backed by you know, commercial contracts with large international enterprises 0:10:12.080 --> 0:10:14.480 that may have triple a credit, right, So you know 0:10:14.520 --> 0:10:15.600 it's it's the framing of the. 0:10:15.679 --> 0:10:17.240 Aid receivables finance. 0:10:17.280 --> 0:10:20.400 Basically it's closer to trade receivables financing than it is Hey, 0:10:20.400 --> 0:10:22.200 we're going to go leverage up a bunch of GPUs 0:10:22.240 --> 0:10:22.959 and see what happens. 0:10:23.080 --> 0:10:26.480 Huh, okay, well walk us through the I guess like 0:10:26.760 --> 0:10:31.440 the sequence in some of these financing agreements. So you know, 0:10:31.480 --> 0:10:33.959 if a customer comes to you and they say, we 0:10:34.040 --> 0:10:37.280 want a certain amount of compute, can you do this 0:10:37.400 --> 0:10:40.800 for us? And you start going down the process of like, okay, 0:10:40.840 --> 0:10:43.320 what do we need to make this happen? What do 0:10:43.440 --> 0:10:47.440 those like financial agreements actually look like. And who's bearing 0:10:47.480 --> 0:10:50.200 the initial risk? Is it the customer? Is it you? 0:10:51.360 --> 0:10:51.880 Good question? 0:10:52.000 --> 0:10:54.080 So when we're approached by a customer, right, you know, 0:10:54.160 --> 0:10:57.360 the ask is typically going to be pretty pretty general, 0:10:57.840 --> 0:11:00.360 and they're going to say, hey, we're looking for facity 0:11:00.400 --> 0:11:02.080 in Q one of next year. What's the largest thing 0:11:02.120 --> 0:11:06.200 you can do? And you know, we take that effectively 0:11:06.240 --> 0:11:08.960 as a mandate of okay, hey, you know this customer. 0:11:08.679 --> 0:11:09.199 We're not business. 0:11:09.240 --> 0:11:11.080 But before you know, we're really comfortable with them, we 0:11:11.120 --> 0:11:12.840 know that we're going to get a contract done. We'll 0:11:12.840 --> 0:11:14.719 go out and we'll try to secure an asset to 0:11:15.080 --> 0:11:16.240 you know, to go build it. And we may have 0:11:16.280 --> 0:11:18.160 it in our portfolio already. We maybe it may have 0:11:18.200 --> 0:11:20.720 been a strategic investment that we made. But once we 0:11:20.760 --> 0:11:22.360 find the data center asset, that's when we go back 0:11:22.360 --> 0:11:24.160 to the customer and say, okay, like we can commit 0:11:24.200 --> 0:11:26.680 to doing this. This is the timeline. We'll structure a 0:11:26.679 --> 0:11:29.600 contract around it. Depending upon who the customer is. There 0:11:29.640 --> 0:11:31.640 may or may not be some credit support associated with 0:11:31.679 --> 0:11:34.280 it around the scaling of the you know, that asset, 0:11:35.080 --> 0:11:38.200 and then we'll get a commercial contract in place, and 0:11:38.760 --> 0:11:42.600 we will initially fund a large portion of that project 0:11:42.720 --> 0:11:44.800 off of our own balance sheet. Right. It's why you 0:11:44.840 --> 0:11:47.160 also see us raising equity, right, is we have to 0:11:47.320 --> 0:11:49.800 have the capital to accelerate the business. And then once 0:11:49.840 --> 0:11:52.439 we have that and we're making progress, you know, think 0:11:52.440 --> 0:11:54.360 about it as you're building real estate. Right, you have 0:11:54.400 --> 0:11:56.520 a construction loan and then you have a stabilized asset loan, 0:11:57.040 --> 0:11:59.439 and we basically fund the construction loan piece off of 0:11:59.440 --> 0:12:01.880 our balance sheet. When we get to a more stabilized asset, 0:12:01.880 --> 0:12:03.679 that's when we go out and kind of do that 0:12:03.800 --> 0:12:07.720 trade financing or trade receivables financing our with our partner lenders. 0:12:08.160 --> 0:12:09.760 You know, they worked with us before, they know that 0:12:09.800 --> 0:12:11.040 these things are going to stand up, They know how 0:12:11.040 --> 0:12:13.480 they perform, and at that point in time, it's it's 0:12:13.480 --> 0:12:21.760 pretty easy for them to underwrite that risk. 0:12:31.160 --> 0:12:35.479 It's funny. Tracy and I had coffee with someone yesterday 0:12:36.160 --> 0:12:39.000 who is sort of in the space I want docs here, 0:12:39.080 --> 0:12:40.560 And I was like, what should we ask Brian? And 0:12:40.600 --> 0:12:42.880 he's like, ask him why he won't let my company, 0:12:43.440 --> 0:12:45.240 why I'm still on the waiting list or something, or 0:12:45.280 --> 0:12:48.480 why he hasn't approved my company to use core weave. 0:12:48.679 --> 0:12:51.679 But what are some of the bars or the threshold? 0:12:51.760 --> 0:12:54.280 So you know, I apparently there's a lot of demand 0:12:54.440 --> 0:12:57.360 for compute these days. What does it take to get 0:12:57.360 --> 0:12:59.880 in the door and get access to some of your 0:13:00.200 --> 0:13:01.319 chips and electricity? 0:13:01.840 --> 0:13:05.000 So it's it's a great question. It's a question that 0:13:05.040 --> 0:13:07.240 we get all the time from our sales teams, right 0:13:07.400 --> 0:13:09.760 is you know, we're faced a lot with a sales 0:13:09.840 --> 0:13:13.400 team that is incredible at delivering product to customer and 0:13:13.440 --> 0:13:15.880 we don't have anything to sell. And it's kind of 0:13:15.880 --> 0:13:19.760 my job. As the strategy organization at Core, We've were 0:13:19.800 --> 0:13:24.360 responsible for two things. It's product and infrastructure. Capacity, and 0:13:24.559 --> 0:13:26.280 you know, I spend most of my time going out 0:13:26.280 --> 0:13:28.160 and finding those data centers and being able to support 0:13:28.200 --> 0:13:31.040 those deals and the growth that we had over the 0:13:31.040 --> 0:13:34.600 past twelve months. The company was pretty flat out right 0:13:34.679 --> 0:13:37.760 in building and delivering this infrastructure. You know, publicly on 0:13:37.760 --> 0:13:40.400 our documentation page it says that we have three regions. 0:13:40.600 --> 0:13:42.360 We'll have twenty eight regions online by the end of 0:13:42.400 --> 0:13:44.439 the year. I think we delivered eleven of them in 0:13:44.559 --> 0:13:47.320 Q one alone, Right, So we're building at a scale, 0:13:48.080 --> 0:13:50.160 you know, i'd say that almost larger than some of 0:13:50.200 --> 0:13:54.120 the three big hyperscalers. But in terms of how do 0:13:54.200 --> 0:13:56.959 you become a customer of Core, it's really relationship driven, 0:13:57.160 --> 0:13:59.040 right is. We want to make sure that we're going 0:13:59.120 --> 0:14:00.880 to be able to be successfu with our customers and 0:14:00.920 --> 0:14:03.679 have an engineering relationship and we're aligned on what they 0:14:03.720 --> 0:14:04.079 need and. 0:14:04.040 --> 0:14:05.120 We can deliver what they need. 0:14:05.800 --> 0:14:07.240 The last thing that we want is for somebody to 0:14:07.280 --> 0:14:09.480 walk in the door and say, hey, I need this 0:14:09.520 --> 0:14:12.080 for three weeks and two weeks into it, they're unhappy 0:14:12.640 --> 0:14:14.720 and we can't give them what they need to be successful. 0:14:14.800 --> 0:14:17.520 Right is, you know, our customers are making such large 0:14:17.600 --> 0:14:20.760 investments in this infrastructure, that we have to have, you know, 0:14:20.960 --> 0:14:23.720 a lot of conviction that we will be successful with 0:14:23.760 --> 0:14:26.880 them and provide a good experience. So it's not that 0:14:26.960 --> 0:14:29.240 we're trying to keep people out, it's we're trying to 0:14:29.320 --> 0:14:31.960 ensure positive experiences for people that we do bring on board. 0:14:32.320 --> 0:14:37.000 Do you build complete housed facilities or is it all 0:14:37.480 --> 0:14:40.720 you're going to bring your chips and expertise into an 0:14:40.760 --> 0:14:44.200 existing Tier one data center and essentially rent floor space 0:14:44.240 --> 0:14:44.640 from them. 0:14:44.760 --> 0:14:47.640 Yeah, so a year ago it was we were effectively 0:14:47.640 --> 0:14:51.000 just a co location tenant, and now we've gone a 0:14:51.040 --> 0:14:54.760 lot more vertical for some strategic builds where we're either 0:14:54.760 --> 0:14:56.640 a partner in the project where we own equity and 0:14:56.680 --> 0:15:00.120 the development company, or we're building the project ourselves. We've 0:15:00.120 --> 0:15:02.040 been scaling that team up over the past six months, 0:15:02.480 --> 0:15:04.800 and we had to at our scale to be able 0:15:04.800 --> 0:15:07.400 to guarantee outcomes. Right, is, we were in a position 0:15:07.440 --> 0:15:09.600 where we had data centers getting delayed with things that 0:15:09.640 --> 0:15:12.560 weren't communicated to us, and you know, we had to 0:15:12.560 --> 0:15:15.920 go build the capability to handle that situation and you know, 0:15:15.960 --> 0:15:17.640 make sure we can still deliver for our customers. 0:15:17.880 --> 0:15:20.400 One of the differentiators that you and some of your 0:15:20.400 --> 0:15:23.800 colleagues have emphasized previously, is this idea that you're designing 0:15:24.320 --> 0:15:27.680 the server clusters kind of from the ground up, whereas 0:15:27.760 --> 0:15:31.880 like other hyperscalers maybe are doing it on a sort 0:15:31.880 --> 0:15:34.800 of different mass scale. But can you walk us through 0:15:34.840 --> 0:15:38.760 like what is the benefit of doing it that way? 0:15:38.840 --> 0:15:42.640 And then secondly, does that end up being an impediment 0:15:43.040 --> 0:15:47.640 to I guess efficiencies or economics of scale and how 0:15:47.680 --> 0:15:49.560 customized Like do you really get here? 0:15:49.960 --> 0:15:54.080 So from a customization perspective, it's aggressive, right, And I 0:15:54.200 --> 0:15:56.960 say that because you know, our customers are involved in 0:15:57.040 --> 0:15:59.200 the design of you know, our network topology of the 0:15:59.200 --> 0:16:02.400 East West fabric for the GPU to GPU communication, for 0:16:02.520 --> 0:16:04.920 things like cooling. You know, I have customers that toward 0:16:04.920 --> 0:16:07.680 the data centers under construction process with me like once 0:16:07.720 --> 0:16:13.640 a week, and it's to the point that they're impacting 0:16:14.120 --> 0:16:17.320 how we build the base level networking products to ensure 0:16:17.400 --> 0:16:20.080 they have enough throughput to you know, meet their use 0:16:20.120 --> 0:16:23.240 case needs. Whereas in you know, what I what we 0:16:23.280 --> 0:16:27.960 call the legacy hyperscaler installations, It maybe they have a 0:16:27.960 --> 0:16:30.080 couple thousand GPUs that are in a data center that 0:16:30.120 --> 0:16:34.160 was really built for CPU computation or to provide services 0:16:34.160 --> 0:16:37.280 to ten thousand customers that is really with a much 0:16:37.360 --> 0:16:40.080 lower base expectation of what they're going to be doing. Right, 0:16:40.160 --> 0:16:43.720 So it's things around connectivity for storage, it's things around 0:16:43.760 --> 0:16:47.480 power and cooling, It's things around how they want to 0:16:47.520 --> 0:16:51.360 be able to optimize their workloads inside of the GPU 0:16:51.400 --> 0:16:53.680 to GPU communication. You know, we have some customers that 0:16:53.720 --> 0:16:57.200 even customize their infiniban fabrics and the size of those 0:16:57.240 --> 0:16:59.200 fabrics and how they connect together. So you know, we 0:16:59.280 --> 0:17:01.240 work with them to really understand what their use case is, 0:17:01.280 --> 0:17:03.440 where they're worried currently and in the future, and then 0:17:03.480 --> 0:17:07.080 design around that. So it's a pretty comprehensive program when 0:17:07.080 --> 0:17:09.080 we're building something from the ground up. 0:17:09.200 --> 0:17:12.240 And how much complexity does that introduce into the business 0:17:12.320 --> 0:17:15.320 and does it end up being a limiting factor on 0:17:15.359 --> 0:17:18.199 your growth or is demand just so strong at the 0:17:18.240 --> 0:17:20.280 moment that it's not really an issue. 0:17:20.440 --> 0:17:22.480 The customization that we do is typically going to be 0:17:22.520 --> 0:17:26.040 above what our base level offering is, meaning the environment 0:17:26.080 --> 0:17:29.040 will be more performant because the customer required it. So 0:17:29.200 --> 0:17:31.640 it's typically not going to be limiting to us from 0:17:31.720 --> 0:17:34.840 a future you know, revenue or resale perspective. It's going 0:17:34.880 --> 0:17:37.560 to make the asset more valuable. But you know, we're 0:17:37.840 --> 0:17:40.440 we're designing our reference builds for ninety nine percent of 0:17:40.520 --> 0:17:43.359 use cases, and we're trying to price it efficiently, and 0:17:43.400 --> 0:17:45.879 then when customer wants something above and beyond, you know, 0:17:45.920 --> 0:17:49.360 it impacts price. But for these installations it's probably deminimus, right, 0:17:49.440 --> 0:17:51.800 So you know, it doesn't really add a lot of 0:17:51.800 --> 0:17:54.919 complexity for us from a business perspective, so we're happy 0:17:54.920 --> 0:17:55.239 to do it. 0:17:55.600 --> 0:17:59.560 You mentioned that some of the hyperscalers, yes they have GPUs, 0:17:59.600 --> 0:18:05.960 but they like built in an environment for like legacy CPUs. 0:18:06.200 --> 0:18:08.960 Can you talk a little bit about a just the 0:18:09.080 --> 0:18:12.879 difference between the legacy architectures and the new one and 0:18:12.920 --> 0:18:15.520 then in the design, like what kind of bottlenecks you 0:18:15.560 --> 0:18:18.320 run into? Is there issues with labor like the types 0:18:18.359 --> 0:18:20.840 of people who know how to string these things together well, 0:18:21.000 --> 0:18:24.960 or other different cooling requirements for this type of compute 0:18:25.080 --> 0:18:28.439 environment that did not exist, Like what are what are 0:18:28.480 --> 0:18:32.040 the challenges in building out these sort of like fundamentally 0:18:32.720 --> 0:18:33.600 different environments. 0:18:33.680 --> 0:18:36.000 Yeah, so that that's changed also in the last twelve 0:18:36.040 --> 0:18:38.640 months in that you used to be able to take 0:18:38.680 --> 0:18:41.560 what was an enterprise data center and you know, creatively 0:18:41.600 --> 0:18:45.080 retrofit it to be capable of supporting the AI workloads 0:18:45.440 --> 0:18:48.000 to a certain density level. Okay, right, Like instead of 0:18:48.040 --> 0:18:49.800 filling up a cabinet, you could put two servers in 0:18:49.840 --> 0:18:52.439 a cabinet and you could meet the power and cooling 0:18:52.440 --> 0:18:55.840 requirements of the installation. It you use a lot more 0:18:55.920 --> 0:18:59.360 floor space, but it was doable. One of the incredible 0:18:59.359 --> 0:19:01.879 things about is that they're always pushing the boundary on 0:19:01.880 --> 0:19:04.919 the engineering side, and their next generation of chips is 0:19:05.040 --> 0:19:08.400 largely dependent upon much more aggressive heat transfer, and they've 0:19:08.440 --> 0:19:11.600 introduced liquid cooling to the reference architectures. So as liquid 0:19:11.600 --> 0:19:14.800 cooling comes in, it changes what type of data center 0:19:14.840 --> 0:19:18.239 is capable of doing this, and it truly requires that 0:19:18.359 --> 0:19:23.200 ground up redesign and almost greenfield only build to support it. 0:19:23.240 --> 0:19:25.640 Is you've gone from an environment where you could take 0:19:25.680 --> 0:19:28.520 an enterprise data center and deploy less servers per cabinet 0:19:28.560 --> 0:19:31.560 and get away with it to hey, nobody's ever built 0:19:31.560 --> 0:19:33.960 this before. It's at an incredible scale and it has 0:19:34.000 --> 0:19:37.399 to happen on a yearly cadence now, so the data 0:19:37.400 --> 0:19:40.280 center industry is in't a full sprint to figure out, Okay, 0:19:40.280 --> 0:19:42.000 how do we do this? How do we do it quickly? 0:19:42.240 --> 0:19:44.840 How do we operationalize it right? And you know that's 0:19:44.920 --> 0:19:46.680 kind of where I've been spending all of my time 0:19:46.920 --> 0:19:48.000 over the past six months. 0:19:48.200 --> 0:19:51.200 Can I ask a really basic question, and we've done 0:19:51.280 --> 0:19:53.959 episodes on this, but I would be very interested in 0:19:54.000 --> 0:19:58.760 your opinion, But why does it feel like customers and 0:19:58.880 --> 0:20:02.719 AI customers in particular, are so I don't know if 0:20:02.720 --> 0:20:05.960 addicted is the right word, but like so devoted to 0:20:06.359 --> 0:20:10.160 in Nvidia chips, Like what is it about them specifically 0:20:10.800 --> 0:20:13.920