Transcript of Ep. 037 - Who's Funding the $11 Trillion AI Buildout? (Capital Markets)
SemiAnalysis
0:01for our first ever advertisement. We're0:03going to be doing a lot more of this on0:04future semi- analysis weekly episodes.0:06Semi- analysis is a readerfunded0:09publication. Over to Dan. What are we0:11advertising this week?0:12>> Well, well, we we noticed that 90% of0:14our viewers aren't subscribed. So, if0:16you could hit the like button and0:17subscribe and be sure to hit the alerts.0:20Seriously though, the real ad we're0:22we're hiring for a new group uh which is0:24called compute capital markets and we're0:26going to look at um the the loan demand0:29how the financing markets are evolving0:31for financing GPUs and we'll look at0:33markets which talks about the uh tools0:36that lenders and everyone in the market0:39need to evaluate how GPU financing is0:42going, how the GPU industry is going.0:44That's uh rental pricing, residual0:46value, um all those things that come0:48with it. Um and so we're looking for0:50credit analysts. We're going to hire um0:52four slots. We're looking for uh0:53Singapore and New York City. Um we're0:56looking for a junior credit analyst. Um0:58about, you know, perhaps one to three0:59years of experience in the credit space,1:02investment banking in in Singapore or1:04the US. We're also looking for a credit1:06specialist. That could be someone who's1:08been uh a little more experienced,1:10probably four or five years credit1:13trading, credit sales, or uh credit1:16investing. you know, IG high yield would1:18be great. Um, and also a credit analyst.1:21Anyone who's been a desk analyst, uh,1:23buy side, sell side. Um, you know, we're1:26looking for people are credit minded.1:27Um, I think that if you enjoy this1:30episode that's going to come and all1:32these structures and, you know, all1:33these Z spreads and T-spreads, I I think1:35you're going to really like this job.1:37So, um, you know, click click the apply1:39button right here.1:41>> All right, roll the credits.1:48Hello everyone, welcome back to semi1:50analysis weekly. In this episode, we got1:52Dan. Uh we are going to talk about1:56everything to do with capital and1:58markets. How we are funding the buildout2:00right now. Dan has been digging in to uh2:03everything credit capital markets2:06compute. Um, specifically we put out an2:09article I mean called Nvidia back up2:12universe heads I win tail2:15uh where we talked about the 11 trillion2:16AI buildout Nvidia's backstock economics2:19and the limits of Nvidia's balance2:20sheet. So, you know, with that as a2:23backdrop, we're we're going to talk2:24through the contents of that article.2:26But at the pace things are moving right2:28now, that's like already a month old.2:30And so, um, yeah, without further ado,2:34Dan, why don't you lead us in talking a2:35little bit about Nvidia's backstop2:38universe, where you came to set up this,2:41you know, article and why you dug into2:43this expanded view of Nvidia's balance2:45sheet.2:47>> Yeah. Yeah. So let me show you um a bit2:50about um Nvidia's backs stop universe2:52and indeed it it was just a month ago2:54but but things are are moving so fast at2:56this point fund and announced is around3:00um um about like a hundred billion um3:04each month um and and this is how we3:07looked at um the balance sheet and the3:09whole the whole premise is um Nvidia's3:12been giving a lot of back stops and the3:14question is is why right um there hasn't3:17a lot of visibility on how much issuance3:20is going to come or how much funding3:21needs there is. Um, so we defined that3:23for the first time. Um, and what we did3:26is we looked at the funding needs and we3:27talked about whether Nvidia's balance3:28sheet would be strong enough to support3:31backto stopping a lot of these funding3:32needs. Um, and we'll talk about that3:34much more later, but I think it's worth3:36going through quickly the constraints3:38and and what what sort of sort of led us3:40to this point. Um, so if I go back to um3:44if we think about the constraints,3:46right, Jordan? And you know you you3:47asked me earlier about the um the3:50trinity right? You remember that?3:52>> Yeah.3:53>> Yeah. We used to talk a lot about the3:55trinity and the trinity the AI project3:58trinity is basically capital offtake and4:01uh data center. Um and and Jordan4:04earlier he was kind of asking me like4:06well what what about compute right?4:08Isn't that the fourth leg? um which4:11which is a good point because it used to4:13be a pretty big short a pretty big uh4:15pinch point and it used to be um kind of4:18a worsening pinch point but what's4:21happened since then right memory was4:22approximately one of one of the the4:24hugest pinch points right so so let's4:26talk about constraints and the4:27constraints have been moving they've4:29been moving from compute constraints to4:31data center constraints and now4:33financing is a new constraint um so so4:37so why why you know why why are those4:39more acute right so what's happened with4:41memory memory has been been been a4:43shortage you can see that AI's percent4:46DM wafers was going to be 70% um and4:50then you can see that on the logic side4:52accelerator is is an ever higher share4:54but what happened um recently is there's4:59been despecking so let me show you what5:01what what happened so if you look at5:02accelerator perspective um this used to5:06be the pattern if you look at GPU ASP it5:08had been going up because of memory and5:10it was set to go up in Ruben Ultra but5:12that that pattern actually stopped in5:15its tracks and what we got instead was5:17dspecing.5:18>> Why? Um for really two reasons, right?5:22Number one, ROI5:25and I'll just kind of flick to a quick5:27uh a quick um5:30uh bridge chart, right? And so with this5:33one, what we saw is the cost was5:35increasing for the uh VRNBL72.5:39Um, and what they did to offset that is5:41is decrease soam. But it's not just5:43about price, it's about availability.5:46The issue here is that there just simply5:47hasn't been enough memory to go around5:49to actually ship racks. Um, so the dspec5:51was necessary to to deconstrain that.5:54Um, so that's that's the chip5:57constraint. Let's move on to the data5:59center constraint. That was that was one6:00of the first data concern the first6:02constraints that that people worried6:03about. Um but we've found a lot of6:07creative solutions around that. We've6:09mustered a lot of capital to build data6:10centers. We've activated more powered6:13sites. We refurbished old sites. We've6:14done a lot of behind the meter and to6:17the point where by 2027 28 it's much6:19more balanced. Um and this is no longer6:21an obstacle. So what are the new6:24constraints? So what what we talk a lot6:25about now is funding. And that's a6:28constraint people have learned about um6:30in just in just the last few months. Um6:35I'll show you how we calculated the6:36overall constraint. Let's take a look.6:38So this is the Camille of total capex.6:40It's 11 trillion of capex that needs to6:43be spent by 2029.6:44>> Yeah. Dan, explain this chart just6:46because a lot of people are throwing6:48around the term of 1 trillion of6:51hyperscaler capex. How do you get to6:5311.3 trillion on this chart?6:56>> Yeah. So 11 11.3 trillion is6:59commumulative capex from 2028 sorry from7:022024 through 2029 and so it's escalating7:06each year. The way we did this um7:08because we track all the GPUs, we track7:10all shipments, we track total cost of7:12ownership of the racks and the build7:14materials um and that's not just across7:16GPUs also XPUs, it's TPUs, um AWS, uh7:20tranium and if you take the sum product7:23of the shipments across the total cost7:25which includes the server, includes7:28networking, includes CPUs, includes7:29everything um you actually get to that7:3211.3 trillion. And by the way um you7:34also mentioned hypers scale capex so7:36that's actually a subset of this if you7:38look at the legend this is AI it capex7:40and data center capex and a lot of that7:42capex is not incurred by the hypers7:44sales themselves it's incurred by7:46independent data center companies so7:48that's part of the equation of of why 17:51trillion in a year of of hyperscaler7:53capex becomes 11 trillion because it's7:55hypers scale neocloud and then it's it7:57data center7:58>> yeah and this is not covering8:00necessarily other things in the supply8:01chain which we do cover which Dylan has8:03talked about which is you know power8:06plants, industrial equipment uh stuff in8:08the fabs that is you know a precursor to8:11AI it in this measurement here like you8:15know equipment or just new8:20>> there's a lot of stuff going into the8:21buildout right now um you said that8:24there's a constraint related to funding8:27talk me through how people are actually8:29funding this buildout and8:32you just how is where's where is8:36everybody finding the capital?8:38>> Yeah. So, so it's a great question. Um8:40and what we what we said is there's8:42going to be 7 trillion of debt8:43outstanding by 2029. Um and so what does8:47that work out to? It works out to um8:50what we did here is I put the the net8:52additions, the net new funding. So it8:54works out to 1 trillion this year. It's8:57a little flat because whenever we8:59forecast stuff forward, um we tend not9:02to add capacity and and deliver it till9:04it's confirmed, right? But and then9:06there'll still be there'll be some debt9:07that rolls off because all amortizing,9:09but you know, I would expect that this9:11estimate will probably creep up. So the9:12green the the blue triangles are our9:14estimates and if you add this up, that9:16gets you your your seven trillion. Um9:19but but Jordan, like where are they9:20getting the money from? Right? To answer9:22that question, let's look at um how much9:24money is put to work in the space. So9:26when the investment grade market, it's9:28about 3 trillion of issuance. That's the9:30the um the the the red box here. Um and9:35investment grade is just anyone with a9:37credit rating of triple B uh triple B9:39minus or higher. Um and this is9:42typically going to be like hyperscalers.9:44SpaceX is also investment grade. Um you9:48know typically like great companies,9:49Nvidia's um investment grade and high9:51yield would be companies are a little9:53more risky. Um you know Cororee is9:55technically high yield um and you know9:58various other companies have a lot of10:00debt or a little bit more of a nent10:02business model um and that high yield10:04issuance is uh denoted down here it's10:06about 400 billion and leverage loans10:08which are loans um uh that are of a kind10:13of riskier nature or often used for10:15acquisition private equity is is also10:16400 billion by the way this this is IG10:19bonds this is high yield bonds and then10:22I know loans are kind of separate is10:24not.10:25>> Yeah. So, everything's growing. Both10:26high yield and investment grade is both10:28growing. But take me through again where10:31is the um where where are some of the10:35individual projects that are that are10:37responsible for10:39>> this much debt being [clears throat]10:42raised?10:44>> Which project? Yeah, absolutely. So, um10:46I can give you a few examples. Um so you10:50know this like just since June since10:53June alone right I mean so to put things10:56in context right so we're saying one10:57trillion this year and we think about10:58probably 500 600 billion was was11:01probably maybe around 500 billion was11:03raised through June so we're pretty much11:04on track um but just since June since11:08June 1 until today is October 7 in Asia11:12at least um there's been about 40011:15billion of capital like loans owns bonds11:18of AI you know either raised or11:22announced. So there's been about call it11:25260 billion worth of major deals that11:28have been raised another 150 billion11:29that are reported but not yet11:30documented. You know some some large11:32examples uh blue sky funding which is a11:3524 billion uh deal which comes with a 611:39billion delayed draw term loan as well.11:42Um11:43this is part of the Broadcom11:46um Google Anthropic TPU deal.11:49>> Yeah.11:50>> Um Soapia is another it's a Meta Black11:53Rockck JV uh for El Paso facility was11:58residual value guarantee for Meta. It's11:5912.5 billion. Hut 8 raised uh 4.2512:02billion of paper.12:03>> Yeah.12:04>> Um you know Core like all all the12:07NeoClouds have been added. So the the12:09the the ones you just mentioned are just12:11hypers scale. So now we're on a12:12neocloud. So we've raised a 4.2 billion12:14convert. They raised 2.6 billion of12:16delayed draw term loan in August. Um12:19NBIS just raised a 3.5 billion convert.12:21Zancor in Asia did a three billion term12:24loan, one of the first in Asia.12:26Anthropic did a 15 billion um and uh12:28Pakio Revolver. I mean I I could I could12:31go on and I haven't even gotten to the12:32hyperscalar bots, right? Um yeah. Does12:36that does that help Jordan? I should12:37talk about hyperscar bonds because12:39that's what kind of kicked off this12:40whole firestorm,12:41>> but I could go through I have pages and12:43pages of this.12:45>> Yeah. Well, I mean that's a nice little12:48teaser taste there. Maybe we I'd love to12:51talk about some of the specifics both in12:52terms of the actual projects but also in12:54terms of what they are backstopping or12:57what they are raising debt for, which is13:00to say this is not like strictly for13:03GPUs. I think a lot of people might get13:05confused by this and think like,13:08>> you know, this is strictly just people13:10raising money to buy Nvidia GPUs, but13:12there's so much more that goes into13:14actually building a data center that can13:15accept the GPUs and means that Nvidia as13:19well as all the hyperscalers are13:20motivated to backs stop people across13:22the whole industry.13:26>> Yeah. Yeah. It's it's it's the data13:27center too. And um even though when when13:31you know those who are used to looking13:32at total cost of ownership um you know13:35let me show you one of the total cost of13:37ownership um charts here. So this is the13:40full stack and and what you notice we13:42always say that most of the total cost13:43ownership is the capital cost versus the13:46data center cost. So if you look here13:48the operating cost per hour is 6613:51and um the the capital cost is is $2 out13:55of out of 280 right? So if I do 213 out13:58of 280 that's around 76% opex. So you14:03might you might think the data center is14:04not a large part14:06right um capex is around for a GB300 is14:10around 31 million per megawatt 31 3314:12million per megawatt and then for data14:14center it's about maybe 10 million plus14:17right um so if it's if it's you know if14:21it's 30 mill if it's like 30 billion 1014:25billion that sounds um you know kind of14:29kind of mismatched right um And and and14:33the thing is right the the kind of trick14:34here is that the um the capital side is14:38a shorter term project right it's a14:395-year depreciation period whereas your14:41cap your data center is a 15 year right14:43so on a on a cost of ownership it's like14:457514:47um it's it's 7525 but on a capex basis14:52it is more like 6040 right so the data14:55center is even more because we're14:57building more upfront and this data14:59center is going to last 15 years it's15:00going to do three projects15:02of the AIT capex. Does that help put in15:05context like how much the need for data15:08centers?15:09>> 100%. Yeah. I think a lot of people15:10don't really understand the breakdown in15:12terms of like the chips are expensive.15:14They make up 3/4 of the capital cost of15:16the project. Um15:19you know if you're if you're depending15:21how you model them, right? Which I think15:23maybe is the critical question that's15:26outstanding. um namely the residual15:29value of some of these GPUs that are15:31coming off their 5year life cycle that15:33people are expecting warranties expired.15:36Do they keep running them for year 6, 7,15:388? And how does that how do you view the15:42cash flows for GPUs after the residual15:44value period or the the the modeled15:48lifespan is expired15:51in terms of how it's enabling them to15:53fund future projects?15:56>> Yeah. Yeah. So I think um that's a15:58question people ask a lot um and and16:00I'll go to the you know we've gone16:03through this many times but um you know16:06this this one way to think about it16:08right to directly answer your question16:10um many people that are doing these16:12projects they actually underwrite it to16:14make money on a 5 to six year period. So16:16they're not actually banking on the16:18residual value at all. Um so so and this16:22gets into um the the lay of the land now16:25because most of the clusters are16:28actually um you know built on long-term16:32optakes because that's the only way you16:34can get funding right remember the16:35trendy you need uptake capital and um16:40data centers. So, if you don't have a16:445-year offtake, you're not getting16:45financing. And that's one of the biggest16:47constraints. And so, because of that16:49constraint, because you don't get16:50financing without a long-term contract,16:52everything's typically like pre16:54pre-ordained. So, typically on a 5-year16:56deal, you're going to get an IR of about17:0020 to 30% at least. It's been higher in17:02the past. Um, and what that means is by17:05the time your project is over in 517:08years, you'll have made on your on your17:11um capital investment uh 25% yield per17:14year without banking on residual value,17:17right? So, so most most things are17:19underwritten actually interestingly17:21without regards to what happens after.17:24It's like a bonus.17:26>> Yeah. So, can we back up a second and17:29and talk a little bit about the broader17:32industry? like we just keep saying data17:34center and17:36>> uh and we you know we kind of talk about17:37GPUs but when it comes to actually um17:44analyzing what is being backed up by17:46Nvidia for example17:49>> uh you have this awesome chart that I've17:51got on screen now Nvidia's offbalance17:53sheet commitments where you summarize17:55the $49717:57billion of offbalance sheet backs stop18:00commitments that they've made at this18:02And you go through18:05seven categories here. Uh AI cloud18:08agreements where they are, you know,18:10signing deals directly with NeoClouds,18:12data center leases where they're signing18:14deals with, you know, providers that are18:16building data centers. uh guarantees and18:20LPS guarantees again with more on the18:23energy side, cloud services agreements18:26which are not the same as um necessarily18:29a neocloud agreement in this uh18:32measurement.18:33uh basically an accounting difference18:36there.18:37>> DS center releases for self use18:39different than the previous18:42>> supply and capacity commitments which18:45are primarily around um component18:47suppliers like memory and then finally18:49this residual value guarantee where we18:51haven't assigned any value to it but you18:53know it's a category of something that18:55they're guaranteeing for people. So just18:58give me a lay of the land of all of this19:00like it's it's such a mess. There's so19:03much going on.19:06>> Yeah. Yeah. So, um yeah, I'm happy happy19:08to do that. Um we can go through each of19:10them. I've got some some nice diagrams I19:13can tell you about, you know, why each19:14of them and how we modeled it and you19:16know what what the point is. Um you19:19know, good thing this is going to be on19:20YouTube because you're going to want to19:21slow this down to 0.75x when I go19:24through it for all you listeners. But19:26but before I do that, let me talk about19:27like the why, right? like like why why19:30all this backs stop? Because the problem19:31is a lot of people just kind of skip the19:34why and and they go to the what, right?19:36Um and then they they they look at the19:40what and they're like, "Okay, this is19:41circular financing." Um now now let's19:44talk about the why. Um why are we going19:47to hit an obstacle? Why is it a19:49constraint? Um other than rates going19:51up, which is the market telling you um19:53that that it's going to be tough, right?19:55And and the numbers I just showed you.19:57Um, but there's something more. Um, like20:00I said, Jordan, most of the optakes are20:03five-year. Most of them are hyperscaler20:05backs stop. And what that means is, um,20:08let's say you've got a a a Microsoft20:12Nscale OpenAI deal. What's happening is20:16NScale is selling capacity to Microsoft,20:19which is then selling it to OpenAI. So20:21Microsoft is effectively backstopping20:24that sale because they're guaranteeing20:27um that that capacity will be used right20:30and some uh hypers scalers also do that20:33directly like meta is directly offtaken20:35from neopods for instance there are many20:37deals like that20:38>> and scale nscale needs a guarantee to go20:40out and take out a construction loan to20:43buy and to plug them in and turn them on20:46>> and the guarantee is not good if it20:48comes from open AAI because they don't20:50have an ID rating20:51And so instead of relying on OpenAI to20:55pay the bill, they're going to say if20:57OpenAI doesn't pay the bill, Microsoft's20:59on the hook.21:00>> Exactly.21:01>> So why does Microsoft sign up for this,21:03Dan?21:05>> Well, it's uh it's a good question. Um I21:08think there's internal and external use,21:11right? I think they also take a cut of21:13it. Um it allows them to expand and21:15access capital without directly using21:17their balance sheet. um because it'll21:19show up uh and and you know as an21:21upcoming uh work right a lot of it shows21:23up as services of like contract which is21:27uh under a certain accounting rule I21:28forgot which one is it Jordan ASC 60621:31>> 606 not 842 yeah21:34>> okay 606 is the uh um yeah I was going21:37about to be ASC 404 if I didn't remember21:39that correctly there's21:41>> yeah there's a number of different uh21:43designations the accounting standards21:45committee or commission or something21:47anyway the Um it's it's it's very21:51material to these deals because the21:54definition of a managed service is21:57different than the definition of an21:58operating lease in that a managed22:01service does not end up on your balance22:02sheet. An operating lease does. And so22:06>> the the22:08>> yeah the technical treatment of uh who's22:11responsible for what in the data center22:12is different than the accounting22:13treatment. And you know who can log into22:16what or enter what you know room in the22:18site or touch what or you know22:21[laughter]22:22this is less relevant than how the22:26accountants want to track or keep track22:29of where these assets sit and how they22:32depreciate. Right.22:33>> Yeah. I think Jordan it's kind of like22:35if it talks like a duck quacks like a22:38duck. It's a duck. So and I don't think22:41rage is like no one's stupid right. Um,22:43but it takes a little more work, right?22:45Because the problem is, uh, they don't22:48do it for you. They don't say like, oh,22:49they add up all this like like Nvidia22:51discloses their obligations. And I think22:53in some of the filings, they will22:54disclose it, but it's not in the balance22:56sheet, right? You have to go looking for22:58it, right? And that's part of what we're23:00going to do at at Compute Capital23:02Market, CCM, our new group, who we're23:04hiring for, by the way, we'll have a PSA23:06later. Um, so if you love all things23:09credit and like and we'll show you some23:11charts, like if you like complicated23:12charts, you're gonna love this job. But23:14anyway, to Jordan's point, it it's it's23:16kind of like, you know, it's something23:18that that's going to be hard for them to23:19walk away from, but my point is even23:21that even if it's not debt, like that is23:23not unlimited, right? It's there's going23:24to be a point where rating agencies23:26where investors say, "Hey, you know,23:27this is just too much for you to backs23:29stop." Like, you know, how are you going23:30to backs stop a three trillion like no23:32one has a crystal ball, right?23:34>> Yeah. you specifically use the term um23:36balance sheet as a service. So23:38>> like right now it seems that all of the23:40hyperscalers when we talk about this one23:42trillion of capex are effectively23:43putting their balance sheet to work as a23:45service um that has limits right.23:51>> Yeah.23:51>> Well, yeah. Yeah. So if you look at um23:54you you you look at Amazon, right, for23:56instance. Um so so they did a 25 billion24:00deal uh across six. This is back in24:02July, right? And and it priced, you24:05know, kind of wellwide of uh where their24:06benchmarks were, right? And this was24:08kind of like the canary in the coal24:09mine. I mean, people were definitely24:11aware, but um you you are seeing um the24:14the hyperscaler spreads blow up by quite24:17a bit and the CDS spreads blow up by24:19quite a bit. I'll actually check for you24:21right now. We can maybe we can edit this24:23a little bit. But let me because I want24:24to say a number if that's okay. You guys24:26don't mind. Give me a sec.24:32Model directory.24:35And I can fire up Bloomberg as well.24:38Oops.24:41>> Take your time, dude. No rush.24:43We can just clip this section.24:46>> Yeah.24:46>> Yeah.24:47>> Okay. So, Amazon So, Amazon spread. So,24:50if I look at um if I look at July, I'm24:54just looking at July, right? And their24:5610-year um you know, Z spread.24:59>> Sorry. Restart, dude. I'm I'm only25:00seeing the financing market obstacles.25:02>> Well, I'm I'm not I'm not sharing it.25:04I'm not sharing it.25:04>> Oh, I saw your cursor moving like you25:06were sharing something.25:07>> I know. No, no, no. It's because it's25:08over it's overlapping the deck. So,25:11another thing I interrupted you.25:12>> Sure. Sure. Sure. Sure. Sure. Yeah. So,25:15so um if I look at Amazon, right, for25:19instance, um remember when they issued25:21that bond, their 10-year uh zspread25:24would have been like around 100 basis25:26points. So, roughly, if you have25:28treasuries, it used to be like four four25:30and a half%. and then you add 100 base25:32points or 1% then that means the yield25:34on Amazon debt is about 5 um you know25:385.5 or 5.75 as the case may be and for25:415year that spread was roughly around um25:45his 65 basis points. Um and then you25:48know what what what sort of happened25:50right after is it is it is it blew out25:51probably by like um in the summer it25:54blew out by about probably 10 20 basis25:56points from that from that issue which25:58is kind of a lot of a move for IG. I26:00mean things are relatively under control26:02for now but um you know you've26:05definitely seen like Oracle has moved up26:08quite a bit um since the summer um their26:11entire curve probably moved up like 5026:14basis point 30 50 basis points um so you26:18know you are seeing the debt market take26:20note um and and and you know rating26:23agencies like will see through this26:25stuff.26:27Yeah, walk me through this Oracle26:29increase of 105 basis points and then26:31compare that to Corewave which is well26:35above26:36investment grade at almost 400 basis26:39points um higher.26:42>> Yeah. Yeah. So, so, so Oracle um their26:45spread is going to be around um like26:48looking currently um 10 year I've got26:51them down for about 250 Z spread um26:54versus an Amazon26:56um I've got them only at a spread of um27:00uh 100 110 for for about 10 years. Um27:03but but this chart is a little bit about27:05like as you said the relative you know27:08like how do we compare credit risk? So27:11um what what what this is illustrating27:13is a different type of structures. So um27:16what I'm showing you here is is let's27:18look at the structure where coreweave is27:21um selling an optay contract to meta and27:24the question is how much am I going to27:26charge to lend to coreweave for that. So27:28someone who's lending money to core will27:29say actually this is meta credit risk27:32plus execution. So um the meta meta27:36curve at the time we did this for 5 year27:38is around 97 basis points zspread. Um27:42so um now now the actual lending to27:46corby for this project was around 20027:48right it was 225 so far plus 225 which27:51you know when you do a little bit of27:52math comes in slightly under a zpre of27:54200. Um and so the difference between27:57that 97 base points and the 200 is27:59actually the execution risk because even28:02though meta is is is there right opt or28:06Microsoft is there you still have to28:07stand up the cluster right which is why28:10you know I we should actually rename28:12this to the cluster max risk right28:14because if they don't have a platinum28:15right you know in fact like there's28:18going to come a day Jordan when when28:20it's like this is your SAS rating and28:22this is your cluster max rating right28:23>> it's not And everybody wants to do make28:26stock picks from cluster max. Now you28:29want to raid people's debt from cluster28:31max. Okay. It is one data point.28:33>> Don't tell Ethan he's going to shut us28:35down.28:37>> God then28:40no. We're not we're not branding Neo28:44cloud debt product pricing as cluster28:47backst.28:48>> Sorry. Sorry audience. Jordan does not28:50give investment advice.28:51>> There's a Okay, look there. There's a28:53lot of things that go into building a28:54good quality cluster. I think execution28:56risk on standing up a given data center28:58on a given timeline and then maintaining29:01a contract such that there isn't a29:03cancellation or termination of the29:04contract is typically bare metal and29:07it's typically a very different uh sort29:10of product that is being sold to people29:12than the managed clusters that we're29:14rating. But actually, I'm thinking this29:16through and I'm I'm not completely29:18convinced in what I'm saying, which is29:19that look, a lot of the experience that29:22applies for building a great bare metal29:25service certainly applies to the quality29:27of the cluster. Like people who build29:30reliable clusters are generally able to29:32maintain and deliver on their contract.29:34So, all right, fair enough. We can use29:36this as a rating of some something29:38towards execution risk.29:40>> Well, you you know, Jordan, you say29:41that, but but I've been on a road show29:43lenders. I did New York. I did Singapore29:45and a lot of them have said, "Yeah, this29:47this has been incredibly helpful to29:48their process." So, I think it's like,29:51you know, I mean, I think it's it's29:53look, you know, if someone's made it29:54onto cluster max, go to platinum, like29:56they're unlikely less likely to mess up,29:59right? Like on execution generally, it's30:01just a stamp of approval, right? Just30:02because, you know, you got good grades,30:05you know, you got good doesn't mean30:07you're a genius, but like, you know,30:08probably is more likely than not, right?30:10>> I think it's a a signal and a data30:13point, but managed clusters and large30:15bare metal data centers are uh very30:18different products. And I think just30:19because somebody's in platinum does not30:21necessarily mean that all their data30:22centers are going to be on time and vice30:24versa. Just because somebody's down the30:26list or maybe ranked as like30:27underperforming or something like that30:29does not mean that their data centers30:32will be late or that they will be low30:33quality in the future. So30:35>> I think that there's if if anything I30:37would be um30:40>> asking people to assess a given project30:42or a given site. I mean, we've seen30:44people structure every single site30:46within a NeoCloud's portfolio as an SPV.30:49And I think that's the right way to30:50think about it where there is a30:51different in some cases very different30:54uh level of execution risk as you're30:57describing it for a site in one location31:00on one continent that happens to be you31:03know owned and operated by a given31:05neocloud even though they can go to a31:07completely different continent for a31:09different project the next week the next31:10month and raise debt in a totally31:12different way have a totally different31:13team working on it and you know31:15therefore things are a little bit31:17different. Um Dan, let's move on to31:20>> Well, well, Jordan, I I mean that that's31:22an interesting point, right? Because,31:23you know, a lot of the a lot of the um a31:26lot of the clusters, a lot of the NEOCs31:28we talked to um that are unavailable,31:31they're not unavailable. It's because31:32they're busy doing bare metal, right?31:34You know, like, you know, it's your31:36point, it's not their focus, right? And31:37so, like, yet they've landed large31:39deals. So, it's not like, you know, it's31:41just one data point and some people just31:43don't focus on it. like some some guys31:44who are doing great have landed lots of31:46deals just like hey sorry guys you know31:48we just don't have the product for you31:49to test right now31:50>> and and the whole market has changed I31:52mean we're um we're constantly going31:53back to this but when cluster started31:58buyers had32:00five different providers offering them32:02twoe PC's on 256 GPUs so that they could32:06make their decision in an informed32:07manner and sign up32:10now you got to pick a dance partner 632:12months before the GPU get delivered and32:14you got to make a prepayment to these32:16people that's worth 30 50% of your32:18contract value in some cases. I mean,32:23you said this earlier on on this32:26episode, effectively the buyers are just32:28paying for people's cost of capital32:30directly.32:32And so you're you're you're32:34taking money that you've raised which32:36has a cost of capital and then you're32:37giving to somebody else and you have to32:38pay for their cost of capital. I mean,32:40it's uh,32:42you know, it's it's stacking on top,32:43which is which is driving a lot of32:44people to try to do self-build work, and32:46I think is what it's it's almost like32:49this oscillation. It's like you start32:50out early and you do rental because you32:52don't want it on your balance sheet. You32:53want to move fast. You want the biggest32:54cluster possible. And then you start to32:56grow and you realize, hey, this is a bad32:57deal. Like, I'm kind of getting ripped32:59off. I should own this stuff myself. And33:01then you start doing these self-built33:03sites. And then you reach this limit of33:05what your balance sheet can handle, how33:06much money you can raise. And then you33:08start going, the Frontier Labs go, "Hey33:11guys, I need help. I need to basically33:12mortgage my future right now, and I I33:15need you to build some stuff, and I need33:16to guarantee you some returns on this33:18work because like, hey, I I don't have33:20the balance sheet to do it, and I I need33:21to put your balance sheet to risk." So,33:23it's an interesting curve where the33:25Frontier Labs, the way they're doing33:27deals right now and trying to get33:28everything off their balance sheet, and33:30even the Hyperscalers are trying to get33:31stuff off their balance sheet. It kind33:33of looks like an early stage startup,33:35right? They want the most that they can33:36get right now.33:39>> And I think you you've touched on like33:41the the why, right? Which is where we we33:43kind of got here, right? Like why is33:44there all this back stop? And you33:46described the problems, right? Like you33:47you can't get you got to book six months33:49in advance. You can't get get the33:51compute you want. You got to put 50%33:53down. Um and and and that's that's you33:57know, kind of why they've done the back33:58stop. I mean, let me share I I did up34:00this slide that describes some of the34:02why. Um, and and it's what you said,34:05broaden compute availability, what you34:07just talked about. Um, now supporting34:10the financing market is another thing34:12because we just talked about how unless34:14you have that IG backs stop, you can't34:16do anything. So, it's very hard for like34:18a neoc it there's no incentive as a34:21business model to say like, hey, I'm34:22going to fill my book with half of like34:24various inference as a service. um34:26because your lenders will will go like34:28hey you know I'm not going to if it's34:29not IG credit like I don't understand34:31this business well enough right which34:33which which kind of um you know goes34:35back to this earlier um slide that I had34:39on the um the willing the the the34:42obstacles right they're still on the34:43learning curve they don't understand the34:44business well enough to underwrite34:46anything other than IG and then they34:48don't have the tools right this is like34:49a lot of what Jordan and I are working34:51on um you know Jordan on the tools to34:53assess the operational quality and and34:55myself unassessing like what's happening34:57in the market, right? That's why we have34:59the GPU index. That's that's that's why35:01we're trying to, you know, create tools35:03that lenders need to monitor risk. And35:05you mentioned residual value. This is35:06one of the huge asks we had in in our35:08road show. Um because banks think in35:10terms of residual value that back35:11they're lending its assets. They're used35:13to lending against property, cars, and35:15they have to know residual value, which35:17is very very hard with with GPUs. It's35:19not like you straight line depreciate. I35:21think the residual value is what you can35:22earn on tokens um or what you can easily35:25rent to that.35:26>> And it's so distorted right now because35:27I think in a in a normal market where35:29you have like this this contention35:31between supply and demand where people35:32produce enough GPUs based on the demand35:34for them and they can always go a little35:36bit higher then you have real um pricing35:40and you have real like settlement. Um35:42but what what's happening right now is35:45nobody has an understanding of what you35:47know the true theoretical H10035:49depreciation curve looks like as GB30035:52comes online as Vera Rubin comes online35:53next year and it outperforms that and35:55then you say how can I run an H10035:58profitably well we already have models36:00with inference X that show that you36:03really can't if if a GB300 is36:05outperforming your hoppers by like seven36:07times why would anybody want to buy a36:10marginal hopper over a marginal GB300,36:12you you'd never want it. You you'd want36:14the latest and greatest. But this isn't36:16the question people are asking about36:18Hopper. Instead, the question is, I have36:21no option to purchase a B300 GPU because36:25it's all sold out or I have to wait 636:27months. I want to do research right now.36:29What option do I have to buy? And the36:31answer is renew my H100 contract with36:33four-year-old GPUs for double the price36:35of what I signed it for previously. And36:38this is on screen. And Dan's got our our36:40pricing chart for the one-year uh term36:42length H100 deals that we know about. I36:45mean, they're approaching the prices36:46that they were signed in the first half36:47of 2023 today, like three bucks an hour36:51for H100 GPUs on a one-year term. And um36:55look, if you believe in AG, frankly, I36:57talk to all of these startups a lot and36:59I basically give them the coaching of if37:01you believe in your startup being37:02successful, you have to imagine that37:04somebody's going to do something OpenAI37:06Anthropic can't and it's going to be37:07you. And so you have to, you know,37:10flippantly you have to like believe in37:11AGI here, right? And uh if you don't,37:16>> you know, Jordan, but like I I had a37:17question because you said you said that37:18no one doing inference would ever rent37:20an H100. So like why are people uh why37:23is the price going up then?37:25>> Well, because I think there's so much37:27demand for it that the price goes up37:28strictly because of it's the only option37:31for a lot of people to do. Second of37:33all, there there are some researchers37:35who prefer H100 because it's familiar.37:38They are more productive by the metric37:40of jobs per day rather than, you know,37:42tokens per second. It's like I don't37:45need to spend time debugging. I don't37:46need to port stuff from BF-16 to FP8 or37:49FP4. I'm doing training and, you know, I37:52need numerical stability. I don't care37:54about these low precision data types37:55that are used for inference. I'm doing37:57some new model where again I care about37:59numerical stability. I don't want this38:01risk. So there's there's reasons why38:03people want to do research with hoppers.38:05H100 and H200, but I think the biggest38:08reason why the prices are increasing38:10today is just because GB300 sold out. So38:13people go to the B200, B300, right?38:16Those are sold cascades down.38:19>> Exactly. And I mean, if your answer is38:23wait 6 months and have nothing or get38:25started with H100s today, a startup that38:28just raised money has a pretty easy38:29decision. they go for the H100s they can38:32get access to right now.38:34>> Yeah. Yeah. That makes so much sense.38:36And that's kind of like why when we do38:37this index like we we we do it survey38:40based because like you know there's a38:41story behind every data point, right?38:43This is not just you know going like you38:45know scraping a bunch like you know38:47there is a story and and and those38:49matter and you know what I did Jordan I38:51actually decomposed. Oh do you want to38:52go or I can talk about this?38:53>> Yeah sorry. Well, just one last thing.38:55If you go back to the previous one, like38:57we would have a more academic38:59theoretical way to forecast the39:01appreciation in the H100 pricing curve.39:03If you look at this chart and sometime39:05around September or October of last39:07year, it just kept going down below39:10$180, below a $140. And we actually I39:13remember some deals being signed at that39:15time for like a $130 for an H100. People39:18were offering stuff that low.39:19>> Yeah, that was the bottom. And but those39:20were not great providers though39:23>> for sellers.39:25>> Uh well no the guys who were selling at39:28130 were not great were not like great39:30providers right39:32>> that fair enough. Yeah. Yeah. Exactly.39:34Um but okay but the the point is like I39:36know some sellers at that time who were39:39evaluating $130 and going should I do a39:42one year at a$130 or should I do a39:44two-year at a$125 or something like39:46that. And man, those guys are kicking39:48themselves for not doing a two-year then39:50because they just renewed at like39:52literally some of them at 250. Like way39:55way more than they were paying this time39:57last year. And so anyway, the point is40:00like we we would have a real40:01understanding for H100 res residual40:03value if we had seen the pricing curve40:05depreciate like some other asset like a40:07car or something, right? But40:09four-year-old H100s going into next year40:12are going to be selling for for likely40:14higher prices than they are today if the40:16trend continues. And I mean I expect it40:19to because I expect demand to continue40:21to rise. Um and let me make two more40:24cases for this that are not um based on40:27like an AGI future. You don't need to40:28believe in that. You need to believe in40:31um data center delays or cancellations40:32due to political posturing and40:34moratoriums and execution risk as you40:36were describing which we see a lot of40:39and um and you also have to believe in40:43uh more demand from consumers on you40:46know anthropic and open API endpoints40:48which I think you can believe in even if40:51you don't believe the models are getting40:52better. You just need to look at things40:54like instinct growing or muse from meta40:57and just imagine a world where more40:59people use AI driven products even if41:03the models stop getting better. Right?41:05So the combination of like data centers41:08might take longer to build politicians41:10may get in the way and stop them that41:12will constrain supply41:14more demand for um the existing models41:17and the models get better are all a41:19recipe for increased prices over time. I41:21mean that's that's generally the case41:22that I've been making to a lot of41:24people.41:25>> Yeah. Yeah. And and and the thing is41:27right if you did the scientific version41:29we we started this in 2024 we we41:31actually have a forecast and that41:33actually looked like um you know41:34downward sloping it actually started41:36detaching. we were like within 2% for41:38like a you know a good year and a half41:39right good 18 months but then in it41:41started dislocating for September41:43because the way we calculate the the41:44theoretical depreciation curve is um the41:47relative power the relative compute41:49power of new new GPUs so for instance a41:52GB300 had seven times more inference uh41:55speed and three times more flops right41:58so it would basically those new capacity42:00would push down the price of your old42:01capacity in theory that's what should42:03happen but what we always said is is you42:06know we understand supply we track price42:09but it is hard to track in real time42:12demand. So what we do is we actually42:14flip it on its head and we say let's42:17observe price and let it derive demand.42:19And so you know in this slide this is42:21exactly what I did. Um what I did is I42:23said you know if if you have a balanced42:25if you have a balanced um neutral supply42:28demand then what you see is it42:30depreciating42:32in value or rental price going down42:35you know in a scientific way right in42:38line with new capacity right but what42:40happened is in October right that that42:42that dislocated and what that implies42:44right because if your demand is higher42:45than supply then your price surprised42:48the upside right and so what I did is I42:50I I reverse it and I said Hey, what does42:52this imply for demand supply balance42:54based on price? And what we saw is it42:56blew out completely in late 2025 and42:59it's never looked back. Does this kind43:00of like tie in with what what you've43:02seen, Jordan? Like I assume so. Like43:06maybe not the theoretical stuff, but43:07like what you have started the market.43:09Well, a couple of things happened around43:11that time last year. I think um43:14basically models stopped getting bigger43:17like the uh it took a while for the43:19Chinese models to catch up but up until43:22around this time last year when um you43:26know reasoning models were were taking43:28off people kind of discovered we can43:31trade flops for parameter count and43:34continue to improve the quality of the43:36model with a lot of innovations around43:39RL and post- training and uh they've43:42just continued to exploit that where we43:44may have gotten bigger models. I think43:45we we likely did get bigger models, you43:47know, with the Astra and the Fable43:49generation or the Mythos generation, but43:51we didn't get significantly bigger43:53models that outpaced the improvement in43:56hardware that came when the GB300 NVL7243:58came online, which is to say44:01significantly more memory, memory44:02bandwidth, and FP4 performance were able44:04to keep up with the increased size of44:05the models. And so what that meant is44:08that these these companies could44:10continue increasing price per token on44:13their APIs without necessarily44:15increasing their costs to serve the44:17model significantly.44:19And um they're you know that like the44:23result is just they're way more44:24profitable like way way more profitable.44:27And you you don't even need to know44:28anything about the anthropic open AAI44:30models to to make this case. You just44:33need to look at the Chinese models and44:34you need to look at what the performance44:36of a Deep Seek V3 was on like an H100,44:39H200 or even a GB200 when you know those44:42were being brought online last year and44:44then you need to compare it to Deep Seek44:46V4 or 4.1 Pro or 4.1 Flash or GLM 5.3 or44:51Kimmy K3 where you just look at all the44:54innovations on sparse attention. You44:55look at the parameter count growing a44:57bit like around two times uh or three45:00times in some cases but you know flops45:03not increasing significantly and45:05performance increasing so much and so45:07they can command uh a a pricing premium45:10over the previous tokens because the45:12tokens are more valuable but it's45:14cheaper to serve. you get more45:15throughput per GPU with these. And so45:18people serving Chinese models are45:20enjoying gross margins of 80% right now45:22on a GB300 NVL72.45:25And so it's it's just like it's like45:29look obvious that um there's going to be45:32more demand for this compute this asset.45:35Now45:37the the the question or or what gets45:40confusing to me here in some cases and45:43maybe you have an answer for this is uh45:47how nobody forecasted this like we did45:51our best we understood right and I think45:54even the frontier labs who really deeply45:56understand how these models work did not45:59expect themselves to be this profitable46:01and to be able to serve this many users46:03and to grow this past and so even they46:06were caught by surprise and had to do46:08crazy things and sign big contracts at46:10very high premiums just to get access to46:12GPUs tomorrow to serve their users and46:14to go for the bigger research jobs. Um I46:18don't view this as like an AI fast46:20takeoff, but it seems to be like a46:22revenue and profitability fast takeoff46:24for a lot of these labs.46:27>> Yeah. Yeah. And and and I think you know46:29we we've kind of in our discussion we've46:31really hit on what what is the debate46:33going forward and and it's the46:35willingness and the ability and I think46:36the ability is the profitability. Uh46:39like a lot of people who are new to this46:41are are are don't realize just how46:44profitable it is. Um and that's one of46:47the things that will determine whether46:48we we we overcome this obstacle this46:51kind of um you know funding wall. Well,46:53I think yeah, it's also confusing which46:56um layers of the cake are profitable,46:59which is to say,47:00>> right? Yeah. Yeah. Look at that. Yeah.47:02>> Nvidia clearly had durable profit47:04margins for a long long time, but47:07>> Mhm.47:07>> your Yeah. your layer cake chart is is47:09great to explain like um clearly there47:13has been uh an increase in the47:16profitability of the neoclouds because47:18the amount that they can rent the GPUs47:20for to people has increased recently47:22when the GPU has not there's nothing47:24changed on the operational costs or47:26anything right the people are selling47:28tokens with more efficient models or47:30higher prices and so therefore their47:32profitability is increasing and then my47:35personal experience and I think a lot of47:36other people's experience is value or47:38utility that you're able to derive from47:40using these AI applications is also47:41increasing over time as they get higher47:43quality. And so I don't look at this and47:46say, you know, we're taking from one to47:50give to another. It seems like every47:52layer of the stack in some ways is47:53increasing profitability except for like47:56the server OEMs where it's still a47:58pricing war and they can't 9% margins.48:00[laughter]48:01>> No, no, no. There there's still a48:02there's there's a bit for everyone, but48:03um but um48:05>> the tin box and the people to plug it48:07in. [laughter]48:07>> Well, I I can get into that, but but I48:10want to I want to you know, just so48:12everyone stays um organized, right? I48:15think again like everything we've talked48:17like Jordan and I have talked to so far,48:18you can you can fit into two buckets.48:20The willingness and the ability. What48:21we're just talking about now is the48:23willingness. You know, what is the48:24return? Will the labs remain high? Like48:26how high are the margins? You know, can48:28end users? And earlier on we were48:30talking about the ability, right? We48:31talked about the balance sheet as a48:33service. We talked about Nvidia backs48:36stop which is trying to address the48:37ability. Um, you know, and who's going48:39to fund them. We talked a lot about48:40that. But let's, you know, why don't we48:42why don't we dip into the the48:43willingness, right? Because that's what48:44we were just talking about. Um, and48:46we've done analysis where you can48:48actually earn up to 100 million per48:49megawatt um, generating Frontier tokens48:52off of GB300. It it may be a little, you48:55know, optimistic in terms of48:56utilization, but when this thing costs48:58about 40 million per megawatt to stand49:00up, it it's like, you know, if you're if49:01you're force to sell fable, you make49:03this back in half a year. Um, so it's49:05incredibly profitable. Um, and let's49:08compare that against what NeoClouds are49:10typically renting for a GB3 at least a49:12few months ago. Um, about about 1249:15million per megawatt or 12 million per49:16gigawatt was how much you can how much49:18neoconsor contract. But if you're doing49:21frontier inference, like I said, you can49:23do 100 million per megawatt. And people49:26were were were were shocked, right, when49:27they saw SpaceX and Google renting at 4849:30million per megawatt. But but but you49:33know that that makes sense, right? Like49:35because they're buying here, selling49:36here. So of course they would, but49:38people were used to seeing this. Um so49:41so the the profit is there. Um and and I49:44think Jordan, you mentioned, you know,49:45where where does the profit stay in each49:47layer? So what we did here is we49:49actually organized um the entire uh kind49:52of um business as an income statement.49:55Um so you can think about this uh you49:57can imagine you're you're an anthropic49:59right? So this is your revenue and we50:01put in a couple different ways. So50:03million token about 185 per million50:05token which is a good blend of of of set50:07opus um with cash hits. This is actually50:10our own blended um spending. Um even if50:14Opus is is five five dollar input um the50:17the cash hits will take it down. Um so50:19about 185 per million token which is50:22equivalent if you're generating um I50:25think we put around I want to say 9,00050:26tokens per second. I can't remember what50:28we used um but uh it's equivalent to $3850:32per hour right and then we t we already50:35talked about the cost per hour of of 28050:37and then a Neocloud. So the cost to own50:40the system is 280, but the cost to rent50:42from a NeoCloud is around it's probably50:44around $4 north of $4 by now. $45050:47almost $5. So if we just simplistically50:51look at that, what is a gross margin?50:52It's 90%, isn't it? Right. Because50:54you're selling at 38, your cost is 38050:57and this is your profit. So it's50:59incredibly profitable.51:01>> Yeah.51:02>> At least for Frontiers. Um, by the way,51:05this was really cool. Everyone everyone51:07should should read this like drop51:09everything and read this now cuz cuz um51:11you know Max and team51:12>> we recorded a pod earlier today you51:14don't know. Yeah this is this is my51:16third pod of the day. I've been cheating51:17on you Dan but yeah Max and Andrew went51:19through51:22>> ligamist.51:23[laughter]51:24>> Come on.51:28>> Okay. Well anyway Max and Andrew went51:31through this on the previous episode.51:32Maybe for people who who missed it, do51:34this one again, which is like for the51:36all-in um consumption plans on some of51:39these uh uh51:43on some of these these Yeah. these these51:45businesses, you can get a lot of value,51:48but um it kind of depitable at this51:51layer or not. Kind of depends on how51:52much people actually use the plans that51:54they let people have access to.51:58Yeah, we peg 90% API margin and then I52:01think in the mall we we use like 50 6052:04um but but the question I have is is how52:07much of these people are really using52:0810%. So I'm kind of you there's a whole52:10cottage industry in in China, right,52:12that like takes apples and makes52:14applesauce, right? They take these $2052:16plans and then just sell them as API,52:18right, on the gray market. And then, you52:20know, me, I'm on a $20 plan. Like, I I I52:23actually plan my sessions around like,52:25you know, utilizing I have loops so52:26that, you know, I I I use it even when52:28I'm not there. Um, so I'm definitely on52:30on on this bucket, you know, but I don't52:32know where everyone else is.52:35>> Yeah. I mean, um,52:38Max made the case earlier that I think52:39there's a lot of people who maybe get52:41these through their business or, you52:43know, like they work for and they're52:45just like not really worried about the52:47value trade. Exactly. and you just kind52:49of blend it over a bunch of users and it52:51works out. And the loud people are the52:54ones pushing for more limits who are the52:56ones you're losing money on. But well,52:58you keep them around because you don't53:00know how to separate the wheat from the53:02chaff and only [laughter]53:05only sell the uh all-in plants to the53:09most profitable segment of users, which53:10is the ones that don't use it. like you53:13know the whole thing about um Planet53:15Fitness is marketing it to people53:17marketing gym memberships to people who53:19come once in January for the new year53:21and then never come back that a53:22profitable gym is the one that's empty53:24they don't want anybody to show up I53:26mean kind of uh cynical there but that's53:30the incentive structure for these guys53:32to do all-in plans whereas the exact53:34opposite is true for the uh for token53:36pricing where they're just trying to get53:38people to consume as much as possible.53:39So53:41>> yeah. Yeah. And and then the next53:42question becomes like okay it's so so53:44profitable now but like what people ask53:46is is that going to be the case in the53:48in the future right um let me find our53:50token pricing53:52>> oh gosh where'd I put the token pricing53:55um and people say like how fast is token53:58price dropping53:59>> sorry54:00>> I mean54:01>> not to cut you off but let me Yeah.54:04Okay. Yeah. Yeah. Do this one. Yeah.54:06>> Yeah. Because because my question for54:07you for you for you uh um uh Jordan is54:10like people ask me like how fast is54:12token price falling like what do you54:14mean it's falling it's going up right54:16and and in fact even if it's going flat54:18that's gross margin going up because um54:20while token price been staying constant54:23they've been generating more tokens54:24right and then they're like okay well54:26great so I see frontier malls been54:28holding up but like how fast are open54:30malls falling I'm like well they're54:31going up right and and the other thing54:33is when you look at you know when you54:35look at token expenditure54:36It's like a changing basket. So, so54:38these indexes that track token54:40expenditure like they don't they don't54:43adjust for that in my opinion. I don't54:44know how they work but they don't adjust54:46for that and they don't adjust for54:46increasing you know with AENTIC it's54:48been all about effective pricing cash54:50but I don't know what do you think of54:51what do you think where do you think54:53token pricing is going in the future54:54Jordan?54:55>> Yeah I mean I had a little tweet thread54:57where I went at David Friedberg from the54:59all-in podcast. He didn't respond of55:01course but um a lot of people are saying55:03that the uh there's like this overtaking55:06from open models overtaking closed55:08models and I think the dynamic is just55:10like people are consuming more of55:12everything and they're consuming it at55:14higher prices but you've got the price55:15of input on screen here which I think is55:17a decent metric but you know that little55:19drop down in the middle you55:20>> do the effective right well I you could55:22it's just different approach55:23>> yeah you can do it for cash read which I55:25think is the bulk of our spend is55:26actually on the cash read um tier55:28because you Aenic coding is just so much55:31cache it hit and then output is like the55:34most expensive tier even though you know55:36you don't actually generate a bunch of55:37output that is what takes up all the55:39time on the GPU when you're memory55:41bandwidth constrained. Anyway, the the55:42point is like um I totally agree that55:46model pricing is going up over time on a55:49blended dollar per million token basis.55:51I think that token efficiency is a55:54question mark because you have to trade55:58off efficient ability to solve simple56:01problems with people's56:03um more grandiose expectations as they56:06try to solve more complicated problems56:08that weren't possible before with56:10unlimited tokens and now are possible56:12with a lot of tokens. I don't know how56:14you measure that trade-off, but look,56:17none of these dynamics to me point to56:19anybody deciding, I've had enough of AI,56:21I'm going to use less of it. I've had56:23enough of these models, let's make them56:25worse and less efficient. I've had56:27enough of these new GPUs, let's make the56:28next one worse and and less performant.56:31It all seems to be going the opposite56:33direction, right? They're going to raise56:34prices over time, more demand for it.56:36The GPUs are going to be more efficient,56:37and the models are going to get higher56:39performance. So, we'll see. Um there's a56:43lot of uh ways in which this can break56:45if people go and decide to yolo for the56:4820 trillion parameter model with the new56:50Nv72 and really use it if they start56:53introducing but you know ultra fast mode56:56and and some of these low batch size56:57optimizations and stuff that can jack56:59prices up. But um I think all of that is57:02just like artificial constraints on what57:05to me feels like continued overwhelming57:07demand for these products that drive the57:09usage of the GPUs underlying it.57:12>> Yeah. And and we talked a bit about the57:14the income statement, the explicit one,57:16but what about the income statement57:18above, right? What about the users?57:19Right? That's part of the equation, too.57:21What value are they getting? So, I don't57:23see why anyone would step back when57:25these are the kind of efficiencies57:26they're getting.57:28um like we built an entire database of57:31you know thousands thousands of debt57:33facilities and all these relationships57:35you know pure public domain and and you57:37know it's been myself Oliver Terrence um57:41you know have been doing this and and57:43you know we're all part of current57:44groups because you know we're hiring57:46like I mentioned PSA um but you know it57:49would have taken me four months and a57:51team of four to do like half as good of57:54a job as as we've done so far in57:55building up this this this database and57:57the you know relationship table so far.57:59Um and so these are you know we we we've58:02showed this slide a lot. We we had a pot58:04on this but you know again I don't see58:06why anyone who who is who is paying $658:09to initiate a company would ever go back58:12you know. No58:12>> would you ever go back Jordan like or58:14would you just stop using fast mode?58:16Like what would you58:19[laughter]58:19>> I don't use fast mode. Um58:21>> me neither.58:23>> You'd have to pry this stuff out of my58:24cold dead hands man. Even in my personal58:26life now with um some of these like uh58:30support agents,58:31>> it's like pick a wine. Pick a wine to58:32bring to dinner with my friends. It's58:35like what I'm saying.58:36>> Okay. Mom's for $3. [laughter]58:39I still have a personal touch, Dan.58:42Okay. Like I'm not not writing birthday58:44cards.58:45>> Speak for yourself.58:46>> I'm not that far off the deep end. But58:48yeah, like I I just got this refund on a58:51flight where I was delayed on for 458:52hours a few, you know, there's no way I58:54would sit on the phone trying to get58:56this refund ever again, right? I'm like58:58checking, you know, times at the gym or59:02I'm checking like like booking a flight59:04or looking for hotels or um checking on59:08the grocery list or just check my email,59:10man. Like what came in last night. I59:12mean59:14it's [laughter]59:16it's a little weird but I talked to some59:19friends who are not enlightened to the59:23possibilities that uh having a personal59:26assistant can provide and uh59:30it does feel like uh everybody can never59:34go back once they get the experience of59:36the amount of time that you save and and59:39the feeling of like how much wasted59:42time.59:43>> Yes. Yes. It's it's it's and I think59:45Muse is going to be eye opening. It's59:47like why did I ever sit there for two59:49hours and browse all these like airline59:51booking sites? Why did I ever59:54you know like59:56>> I I think there's going to be that59:57moment,59:57>> right?59:58>> It gets a little egotistical when you1:00:00talk about it this way. I think because1:00:02you know you kind of anthropomorphize1:00:04the agent then you you view people who1:00:06have personal assistance today as kind1:00:07of exploiting like this thing. But1:00:09anyway, my personal view is kind of like1:00:12um1:00:13it's uh1:00:16it's shocking how much time we waste on1:00:21this monotonous stuff. And it's it's1:00:23shocking how,1:00:25you know, some people that I'm close1:00:26with do not value their own time.1:00:31Um, you know, I like Yeah, I'm not going1:00:34to go into all the details there, but uh1:00:37you can1:00:37>> Well, I think I think we're sold on1:00:39willingness, right? We're sold on the,1:00:41you know, AI is like useful, right?1:00:43>> Yeah.1:00:44>> Well, look, I [laughter] felt like we're1:00:47I felt since working at SM analysis, I1:00:48felt like we've been at the tip of the1:00:50spear for adopting a lot of this stuff1:00:52and just like feeling the ability to use1:00:54it, which both makes our research better1:00:56and informs the underlying research that1:00:58motivates some of the data that we see.1:01:00And uh it's really nice to see other1:01:03people that we work with and in my1:01:05personal life adopt these things and1:01:07then start to like regain control over a1:01:10lot of things in their life. It's it's1:01:12it's really cool really cool to see. And1:01:15um I I yeah I just I get really happy1:01:18when I see uh people have these good1:01:20experiences using it and start to get1:01:23motivated to like build a more positive1:01:25and1:01:27you know high quality future together. I1:01:30don't know how to rap, but yeah, talk to1:01:32me about the willingness to finance. Do1:01:33you think1:01:34>> Yeah. Yeah. Well, we we just we just1:01:35talked about that. I think we're sold.1:01:37Like, what I got to convince you on is,1:01:39you know, again, it's willingness and1:01:40ability, right? And and I got to I got1:01:42to sell you on ability, right? Like I I1:01:44I might have scared you with this, but1:01:47>> talk Yeah. Talk to me about the human1:01:48element here, right? Because I think1:01:50it's one thing I've also realized1:01:51working on SM analysis is how much of1:01:54the this stuff that seems so data driven1:01:56all comes down to some human being1:01:58making a decision like Satcha said pause1:02:01therefore pause or Jensen said build1:02:04therefore build or backs stop or1:02:06whatever right so where does this break1:02:08down to1:02:10>> genuinely one of the leaders of these1:02:12companies having this experience using a1:02:14personal assistant or an AI or just like1:02:15having that like personal moment where1:02:17you're like oh my god I can see the math1:02:19discovery before it comes out and gets1:02:20published or I can see this progress1:02:22we're making on this cancer drug or1:02:23whatever and like I just really want to1:02:26bet the company on this and I want to do1:02:28this work because it means something to1:02:29me and this is like my life's work as1:02:31opposed to rationalizing all of the1:02:33financial details to make it work like1:02:35like can you even contend with1:02:38explaining that balance between those1:02:39two things?1:02:41Yeah, but you know like I I I that that1:02:44is one of the big questions because we1:02:46we we covered this is still under1:02:47willingness, right? Because we covered1:02:49we're we're sold, right? And I think1:02:51what you're saying is are other people1:02:53sold too? And and that's kind of what1:02:55this chart is a little bit about, right?1:02:57And I think the question like I I would1:02:59I would make one that's slightly1:03:00different because what I would do is I1:03:02stratified consumer like API tip of the1:03:04spear people like AI natives and then1:03:06the rest of the enterprise, right? So I1:03:08think we're convinced on the sphere tip1:03:09of the spear. I feel like Muse is going1:03:11to prove the case in consumer but I1:03:13don't know in enterprise because I sit1:03:15at these t you know I I I do like a lot1:03:17of these events and I meet a lot of you1:03:18know folks uh you know like we do talks1:03:21and dinners on AI and then you know kind1:03:23of fortune 500 they're like oh like what1:03:26would I use it for like just summarizing1:03:27stuff like no it's so much so you know I1:03:30feel like there's just just a lack of1:03:31discovery in Fortune 500 and lack of1:03:33imagination and a lot of people they1:03:36actually kind of don't come out this1:03:38with an open mind they come out this1:03:39like oh great AI like how do I fire1:03:41people, right? You know, like that's1:03:43that's all they think about, right? And1:03:45it's and it's like so I just think the1:03:46enterprise is it's it's it's of the1:03:49three it's it's the one I don't know the1:03:52most. I don't know because it's it's1:03:53it's so much to like does the C say hey1:03:56let's embrace it let's trust Amazon1:03:57let's go full blast or they go like no1:03:59AI in your work like like what do you1:04:01have you talked to fortune like where do1:04:03you think they are on this? Well, I1:04:05think the dynamic that I imagine happens1:04:07with the Fortune 500 is that the Fortune1:04:10500 just slowly gets replaced with1:04:12companies that are all in on AI.1:04:16>> Well, look, look at the top 10.1:04:18>> Possibly true.1:04:20>> Look at the top 10 companies in the1:04:21world by market cap, right? I mean,1:04:23these are all like nine out of 10 of1:04:25them are AI or semiconductor related or1:04:27data center related at this point,1:04:28right? Like when we say hypers scale,1:04:30all five of them are in there. Then1:04:31you've got Nvidia, Broadcom, TSMC, the1:04:34memory guys are are getting close to the1:04:36top 10. So Fortune 500, I think we use1:04:39that term to describe if you if you1:04:42actually mean the 500 biggest businesses1:04:44in the world measured by revenue. I1:04:45think a lot of them are just going to1:04:48>> maybe I should say corporate America,1:04:50you know,1:04:51>> legacy businesses, corporate America,1:04:53>> legacy business. Yeah.1:04:54>> Government, just like people that are1:04:56slow to adopt AI and and fair enough. I1:04:58think there are lots of people who will1:05:00be a lot slower to adopt AI than the uh1:05:03biggest companies in the world which I1:05:05would lump in all the hyperscalers into1:05:07their soon to be the AI labs if1:05:09anthropic goes and IPOs at $2 trillion1:05:11they're going to be you know a top 251:05:13company in the world in market cap or1:05:15something like that right so I mean1:05:17that's that's the the proof point right1:05:19there but um you have to make this case1:05:23of like going out two three years do you1:05:25expect that this next crop of NEO labs1:05:27that pursuing all of these new um1:05:31innovations just like completely all get1:05:34swallowed up and acquired and none of1:05:35them pan out. I I actually expect many1:05:37of them to do quite well across a number1:05:40of different industries continue to grow1:05:42and if they do get acquired to probably1:05:44help those businesses grow and and those1:05:46you know take over and so I you know I1:05:48don't view like um a future where uh1:05:52Visa and Bank of America and AT&T are1:05:56just like so critical to uh the success1:05:59or failure of the AI uh native companies1:06:03that are uh trying to grow. Maybe what1:06:06I'm most interested in is the unlock of1:06:08capital that uh the companies that may1:06:11have been funding those companies in the1:06:13in the past may turn and shift over and1:06:16say, "Hey, I I work in infrastructure1:06:18finance. I work in credit and for the1:06:19next round of this, I'm not actually1:06:21going to go um do another uh you know,1:06:26project for another bridge or for um you1:06:30know, some debt for some credit card1:06:32payments provider or something like1:06:33that. I'm actually going to go into this1:06:35AI thing because I think that's upstream1:06:37of all of those other businesses I've1:06:38done and and it actually is1:06:40infrastructure and I'm going to get1:06:41these returns and I'm going to have1:06:42these guarantees and things like that1:06:44and and that's where I ask you because1:06:47you know these guys better than me about1:06:49the human element of like for people who1:06:51really just don't understand this and1:06:52don't have the experience of using AI1:06:55who represent so much you know capital1:06:58that's still sitting on the sidelines.1:07:00We we joke about stuff like Nvidia's1:07:03revenue1:07:05at if it hits $500 billion1:07:08uh is going to be like 2.5% of the M21:07:11money supply in the US, right? Uh just1:07:14unbelievable numbers. We're playing with1:07:16trillions of dollars of capex we talked1:07:18about earlier in this episode. But1:07:20there's still so much left over like1:07:22what's the other 97.5% of the M2 money1:07:25supply? So,1:07:26>> well, um, like let me, yeah, I can I can1:07:29share some some stuff because now we're1:07:31on kind of the the ability, right? Um,1:07:34and and and again, one of the obstacles1:07:36I talked about earlier was um, you know,1:07:39getting people like I guess I'll show1:07:40you. Oops, it's up here. Um, let me go.1:07:45One of the key things on the ability is1:07:48is like understanding and one of the key1:07:50obstacles is is lenders on the learning1:07:51curve like so you know the people making1:07:53decisions on lending like right now1:07:55they're doing it because like okay it's1:07:57Nvidia credit risk it's Microsoft credit1:07:59risk like I don't have to know but1:08:00that's going to change and that's why1:08:02all these backups are there so they have1:08:04time to understand this industry but but1:08:06let me let me size the problem for you a1:08:08little bit um you know let's let's talk1:08:11about um you know what where where where1:08:13this is going right like Obviously, it's1:08:15going to be the second biggest um debt1:08:18financing market, right? It's already1:08:20surpassed auto loans, surpassed1:08:21everything1:08:23>> um this year. Um but everyone's done it1:08:25without having to have residual value1:08:27model for GPUs because it's in back stop1:08:29and it's only second to the US mortgage1:08:31back market. Um1:08:33>> so move on from that chart. We've got1:08:35some audio only listeners, but I I just1:08:38want to describe it. Uh it's such a1:08:40compelling chart. auto loans, student1:08:42loans, credit cards, non- agency, CNBS,1:08:45HELOC, these um1:08:49US assetbacked credit markets are just1:08:52like flat lines like slowly going up to1:08:54the right as people1:08:55>> you can't you can't even see them going1:08:57up actually.1:08:58>> Yeah. And it it they're just five flat1:09:01lines at the bottom of the chart and1:09:02then it's just this massive exponential1:09:05[laughter]1:09:06line up to the right. more of a straight1:09:07line, but yeah, straight line up to the1:09:09right called AI debt financing. And just1:09:12explain to me the implications of AI1:09:14debt financing in 2026 crossing over1:09:16auto loans and student loans on this1:09:18chart.1:09:20Um the implication is and we already see1:09:22it. We already see long long end we1:09:25already see 10ear rates going up. We1:09:26already see credit spreads going up. we1:09:28already see um an industry and and the1:09:30reason the credit market um it it's1:09:33going to be hard because they don't have1:09:34the tools to understand like because1:09:35it's it's all coming from like uh1:09:38different fronts. It's a war on so many1:09:39fronts because you're getting hit by1:09:41hypers scale or neocloud data center1:09:43everyone at the same time and it's like1:09:46it's like a zombie apocalypse but you1:09:48don't know how many waves of zombies are1:09:49coming right you know1:09:52[laughter] it's like a terminator you1:09:54know how many terminators1:09:57give me your money [laughter] anyway1:10:00they're money zombies coming right and1:10:02and you don't know whether whether1:10:03you're on wave five or wave 50 right um1:10:06so So, so, um, you know, and and to put1:10:09things in in in context as well, I've1:10:12got a slide here that shows you, um,1:10:16what this is, like I told you, I showed1:10:18you what it is in terms of like the1:10:19corporate bond market, right? And that's1:10:22that's just up here. Um, but on the next1:10:24slide, let's talk about overall US1:10:26issuance. Um, so it's 1 trillion. Um,1:10:29and it's going to cross 1 trillion. So,1:10:32so if you think about it, right, the1:10:33total US corp, the US issuance, this is1:10:36federal, which is in the blue. Um, you1:10:38know, for those of you listening, that's1:10:39about five or six trillion a year. You1:10:41got MBS, RMBBS. Um, just looking1:10:43ballparking from that, it's looks like 31:10:46or 4 trillion a year of issuance. Um,1:10:48maybe two to three corporate bonds, we1:10:51already talked about that about 3 to 41:10:53trillion. Um, total 12.7 trillion. and1:10:55and and 2026 is the last year that AI1:10:58debt is going to be less than 10% of all1:11:00issuance, right? All issuance in the US.1:11:04Um so so it's going to take believers.1:11:06Um and and how how do we handicap that,1:11:09right? Um you know, well, this is maybe1:11:14what we can do is we can we can jump1:11:16into like, you know, back to the1:11:17question you asked earlier. It's like1:11:18why why is Nvidia doing these back1:11:20stops, right? Um, and let's just look at1:11:24um, well, sorry, I'm going to go. Okay,1:11:27here. Yeah, like why do they have all1:11:28these different different back stops?1:11:30And again, it's it's they see this gap1:11:32coming and and and they're figuring out1:11:33how to plug it. Um, let's talk a bit1:11:36about their um, AICP, right? So, by the1:11:39way, you know, to start off, we1:11:41projected Nvidia's balance sheet and we1:11:43were looking at their ability to1:11:44actually continue doing these back stop1:11:46and how much they can do. Um but let's1:11:48talk about like an example of one of the1:11:49backs stops and how it accomplishes1:11:50those goals. This is the AICP program.1:11:53Um you know this actually started in1:11:55Asia. The first one was firm well1:11:57actually first was sharing AI then firm1:11:58has followed with another one. Um you1:12:00know there's been Alani there's been1:12:02many others. GMI did one. And the whole1:12:04idea here is the way this works is um1:12:08Nvidia goes to Neocon and says, "Hey,1:12:10I'll give you a guarantee." Like if you1:12:12can't rent your cluster, I'll rent it1:12:14from you. Um and I'll give you that that1:12:17guarantee for 6 years on on 325 on1:12:19average. Um it's the the actual um table1:12:22is different. Each year is a different1:12:24uh minimum, but if you can't find1:12:26clients, you can always go to Nvidia and1:12:28rent to them at at this price. Now, the1:12:31intention is not for that to actually1:12:32happen. The idea is you use this as a1:12:35guarantee to then get financing.1:12:37Remember the trendy right? So now you1:12:39have an offtake and now you can get1:12:40financing and then with that you can get1:12:42data center. Um with this in mind1:12:44lenders lenders you see this this little1:12:46thing that it's priceless Nvidia risk.1:12:48So it says hey you know I'm really1:12:50lending to Nvidia plus execution risk.1:12:52Um, and then what that allows allows the1:12:55NeoCloud to do is, oh, now I don't have1:12:57to rent a five-year offtakes to whoever.1:13:00I can do um, you know, tokens of1:13:03service. I can do all these things. It1:13:04really opens up the market. That's the1:13:06whole idea. Um, and they don't need to1:13:08be IG. That's the important part. Um but1:13:11this is only like an intermediate1:13:13intermediate step because what it does1:13:15is lenders have to um understand that1:13:18they're they're also taking the platform1:13:20risk whether neo can build a good book1:13:21clients and they'll get to see it as a1:13:23lenders neoc. like, hey, can they build1:13:25an independent business outside the1:13:26optic? They get to understand the1:13:28customer profile, the churn rates, the1:13:30cancellation rates, and they get to, you1:13:32know, kind of do like a free look or a1:13:34dry run, uh, practice run until the real1:13:37thing. And that's why this is kind of1:13:38really important why they're doing it.1:13:40Does that make sense, Jordan? Any1:13:41questions?1:13:41>> Makes perfect sense. Yeah, man. Uh,1:13:43balance sheet as a service. Maybe the1:13:45the one thing to talk about is like they1:13:48at some point introduced a um a take on1:13:52the percentage of revenue that the1:13:53NeoCloud1:13:54>> I forgot about that1:13:55>> above the minimum uh that they promised1:13:58>> and then they took that away and1:14:00>> you know1:14:00>> No, no, no. It's still there, right?1:14:02>> Okay. Explain that. Yeah.1:14:04>> Yeah. Yeah. So, so it it's also a1:14:06revenue model as well. Um and and it's1:14:09fair, you know, we're not communists1:14:11after all. Um, so in exchange for the1:14:13backs stop, Nvidia gets a 50% revenue1:14:16share above the backs stop. Um, and so1:14:18for example, let's say you rent at $6 or1:14:21let's make it easy, $668. Then you the1:14:24difference is $3. So you got to pay uh1:14:26150 to Nvidia out of out of the six in1:14:29exchange for that. So it's actually a1:14:31meaningful um uh revenue share for1:14:34NVIDIA. And and if you look at the other1:14:35structures they've done, this is the1:14:36Lambda structure. um they're also taking1:14:39a revenue share out of Lambda in1:14:41exchange for that that that guarantee.1:14:43Um so you know and another third reason1:14:46Jordan by the way one of these back1:14:48stops do is is not just facilitate um1:14:50these deals happen and allow you know in1:14:53this instance they're allowing lambda to1:14:54to get financing because anthropics not1:14:56yet IPO is not IG right but with Nvidia1:14:59supporting the lease um for the data1:15:02center they can actually get this done1:15:04but there's another reason right um1:15:05let's look at ports pike this is a1:15:07little complicated um but the the bottom1:15:10line here is Nvidia is is giving 1051:15:13billion dollars of residual value1:15:14guarantees um to the ports Pike data1:15:17center complex in order to support this.1:15:19But the catch is it's going to so what1:15:21they've done is they've guaranteed that1:15:234.25 gigawatts or up to nine um almost1:15:26nine across two phases will be running1:15:29Nvidia GPUs. So it also locks in. It's1:15:32like it's like a kind of a land grab is1:15:34another another reason for it. Um, and1:15:37so we looked at we looked at all these1:15:38different models and I'll talk about one1:15:40last one and this is the kind of the1:15:42most interesting one is a 500 billion1:15:44capital partnership. Um, because this1:15:46gets into how can we actually you know1:15:48to your point right you talked about1:15:50lenders as as one but it's actually very1:15:53different. There's bank lenders,1:15:54insurance lenders, they don't want to1:15:55take a lot of risks. You've got private1:15:56equity they want to take a lot of risk.1:15:58You got private credit they're in the1:15:59middle.1:15:59>> They all syndicate into these things as1:16:01well. Right. Well, well, it's creating1:16:04different products for different um u1:16:07different types of investors. So, so the1:16:09500 billion capital partnership, this is1:16:11what we think it could look like. And1:16:12what you do is you can have an AI lab.1:16:14They don't have to be IG, but what you1:16:16do is you trunch the risk. So, you have1:16:18equity and then you've got a B piece1:16:20which which doesn't have any residual1:16:22value guarantee and then you have an A11:16:24A2. You can put one senior to the other1:16:26and then Nvidia is giving a 25% residual1:16:28value guarantee in this 500 billion1:16:30partnership. maybe you can attach this1:16:3125% to this trunch or this trunch but1:16:34because let's say there are losses in1:16:36this program you know in this business1:16:38then the equity will take the loss first1:16:40then the BPS and then part of the A2 and1:16:43so it will have to go all the way down1:16:45to to to this the loss have to be so1:16:47severe to affect the senior trunch which1:16:49is how you can create different risk1:16:51profiles and open up the market um but1:16:54anyway all this depends right on Nvidia1:16:56having the capacity to do this so we1:16:59analyze the balance sheet1:17:00you know, um and and you know, if you1:17:03you kind of love talking about this, you1:17:05know, um this is all we all we're going1:17:06to do. Um but you know, here's the kind1:17:09of conclusion of this, Jordan, right? Um1:17:11they we think they have the capacity,1:17:13this is not our forecast, but we think1:17:14they have the capacity to support up to1:17:1646 gawatt of capacity. We haven't done1:17:19it for the hyperscalers yet, but that's1:17:21still not enough. There's 240 gawatts is1:17:24going to come in. And so what I'm saying1:17:26here is is the lending market has to1:17:29evolve to take risk outside of back1:17:31stops. And that's what you know we're1:17:34all about understanding that question1:17:36and it's it's really the biggest1:17:37question of our time to be honest which1:17:39is why we're starting an entire group1:17:41around this question.1:17:44>> Makes sense man. Yeah. If there's two1:17:46tangible things I'm taking away from1:17:47this it's one more education to the1:17:49lenders on what they're actually1:17:51>> getting into here. um you know what the1:17:53real deal structures are like and who1:17:56the end user customers are and why1:17:57they're structuring things this this1:17:59way. And then two um1:18:03well I guess two is is twofold. Uh1:18:07it's it's always incredible to see the1:18:09size of of these numbers and how much1:18:11has already been done but also how much1:18:14bigger it needs to get from here. Uh1:18:17>> this is incremental capacity ads. So1:18:20>> incremental 60 some odd gigawatts of1:18:23capacity hats in 2029. Yeah. Let's see.1:18:26>> When we first start this journey, global1:18:28data center capacity is like 40 gawatt1:18:30maybe. And now we're adding like 40 a1:18:33year. Can you imagine that?1:18:38>> I can. Yeah. [laughter]1:18:41>> Yeah.1:18:42>> Excited for1:18:43>> your usage. Yeah. I know. I know. I1:18:45know. I was looking at the numbers.1:18:47>> Let's let's wrap up one one little funny1:18:49story. When I got started working at1:18:50semi analysis, I did not understand like1:18:52anything about financing and how this1:18:54stuff works. I was trying to figure out1:18:56who a lot of these lenders were, what1:18:58private equity was. I I'd known these1:19:00names, things like that. And I I1:19:02remember working on this one deal. Uh I1:19:04won't say, you know, who was setting up1:19:06the deal, but it was like private1:19:06equity.1:19:07>> Do you think DA was like a musical at1:19:09some point? [laughter]1:19:17>> [laughter]1:19:18>> Yeah. Yeah. Um,1:19:21yeah. I still get confused when you talk1:19:23about project versus workbook IRRa. I I1:19:25don't know. I just know the project1:19:27versus project versus equity IRR.1:19:30>> Equity IR. Yeah. Okay. Okay. I I I just1:19:33know irr is you spend enough time with1:19:35Dan, you know, irr is important.1:19:37>> And it doesn't get easier to say IRR.1:19:40[laughter]1:19:41>> IRR. Yeah,1:19:42>> that's that was the pirate, right? Dan1:19:45when he started like he needs an eye1:19:46patch because he's talking about his eye1:19:48or RR.1:19:49>> Anyway,1:19:51>> um okay, let me tell this quick story1:19:52before we wrap. So, this private equity1:19:55company was setting up this deal. We had1:19:56to go do this uh technical DD on it. We1:19:59had to present to some of the lenders1:20:00and the lenders were like, you know,1:20:02from all sorts of different institutions1:20:05and uh I remember at the beginning1:20:07learning like, oh, private equity1:20:08doesn't have any money for themselves.1:20:11They have lots of ass management, but1:20:13they set up the deals and then the1:20:14lenders come in and they actually like1:20:15fund the deal and private equity like1:20:17makes carry off of the the top of that,1:20:20right? Anybody in PE who's listening to1:20:22this, it's like, yeah, of course that's1:20:23the way. I don't know this. [laughter]1:20:25Anyway, um and so I had this meeting1:20:28with this pension fund uh to explain1:20:30like the work that we did and I remember1:20:32sitting on the call. It was the Ontario1:20:34Teachers Pension Fund [laughter]1:20:37and I was like, "Oh,1:20:40it's my mom's money." [laughter]1:20:43Like [gasps]1:20:44this is what's you know getting in here.1:20:46So anyway, the kind of this whole like1:20:50loop back to thinking, okay, yeah, yeah,1:20:52like Canadian pension, Ontario teachers1:20:55pension, lots of stuff in Quebec. Like1:20:57I'm Canadian. So anyway, that it it like1:21:00I ended up feeling like kind of proud of1:21:01the work that we did. Not like I was1:21:03helping these people who were just like1:21:05rent seekers trying to make a little bit1:21:06of money off of this. Like it actually1:21:08is pretty hard to set this things these1:21:09things up. Make people comfortable1:21:11understanding what they're investing in,1:21:12explaining the risks, uh understanding1:21:14the technology, what's the return, blah1:21:16blah blah. And uh yeah, there's going to1:21:18be more to do in the future because lots1:21:20of people just like that are probably1:21:21getting in and evaluating their first AI1:21:24cloud opportunity or neo cloud uh deal.1:21:27and it's going to be a big part of the1:21:29future.1:21:30>> We're we're on a learning curve and1:21:31that's, you know, I think I think the1:21:33more we we help people get up there, uh,1:21:35you know, talk about everything, I think1:21:37the more it it helps everyone um make1:21:40make great decisions.1:21:42>> Yeah. Cool, man. Okay.1:21:43>> You know, one one last joke. You know,1:21:45why why it's IRR?1:21:47>> No. Why?1:21:48>> Cuz pirate equity.1:21:51Pirate equity.1:22:02Hey everyone, if you're still listening1:22:04after this hour and 15 minutee long1:22:06conversation about credit and markets,1:22:07you should consider applying to work at1:22:09semi analysis on credit and markets. Dan1:22:11is a very fun guy to work with. Dan, who1:22:13are you hiring for?1:22:15>> Well, we're we're we're hiring credit1:22:17analysts in New York and Singapore. So,1:22:19we're hiring junior credit analysts. So,1:22:21anyone who's done um one to three years1:22:24of experience could be investment1:22:25banking uh e GC well DCM uh you know1:22:30anyone who's focused on that. Um, also,1:22:33uh, we're looking for a more senior1:22:35credit analyst, anyone who's been, um,1:22:37with desk analyst or investment analysts1:22:40on the buy side. And also a credit1:22:42specialist, someone who's quite in touch1:22:43with markets, has a great rolodex, um,1:22:46worked in credit trading, credit sales,1:22:49um, you know, buy side or sell side. So,1:22:52um, love to have you apply. Just click1:22:54here to find out more. I also need a1:22:55video editor. If anybody knows about a1:22:57video editor, shoot me a text.1:23:01Okay, this is your third interruption.1:23:03We're experimenting with ads on the Semi1:23:05analysis weekly podcast. We also need a1:23:07video editor. A cautious hiring. Yeah.1:23:09So, if you know how to edit videos,1:23:11please hit me up. We're looking for1:23:13people. [laughter]1:23:15Please hit me up.
15,932 words · 2267 lines
