Ep. 037 - Who's Funding the $11 Trillion AI Buildout? (Capital Markets)

For our first ever advertisement, we're going to be doing a lot more of this on future SemiAnalysis Weekly episodes. SemiAnalysis is a reader-funded publication. Over to Dan. What are we advertising this week?

Well, we noticed that 90% of our viewers aren't subscribed. So if you could hit the like button and subscribe and be sure to hit the alerts. Seriously though, the real ad, we're hiring for a new group, which is called Compute Capital Markets. And we're going to look at the loan demand, how the financing markets are evolving for financing GPUs. And we're going to look at markets, which talks about the tools that lenders and everyone in the market need to evaluate how GPU financing is going, how the GPU industry is going. That's rental pricing, residual value, all those things that come with it. And so we're looking for credit analysts. We're going to hire 4 slots. We're looking for Singapore and New York City. We're looking for a junior credit analyst, about perhaps 1 to 3 years of experience in the credit space, investment banking in Singapore or the US. We're also looking for a credit specialist. That could be someone who's been a little more experienced, probably 4 or 5 years credit trading, credit sales, or credit investing. IG, high yield would be great. And also a credit analyst. Anyone who's been a desk analyst, buy side, sell side, you know, we're looking for people who are credit-minded. I think that if you enjoy this episode that's going to come and all these structures and, you know, all these Z spreads and T spreads, I think you're going to really like this job. So, you know, click the apply button right here.

All right, roll the credits. Hello everyone. Welcome back to Semi Analysis Weekly. In this episode, we got Dan. We are going to talk about everything to do with capital and markets, how we are funding the buildout right now. Dan has been digging into everything credit, capital markets, compute. Specifically, we put out an article on, I mean, it's called NVIDIA's Access to the Universe: Heads I Windtail to Where we talked about the $11 trillion AI buildout, NVIDIA's backstop economics, and the limits of NVIDIA's balance sheet. So, you know, with that as a backdrop, we're going to talk through the contents of that article, but at the pace things are moving right now, that's like already a month old. And so, yeah, without further ado, Dan, why don't you lead us in talking a little bit about NVIDIA's backstop universe? set up this article and why you dug into this expanded view of NVIDIA's balance sheet.

Yeah. Yeah. So let me show you a bit about NVIDIA's backstop universe. And indeed it was just a month ago, but things are moving so fast at this point. Funded announced is around about like $100 billion each month. And this is how we looked at the balance sheet. And the whole premise is NVIDIA's been giving a lot of backstops and the question is why, right? There hasn't been a lot of visibility on how much issuance is going to come in or how much funding needs there is. So we defined that for the first time. And what we did is we looked at the funding needs and we talked about whether NVIDIA's balance sheet would be strong enough to support backstopping a lot of these funding needs. And we'll talk about that much more later, but I think it's worth going through quickly the constraints and, and what, what sort of, sort of led us to this point. Um, so if I go back to, um, if we think about the constraints, right, Jordan, and, you know, you, you asked me earlier about the, um, the Trinity, right?

You remember that? Yeah.

Yeah. We used to talk a lot about the Trinity and the Trinity, the AI Project Trinity is basically capital, offtake, and, uh, data center. And Jordan earlier, he was kind of asking me like, well, what about compute, right? Isn't that the 4th leg? Which is a good point because it used to be a pretty big short— a pretty big pinch point. And it used to be kind of a worsening pinch point. But what's happened since then, right? Memory was approximately one of the hugest pinch points, right? So let's talk about constraints and the constraints have been moving. They've been moving from compute constraints to data center constraints, and now financing is a new constraint. So why are those more acute, right? So what's happened with memory, memory has been a shortage. You can see that AI's percent of DRAM wafers was going to be 70%. And then you can see that on the logic side, accelerator is an ever higher share. But what happened recently is there has been despeccing. So let me show you what happened. So if you look at accelerator perspective, this used to be the pattern. If you look at GPU ASP, it had been going up because of memory and it was set to go up on Ruben Ultra, but that pattern actually stopped in its tracks. And what we got instead was despeccing.

Why?

For really 2 reasons, right? Number one, ROI, and I'll just kind of flick to a quick bridge chart, right? And so with this one, what we saw is the cost was increasing for the VRNVL72, and what they did to offset that is decrease SoC count, but it's not just about price, it's about availability. The issue here is that there just simply hasn't been enough memory to go around to actually ship racks. So the despec was necessary to deconstrain that. So that's the chip constraint. Let's move on to the data center constraint. That was one of the first data constraints, the first constraints that people worried about. But we've found a lot of creative solutions around that. We've mustered a lot of capital to build data centers. We've activated more powered sites, we've refurbished old sites, we've done a lot of behind the meter to the point where by 2027, 2028, it's much more balanced and this is no longer an obstacle. So what are the new constraints? So what we talk a lot about now is funding, and that's a constraint people have learned about in just the last few months. I'll show you how we calculated the overall constraint. Let's take a look. This is the cumulative total CapEx. It's $11 trillion of CapEx that needs to be spent by 2029.

Yeah, Dan, explain this chart just because I think a lot of people are throwing around the term of $1 trillion of hyperscaler CapEx. How do you get to $11.3 trillion on this chart?

Yeah, so $11.3 trillion is cumulative CapEx from 2028, sorry, from 2024 through 2029. And so it's escalating each year. The way we did this, because we track all the GPUs, we track all shipments, we track total cost of ownership of the racks and the build materials. And that's not just across GPUs, it's also XPUs, it's TPUs, AWS Trainium. And if you take the sum product of the shipments across the total cost, which includes the server, includes networking, includes CPUs, includes everything. You actually get to that $11.3 trillion. And by the way, you also mentioned hyperscaler CapEx. So that's actually a subset of this. If you look at the legend, this is AI IT CapEx and data center CapEx. And a lot of that CapEx is not incurred by hyperscalers themselves. It's incurred by independent data center companies. So that's part of the equation of why $1 trillion in a year of hyperscaler CapEx becomes $11 trillion because it's hyperscaler and NeoCloud, and then it's IT and data center.

Yeah. And this is not covering necessarily other things in the supply chain, which we do cover, which Dylan has talked about, which is, you know, power plants, industrial equipment, uh, stuff in the fabs that is, you know, a precursor to AI, IT in this measurement here, like, you know, fab equipment or just new— like there's a lot of stuff going into the buildout right now. Um, you said that there's a constraint related to funding.

Yeah.

Talk me through how people are actually funding this buildout and how is that, where is everybody finding the capital?

Yeah, so it's a great question. And what we said is there's going to be $7 trillion of debt outstanding by 2029. And so what does that work out to? It works out to, what we did here is I put the net additions, the net new funding, So it works out to $1 trillion this year. It's a little flat because whenever we forecast stuff forward, we tend not to add capacity and deliver it till it's confirmed, right? But, and then there'll still be, there'll be some debt that rolls off because it's all amortizing. But, you know, I would expect that this estimate will probably creep up. So the green, the blue triangles are our estimates. And if you add this up, that gets you your $7 trillion. But Jordan, like, where are they getting the money from, right? To answer that question, let's look at how much money is put to work in this space. So in the investment-grade market, it's about $3 trillion of issuance. That's the red box here. And investment-grade is just anyone with a credit rating of BBB, BBB- or higher. And this is typically going to be like hyperscalers. SpaceX is also investment-grade. typically like great companies, Nvidia's investment grade, and then high yield would be companies that are a little more risky. CoreWeave is technically high yield and various other companies have a lot of debt or are a little bit more of a nascent business model. And that high yield issuance is denoted down here. It's about $400 billion. Then leveraged loans, which are loans that are of a kind of riskier nature or often used for acquisition, private equity is also $400 billion. By the way, this is IG bonds. This is high yield bonds and then high yield loans are kind of separate.

Yeah. So everything's growing, both high yield and investment grade is both growing. But take me through again, where is the, where are some of the individual projects that are responsible for this much debt being raised?

Which project? Yeah, absolutely. So I can give you a few examples. So, you know, this, like, just since June, since June alone, right? I mean, so to put things in context, right? So we're saying $1 trillion this year. And when we think about probably $500 or $600 billion was, was probably maybe around $500 billion was raised through June. So we're pretty much on track. Um, but just since June, since June 1st until today's October 7th in Asia, at least, um, there's been about $400 billion. Of capital, like loans, bonds of AI, you know, either raised or announced. So there's been about, call it $260 billion worth of major deals that have been raised, another $150 billion that are reported but not yet documented. You know, some, some large examples, uh, Bluesky Funding, which is a $24 billion, uh, deal, which comes with a $6 billion delayed draw term loan as well. Um, this is part of the Broadcom, um, Google Anthropic TPU deal. Um, Sopevia is another, it's a Meta BlackRock JV, uh, for El Paso facility with residual value guarantee for Meta. It's $12.5 billion. Hut 8 raised, uh, $4.25 billion of paper. Um, you know, Qoriv, like all, all the NeoClouds have been added. So the ones I just mentioned are just hyperscalers. So now we're on a NeoCloud. So Corvi raised a $4.2 billion convert. They raised $2.6 billion of delayed draw term loan in August. Nebius just raised a $3.5 billion convert. Zancor in Asia did a $3 billion term loan, one of the first in Asia. Anthropic did a $15 billion pre-IPO revolver. I mean, I could go on and I haven't even gotten to the hyperscaler bonds, right? Yeah. Does that help, Jordan? I should talk about the HyperScout bonds because that's what kind of kicked off this whole firestorm. But yeah, I could go through, I have pages and pages of this.

Yeah. Well, I mean, that's a nice little teaser or taste there. Maybe we, I'd love to talk about some of the specifics, both in terms of the actual projects, but also in terms of what they are backstopping or what they're raising debt for, which is to say, this is not like Strictly for GPUs. I think a lot of people might get confused by this and think like, you know, this is strictly just people raising money to buy NVIDIA GPUs, but there's so much more that goes into actually building a data center that can accept the GPUs. And I think that NVIDIA, as well as all the hyperscalers, are motivated to backstop people across the whole industry.

Yeah. Yeah. It's, it's, it's the data center too. And even though when, you know, those who are used to looking at total cost of ownership, you know, let me show you one of the total cost of ownership charts here. So this is the full stack. And what you'll notice, we always say that most of the total cost of ownership is the capital cost versus the data center cost. So if you look here, the operating cost per hour is 66 cents and the capital cost is $2 out of out of 280, right? So if I do 213 out of 280, that's around 76% OpEx. So you might, you might think the data center is not a large part, right? Um, CapEx is around, for a GB300 is around $31 million per megawatt, $31, $33 million per megawatt. And then for data center, it's about maybe $10 million plus, right? Um, so if it's, if it's, you know, if it's $30 million, if it's like $30 billion, $10 billion, that sounds, um, you know, kind of, kind of mismatched, right? Um, and, and the thing is, right, the, the kind of trick here is that the, um, the capital side is a shorter-term project, right? It's a 5-year depreciation period, whereas your cap— your data center is a 15-year, right? So on a cost of ownership, it's like 75, um, It's, it's 75-25, but on a CapEx basis, it is more like 60-40, right? So data centers even more because we're building more upfront and this data center is going to last 15 years. It's going to do 3 projects of the AI IT CapEx. Does that help put it in context? Like how much the need for data centers?

100%. Yeah. I think a lot of people don't really understand the breakdown in terms of like The chips are expensive. They make up 3 quarters of the capital cost of the project, depending how you model them, right? Which I think maybe is the critical question that's outstanding, namely the residual value of some of these GPUs that are coming off their 5-year lifecycle that people are expecting, warranties expired. Do they keep running them for year 6, 7, 8? And how does that— How do you view the cash flows for GPUs after the residual value period or the modeled lifespan is expired in terms of how it's enabling them to fund future projects?

Yeah. Yeah. So I think that's a question people ask a lot and I'll go to the, you know, we've gone through this many times, but You know, this, this, the one way to think about it, right, to directly answer your question, um, many people that are doing these projects, they actually underwrite it to make money on a 5 to 6 year period. So they're not actually banking on the residual value at all. Um, so, so, and this gets into, um, the, the lay of the land now, because most of the clusters are actually, um, you know, built on long-term offtakes. Because that's the only way you can get funding, right? Remember the trendy, you need offtake, capital, and, um, data centers. So if you don't have a 5-year offtake, you're not getting financing. And that's one of the biggest constraints. And so because of that constraint, because you don't get financing without, without a long-term contract, everything's typically like pre, preordained. So typically on a 5-year deal, you're going to get an IRR of about 20 to 30%, at least. It's been higher in the past. And what that means is by the time your project is over in 5 years, you'll have made on your capital investment a 25% yield per year without banking on residual value, right? So most things are underwritten actually, interestingly, without regards to what happens after. It's like a bonus.

Yeah. So can we back up a second and talk a little bit about the broader industry? Like we just keep saying data center and we kind of talk about GPUs, but when it comes to actually analyzing what is being backstopped by Nvidia, for example, you had this awesome chart.

Yeah.

That I've got on screen now, Nvidia's off-balance sheet commitments, where you summarize the $497 billion of off-balance sheet backstop commitments that they've made at this point. And you go through 7 categories here. AI cloud agreements, where they are signing deals directly with NeoClouds. Data center leases, where they're signing deals with providers that are building data centers. Guarantees and LPS guarantees, again, with more on the energy side. Cloud services agreements, which are not the same as necessarily a NeoCloud agreement in this measurement. Basically an accounting difference there. Data center leases for self-use, which is different than the previous category.

I know.

Supply and capacity commitments, which are primarily around component suppliers like memory. And then finally, this residual value guarantee where we haven't assigned any value to it, but it's a category of something that they're guaranteeing for people. So just gimme a lay of the land of all of this. Like, it's such a mess. There's so much going on.

Yeah. Yeah. So yeah, I'm happy to do that. We can go through each of them. I've got some nice diagrams I can tell you about. you know, why each of them and how we modeled it and, you know, what the point is. You know, good thing this is going to be on YouTube because you're going to want to slow this down to 0.75x when I go through it for all you listeners. But before I do that, let me talk about like the why, right? Like why all this backs up, because the problem is a lot of people just kind of skip the why and they go to the what, right? And then they look at the what and they're like, okay, this is circular financing. Now let's talk about the why. Why are we going to hit an obstacle? Why is it a constraint other than rates going up, which the market telling you that it's going to be tough, right? And the numbers I just showed you, but there's something more. Like I said, Jordan, most of the offtakes are 5-year, most of them are hyperscaler backstop. And what that means is Let's say you've got a Microsoft Nscale OpenAI deal. What's happening is Nscale is selling capacity to Microsoft, which is then selling it to OpenAI. So Microsoft is effectively backstopping that sale because they're guaranteeing that that capacity will be used, right? And some hyperscalers also do that directly, like Meta is directly Uh, offtaking from Neopods, for instance. There are many deals like that.

So, so, so, Nscale, Nscale needs a guarantee to go out and take out a construction loan to buy GPUs and to plug them in and turn them on. And the guarantee is not good if it comes from OpenAI because they don't have an IG credit rating. And so instead of relying on OpenAI to pay the bill, they're going to say, if OpenAI doesn't pay the bill, Microsoft's on the hook.

Exactly.

So why does Microsoft sign up for this, Dan?

Well, it's a good question. I think there's internal and external use, right? I think they also take a cut of it. It allows them to expand and access capital without directly using their balance sheet because it'll show up and, you know, as an upcoming work, right? A lot of it shows up as services of like contract which is under a certain accounting rule. I forgot which one is it, Jordan? ASC 606?

606, not 842. Yeah.

Okay. 606 is the— yeah, I was kind of about to be ASC 404 if I didn't remember that correctly.

There's, yeah, there's a number of different designations. Yeah. The Accounting Standards Committee or Commission or something. Anyway, it's very material to these deals because the definition of A managed service is different than the definition of an operating lease in that a managed service does not end up on your balance sheet and operating lease does. And so the technical treatment of who's responsible for what in the data center is different than the accounting treatment and who can log into what or enter what room in the site or touch what or This is less relevant than how the accountants want to track or keep track of where these assets sit and how they depreciate, right?

Yeah. I think, Jordan, it's kind of like if it talks like a duck, quacks like a duck, it's a duck. So, and I don't think Ray is like, no one's stupid, right? But it takes a little more work, right? Because The problem is, uh, they don't do it for you. They don't say like, oh, they add up all this. Like, like NVIDIA discloses their obligations. And I think in some of the filings, they will disclose it, but it's not in the balance sheet, right? You have to go looking for it. Right. And that's part of what we're going to do at Compute Capital Markets, CCM, our new group. Who we're hiring for, by the way, we'll have a PSA later. Um, so if you love all things credit and like, and we'll show you some charts, like if you like complicated charts, you're going to love this job. But anyway, to Jordan's point, it's kind of like, you know, it's something that's going to be hard for them to walk away from. But my point is, even that, even if it's not debt, like that is not unlimited, right? It's, there's going to be a point where rating agencies, where investors say, hey, you know, this is just too much for you to backstop. Like, you know, how are you going to backstop a $3 trillion? Like no one has a crystal ball, right?

Yeah. You specifically used the term balance sheet as a service. So like, Right now it seems that all of the hyperscalers, when we talk about this $1 trillion of CapEx, are effectively putting their balance sheet to work as a service. That has limits, right?

Yeah. Well, yeah. Yeah. So if you look at Amazon, right, for instance, so they did a $25 billion deal across 6 tranches. This is back in July, right? And it priced You know, kind of well wide of where their benchmarks were, right? And this was kind of like the canary in the coal mine. I mean, people were definitely aware, but you are seeing the hyperscaler spreads blow up by quite a bit and the CDS spreads blow up by quite a bit. I'll actually check for you right now. We can, maybe we can edit this a little bit, but let me, 'cause I want to say a number if that's okay. You guys don't mind? Give me a sec. Model directory, and I can fire up Bloomberg as well. Oops.

Take your time, dude. No rush. We can just clip this section.

Yeah. Okay. So Amazon, so Amazon spread. So if I look at, if I look at July, I'm just looking at July, right? And their 10-year. Your Z spread.

I'm sorry, restart, dude. I'm only seeing the financing market obstacles.

Well, I'm not sharing it. I'm not sharing it.

Oh, I saw your cursor moving like you were sharing something.

I know. No, no, no. It's because it's overlapping the deck. So I'm moving to another thing.

Sorry, I interrupted you.

Sure, sure, sure, sure, sure, sure. Yeah. So if I look at Amazon, right? For instance, remember when they issued that bond? Their 10-year Z spread would've been like around 100 basis points. So roughly, if you have treasuries, it used to be like 4, 4.5%, then you add 100 basis points or 1%, then that means the yield on Amazon debt is about 5.5 or 5.75 as the case may be. And for 5-year, that spread was roughly around 65 basis points. And then You know, what sort of happened right after is it blew out probably by like, in the summer it blew up by about probably 10, 20 basis points from that issue, which is kind of a lot of a move for IG. I mean, things are relatively under control for now, but you know, you've definitely seen like Oracle has moved up quite a bit since the summer. Their entire curve probably moved up like 50 basis points, 30, 50 basis points. So, you know, you are seeing the debt market take note and, you know, rating agencies like will see through this stuff.

Yeah. Walk me through this Oracle increase of 105 basis points and then compare that to CoreWeave, which is well above investment grade at almost 400 basis points.

Higher. Yeah. So Oracle, their spread is going to be around, like looking currently, 10-year, I've got them down for about 250 Z spread versus an Amazon. I've got them only at a spread of 100, 110 for about 10 years. But this chart is a little bit about, like, as you said, the relative You know, like, how do we compare credit risk? So, um, what, what, what this is illustrating is a different type of structures. So, um, what I'm showing you here is, is let's look at the structure where CoreWeave is, um, selling an offtake contract to Meta. And the question is, how much am I going to charge to lend to CoreWeave for that? So someone who's lending money to CoreWeave will say, actually, this is Meta credit risk plus execution. So the MetaCurve at the time we did this for €5 is around 97 basis points Z-spread. So now the actual lending to Corvi for this project was around 200, right? It was 225, so far plus 225, which, you know, when you do a little bit of math comes in slightly under Z-spread of 200. And so the difference between that 97 basis points and the 200 is actually the execution risk. Because even though Meta is there, right? Optaking, or Microsoft is there, you still have to stand up the cluster, right? Which is why, you know, we should actually rename this to the cluster max risk, right? Because if they don't have a platinum, right? You know, in fact, like, there's going to come a day, Jordan, when it's like, this is your S&P rating and this is your cluster max rating, right?

It's not, man. Everybody wants to do Make stock picks from ClusterMax. Now you want to rate people's debt from ClusterMax.

Okay.

It is one data point.

Don't tell Ethan he's going to shut us down.

Oh God, man. No, no, no. We're not, we're not branding NeoCloud debt product pricing as ClusterMax, Dan.

Sorry, sorry, audience. Jordan does not give investment advice.

There's a, okay, look, there's a lot of things that go into building a good quality cluster. I think execution risk on Standing up a given data center on a given timeline and then maintaining a contract such that there isn't a cancellation or termination of the contract is typically bare metal. And it's typically a very different sort of product that is being sold to people than the managed clusters that we are rating. But actually, I'm thinking this through and I'm not completely convinced in what I'm saying, which is that, look, a lot of the experience that applies for building a great bare metal service certainly applies to the quality of the cluster. Like people who build reliable clusters are generally able to maintain and deliver on their contracts. So, all right, fair enough. We can use this as a rating of some, something towards execution risk.

Well, you know, Jordan, you say that, but I've been on a roadshow with lenders. I did New York, I did Singapore, and a lot of them have said, yeah, this has been incredibly helpful to their process. So I think it's like, you know, I mean, I think it's, look, you know, if someone's made it onto ClusterMax Gold or Platinum, Like they're unlikely, less likely to mess up, right? Like on execution generally, it's just a stamp of approval, right? Just because, you know, you got good grades, you know, you, you got good, doesn't, doesn't mean you're a genius, but like, you know, probably is more likely than not, right?

I think it's an A signal on a data point, but managed clusters and large bare metal data centers are, uh, very different products. And I think just because somebody's in platinum does not necessarily mean that all their data centers are going to be on time. And vice versa, just because somebody's down the list or maybe ranked as like underperforming or something like that does not mean that their data centers will be late or that they will be low quality in the future. So true. I think that there's, if anything, I would be asking people to assess a given project or a given site. I mean, we've seen people structure every single site within a NeoClouds portfolio as an SPV, and I think that's the right way to think about it. it, where there is a different, in some cases very different level of execution risk as you're describing it for a site in one location on one continent that happens to be owned and operated by a given NeoCloud, even though they can go to a completely different continent for a different project the next week, the next month, and raise debt in a totally different way, have a totally different team working on it. And therefore things are a little bit different. Dan, let's move on to—

Well, Jordan, I mean, that's an interesting point, right? Because a lot of the clusters, a lot of the NeoClouds we talk to that are unavailable, they're not unavailable because they're busy doing bare metal, right? To your point, it's not their focus, right? And so yet they've landed large deals. So it's not like, it's just one data point and some people just don't focus on it. Like some guys who are doing great have landed lots of deals, just Like, hey, sorry guys, you know, we just don't have the product for you to test right now.

And the whole market has changed. I mean, we're, we're constantly going back to this, but when ClusterMax started, buyers had 5 different providers offering them 2-week POCs on 256 GPUs so that they could make their decision in an informed manner and sign up. Now you got to pick a dance partner 6 months before the GPUs get delivered. And you got to make a prepayment to these people that's worth 30, 50% of your contract value in some cases. I mean, you said this earlier on this episode, effectively the buyers are just paying for people's cost of capital directly. And so you're taking money that you've raised, which has a cost of capital, and then you're giving to somebody else and you have to pay for their cost of capital. I mean, it's It's stacking on top, which is driving a lot of people to try to do self-build work. And I think it's almost like this oscillation. It's like you start out early and you do rental because you don't want it on your balance sheet. You want to move fast, you want the biggest cluster possible, and then you start to grow and you realize, hey, this is a bad deal. Like, I'm kind of getting ripped off. I should own this stuff myself. And then you start doing these self-build sites and then you reach this limit of what your balance sheet can handle, how much money you can raise. And then you start going, the Frontier Labs go, hey guys, I need help. I need to basically mortgage my future right now and I need you to build some stuff and I need to guarantee you some returns on this work because like, hey, I don't have the balance sheet to do it and I need to put your balance sheet to risk. So it's an interesting curve where the Frontier Labs, the way they're doing deals right now and trying to get everything off their balance sheet, and even the hyperscalers are trying to get stuff off their balance sheet. It kind of looks like an early-stage startup. Right. They want the most that they can get right now.

And I think you've touched on like the why, right? Which is where we kind of got here, right? Like, why is there all this backstop? And you described the problems, right? Like, you can't get, you got to book 6 months in advance. You can't get the compute you want. You got to put 50% down. Um, and, and, and that's, that's, you know, kind of why they've done the backstop. I mean, let me share, I did up this slide that describes some of the why. And it's what you said, broaden compute availability is what you just talked about. Now, supporting the financing market is another thing because we just talked about how unless you have that IG backstop, you can't do anything. So it's very hard for like a NeoCloud. There's no incentive as a business model to say like, hey, I'm going to fill my book with half of like various inferences service because your lenders will go like, hey, you know, I'm not going to if it's not IG credit, like, I don't understand this business well enough, right? Which kind of, you know, goes back to this earlier slide that I had on the obstacles, right? They're still on the learning curve. They don't understand the business well enough to underwrite anything other than IG, and then they don't have the tools, right? And this is like a lot of what Jordan and I are working on, you know, Jordan on the tools to assess the operational quality and myself On assessing like what's happening in the market, right? That's why we have the GPU index. That's why we're trying to create tools that lenders need to monitor risk. And you mentioned residual value. This is one of the huge asks we had in our roadshow because banks think in terms of residual value. They're lending against assets, they're used to lending against property, cars, and they have to know residual value, which is very, very hard with GPUs. It's not like you straight line depreciate. I think the residual value is what you can earn on tokens. Or which you can reasonably rent it at.

And it's so distorted right now because I think in a normal market where you have like this contention between supply and demand, where people produce enough GPUs based on the demand for them and they can always go a little bit higher, then you have real pricing and you have real like settlement. But what's happening right now is nobody has an understanding of what the true theoretical H100 depreciation curve looks like as GB300 comes online, as Vera Rubin comes online next year. And it outperforms that. And then you say, how can I run an H100 profitably? Well, we already have models with InferenceX that show that you really can't. If a GB300 is outperforming your Hoppers by like 7 times, why would anybody want to buy a marginal Hopper over a marginal GB300? You'd never want it. You'd want the latest and greatest. But this isn't the question people are asking about Hopper. Instead, the question is, I have no option to purchase a B300 GPU because it's all sold out or I have to wait 6 months. I want to do research right now. What option do I have to buy? And the answer is renew my H100 contract with 4-year-old GPUs for double the price of what I signed it for previously. And this is on screen. Dan's got our pricing chart for the 1-year term length H100 deals that we know about. I mean, they're approaching the prices that they were signed in the first half of 2023 today. Like $3 an hour for H100 GPUs on a 1-year term. And look, if you believe in AGI, frankly, I talk to all of these startups a lot and I basically give them the coaching of, if you believe in your startup being successful, you have to imagine that somebody is going to do something OpenAI and Anthropic can't, and it's going to be you. And so you have to, you know, flippantly, you have to like believe in AGI here. Right. And If you don't—

Jordan, but like I had a question because you said that no one doing inference would ever rent an H100. So like, why are people, why is the price going up then?

Well, because I think there's so much demand for it that the price goes up strictly because of it's the only option for a lot of people to do. Second of all, there are some researchers who prefer H100 because it's familiar. They are more productive by the metric of jobs per day. Rather than tokens per second. It's like, I don't need to spend time debugging. I don't need to port stuff from BF16 to FP8 or FP4. I'm doing training and I need numerical stability. I don't care about these low precision data types that are used for inference. I'm doing some new model where again, I care about numerical stability. I don't want this risk. So there's reasons why people want to do research with Hoppers, H100 and H200. But I think the biggest reason why the prices are increasing today is just because GB300 sold out. So people go to the B200, B300, right? Those are sold out.

Cascades down.

Exactly. And I mean, if your answer is wait 6 months and have nothing or get started with H100s today, a startup that just raised money has a pretty easy decision. They go for the H100s they can get access to right now.

Yeah, yeah, that makes so much sense. And that's kind of like why when we do this index, like we do it survey-based because like, you know, there's a story behind every data point, right? This is not just, you know, going like, you know, scraping a bunch. Like, you know, there is a story and those matter. And, you know, what I did, Jordan, I actually decomposed. Oh, do you want to go or I can talk about this?

Yeah, sorry. Well, just one last thing. If you go back to the previous one, like we would have a more academic theoretical way to forecast the depreciation in the H100 pricing curve. If you look at this chart and time around September or October of last year, it just kept going down below $1.80, below $1.40. And we actually, I remember some deals being signed at that time for like $1.30 for an H100. People were offering stuff that low.

Yeah, that was the bottom. And, but those were not great providers though.

For sellers?

Well, no, I mean, the guys who were selling at $1.30 were not great. We're not like great providers, right?

Fair enough. Yeah, yeah, exactly. But okay, but the point is like, I know some sellers at that time who were evaluating $1.30 and going, should I do a 1-year at $1.30 or should I do a 2-year at $1.25 or something like that? And man, those guys are kicking themselves for not doing a 2-year then because they just renewed at like literally some of them at $2.50, like way more than they were paying this time last year. And so anyway, the point is like we would have a real understanding for H100 residual value if we had seen the pricing curve depreciate like some other asset, like a car or something, right? But 4-year-old H100s going into next year are going to be selling for likely higher prices than they are today if the trend continues. And I mean, I expect it to because I expect demand to continue to rise. And let me make 2 more cases for this that are not Based on like an AGI future, you don't need to believe in that. You need to believe in data center delays or cancellations due to political posturing and moratoriums and execution risk, as you were describing, which we see a lot of. And you also have to believe in more demand from consumers on Anthropic and OpenAI's API endpoints, which I think you can believe in even if you don't believe the models are getting better. You just need to look at things like Instinct growing or Muse from Meta and just imagine a world where more people use AI-driven products, even if the models stop getting better, right? So the combination of like data centers might take longer to build, politicians may get in the way and stop them, that will constrain supply, more demand for the existing models and the models get better are all a recipe for increased prices over time. And that's generally the case that I've been making to a lot of people.

Yeah. And the thing is, right, if you did the scientific version, we started this in 2024, we actually have a forecast and that actually looked like downward sloping. It actually started detaching. We were like within 2% for like a good year and a half, right? Good 18 months. But then it started dislocating for September because the way we calculate the theoretical depreciation curve, is the relative power, the relative compute power of new GPUs. So for instance, the GV300 had 7 times more inference speed and 3 times more FLOPS, right? So it would basically, those new capacity would push down the price of your old capacity. In theory, that's what should happen. But what we always said is, we understand supply, we track price, but it is hard to track in real-time demand. So what we do is we actually flip it on its head and we say, let's observe price and let it derive demand. And so in this slide, this is exactly what I did. What I did is I said, if you have a balanced, neutral supply-demand, then what you see is it depreciating in value or rental price going down in a scientific way, right? In line with new capacity, right? But what happened is in October, right, that dislocated. And what that implies, right, because if your demand is higher than supply, then your price surprises to the upside, right? And so what I did is I reversed it and I said, hey, what does this imply for demand-supply balance based on price? And what we saw is it blew out completely in late 2025 and it's never looked back. Does this kind of like tie in with what you've seen, Jordan? I assume so. Like maybe not the theoretical stuff, but like what you observed in the market.

Well, a couple of things happened around that time last year. I think basically models stopped getting bigger. Like the, it took a while for the Chinese models to catch up, but up until around this time last year when reasoning models were taking off, people kind of discovered we can trade FLOPS per parameter count and continue to improve the quality of the model with a lot of innovations around RL and post-training. And they've just continued to exploit that where we may have gotten bigger models. I think we likely did get bigger models, you know, with the Astra and the Fable generation or the Mythos generation, but we didn't get significantly bigger models that outpaced the improvement in hardware that came when the GB300 and VL72 came online. Which is to say significantly more memory, memory bandwidth, and FP4 performance were able to keep up with the increased size of the models. And so what that meant is that these companies could continue increasing price per token on their APIs without necessarily increasing their costs to serve the model significantly. And they're, you know, that like the result is just they're way more profitable. Like way, way more profitable. And you don't even need to know anything about the Anthropic or OpenAI models to make this case. You just need to look at the Chinese models and you need to look at what the performance of a DeepSeek V3 was on like an H100, H200, or even a GB200 when, you know, those were being brought online last year. And then you need to compare it to DeepSeek V4 or 4.1 Pro or 4.1 Flash or GLM 5.3 or Kimi K3, where you just look at all the innovations on sparse attention. You look at the parameter count growing a bit, like around 2 times or 3 times in some cases, but FLOPS not increasing significantly and performance increasing so much. And so they can command a pricing premium over the previous tokens because the tokens are more valuable, but it's cheaper to serve. You get more throughput per GPU with these. And so people serving Chinese models are enjoying gross margins of 80% right now on a GV300 NVL72. And so it's just like, it's like, look, obvious that there's going to be more demand for this compute, this asset. Now, the question or what gets confusing to me here in some cases, and maybe you have an answer for this, is how nobody forecasted this. Like we did our best, we understood, right? And I think even the Frontier Labs who really deeply understand how these models work did not expect themselves to be this profitable and to be able to serve this many users and to grow this fast. And so even they were caught by surprise and had to do crazy things and sign big contracts at very high premiums just to get access to GPUs tomorrow to serve their users and to go for the bigger research jobs. I don't view this as like an AI fast takeoff, but it seems to be like a revenue and profitability fast takeoff for a lot of these labs. Yeah.

Yeah. And I think, you know, we've kind of in our discussion, we've really hit on what is the debate going forward and it's the willingness and the ability. And I think the ability Is the profitability. Like a lot of people who are new to this are, are, are, don't realize just how profitable it is. Um, and that's one of the things that will determine whether we, we, we overcome this obstacle, this kind of, um, you know, funding wall.

Well, I think it's also confusing which, um, layers of the cake are profitable, which is to say, right.

Yeah.

Yeah. Nvidia clearly had durable profit margins for a long, long time, but Your, yeah, your layer cake chart is great to explain. Like, clearly there has been an increase in the profitability of the NeoClouds because the amount that they can rent the GPUs for to people has increased recently when the GPU has not, there's nothing changed on the operational costs or anything, right? The people are selling tokens with more efficient models or higher prices. And so therefore their profitability is increasing. And then my personal experience, and I think a lot of other people's experiences, the value or utility that you're able to derive from using these AI applications is also increasing over time as they get higher quality. And so I don't look at this and say, you know, we're taking from one to give to another. It seems like every layer of the stack in some ways is increasing profitability, except for like the server OEMs where it's still a pricing war and they get 9% margins.

No, no, no. There's still a bit for everyone, but, um, But, um, you know, the tin box and the people to plug it in. Well, I, I can get into that, but, but I want to, I want to, you know, just so everyone stays, um, organized, right? I think again, like everything we've talked, like Jordan and I've talked to so far, you can, you can fit into 2 buckets, the willingness and the ability. What we're just talking about now is the willingness. You know, what is the return? Will the labs remain high? Like how high are the margins? You know, can end users— and earlier on we were talking about the ability. We talked about the balance sheet as a service. We talked about NVIDIA Backstop, which is trying to address the ability, you know, and who's going to fund them. We talked a lot about that, but let's, you know, why don't we dip into the willingness, right? Because that's what we were just talking about. And we've done analysis where you can actually earn up to $100 million per megawatt generating Frontier tokens off of GB300. It may be a little optimistic in terms of utilization, but When this thing costs about $40 million per megawatt to stand up, it's like, you know, if you're fortunate enough to sell Fable, you make this back in half a year. So it's incredibly profitable. And let's compare that against what NeoClouds are typically renting for a GV3 at least a few months ago, about $12 million per megawatt or $12 million per gigawatt was how much NeoClouds were on a fiber contract. But if you're doing frontier inference, Like I said, you can do $100 million per megawatt. And people were shocked, right? When they saw SpaceX and Google renting at $48 million per megawatt. But, but, but, you know, that, that makes sense, right? Like, because they're buying here, selling here. So of course they would, but people were used to seeing this. So, so the profit is there. And, and I think Jordan, you mentioned, you know, where, where does the profit stay in each layer? So what we did here. Is we actually organized the entire kind of business as an income statement. So you can think about this, you can imagine you're an Anthropic, right? So this is your revenue and we put it in a couple different ways. So million token, about $1.85 per million token, which is a good blend of Sonnet, Opus with cash hits. This is actually our own blended spending. Even if Opus is $5 input, The cash hits will take it down. So about $1.85 per million token, which is equivalent if you're generating, I think we put around, I wanna say 9,000 tokens per second. I can't remember what we used, but it's equivalent to $38 per hour, right? And then we already talked about the cost per hour of $2.80 and then a NeoCloud. So the cost to own the system is $2.80, but the cost to rent from a NeoCloud is around, it's probably around $4, north of $4 by now, $4.50. Almost $5. So if we just simplistically look at that, what is a gross margin? It's 90%, isn't it? Right? Because you're selling at $38, your cost is $3.80, and this is your profit. So it's incredibly profitable.

Yeah.

At least for Frontiers. Um, by the way, this was really cool. Everyone, everyone should, should read this, like drop everything and read this now. 'Cause, 'cause, um, you know, Max and team.

We recorded a pod earlier today. You don't know. Yeah, this is. This is my third pod of the day. I've been cheating on you, Dan. But yeah, Max and Andrew went through this one. Oh my God.

You're a pod lickerist. Come on, man.

Okay. Well, anyway, Max and Andrew went through this on the previous episode. Maybe for people who missed it, do this one again, which is for the all-in consumption plans on some of these On some of these businesses, you can get a lot of value, but it kind of depends if they're going to be profitable at this layer or not. Kind of depends on how much people actually use the plans that they let people have access to.

Yeah, we peg 90% API margin, and then I think in the mall we use like 50, 60. But the question I have is how much of these people are really using 10%? So I'm kind of, And there's a whole cottage industry in China, right? That like takes apples and makes applesauce, right? They take these $20 plans and then just sell them as API, right? On the gray market. And then, you know, me, I'm on a $20 plan. Like I actually plan my sessions around like, you know, utilizing, I have loops so that, you know, I use it even when I'm not there. So I'm definitely on this bucket, you know, but I don't know where everyone else is. Yeah.

I mean, Max made the case earlier that I think there's a lot of people who maybe get these through their business or, you know, like the company they work for, and they're just like not really worried about the value trade exactly. And you just kind of blend it over a bunch of users and it works out. And the loud people are the ones pushing for more limits who are the ones you're losing money on. But while you keep 'em around because you don't know how to separate the wheat from the chaff and only sell the all-in plans to the most profitable segment of users, which is the ones that don't use it. Like, you know, the whole thing about Planet Fitness is marketing it to people, marketing gym memberships to people who come once in January for the new year and then never come back. That a profitable gym is the one that's empty. They don't want anybody to show up. I mean, Kind of cynical there, but that's the incentive structure for these guys to do all-in plans. Whereas the exact opposite is true for the per-token pricing where they're just trying to get people to consume as much as possible.

So yeah, yeah. And then the next question becomes like, okay, it's so profitable now, but like what people ask is, is that going to be the case in the future, right? Let me find our token pricing. Oh gosh, where'd I put the token pricing? And people say like, how fast is token price dropping? Sorry?

Yeah. I mean, not to cut you off, but let me, yeah. Okay. Yeah.

Yeah.

Do this one. Yeah.

Yeah. 'Cause my question for you, for you, for you, Jordan, is like, people ask me like, how fast is token price falling? Like, what do you mean it's falling? It's going up. Right. And in fact, even if it's going flat, that's gross margin going up because While token price has been staying constant, they've been generating more tokens, right? And then they're like, okay, well, great. So I see Frontier Mall's been holding up, but like, how fast are Open Mall's falling? I'm like, well, they're going up, right? And the other thing is when you look at the token expenditure, it's like a changing basket. So these indexes that track token expenditure, like they don't adjust for that, in my opinion. I don't know how they work, but they don't adjust for that. And they don't adjust for increasing, you know, with Agentic, it's been all about effective price and cash hits. But I don't know, what do you think of, what do you think? Where do you think token pricing is going in the future, Jordan?

Yeah, I mean, I had a little tweet thread where I went at David Friedberg from the All-In podcast. He didn't respond, of course, but a lot of people are saying that there's like this overtaking from open models overtaking closed models. And I think the dynamic is just like people are consuming more of everything and they're consuming it at higher prices. You've got the price of input on screen here, which I think is a decent metric, but You know, that little dropdown in the middle, you should do the effective, right?

Well, I mean, you could, it's just a different approach.

Yeah, you can do it for cache read, which I think is the bulk of our spend is actually on the cache read tier because, you know, eugenic coding is just so much cache hit. And then output is like the most expensive tier, even though, you know, you don't actually generate a bunch of output. That is what takes up all the time on the GPU when you're memory bandwidth constrained. Anyway, the point is like, I totally agree that model pricing is going up over time on a blended dollar per million token basis. I think that token efficiency is a question mark because you have to trade off efficient ability to solve simple problems with people's more grandiose expectations as they try to solve more complicated problems that weren't possible before with unlimited tokens and now are possible with a lot of tokens. I don't know how you measure that trade-off. But look, none of these dynamics to me point to anybody deciding, I've had enough of AI, I'm going to use less of it. I've had enough of these models, let's make them worse and less efficient. I've had enough of these new GPUs, let's make the next one worse and less performant. It all seems to be going the opposite direction, right? They're going to raise prices over time, more demand for it. The GPUs are going to be more efficient and the models are going to get higher performance. So We'll see. Um, there's a lot of, uh, ways in which this can break if people go and decide to YOLO for the 20 trillion parameter model with the new NVL72 and really use it. If they start introducing, but, you know, ultra fast mode and some of these low batch size optimizations and stuff that can jack prices up. But, um, I think all of that is just like artificial constraints on what to me feels like continued overwhelming demand for these products that drive the usage of the GPUs underlying it.

Yeah. And we talked a bit about the income statement, the explicit one, but what about the income statement above, right? What about the users, right? That's part of the equation too. What value are they getting? So I don't see why anyone would step back when these are the kind of efficiencies they're getting. Like we built an entire database of thousands, thousands of debt facilities and all these relationships purely from public domain. And it's been myself, Oliver, Terence, have been doing this and we're all part of current groups because we're hiring, like I mentioned, PSA. But it would've taken me 4 months and a team of 4 to do like half as good of a job as we've done so far in building up this database and relationship table so far. And so these are, we've showed this slide a lot. We had a pod on this, but Yeah, again, I don't see why anyone who is paying $6 to initiate a company would ever go back. Would you ever go back, Jordan? Or would you just stop using fast mode? What would you—

I don't use fast mode.

Me neither.

You'd have to pry this stuff out of my cold dead hands, man. Even in my personal life now with some of these support agents and stuff.

It's like, pick a wine, pick a wine to bring to dinner. with my friends. It's like, okay, moms for $3. I still—

some stuff still has a personal touch, Dan. Okay, like, I'm not writing birthday cards.

Speak for yourself.

I'm not that far off the deep end. But yeah, like, I just got this refund on a flight where I was delayed on for 4 hours. A few, you know, there's no way I would sit on the phone trying to get this refund ever again, right? I'm like checking, you know, times at the gym, or I'm checking like booking a flight or looking for hotels or checking on the grocery list or just check my email, man. Like what came in last night. I mean, it's a little weird, but I talked to some friends who are not enlightened to the possibilities that having a personal assistant can provide. And it does feel like everybody can never go back once they get the experience of the amount of time that you save and the feeling of like how much wasted time.

Yes. Yes. It's, it's, it's, and I think Muse is going to be eye-opening. It's like, why did I ever sit there for 2 hours and browse all these like airline booking sites? Why did I ever, you know, like, I think there's going to be that moment, right?

It gets a little egotistical when you talk about it this way, I think, because, you know, you kind of anthropomorphize the agent and then you view people who have personal assistants today as kind of exploiting like this thing. But anyway, my personal view is kind of like, it's shocking how much time we waste on this monotonous stuff. And it's shocking how, you know, some people that I'm close with do not value their own time. You know, I like, yeah, I'm not going to go into all the details there, but you can—

Well, I think we're sold on willingness, right? We're sold on the, you know, AI is like useful, right? Yeah.

Well, look, I felt since working at Semm Analysis, I felt like we've been at the tip of the spear for adopting a lot of this stuff and just feeling the ability to use it, which both makes our research better and informs the underlying research that motivates some of the data that we see. And, it's really nice to see other people that we work with and in my personal life adopt these things and then start to regain control over a lot of things in their life. It's really cool, really cool to see. And I just, I get really happy when I see people have these good experiences using it and start to get motivated to like build a more positive and high-quality future together. I don't know how to wrap that up, but yeah. Talk to me about the willingness to finance.

Do you think— Yeah. Well, we just talked about that. I think we're sold. Like what I got to convince you on is, again, it's willingness and ability, right? And I got to sell you on ability, right? Like I might've scared you with this, but—

Yeah. Talk to me about the human element here, right? Because I think it's one thing I've also realized working at SemiAnalysis is how much of this stuff that seems so data-driven comes down to some human being making a decision. Like Satya said pause, therefore pause, or Jensen said build, therefore builds, or backstop or whatever, right? So, where does this break down to genuinely one of the leaders of these companies having this experience using a personal assistant or an AI or just having that personal moment where you're like, oh my God, I can see the math discovery before it comes out and gets published, or I can see this progress we're making on this cancer drug or whatever, and I just really want to bet the company on this and I want to do this work because it means something to me and this is my life's work, as opposed to rationalizing all of the financial details to make it work. Like, like, can you even contend with explaining that balance between those 2 things?

Yeah, but you know, like, I, I, I, that, that is one of the big questions because we, we, we covered, this is still under willingness, right? Because we covered, we're, we're sold, right? And I think what you're saying is, are other people sold too? And, and that's kind of what this chart is a little bit about, right? And I think the question, like, I would, I would make one that's slightly different. 'Cause what I would do is I stratify into consumer, like API tip of the spear, people like AI natives, and then the rest of the enterprise, right? So I think we're convinced on the sphere, tip of the spear. I feel like Muse is gonna prove the case in consumer, but I don't know what enterprise, 'cause I sit at these, I do like a lot of these events and I meet a lot of folks. We do talks and dinners on AI, and then kind of Fortune 500, they're like, oh, what would I use it for? Like just summarizing stuff? Like, no, it's so much. So, You know, I feel like there's just a lack of discovery in Fortune 500 and lack of imagination. And a lot of people, they actually kind of don't come at this with an open mind. They come at this like, oh, great AI. Like, how do I fire people? Right. You know, like that's all they think about. Right. And it's like, so I just think the enterprise is, it's of the 3, it's the one I don't know the most. I don't know it because it's so much to like, Does the CEO say, hey, let's embrace it. Let's trust Amazon. Let's go full blast. Or they go like, no AI in your work. Like, like, what do you think? Have you talked to Fortune? Like, where do you think they are on this?

Well, I think the dynamic that I imagine happens with the Fortune 500 is that the Fortune 500 just slowly gets replaced with companies that are all in on AI. Well, look, look at the top 10 companies.

Possibly true.

Look at the top 10 companies in the world by market cap, right? I mean, these are all. Like 9 out of 10 of them are AI or semiconductor related or data center related at this point, right? Like when we say hyperscaler, all 5 of them are in there. Then you've got NVIDIA, Broadcom, TSMC, the memory guys are getting close to the top 10. So Fortune 500, I think we use that term to describe, if you actually mean the 500 biggest businesses in the world measured by revenue, I think a lot of them are just going to—

Maybe I should say corporate America, you know?

Legacy businesses, corporate America, legacy business, yeah, government, just like people that are slow to adopt AI and fair enough. I think there are lots of people who will be a lot slower to adopt AI than the biggest companies in the world, which I would lump in all the hyperscalers into their soon to be the AI labs. If Anthropic goes and IPOs at $2 trillion, they're going to be a top 25 company in the world in market cap or something like that, right? So, I mean, that's the proof point right there. But you have to make this case of like going out 2, 3 years. Do you expect that this next crop of Neo Labs that are pursuing all of these new innovations just like completely all get swallowed up and acquired and none of them pan out? I actually expect many of them to do quite well across a number of different industries, continue to grow. And if they do get acquired to probably help those businesses grow and those take over. And so I don't view like a future where Visa and Bank of America and AT&T are just like so critical to the success or failure of the AI native companies that are trying to grow. Maybe what I'm most interested in is The unlock of capital that the companies that may have been funding those companies in the past may turn and shift over and say, hey, I work in infrastructure finance, I work in credit, and for the next round of this, I'm not actually going to go do another project for another bridge or for some debt for some credit card payments provider or something like that. I'm actually going to go into this AI thing because I think that's upstream. of all of those other businesses I've done. And it actually is infrastructure and I'm going to get these returns and I'm going to have these guarantees and things like that. And that's where I ask you, because you know these guys better than me, about the human element of like, for people who really just don't understand this and don't have the experience of using AI, who represent so much capital that's still sitting on the sidelines. We joke about stuff like Nvidia's Revenue at, if it hits $500 billion, is going to be like 2.5% of the M2 money supply in the US, right? Just unbelievable numbers we're playing with. Trillions of dollars of CapEx we talked about earlier in this episode, but there's still so much left over. Like what's the other 97.5% of the M2 money supply?

So, well, like let me, yeah, I can share some stuff because now we're on kind of the ability, right? And again, one of the obstacles I talked about earlier was getting people, like, I guess I'll show you, oops, it's up here. Let me go to one of the key things on the ability is like understanding. And one of the key obstacles is lenders on the learning curve, like such, you know, the people making decisions on lending, like right now they're doing it because like, okay, it's immediate credit risk. It's Microsoft Credit Risk. Like, I don't have to know, but that's going to change. And that's why all these backstops are there so they have time to understand this industry. But let me size the problem for you a little bit. You know, let's talk about, you know, where this is going, right?

Absolutely.

Like, obviously it's going to be the second biggest debt financing market, right? It's already surpassed auto loans, surpassed everything this year. But everyone's done it without having to have a residual value model for GPUs because it's been backstopped and it's only second to the US mortgage-backed market.

Stan, sorry, move on from that chart. We've got some audio-only listeners, but I just want to describe it. It's such a compelling chart. Auto loans, student loans, credit cards, non-agency CMBS, HELOC, these US asset-backed credit markets are just like flat lines, like slowly going up to the right as people You can't even see them going up actually. Yeah. And it's just 5 flat lines at the bottom of the chart and then it's just this massive exponential line up to the right, more of a straight line, but yeah, straight line up to the right called AI debt financing. And just explain to me the implications of AI debt financing in 2026 crossing over auto loans and student loans on this chart.

The implication is, and we already see it, we already see long end, we already see 10-year rates going up. We already see credit spreads going We already see an industry and the reason the credit market, it's going to be hard because they don't have the tools to understand, like, because it's all coming from like different fronts. It's a war on so many fronts because you're getting hit by hyperscaler, neo-cloud, data center, everyone at the same time. And it's like a zombie apocalypse, but you don't know how many waves of zombies are coming, right? You know, it's like Terminator, you know, how many Terminators? Give me your money. Anyway, they're money zombies coming, right? And you don't know whether you're on wave 5 or wave 50, right? So, you know, and to put things in context as well, I've got a slide here that shows you what this is. Like I told you, I showed you what it is in terms of like the corporate bond market, right? And that's That's just up here. But on the next slide, let's talk about overall US issuance. So it's $1 trillion and it's going to cross $1 trillion. So if you think about it, right, the total US issuance, this is federal, which is in the blue. You know, for those of you listening, that's about $5 or $6 trillion a year. You've got MBS, RMBS, just looking ballparking from that, it looks like $3 or $4 trillion a year of issuance. Maybe $2 to $3 trillion. Corporate bonds, we already talked about that, about $3 to $4 trillion. Total $12.7 trillion. And 2026 is the last year that AI debt is going to be less than 10% of all issuance, right? All issuance in the US. So it's going to take believers. And how do we handicap that, right? Well, this is Maybe what we can do is we can jump into like, you know, back to the question you asked earlier. It's like, why is NVIDIA doing these backstops? Right. And let's just look at, well, sorry, I'm going to go. Okay. Here. Yeah. Like, why do they have all these different backstops? And again, it's, they see this gap coming and they're figuring out how to plug it. Let's talk a bit about their AICP, right? So by the way, You know, to start off, we projected Nvidia's balance sheet and we were looking at their ability to actually continue doing these backstops and how much they can do. But let's talk about like an example of one of the backstops and how it accomplishes those goals. This is the AICP program. You know, this actually started in Asia. The first one was Fermi. Well, actually, first one was Sharon AI, then Fermi followed with another one. You know, there's been Aulani, there's been many others. GMI did one. And the whole idea here, Is the way this works is NVIDIA goes to NeoCon and says, hey, I'll give you a guarantee. Like, if you can't rent your cluster, I'll rent it from you. And I'll give you that guarantee for 6 years on $325 on average. It's the actual table is different. Each year is a different minimum. But if you can't find clients, you can always go to NVIDIA and rent to them at this price. Now, the intention is not for that to actually happen. The idea is you use this as a guarantee to then get financing. Remember the trendy, right? So now you have an offtake and now you can get financing. And then with that, you can get data center. With this in mind, lenders, lenders, you see this, this little thing that it's priced as NVIDIA risk. So it says, hey, you know, I'm really lending to NVIDIA plus execution risk. And then what that allows, allows the NeoCloud to do is, oh, now I don't have to rent a 5-year offtake to whoever. I can do, you know, token as a service. I can do all these things. It really opens up the market. That's the whole idea. And they don't need to be IG. That's the important part. But this is only like an intermediate step because what it does is lenders have to understand that they're also taking the platform risk, whether Neocredit can build a good book of clients and they'll get to see it as they lend to Neocredit. They say like, hey, can they build an independent business? Outside the optic, they get to understand the customer profile, the churn rates, the cancellation rates, and they get to, you know, kind of do like a free look or a dry run, a practice run until the real thing. And that's why this is kind of really important, why they're doing it. Does that make sense, Jordan? Any questions?

It makes perfect sense. Yeah, man. Balance sheet as a service. Maybe the one thing to talk about is like they at some point introduced a take On the percentage of revenue that the Neo Cloud— I forgot about that— above the minimum, uh, that they promised.

And then they took that away and, you know, no, no, no, it's still there. Right.

Explain that. Yeah.

Yeah. Yeah. So, so it's also a revenue model as well. Um, and it's fair, you know, we're not communists after all. Um, so in exchange for the backstop, NVIDIA gets a 50% revenue share above the backstop. And so for example, let's say you rented $6, or let's make it easy, $6.68. Then you have the difference is $3. So you got to pay $1.50 to NVIDIA out of the $6 in exchange for that. So it's actually a meaningful revenue share for NVIDIA. And if you look at the other structures they've done, this is the Lambda structure. They're also taking a revenue share out of Lambda in exchange for that guarantee. So, you know, and another third reason, Jordan, by the way, one of these backstops do is not just facilitate these deals happen and allow, you know, in this instance, they're allowing Lambda to get financing because Anthropic's not yet IPO'd, is not IG, right? But with NVIDIA supporting the lease for the data center, they can actually get this done. But there's another reason, right? Let's look at Ports Pike. This is a little complicated. But the bottom line here is NVIDIA is giving $105 billion of residual value guarantees to the Ports Pike data center complex in order to support this. But the catch is it's going to— so what they've done is they've guaranteed that 4.25 gigawatts or up to 9, almost 9 across 2 phases will be running NVIDIA GPUs. So it also locks in. It's like, it's like a kind of a land grab is another, another reason for it. And so we looked at it, we looked at all these different models, and I'll talk about one last one. And this is the kind of the most interesting one is a $500 billion capital partnership because this gets into how can we actually, you know, to your point, right? You talked about lenders as one homogen, but it's actually very different. There's bank lenders, insurance lenders. They don't want to take a lot of risks. You've got private equity. They want to take a lot of risks. You got private credit. They're in the middle.

They all syndicate into these things as well, right?

Well, it's about creating different products for different types of investors. So the $500 billion capital partnership, this is what we think it could look like. And what you do is you can have an AI lab. They don't have to be IG, but what you do is you tranche the risk. So you have equity and then you've got a B piece, which doesn't have any residual value guarantee. And then you have an A1, A2, you can put one senior to the other. And then NVIDIA is giving a 25% residual value guarantee in this $500 billion partnership. Maybe you can attach this 25% to this tranche or this tranche. But because let's say there are losses in this program, you know, in this business, then the equity will take the loss first, then the BPs, and then part of the A2. And so it will have to go all the way down to this. The loss has to be so severe to affect the senior tranche, which is how you can create different risk profiles and open up the market. But anyway, all this depends, right, on NVIDIA having the capacity To do this. So we analyzed the balance sheet, you know, and, you know, if you kind of love talking about this, you know, this is all we're going to do. But, you know, here's the kind of conclusion of this, Jordan, right? They, we think they have the capacity. This is not our forecast, but we think they have the capacity to support up to 46 gigawatts of capacity. We haven't done it for the hyperscalers yet, but that's still not enough. There's 240 gigawatts is going to come in. And so what I'm saying here is, is the lending market has to evolve to take risks outside of backstops. And that's what, you know, we're all about understanding that question. And it's really the biggest question of our time, to be honest, which is why we're starting an entire group around this question.

Makes sense, man. Yeah. If there's 2 tangible things I'm taking away from this, it's one, more education to the lenders on what they're actually Getting into here, you know, what the real deal structures are like and who the end user customers are and why they're structuring things this way. And then 2, well, I guess 2 is twofold. It's always incredible to see the size of these numbers and how much has already been done, but also how much bigger it needs to get from here.

Uh, this is incremental capacity adds.

So incremental 60-some-odd gigawatts of capacity adds in 2029.

Yeah.

Let's see.

When we first started this journey, global data center capacity is like 40 gigawatts maybe. And now we're adding like 40 a year. Can you imagine that?

I can.

Yeah.

Yeah.

Excited for your usage. Yeah, I know, I know, I know. I was looking at the numbers.

Let's wrap up one little funny story. When I got started working at SemiAnalysis, I did not understand like anything about financing and how this stuff works. I was trying to figure out who a lot of these lenders were, what private equity was. I'd known these names, things like that. And I remember working on this one deal. I won't say, you know, who was setting up the deal, but it was like private equity.

Did you think EBITDA was like a musical at some point?

Yeah. Yeah. I still get confused when you talk about project versus workbook IRR. I don't know. I just know that it's really important.

Project versus equity IRR.

Equity IRR. Yeah.

Okay.

Okay. I just know IRR is— you spend enough time with Dan, you know IRR is important and it doesn't get easier to say IRR. IRR.

Yeah.

That's the, that was the pirate, right? Dan, when he started, it's like he needs an eye patch because he's talking about his IRR.

Anyway.

Okay. Let me tell this quick story before we wrap. So this private equity company was setting up this deal. We had to go do this technical DD on it. We had to present to some of the lenders and the lenders were like, you know, from all sorts of different institutions and I remember at the beginning learning like, oh, private equity doesn't have any money for themselves. They have lots of assets under management, but they set up the deals and then the lenders come in and they actually fund the deal and private equity makes carry off of the top of that. If anybody in PE is listening to this, it's like, yeah, of course that's the way.

I don't know.

Anyway, and so I had this meeting with this pension fund to explain the work that we did. And I remember sitting on the call, it was the Ontario Teachers' Pension Fund. And I was like, oh, it's my mom's money. This is what's getting in here. So anyway, this whole loop back to thinking, okay, yeah, yeah, Canadian pension, Ontario Teachers' Pension, lots of stuff in Quebec. I'm Canadian. So anyway, I ended up feeling kind of proud of the work that we did, not like I was helping these people who were just like— rent seekers trying to make a little bit of money off of this. Like, it actually is pretty hard to set these things up, make people comfortable understanding what they're investing in, explaining the risks, understanding the technology, what's the return, blah, blah, blah. And yeah, there's going to be more to do in the future because lots of people just like that are probably getting in and evaluating their first AI cloud opportunity or neo-cloud deal, and it's going to be a big part of the future.

We're on a learning curve and that's, you know, I think the more we help people get up there, you know, talk about everything, I think the more it helps everyone make great decisions. Yeah.

Cool, man. Okay.

You know, one last joke. You know why it's IRR?

No, why?

Because pirate equity.

Pirate equity. Hey everyone, if you're still listening after this hour and 15-minute long conversation about credit and markets, you should consider applying to work at SemiAnalysis on credit and markets. Dan is a very fun guy to work with. Dan, who are you hiring for?

Well, we're hiring credit analysts in New York and Singapore. So we're hiring junior credit analysts. So anyone who's done 1 to 3 years of experience, could be investment banking, Well, DCM, you know, anyone who's focused on that. Also, we're looking for a more senior credit analyst, anyone who's been a desk analyst or investment analyst on the buy side, and also a credit specialist, someone who's quite in touch with markets, has a great Rolodex, worked in credit trading, credit sales, you know, buy side or sell side. So love to have you apply. Just Click here to find out more. I also need a video editor. If anybody knows about a video editor, shoot me a text. Okay.

This is your 3rd interruption. We're experimenting with ads on the Semi Analysis Weekly podcast. We also need a video editor, a cautious hiring.

Yeah. So if you know how to edit videos, please hit me up. We're looking for people. Please hit me up.