Ep. 021 - The AI Project Trinity: Capital, Offtake, Data Center (Datacenter, Energy)

Hello everyone. Welcome back to Semi Analysis Weekly. We're here with episode number 21. I've got the crew from Singapore to talk about everything Nvidia backstops, including the holy trinity: capital, offtake, and data centers. Starting off with Dan chewing away on his lunch. How's it going, man?

That was dessert. That was apples. I'm meaning healthy dessert.

Good man. Okay. With Dan, we've got Kong Wen. How you doing?

Doing great. Doing great.

Okay. And Zane, how are you, man?

Yep, doing well. Thank you for asking.

Okay, awesome. So in this episode, we're going to talk about NVIDIA backstops. We'll start by discussing the problem statement, explain what a backstop is, how it's structured, how people are pricing GPU loans, as well as other things to backstop these data center buildouts. Go through a couple of examples in the Asia-Pacific region, some of the tools that lenders are going to need in the future. And finally, the impact on Nvidia's financials. So no better team to dig into this than our, our team that works on Cloud TCO every day, including Mr. Dan Nishball. Dan, can you start us off with the problem statement? What does that trinity capital offtake data centers actually mean and why are these backstops so important right now?

Yeah, let me— let's start off with You know, how we look at CapEx, right? So if you look at AI and data center CapEx, it's soon gonna reach, you know, very large levels, right? So cumulatively about $11 trillion from 2024 to 2029, and we're pretty soon gonna cross $1 trillion of annual CapEx, right? So the debt's already been building up, you know, hundreds of billions per year. Soon going to reach over a trillion per year. So the question is, okay, I think we well understood, we well understand the amount of CapEx coming through. And I think markets are starting to understand given how much debt hyperscalers are issuing. There's a bit of angst over whether they continue. But, you know, the point is the demand is there, right? The demand and the production is there and everyone's gearing up. For $11 trillion of CapEx. And the question is, how do we fund it? So how much funding will be needed is about $7.1 trillion. And the way we look at that is we look at about 75% debt financing, and you can think of that as amortizing. So it's gonna roll off. So you have a bunch of stacks of debt that then roll off, and typically it's about 5 to 6 years. It'll match the contract duration or the expected lifetime. of the GPU. And of course, all of these are very debt financed. It makes sense, it's very capital intensive, it should be debt financed. And we'll sort of go through some examples later of what the pricing could look like, but this really is a problem statement. And the other important problem statement to understand is how big this market is going to be in relation to all other markets. So here what we're showing is AI debt financing compared to all other markets, all other US asset-backed markets. So AI debt financing, including data center and GPU IT CapEx, again, what we said is it's going to reach $7 trillion by 2029. Auto loans, student loans, these are all in like the single digit, you know, low, like $1 or $2 trillion. The only market that's bigger is the US mortgage market at $13 trillion. So the point is there's going to have to be a lot more capacity quickly, and there's going to have to be a lot more understanding. So one of the biggest problems so far, and we'll sort of talk through the objective of like why this exists, right, is what can get financed, right? And what can get financed today is mostly if you have a NeoCloud that has a 5-year offtake with an IG hyperscaler, that can be financed. Anything else, it's very, very hard to finance. So I'll take a step back and just introduce what we call the AI Project Trinity. That's capital offtake and data center. So any NeoCloud that has to build a project has to figure out how to solve these 3 things. So capital, in order, like what I just said, in order to get capital, in order to get lending, you need to have a 5-year offtake from an IG hyperscaler. But to win that from a hyperscaler, a 5-year IG, you first have to demonstrate, you know, you're a going concern. You can find equity to place those deposits. And you can definitely secure a data center. But on the other hand, to get data center, you got to convince a data center operator to want to rent from you or you build yourself. So, you know, these are, these are 3 legs. And, um, the, the easiest solution to this has been, um, let me find a hyperscale off-taker, 5 years, get the capital. You know, I have a small pilot project to sort of convince everyone. Um, so Jordan, yeah, I think that's the, that's a quick summary. Any questions so far?

So Dan, the thing that jumps out to me is that Basically, this has changed. Like, there's been a significant change where off the bat at the beginning, when we saw a bunch of these neo clouds getting started, they could raise equity. And that was just because to me, the scale allowed them to raise $10 million to $100 million maybe to get their first data center off the ground. But this is like, enter private equity infrastructure investors, totally different return profile that they expect, and just like a totally different amount of capital you need to raise when these projects are measured in the billions. So that's like an obvious fundamental thing that I think has changed and why you now need investment-grade hyperscalers signing 5-year contracts in order to even be able to raise the money because people don't want to take that amount of risk on things measured in the billions. But maybe are there other things other than the sheer scale that you're seeing change as people start Trying to fund all of these projects?

Well, I think it's, it's, it's not necessarily the scale, it's, it's the volume, right? So it's not just, um, can I do a $10 billion, a $5 billion? Um, can I do 100 megawatts, 200 megawatts? It's more, can I do 20 of these? Um, so when this first started, um, there were a couple private equity pioneers, right? Very famously, um, Blackstone and CoreWeave. Um, and at the time it was actually like a 5-year backstop. But at the time, you know, it was really only these private equity guys that were willing to look at that. And eventually their conclusion was effectively that, okay, well, it's, you know, Microsoft risk or whichever risk. And so yeah, the loan was a pretty good rate, I think, but it was effectively, if the execution went well, then it effectively turned into hyperscaler risk, right? Now, I think the market has matured and many more lenders are willing to do this, which has driven the rates down. But then to answer your question, what is the limiting factor? It is the sheer quantity. Because even if some lenders are willing to do this and it's more than just private equity, it's actually, you know, it's actually getting a lot of lenders to do this, even if it's the 5-year backstop.

Makes sense. Yeah. So we've described backstops at a high level. Zane, do you mind taking us through how these backstops are actually structured?

Well, before that, actually, before we go to that, I think let me talk about, you know, the, like, other problems, right? So, like, you know, I talked about, like, why I talked about where the market is today, right? Today you can actually only in large quantities finance a 5-year backstopped IG, right? But what are the problems to that, right? And what we said a moment ago is to get to $11 trillion of cumulative CapEx, right? We can't just— that's not going to happen just through a bunch of hyperscaler backstops, right? So that's the first problem with this market structure is that hyperscaler backstops are not infinite. So they're not gonna be able to absorb $2, $3, $4, $5 trillion. And, you know, like not everything can be backstopped by hyperscalers. Secondly, it's what I mentioned to you a moment ago, Jordan, that your private equity and private credit have led the charge, but as spreads compress over time, the returns are gonna be too low for them. The capital needs will increase. It's gonna need a broader set of capital providers. The other problem is it's really like a one-sided market. It's really only supplying compute to 5-year off-takers and no one else. So inference providers can't get the GPUs they want. Jordan, you could probably talk much better about this than me, but there's this total lack of variety, like 1-year, 2-year, people can't just get what they need.

It's all the startups. I mean, everybody who's trying to build, who's left one of the frontier labs and they're trying to build a new model in a new modality, like let's say they're working on material science, or they're working on drug discovery, or they're working on video generation or robotics or world models or anything that the frontier labs are not like exclusively focused on. They raise money from VCs on the promise that they're going to go rent a ton of GPUs, probably spend 80% or more of the money that they've raised on GPUs. And they really want the biggest cluster they can get for about 6 months to a year. But the answer that they're getting from NeoClouds right now Is that your $80 million isn't good if you can only commit to a year. Instead, give me $15 to $20 per year over the next 5, and the size of the cluster that you expected you were going to get has been divided by 4 or 5, and that's how we can actually get a deal done. And that wasn't really the promise of NeoClouds upfront because at that point, you know, you're waiting 6 months and you're spending all this extra money and you're not getting a big enough cluster. It's not really a venture. style approach where you just like take a big shot, try to train a great model, and then raise the next round of funding and go even bigger. You're basically ending up in a world where you have to run a long-term sustaining business, which is again, not what you raised money for. So it looks very different than the way it did 3, 4 years ago when people were starting these companies and looking for compute because really these neo clouds cannot— I think it's due to the scale that they're trying to operate at now. There's just so many of these companies, like you said, Dan. But nobody's willing to take the risk of speculatively buying a billion dollars of GPUs and then hoping that they can, that everybody gets another round of funding in a year or two. However, obviously Nvidia has recognized that this is a need in the industry and they're going to come in and do something about it. So maybe you guys can take us through what that actually, what that structure looks like.

Enter the backstop, right? Enter the backstop, right?

Exactly.

And this is, you know, Zane will take us through what exactly it is and how it's going to solve that problem. Zane, you want to explain?

Yeah, for sure. So this, this, the way that this backstop is structured, typically we see it being structured at like a 5, 6-year duration. Here we have it at 6 years for a GP300. And what Nvidia does is basically they're ready to purchase at these pre-agreed levels from the NeoCloud and then anything above the NeoCloud. So let's say on a year 1 basis, right? $3.68. If the NeoCloud's able to go ahead and rent it out for something like $6 instead, then everything like $3.68, that goes straight to the NeoCloud and anything above the NeoCloud. Like that is split between the NeoCloud and Nvidia. And the reason why we have this at a declining like rent profile is because of the naturally expected decay in GPU rental prices, right? But you'll notice the average backstop price here we have is only $236, which is very, very low for a GV300. And we'll delve a bit into the rationale for that later on. But for now, suffice it to say that. NVIDIA is not doing this to guarantee this NeoCloud a really high rate of return, right? They're more of just doing this to make the NeoCloud, to make all these clusters like financeable in the first place. So that's the overall structure of RackStop. And we've devised a few scenarios in which this RackStop could be used. So the first scenario here is we are running a short-term rental book for GP300s and Obviously, that's going to get you pretty high rental rates here. So we've plotted that out in a 6-year average price of $427, starting from $675 in the first year. So once again, the customer would pay the NeoCloud $675. The NVIDIA backstop price here for the first year would be $368. And then, so at $368, the NeoCloud takes, and then anything above $368, we would assume a certain revenue share between NVIDIA and NeoCloud. So we've since that in the various scenarios here. But essentially it's pretty intuitive as you would expect. If we just jump over to the right here on the IRR section, of course the more of that percentage revenue share goes to Nvidia, then the lower the IRR for NeoCloud. But essentially that's, this is what we see the The backstop meaning to do, it enables the NeoCloud to take a bit more like speculative, like risk renting to a broader bunch of customers, like we just mentioned, renting on a shorter year rental book. And in this case, the NeoCloud doesn't necessarily have to go and find like a 5, 6-year investment-grade hyperscale off-taker and is able to rent to, you know, a broader bunch of consumer bases, right?

Yeah.

So go ahead, Jordan.

Yeah, I'm curious, obviously Nvidia's kind of double dipping here, right? Because in one sense it's the backstop, but they're gonna make money off of that loan. And on the other side, there's the potential to make upside for them when they can actually go and connect buyers and sell above the backstop price. So maybe can you talk about the other side in terms of how the loans are getting priced, I guess?

Well, I think one thing might be interesting, like a small detour, right? Is if you go through the DSCR analysis, because I think the important thing is like, you know, why does the structure solve for banks, right? Like, why does it solve for them being able to lend without an off-taker, right? So this is how they analyze it, Zane, if you want to. Explain like how banks are able to take this backstop and lend against it.

Yeah, yeah, sure. No problem. So we've run a DSCR analysis of the same, of actually just the backstop price itself. And what we've been hearing from banks is that in order to finance like this kind of an arrangement, they would need the DSCR ratio to be above 1.3, at least for the first couple of years. So here we've put in our backstop price and just on a cluster of 100 megawatts for GB300. Of course, this is illustrative. Then you get the following DSCR ratios, which are quite healthy on the first few years. So in other words, it's not sufficient, like this Nvidia backstop price is not sufficient for the Neo Cloud to earn a very high rate of return on their GPU rentals, but it is sufficient to give lenders the kind of comfort that they need to be able to finance projects like these. So that's where Nvidia aims to land with this kind of backstop pricing.

So effectively they're lending against the backstop and what they're thinking about is, hey, worst case Nvidia will take the compute. So they're thinking in the worst case, what is the cash flow I'm going to get? And then If that's my worst case, I can lend against that. And then if the NeoCloud wants to go and, you know, not avail the backstop, because again, right, they don't have to, right? The whole idea, by the way, is that they don't ever use the backstop. Like no one actually wants to use the backstop. Everyone wants to be above the backstop. NeoClouds want to do 1-year, 2-year business. But for the banks, they're like, okay, what is my worst case scenario? And that's how they, that's how they Yeah.

I mean, so eventually Nvidia, you know, needed to use the backstop. What kind of, you know, what would they even use the GPUs for? You know, are they just paying money just for the sake of backstopping these GPUs?

Plenty of research. I mean, you've seen the Neutron models come up from Nvidia. They do research on all sorts of different smaller things and they've got a Massive CI fleet to make sure that anybody's code in PyTorch, VLLM, SGLang is actually going to run well out of the box. I mean, in part, any GPUs that NVIDIA produces that they rent for themselves and use is going to improve the user experience for other people using those GPUs. And in some ways it contributes to the CUDA moat, let's say, when they have GPUs for themselves to do research and engineering against. Now there's some limit, right? They don't need hundreds of megawatts to do research or, you know, they, maybe they do, they have plenty of ideas, but what we've seen from them is that they have then taken those GPUs possibly And started to form consortiums, namely the Nimotron Training Collective, where you bring together a whole bunch of different startups and organizations like Thinking Machines and Mistral and things like that and say, hey, you guys can use our GPUs, you can contribute if you're willing to work on open source. And so I think they have plenty of uses for these GPUs. It's very rare that you're going to see any GPU cluster that's just sitting around idle collecting dust. But clearly the point of NVIDIA's business is to rent this to other people who are interested in paying something beyond what the backstop would be at the base level. And anybody at NVIDIA would prefer to rent it out to a hot startup that's interested in training their model on their specific domain with all their expertise for $5 an hour in year 1, instead of having NVIDIA research engineers using it for $3.68 or something like that. And anytime that they can support a Neo Cloud actually renting these GPUs to these startups will guarantee that the Neo Cloud is going to be able to grow their business and then buy more NVIDIA GPUs in the future. So it's a, I mean, I'm making the best case for what they're doing here, I think. But generally speaking, Dan, I think you called them the central bank in the article of AI, right?

Yeah.

And certainly for startups and certainly for Nvidia GPUs, that's true. They're acting as the central bank who in some ways is a market participant, but in other ways is like an external party that is just watching everybody else buy and sell from each other and trying to grow the market, grow the pie in total rather than pick winners.

And it's not just that, right? They're also taking data center leases. There was a report by which they may be buying fiber in order to then reallocate it to NeoClouds. There's a lot of areas that they realize this trinity, like AI is just growing way faster than the trinity can organically develop. So the central bank's purpose is to step in and support the economy, When private credit can't, which is the case here, right? It's just growing way too fast. Yeah.

And let me maybe refine something that I said a second ago where I said that they're not picking winners. They're absolutely picking winners here, and much more so than the central bank does, even though the central banks can sometimes pick winners of who they're going to bail out. Nvidia's not waiting for bailouts. They're actively going and picking the winners, and those winners are the people that they like the best. And who do they like the best? The people that buy the most of their So this is a tool that they're using as an active market participant to support people, not necessarily the same way that a central bank just kind of sits back and sets rates.

Yeah. Why don't we, Zane, do you want to jump down to the bullseye diagram? I think that'll illustrate Jordan's point really, really well because Nvidia can decide who are they going to sell to. So Zane, do you want to explain our bullseye concept?

Yeah.

That plays into what Jordan said. Actually, now that I realize it, it's like, I didn't think about that, but it is a bullseye, right? It's their target, right?

Yeah, in a way.

In a way.

Yeah. This is, shout out to probably the least technical chart in the whole article, but I hope at least we understand this. So I think the way that we kind of think of it is is a few concentric pools of demand, right? Like these are basically people that can buy NVIDIA's hardware, but in the first, in the widest pool, you have just all buyers generally. And here the pool is just for NVIDIA to sell hardware to them and get the margin on top of that, right? So this is the first layer, the first circle. Of course, the inner circle is, is NeoClouds, that's the blue one. And they take those chips and they turn them into a rental business. And they are repeat buyers of NVIDIA hardware. They continue to buy them, continue to rent them out, make them available to the ecosystem. The red circle is NeoClouds that are NCPs. So here is your NVIDIA Certified Cloud Partner tier. And then NVIDIA gives, gives them reference designs and priority allocation and so on. But also in return, NVIDIA gets standardization and sort of stickiness with the NCPs, right? And lastly, you also have the innermost circle in green, which is the NeoClouds, their NCPs with a backstop. So here it's probably the stickiest, right? Because NVIDIA, they help to credit support that cluster via the backstop that we just mentioned, but they also take a cut of the revenue above the backstop. And so like this one-time hardware sale ends up being a recurring margin business because of the revenue share as well. So that's, this is our bullseye diagram. And of course in the city you have the NCPs with the backstop.

Makes perfect sense. Yeah.

They're not—

who are they going to sell to? People where they can double dip, right? To your point, Jordan.

Yeah. They're not exactly, throwing darts at this, or if they are, they're hitting the 180 every time. So yeah, bullseye. Anyway, there's lots of people in this industry right now that's pursuing a backstop, let's say, and trying to maneuver so that they can get this. It's definitely sought after. There's only so much to go around. Maybe the one comment I can make is that neoclass that are NCPs in the green circle in the middle, like you said, they get standardization stuff. People can't just choose to be an NCP. They're granted the privilege by NVIDIA, and that privilege represents itself as a SKU, like a part number on a BOM that represents an increased percentage that you are paying Nvidia for your GPU servers and networking. And you're also committing to purchasing all of the Nvidia switching and all the Nvidia software. And, you know, like you said, standardization also means no competition in these designs. You're not going to be able to just put Arista switches in your backend network here, right? You're going to be using the Spectrum X switches if you want to backstop. You can't use somebody else's transceivers. You got to use the NVIDIA-branded LinkX transceivers that are the exact same as the cheaper ones that are 4 times less expensive, right? But, um, I mean, it's a way to keep the, keep the ball moving from— for them. So, uh, good segue. Let's keep the ball moving on our, our side as well. We can talk about how these GPU loans are actually being priced by the banks, and then maybe we can give some examples of backstops that you guys are most familiar with. So we can talk about like the actual projects instead of just this at a, at a high level saying just NeoCloud over and over.

Yeah. And we also have a view on where, where the backstop lending might price. So yeah, it'd be interesting.

Yeah. So I guess some context into thinking about credit spreads is that the way that you price a credit bond is you have the underlying, you know, whatever your government risk is, right? So US Treasuries, maybe at 10 years now, they're like priced at 4.3%. And then you have to think, if I'm gonna lend money to a company, how much riskier are they than the US government, for example? So maybe if, let's say, I think that Microsoft is only, you know, just a tiny bit riskier than the US government. I might charge them an extra 50 basis points or extra 0.5% to lend them money at the rate that I'll lend the US government. It's sort of this interesting idea where you can decompose the level of risk when you're lending to a company where you can sort of understand why whatever credit is priced the way It is. So for example, if you take a look at this CoreWeave chart here, the context to this is that CoreWeave has priced multiple kinds of debt that's available in the market today. The line that you see at the very top, the CoreWeave in the blue or purple line, that's CoreWeave unsecured bonds. So that's trading the market today at just probably over a 9% yield. And you can see the brown dotted line below what CoreWeave would receive, what CoreWeave will pay on interest if they manage to project finance the debt. So in this situation, the debt is secured by the GPUs and they have an investment-grade customer. And you can see below all the different multicolored lines representing the different investment-grade hyperscalers, you know, showing the kind of customers that CoreWeave could lend to. So you can see the 3, you know, big brackets there. That's how you would decompose the credit risk of a CoreWeave debt, for example. So you can see illustratively, if we think about CoreWeave Meta DDTL 4.0, so that was priced at about 225 bps over SOFR. Or what that means is it was priced as 2.25% in interest, more risky than what the US government debt illustratively would be. And if you would try to decompose that 225 basis points, obviously some of it will be the risk of Meta. So Meta obviously has to pay CoreWeave. If not, CoreWeave is not gonna get money to pay the bank. So that 97 basis points is what Meta bonds are trading in the market over the US government bond. And that is the risk that is implicitly taken by Corvi when they have Meta as a customer. Then that begs the question, right? So if we know Meta is good for this 97 basis points of risk, why is there additional, call it 105 basis points on top of that? Well, that 105 basis points would then all be reflected as CoreWeave execution risk. So what happens if CoreWeave doesn't stand up the cluster on time? Now, what happens if CoreWeave, you know, misses that SREs constantly? What happens if CoreWeave has issues with procuring or installing the GPUs? That's the kind of risk that banks are taking off when they price this additional 105 basis points. And then you know that comes into an all-in rate of 225, and then we see the unsecured bonds an additional 400 basis points over that execution risk that we talked about. So these bonds are not secured by the GPUs, but they're just merely you know debt that subordinate to the specific GPU that that's being financed. So what this means is that. Okay. Let's say if CoreWeave goes into bankruptcy, you know, the first debt they'll pay off is the GPUs that are secured with the loans that are already backing them. And then whatever money that's left over, you know, then you go towards the unsecured bonds held by debt investors in CoreWeave unsecured bond. So this is how we think about decomposing the various spreads within different NeoCloud debts. And I think that's quite interesting because that lets us think Okay. In the future, if there are more and more of this kind of debts being raised, what would be the appropriate rate they should lend at?

Yeah, it makes perfect sense. I mean, in my view, like it's quite simple to see that hyperscalers also have an entire other business that's a free cash flow monster. And Corey's entire business is this GPU rental thing. So they're obviously going to have different Credit risk, but the spread is fascinating when you kind of break it down like that in the chart.

Yeah, exactly. Yeah. You know, the hyperscalers are good for the money. They, you know, they're getting unlimited money from different, all different kinds of businesses. But for core REITs, you know, the second method doesn't pay, you know, you know that something's gonna go wrong.

Yeah. Yeah. Let's, and then I think, Hanwen, you had a pretty good chart. where we did the 5-year pricing by deal structure where we kind of, yeah, this one. So I think in this one we a little more explicitly try and decompose these different things and then suggest where we think it would be. So on the right one, you can think this is a bit more comparable to unsecured lending to Corey where you're taking platform risk and execution risk, and then you've got the base rate, which is just risk-free. On the left, which is the other extreme, you're not taking platform risk as much. You're just taking execution risk, which is the backstop, or I should say the, not the backstop, the hyperscaler offtake, 5-year offtake contract. And then I think Nvidia is somewhere in the middle where you gotta believe in their execution. You're taking Nvidia credit risk, but then there's also an element of do they manage to create a good pool of customers? And I think the whole idea is, We want to get the market comfortable with something in the middle. And you can think of this as like a trial period, an initial period for them to really understand how the NeoCloud runs its business. And banks are not just going to look at the NVIDIA backstop and call it a day. They're going to start asking questions about, okay, well, how are you going to build a book of customers? How are you going to build this business? Because the NVIDIA backstop is not going to be around forever. The point is it's only meant to be there For the time being to allow banks and everyone the time to really understand the NeoCloud business model, really understand the risks. And then one day they'll have to lend on a standalone basis, just like they lend to every other business, right? Just like they lend to semiconductor fabs, just like they lend to watch factories. Like every business has duration, you know, it's like a bit ridiculous when people say, oh, but there's duration risk and depreciation. It's like, That's every business on earth. Every business on earth doesn't know if they're going to have a sale the next day, like most businesses.

Yeah. Most people aren't building bridges, I guess.

Yeah. Yeah. Even Hermes doesn't know next month they're going to sell bags. So what's the difference?

But actually, I was thinking about that point you raised, and an interesting counterpoint is that, okay, each Hermes bag is small relative to the overall cost of the factory So for example, on the contrary, if you think of WeWork, you know, they can't fill like in a single office lease, that's a huge proportion of the cost that they have paid for the building. So it would be similar in this situation, right? If let's say I stand up a 200-megawatt cluster and I can't find customers, that's obviously a way bigger issue than, you know, if I can't sell one of these bags.

Yeah, totally. Um, and, and that, that's like, there's a whole spectrum. You've got businesses which are fortunate enough to Have like a Foundry Mall, so then they're not the ones who wear the depreciation. You know, WeWork's a great example. You got property. I think the like to your point, there is there is a spectrum, but it doesn't mean no lending at all. You know, there are there are kind of like worse things which which are super variable, like airlines. But there's a ton of lending for airlines. You know, there could be a pandemic any any time. There could be travel disruptions. That that's one of the hardest. things to get right. And there's still plenty of lending to airlines. And it's even worse because it's like 15-year and then you don't even know, like you're selling tickets like 2 weeks in advance. But I think one thing you do have a point on is it's a very brisk depreciation. So if you do get utilization holes, then there's just not a lot of time to fix it, right? If you're not executing, then time ticks very quickly, right? Do you think that's fair, Kangwen?

Yeah, exactly. I think that was what I was thinking of. Like, you know, you have little time to make it right, right? If something goes wrong. Yeah.

Yeah. Like if you're, if you're commercial property, you have 99 years to figure it out.

I think the other point is that like banks also, they're a bit scared to underwrite, you know, what happens if something goes wrong, right? So if let's say like a building lease, you know, you have that, the building developer defaults on the debt. You know, they do work, they're thinking like, okay, how much can I sell this building for in 2 years, 3 years' time? But for banks, you know, when they approach GPUs, they don't have that same mindset, right? It's more like, you know, I have no clue what these GPUs are going to be worth in 2 years, 3 years. So, you know, I'm just going to take it as zero.

Yeah.

Interesting. I think, but you're right. Like residual value of offices is also a lot easier. Residual value of airplanes is a lot easier. Because it's not like every every year Boeing has like something that flies twice as fast, right? Or three times as fast, you know, except in like in tokens. In nine years, although the seats get twice as twice you know half as wide every year.

So I wish we had more laws for for the internet.

In nine years, we just start teleporting. Yeah, we just double every year. How fast this is.

Okay.

Can you guys give a real example here? I mean, we've talked about it conceptually across these other examples and comparisons, but you walk through a couple of real examples in the article, specifically in Australia and Indonesia with a couple of NeoClouds. Do you mind talking through those examples?

Yeah, I can talk about— Farmus is the largest one in the world. So it's a 360-megawatt project in Batam. And, you know, Sharon actually announced theirs first and then Firmus followed shortly thereafter, just, I think, a week before we wrote this article. Gosh, I don't remember. But anyway, you know, like I said, you know, Firmus is kind of interesting because they started off with actually a cluster in Singapore that was very much like NeoCloud V1.0, raised a bunch of capital and they solved for the trinity originally because they had SCTGDC invested in them. So they had the data center sorted out and they had good investors and they could take a bit of risk. And then they parlayed that into the large cluster in Melbourne. And then also they have one in Tasmania. And there, what they did to solve the trinity is they actually self-built the data centers. And it sidesteps the whole conversation around data centers. So we can talk a little bit more about that. And I believe the next one is actually not going to sidestep that. They're going to work with Day One on that facility. And again, Yeah, that's kind of the, you know, we think if that's the objective as we outlined, if we're right about the objectives, then it's going to be pretty interesting because it's going to be one of the largest sources of compute in the region and it's going to be broadly available to all sorts of folks. You know, assuming that, again, you know, our assessment of the spirit of this thing is that the spirit is not to take the backstop and do a 5-year offtake, right? You probably don't need to pay Nvidia half the revenue, right? So it almost stands to reason anyone doing this is probably going to want to do something a bit broad-based and share in 2. So I think we'll see more in the region. I really feel that this region is somewhat underserved by capacity in general, and there's certainly a lot of capacity in data center in general that's coming up.

So Yeah, that's where, not in Singapore though.

Yeah, it's like high level here. It's unbelievable to see Australia, not the biggest region in the world, and Indonesia, Australia getting 72 megawatts with a planned 102 to scale to, that's 55,000 GPUs by the middle of next year. And Firmus in Indonesia, they're 3rd major project, let's say, after Melbourne and Tasmania is Batam, which is 360 megawatts. That's tens of billions of dollars of capital expense, all going to Indonesia to build out GPU capacity, which again is just a massive amount of investment in a single area that not a lot of people think about when it comes to data centers. And Dan, to go back to your point at the very beginning of this conversation, that's how we add up to trillions of dollars is projects like this all around the world that need to get financing.

Yeah.

And it's interesting because we're having, there's lots of conversations happening right now between banks, from banks that are asking, how do we do this for the first time? Even with a backstop, this is the first time they're really thinking seriously about understanding NeoCloud. And so, you know, there's a bunch of tools that we think they're going to need, and I can talk quickly about that. I guess if you scroll, let's see, I think scroll up, right? Oh yes. Yeah, tools that GPU lenders need. Yeah. They're going to need a bunch of tools if they're going to put trillions to work. They're going to need a few things. They're going to need to know what is a fair market price for rental. So we've actually been working on that for almost 2 years, almost 3 years in fact. We've been tracking bilateral contract prices across the entire term structure. And these are, this is deal information that is basically impossible. You can't get it by scraping. You're not going to find it posted. You really just have to gather it by just surveying a lot of NeoClouds and asking them like, hey, what do you think a B200 could go for? And then Jordan also helps out a lot in terms of connecting people to compute. So him and the Sams have a pretty good sense of where things will transact. And this all goes into composing this index. It's a very different approach than others. It's not very scraping-focused, but we really believe that every data point, there's a story, and we try and recount those stories. We try and explain that in a lot of the research we do and get a sense for the market. The second thing is which NeoClouds are going to execute well, which ones are going to have the lowest execution risk. So if we scroll down, this is what Jordan does. Maybe Jordan, you should explain why this is so important to whether you would want to lend.

Yeah, I mean, to us, like, there's basically a ratings agency that's missing here when it comes to the quality of these, the product that these people provide, and then therefore the quality of the debt. And the only way that you can get an understanding of that is to do hands-on testing and to talk to the existing customers in a real, You know, unbiased way. And that's exactly what we do with ClusterMax. So for everybody listening, stay tuned for ClusterMax 3.0. Testing deadline is August 1st, so we're almost there. And we'll be putting out an article in mid-August to update these rankings based on our hands-on testing and interviews with over 150 customers now of these NeoClouds. You know, in some sense it's hard because you've got to go and talk to over 200 of these NeoCloud providers now to get a sense of what everybody offers. And like I said, over 150 customers that we're connected with, we're always trying to connect with more. So that number's growing. But in another sense, it's easy to get a sense for who's got their stuff figured out and who doesn't. We have a whole checklist of criteria that's public on our website. We've told many of these NeoClouds for months or even years about certain things that they're missing that are big differentiators for why people would want to buy from them and would improve the quality of service that they can provide. And many people just don't do it. Really simple stuff. So in that sense, it's pretty easy to differentiate between people who have health checks and people who don't, people who have monitoring dashboards and people who don't, people who offer managed Slurm and managed Kubernetes and people that don't. And as they expand into other businesses, many of these people serving tokens as a service or serving RL environments and just going up the stack or Many of them going down the stack where they do self-built data centers like you're referring to with Firmus, who need to be able to assess their competency across all of those different facets of the business. And that's exactly what we do with ClusterMax. So anyway, it's a fun project to work on because we just get to learn so much. And Dan, to your point, when you're trying to draw a curve to show the pricing of stuff and how it changes over time, some of these individual stories are worth so much more. And should be weighted so much more heavily when compared to, you know, one spot price that anybody can get from hitting a public API to see that that vendor has chosen to price that GPU instance type at a certain price on a certain day. Talk to somebody else who's rented, you know, who signed a $20 million GPU contract and talk to them about the decisions that they were making at that time. And you're going to get a much better sense for First of all, what that contract is priced at. Second of all, what they care about when they're choosing a provider and how they're going to discount certain providers that can't provide a high level of service versus pay a premium for those who can. Um, and then compare that to some on-demand price that somebody is paying for a single H100 that they're going to rent for 6 hours to train their anime bot or something on. Very, very different business. So. Anyway, yeah, I think in summary, the ratings agency is basically missing here and we're doing our best to help people solve that.

Yeah, I thought this section of the note was pretty prescient because I think, call it 2 weeks after we released this note, Navius announced they raised their first debt facility, right? And then, you know, turned out that it was at a similar rate to Corvi, suggesting that banks thought that execution risk, you know, was similar to Corvi, which, you know, if you look at the Table we have here, we have Coweave and Iblis just side by side.

Makes sense.

And it's pricing of debt, but also pricing of their products. And it's going to be important that if you're lending, you know, the question is, are you lending to like a Hermes? Are you lending to like a, I don't know, what's it like an H&M, right? Are they going to be priced high or low? And, but that's a real thing because If they're doing, you know, Kubernetes, Slurm, right? Versus like just bare metal or like whatever, like there's a price difference, right, Jordan? Yeah, exactly. There's a price difference between what you can deliver on SLA. And so, you know, it's important for lenders to know what they're getting into.

Yeah, look, I mean, you guys have called this out that there's this backwardation curve where people who have GPUs for sale that are available right now are charging a massive premium and those who, you know, have delivery in 3 months or 6 months, provide a discount just to lock in the customers. And the big variable that people play with right now is the prepay amount. So having those milestone payments as you're going there so that these NeoClouds don't need to raise as much bridge financing or take out a construction loan or just change the structure of the debt that they're raising because they can get these prepayments from these customers that trust them. I mean, that changes their financials completely. So that's a big thing on the pricing. But after you get past that, then there are, you know, 2 people offering the exact same GPUs in this similar location with the exact same headline price are actually going to deliver very, very different things, which is what we've talked about on previous episodes to the tune of people willing to pay in some cases 20 or a 30% premium for the exact same quantity of GPUs just because they stay online and they perform compared to some other providers who are less trustworthy. And so That's represented in this chart. The high level is represented there, but for anybody who's looking to dig into more details, please get in touch with us because we do a lot of this consulting work for people who are making these critical decisions, right? When you're making a decision on where to send $50 million of the VC money that you've raised and kind of bet your company on some new cloud, you really need to be able to trust them. And we're just providing people with that peace of mind. to know what they're getting or to just think through the decision, right? How to negotiate, how to get through the SLA discussion, how to write a contract, what sort of things to get the NeoCloud provider to commit to. I mean, these are big questions that we help people answer.

Yeah.

The last piece of the puzzle is, if you scroll down a little bit, is InferenceX. A lot of people ask, what is the residual value of a GPU? It gets into what's a GPU worth. You could argue what a GPU is worth is what it can generate, and that's how many tokens. So InferenceX tells you how many tokens per second you can generate per GPU. Folks can make assumptions on how much the tokens are going to cost, but this is really important because it's the fundamental revenue-generating capacity, just as much as how much bags cost, how much flight tickets cost, how much nuts and bolts cost. This is really the—

Yeah.

The unit cost, the atomic unit of revenue for the entire industry.

Yeah, everybody pays for tokens right now. Some special customers rent by the GPU hour when they're the ones training the models, but the whole ROI on the whole industry depends on people buying tokens. And that's this chart right here.

I thought inference is taking over from training though, Jordan. I thought training is going down.

Yeah.

Okay. You want to have a whole nother podcast right now where I rant about the shift from training to inference?

For the record, we don't I actually think that, for the record.

No, but did you see the 2 charts that I found from back in 2018?

Yeah.

A couple of our favorite forecasters who are linear extrapolators instead of exponential extrapolators, the way we are at Semianalysis, in 2018 were projecting that in 2025, one day, inference might be a $10 billion market.

Yeah. Wow. You're right. One day.

They're right.

Yeah. So we've— and they had inference at double training at that point. And actually edge inference was going to be, I think, $5 billion and data center training was only going to be $4 billion. This was our peers at McKinsey, let's say, who are making these forecasts. I mean, I'm not sure anybody in 2018 could have gotten the order of magnitude Right. And seen the explosion like this, but the ratios were just never making sense that people would someday decide this is the final model we've ever trained. We're not going to invest in any more training. We're only going to buy GPUs for inference. This doesn't make sense. The frontier labs believe they're building the machine god. All the evidence to me points to the fact that they are, and they're going to try to invest as much as they possibly can in training. Inference is a means to buy more GPUs to train the next model.

Wow. Feels like an MMORPG. It's like you buy items to kill monsters, to buy items to kill monsters, right?

Never-ending cycle.

That, that probably describes life, right? You know, it's like we're all just killing time.

Okay guys, anything you, you think that we haven't covered on this episode? I think we've gone off the rails on quite enough tangents on our tour.

Yeah.

Tour of the Holy Trinity, right? I think we've subjected the listeners who are still hanging on to enough talk about GPU debt and financing rates. If you guys are interested in seeing anything fun or clever or witty, we can do it right now. I'll give you a 3 count. Two, three. Okay, that's a no from the Singapore team. We've had enough wit.

We are not smart, no witty.

For today's episode, have our wit drained out.

Sorry.

Yeah. Well, appreciate everybody taking the time to listen in here, guys. Good job walking us through everything to do with the Trinity capital offtake in data centers and what what Nvidia is doing with these GPU debt backstops. Hopefully it was fun.