You're Not Financing the GPU. You're Financing NVIDIA's Margin: Bernie Margulies (American Compute)

0:00 We discount all the hype and mania. So right now there's huge increases in memory prices. if you're seeing your equipment prices go up, a lot of people are saying it's because of memory. Some of that is just because they're adding most more margin and they're just squeezing people. But a lot of it is memory increases. And so the memory price increases, we have to somehow discount that. So we look at memory prices before the excitement. So more like 2023, Mm-hmm.

0:23 2024. What's the memory price? price then Hello everyone. Welcome back to Where Capital Meets Compute podcast. I'm your host, Kony today. We're very happy to have Bernie, the founder of American Compute with us to introduce more about what they do on the insurance products on compute. Welcome, Bernie. Yeah, thanks for having me, Kony. So just a quick intro. I'm the founder of American Compute. What we do is risk management and credit backstops for GPU finance.

0:57 So we help transfer risk out in these structures that makes hopefully the lenders, equity investors, leasing companies, comfortable to do these deals. Perfect. I think this is where the market is needed and also people talking about constantly lately. So why don't you ~ help us understand a bit more on what exactly are you offering and a little bit more information about product if you could. And how do you come up with Yeah, that that ~ idea? what were you doing before this?

1:27 Yeah, definitely. So the context on the products, right, is A lot of these projects are not the most credit worthy. That's the that's the truth of it. Like a lot of these projects are by startups, they're by younger companies that don't have 20 years of financial history. And so underwriting these projects is very hard for the lenders. And so we develop these different products to help get the lenders comfortable and help get the lenders in a more safe position to do these deals.

1:54 ~ And so the way we got here originally was my background is more in investments. so I spent the first few years of my career in investing. I was actually based in Singapore. when I was doing that I ended up encountering a lot of technology companies and getting really into the like the tech scene and I was like, okay, I want to go join a startup, I want to be involved in the startup scene. And so I moved to the Bay Area, I moved to SF, and I got more involved with startups there.

2:19 and eventually I I somehow got into this rabbit hole of insurance. when we were starting American Compute, it was just the idea of, okay, compute is this big asset class. Insurance is this interesting financial mechanism. Are there cool ways, are there interesting ways we can use that tool? And quickly lenders, neo clouds, just people in this industry brought me into hopefully de-risk certain aspects. And so we started building out these products, like, hey Bernie, can you insure this risk?

2:45 Can you insure that risk? And we started building out these products. And really all the products are geared towards how do we make this safer for the lender or investor so they can move forward with the project. They can move forward with the financing. essentially the most famous product we do, like what we're most well known for is residual value insurance. So we are arguably the only provider of residual value insurance for GPUs in the market. We're the only ones who can structure it.

3:10 ~ and that took a lot of effort. And there are other products that are newer from us that are also popular. So we also do appraisals, we also do diligence reports, we also do something we call an off-take interruption cover, which guarantees stable cash flows and stable debt service. So we have other products as well, but what we're most well known for is residual value insurance, which sets a resale for future values. It sets a resale floor for future value of GPUs.

3:38 Very great, very great. So let's touch a little bit about on that. So like on our side, we as underwriter or lender, we definitely care about the residual value of the GPU. Especially, you know, when we have entered into such different generations of mining, including Bitcoin mining there, where quite a lot of people compare like GPUs and Bitcoin mining machines where those machines were usually going to be obsolete very soon, especially in the mining days. Of course, GPU is a different scene.

4:04 And that's why when we as a lender we care the most is if anything goes south, can we recover on these asset value? And what exactly is the asset value there? And ~ there was a precedent in the past around like 1970s you know, the Lloyd story that they got a failure, ~ they couldn't cover ~ these assets residual value. because of the technology obsolescence and also the new embedded pricing machines get came out. And on your side, when you are pricing these assets, especially such young age asset we were talking about, you know, enterprise create GPU, A100, H100, were only released more than five, six years ago, how are Mm-hmm.

4:44 you really coming to the reasoning or how are you really coming to the pricing, forming the benchmark, forming the thesis? on these GPU residual value there. what data are you relying on? Yeah, so you know, so a couple things. One is it's important to understand why there's been previous failures. Like people have tried residual value for many asset classes. Not just you know, what we're doing with GPUs, but people have tried it for laptops, people have tried it for boats, for airplanes, people have tried it for many asset classes.

5:16 And In many asset classes, it's failed horribly. It's actually a very hard product to underwrite. ~ so the IBM one is a good example where with these IBM mainframes, ~ there was a lot of frenzy, there was a lot of hype, there was a lot of excitement. Just like today with the GPUs with AI adoption, that everybody then was like, We're gonna adopt computers. And they were replacing humans with computers for the first times. And at that that 1970s, like they still called calculators were referring to humans.

5:43 Humans would calculate, and they were there's a job known as a calculator, and then they got replaced by these little rectangles. so that transition was really devastating for residual value insurance. There were a large amount of losses out of Lloyds of London. ~ same with like a more recent example, people were doing residual value insurance for Teslas, for EVs. Mm-hmm. that was actually pretty bad too. it wasn't the greatest performance. And the reason it usually ends up being painful is Everybody always obsesses over technology, right? They always think it's about technology obsolescence.

6:15 I would argue that that is a smaller risk compared to margins. So I'll explain a little bit more. If you look at NVIDIA's gross margin on their hardware today, they have a 70-75% gross margin. So when you're financing these projects, In a way you're not even paying for the hardware cost. You're literally just paying NVIDIA seventy five percent. You're you're financing Yeah. NVIDIA. You're not financing hardware, you're financing NVIDIA. And so that's scary because if NVIDIA ever does a pricing change and it's a pricing change in a downward direction, they they cut the their GPU prices in half, which they can do, right? So if their their GPUs today are five hundred K a server, they cut it in half to two hundred fifty K.

6:59 They're still printing money, they're still making tons of profit, but that would be devastating. to the residual prices of all equipment out there. ~ so the the the NVIDIA NVIDIA gross margins are actually very important. And so y we you can study NVIDIA gross margins. Like historically what have they been like? For other ~ silicon manufacturers and designers, chip designers, what have their gross margins been like historically? So I can tell you NVIDIA's gross margin historically was like forty percent.

7:28 Mm-hmm. That's almost half. So 75% to 40%. So there's a lot of headroom to c to cut there. ~ I I don't I'm not necessarily gonna say that's gonna happen. ~ there's a lot of reasons why that shouldn't happen. Nvidia has a very strong dominant position in the market. They would only cut prices and and lose gross margin if they have competition. But that's something we have to consider. So there's a lot of work we do that's kind of qualitative and really understanding what drives residuals down.

7:54 So that's one side of the business. The other side of like the underwriting. is kind of what you said where we have data. So what we do is we look at the historical data for ~ all this kind of data center enterprise IT equipment. So the CPUs, the storage nodes, the memory, ~ the switches, we we collect data on that. We work with different ~ equipment resellers, equipment remarketers, we collect the data from them, and with that data we have a sense of okay, this is what a two-year-old memory of this configuration ~ has a residual for like this is what it transacts for.

8:28 And so we collect all those data points and what you you end up getting is like a distribution. You get a spread of data points. It's not one number. You'd think it's like it's organized, it's transparent, it's one number. No, it's a mess. Everybody's like quoting different numbers and so you have a distribution. And so with that distribution we can model what we're what we expect losses to look like. We can model where we think resales could happen for future equipment.

8:53 and we also discount this is another step we do, it's very important. We discount all the hype and mania. So right now there's huge increases in memory prices. if you're seeing your equipment prices go up, a lot of people are saying it's because of memory. Some of that is just because they're adding most more margin and they're just squeezing people. But it a lot of it is memory increases. And so the memory price increases, we have to somehow discount that.

9:15 So we look at memory prices before the excitement. So more like 2023, Mm-hmm. 2024. What's the memory price? price then and that's kind of like our reference and that we look at distributions from that point rather than distributions from today because the distributions today are completely different and like are skewed much higher. So that's how we approach it. We we do a lot of qualitative analysis on in understanding the margin profiles of these different silicon manufacturers and we do a more quantitative analysis with secondary market data points and we model these distributions and we look at the data before the mania.

9:49 Understand that. And ~ what l let's dig dig a little deeper into the the data side because that actually forms the foundational thesis of how you do the pricing for these entrance products. Like in terms of ~ the data, like how many years of data do have you got or what like how many data points do you think is sufficient enough for you to make a judgment? For instance, if we have a new type of GPU that comes out, let's say a Vera Rubin that will be soon out in productions.

10:20 Like, are you gonna do the pricing or resale value insurance for them from day one? Or are you gonna wait in order to get enough data points for these like resale prices, GPU prices, and even like the whole the whole prices change there. Like when would you do it and how will you do it if if this is like immediately? So ~ you know we we we can price it, ~ but we won't price it day one, we'll wait a few months. So I'll try to explain.

10:52 So we're scared of of insuring assets or underwriting assets that are not going to be popular for some reason or another. And I'll try to explain. Mm-hmm. ~ So imagine like every year there's a there's a Honda Civic that gets released. Every year, it's you know, just every year, twenty twenty, twenty twenty-one, twenty twenty-two. And you talk to car people, they will swear twenty twenty-four or something like that. They'll like they'll tell you like a specific year, yeah, you gotta get th this this model, this year. It's the good one. The next year is no good.

11:24 You gotta get this year. That's the best one. They they the people will swear by that. And so there's just some years where it's just not as good. And so we're Mm-hmm. scared that we we would insure an asset that's just It's not gonna be popular, but there's no way to know if it's a popular asset or not until you give it a few months. So we want to give it a few months and in a way you can call that infant mortality.

11:45 ~ Mm-hmm. it's kind of one way to to frame it, is like, is this is this gonna just die and be unpopular in the first few months? As long as it survives a few months, we'll probably be willing to insure it and we'll probably be able to ~ break it down. The the other part that's hard, and this is more specific to the newer equipment. There's a growing trend of doing fully configured systems. So before, you know, you you would buy them, you could buy GPUs as cards and you'd you'd buy them as servers as nodes.

12:15 Now you're buying them as racks and you're buying them as superpods. And so those are fully configured. How do you resell that? How do you break that up and resell it? Can you break it up? Will it lose a lot of value when you break it up? So there now we're at this area where there's some unknowns because the systems are being configured differently. Now if we broke it down and looked at the components, let's say the memory, the CPUs, yeah, we could have a view. But Yeah, they're they're they're doing some stuff with the packaging that's getting more complicated. Like, you know, with the the new GPUs it's it's not individual cards, it's like they call them packages and they they glue two together at the same time.

12:51 There's stuff like that. You can't break it and and sell them separately. So how do we model that? So that that that's that area gets a bit more tricky. but I would say in general we're comfortable doing it as long as we've seen a few months of the market accepting it and liking the asset. If the if the market, you know, ends up not liking it, we'll probably know within a few months once there's like users complaining, developers complaining, just th there'll be some market feedback.

13:16 But so it's not day one, but maybe day one eighty. I see. I see. ~ okay. You touched about two points. ~ I think popularity is a key where you think ~ which entails several things, right? If it's popular, then you have a lot more buyers on both ends if you want to sell it out ~ and and later if if things does happen, you can still have a market for it. And the other thing as you said, it's a good machine so that you get the usage is high, the cash flow wise later can potentially like you know, can you equip equip the asset value there.

13:50 Okay, that's very cool. ~ in terms of the methodology here, right? Are you you you you that's a that's a one one there's one thing you mentioned very very very very importantly is that the recent products for GPUs has been going into even more heterogeneous. Like we don't we used to have only cards, right? PC PCL E cards there only, and now we have like notes, we have racks, we have super pots, different types of that.

14:18 And in in these different types of products, how are you pricing, for example, the premium in terms of packaging them together? Do you actually take that into account or do you actually look into the components as you set in order to give them a pricing? And other other than the ~ GPU card level, are you also ensuring, for example, the the memory side and then the network switches, these type of equipments together as a whole? Yeah, so ~ to to I'll answer your last question first.

14:46 We do everything in the rack. that's what I tell clients is we' we're we're we'll we'll insure everything in the rack. Now there'll be some exclusions, like probably not gonna insure our your PDUs. ~ we're not gonna insure your coolant. Yeah. ~ it's just not not worth the time. But ~ you know, most of the equipment we're we're we're willing to insure because most of the equipment does hold value. but that includes memory and networking, definitely memory right now. ~ the the previous question about like how do we look at the fully configured systems and how do we think about breaking them down?

15:17 So what we do on a fundamental level is we look at components. There's a couple reasons we look at components. One is we think that's a good worst case scenario. Worst case scenario We'll hire someone to break it apart and we'll sell them, we'll rip off the memory and like that's like the worst case. So that kind of gives us this idea of like a in in the in in residual value insurance, they call it salvage value.

15:41 So this is like a salvage scenario where like the worst case, we're gonna break it apart. So for example, for boats, if you do residual value for ~ boats, it's literally what if we just took the steel from the boat, the hull of the boat, and we sell the raw steel. So like that's salvage Mm-hmm. value. so here is like what if we just take out the raw components? That kind of gives Use a sense of salvage value. The second reason we like we like components is there's more data points. Like if you're just looking at fully configured systems, There's very few enterprise transactions of fully configured systems in a secondary market right now.

16:13 They just don't exist. It wasn't a thing six years ago, so they don't exist in the in in our data sets. And so it's hard for us to price that because it's it's just unknown. So we'd rather look at components because we just have more data. ~ and so we like to look at the components and then w when we look at the rack or the the superpod or like this the fully configured system. What we're looking for, and we we make we make adjustments, what we're looking for is a system that will we call evergreen.

16:41 So the idea of evergreen is it'll always be popular and it'll be popular with anybody. For example, H-100s. H-100s are a great example of evergreen. Everybody's still using them, everybody's happy with them. There is are there more powerful stuff? Yes, but ~ just use H100s. It's like like a lot of people just measure their workloads as it takes eight ec X number of H100s. It's just like a good standard. Exactly. Yeah. ~ so though those those kind of those kind of things are evergreen.

17:10 And it's hard to say what will be evergreen, but we try to take like a qualitative lens. Like for example. Air cooled stuff. Air cooled stuff is more evergreen because if you have liquid cooling, you cannot put it in every single data center. A lot of data centers don't support your liquid-cooled chips. So that's not as evergreen as air-cooled chips. So memory is also another one. If you have a high memory configuration, that is more evergreen. So that's what we're looking for.

17:36 We're looking for these things that are ~ more evergreen. And this might not be intuitive, but The interesting thing is, and this is very clear in our data, when you have something that anybody can use, it has a higher residual than something only one person or a small set of people can use. So if you have a very specific configuration configuration, yeah, a hundred percent. Yeah. Like like exec versus it GPU. Yeah. Exact exactly. So the the more generic it is, the better. And that's why I think, you know, NVIDIA has a really good positioning because they're used for both training and inference.

18:11 They're not training only, they're not inference only. They kinda do everything, general purpose. Those hold residuals a lot better. it's very clear in our data that that holds residuals better. But it's not necessarily intuitive because sometimes, you know, you think, it's like limited edition, a special edition should have a hold value better. Like it's like if I have a jersey from Kobe Bryant, it does better than a generic jersey or something like that. But that's not true. It's like y Here it's like you don't want to be super specialized.

18:35 That doesn't hold value as much as just being very general purpose. Anybody can use you, anybody can put you into their workloads, anybody can buy you. That holds a better residual. I just said you touched upon about the defin the the interesting point about evergreen. I think to me the evergreen definition actually changes ~ according to time. For instance, like high memory ~ notes was not so popular until the recent models get super big. And then sometime like when you know B200 was not so popular until recently, you know, B200 like after H200 are getting a little bit more popular lately.

19:12 So that ~ the technology maturity and also the adoption for the market actually dictates, as you said, like the the the the definition of evergreen, whether people want it to buy it. And that trend that that and that trend will not come like so early. And that makes sense why you are you know offering these products and insurance. After a little bit you have data and get a sense of where the trend is going going to. And can I jump ask a bit more deeper about the methodology here?

19:39 Like ~ Mm-hmm. you mentioned that you will look at the components prices. in order to give a visual value there. Would you care more about the GPU price index or like how much they are being rented out or leased out at in terms of using the other angle, like the cash perspective to to ~ get to the present value of the asset as another reference? What are the ways that how you value it other than just his historical numbers on these like memory chips costs when they were not so high?

20:11 Yeah, actually, you know, we're we're very strict about this. Almost all of our work is based on real transactions we can get in the market. I would rather trust four resale data points ~ on the in the secondary market for used equipment than the rental prices. And I can Mm-hmm. explain why. We went and did a correlation study. We compared, okay, these are the the resale prices of this equipment. What were the GPU rental prices at in the same period?

20:40 They literally went in like different directions. They were like not matched at all. I see. So when rental prices went up, it doesn't mean necessarily that the GPU resale prices ~ go up at the same time. It can lag and like before it like it it's not always one to one. And so it's not actually the best indicator. ~ I think it's it's it's generally positively correlated. Like if rent rental prices go up, yeah, GPU prices will depreciate slower, but it's not one to one and it's not as strongly correlated as you'd expect.

21:08 ~ And I think a large reason for that is, you know, a lot of these rental indices aren't tracking the same things as ~ what secondhand buyers care for, right? A lot of these rental indices Mm-hmm. are they suck in a lot of like consumer, not consumer, but like more hobbyist grade data where it's like s smaller developers, etc. ~ but you know the sec yeah, On demand data bringing. Yeah. yeah, on demand data is like versus like, you know, enterprise long term contracts, et cetera. So it just it ends up being not correlated.

21:43 I don't fully know why, but I can tell you it's not correlated. And so what we do is we go off almost purely off resale transactions. I mean I I'll look at the s the the indices and I'll I'll check it out because I'm just curious. But it doesn't play a big role in our methodology. and it it's because it's That's the main reason. Well, that's kind of interesting. Like, you know, a lot of the not only lenders, but also investors and the cloud companies, when they look into buying what what type of GPU, they definitely go for the index.

22:13 And then now you're telling them, hey, you know, that type those subtypes index doesn't really dictate anything because anything on the residual value. Like you can't rely on that and say, Hey, we are buying an asset that can retain this value as much and earn this X amount of X dollar as much as well. So that's a very interesting observation there that ~ yeah, I think the market should pay attention to. One last thing before we jump to the next session is that ~ you argue very, very high ~ highly ~ that compute is not a commodity.

22:45 I would like to like like you to Sure. actually explain on this hot take there. Yeah, no, so I mean it really depends on how you define commodity. But what I I I would say is I I hear a lot of people say compute is oil. Compute is electricity. And you know, I I I don't know enough about oil or electricity, so I I don't have the full knowledge. But I would argue that compute is like real estate.

23:13 It's a very expensive asset. You park it in one place and it does it generates income. It generates some kind of output. But it's it's like a hard asset that you own and park. ~ so imagine like treat it like an apartment building or a factory. it's not something you can easily swap. One cluster here, one GPU cluster here in Washington, is not equivalent to another cluster here in Ohio, right? With a different set of different, you know debt schedule behind it. It's like a it's like an asset.

23:44 And I think that's a better way to think about it. And when you look at the real estate market, I don't think people say real estate is a commodity. now is it super differentiated? Is real estate super differentiated? One apartment building is it really differentiated from another apartment building? Not necessarily. They can kind of do the same stuff. It's they're both apartment buildings. But Are they easily fungible? Are they easily swappable? Can you easily design hedging products around them?

24:09 I think it's quite tricky. What I would argue is that tokens are a commodity. Tokens you can easily swap. One, like if I have these ~ Kimi tokens or GLM tokens from this provider and I have the equivalent models tokens from another provider, those are kind of exchangeable. ~ and those I think you would actually design like products around where I can say, Hey, I promise you X amount of tokens at this token speed and this kind of like parameters.

24:37 And y anybody could serve that contract. And so then you can ha kinda have that commodity feel where you have all these derivatives. I think it I think they should be focused on tokens. ~ it's just that right now the market isn't fully on inference yet, and the market's still on training. And so that that's why the focus is on compute. But as the market matures and we move on to like an inference-based world, I actually can imagine you can do all kinds of weird derivatives ~ for token.

25:03 Yeah. Hundred percent, hundred percent. Yeah. We we yeah, hundred percent. We're cooking something there too. ~ so I totally agree. Okay. ~ Yeah. And then ~ regarding the the the the products, right? Regarding the insurance products that we're talking about. Like in usually in GPU or GPU financing and in the things we do, we care about several things. We care about the terminal value of the hardware, we care about the ~ continuity of the revenue, like whether there's cash constant cash flow streams to pay us back.

25:34 care about the credibility of the borrower and also the market price of the computer itself, which is ~ the like the the market the the market price, et cetera. And a lot of different Mm-hmm. people would ~ would would would like combine all the risks and talk about them as one thing, but you actually trying to separate them first. Let's you separate them into still value insurance product. And then now you you have the offtake interruption ~ insurance product there.

26:02 Like I want I would like to ~ understand more from you, like how and why you you're doing this and how, okay, what are the risks that you think ~ that can be insured? The second thing is how are you like ~ approaching them, like starting with residual residual then interruption? And also are there any risks that you think out of the four, even more that you think is they're not insurable and they cannot have any products being built on there?

26:28 Yeah. So to be very straight ver like very blunt, there's only really one risk in town. There's only one risk that anybody really cares about. It's credit risk, right? That at the end of the day, it's credit risk. Like you you talk to any lender, any investor, the reason they're worried about the project is the the off taker or like the people involved, all these AI startups, will run out of money. The the clock will stop. ~ there's a bubble and it's gonna pop.

26:55 That's that's what people are scared of. And so that I would argue is very hard to insure Because you know i if you n if you're confident or you think there's a high chance of a bubble popping then why would you insure that, right? If you if you know there's gonna be some credit issues in the market and the market's really hot right now, why would you insure that? So like that I think is impossible to insure.

27:16 But can we kind of mitigate some of that? From other angles. So instead of doing credit insurance, can we do like these credit backstops? Like the synthetic way to approach credit insurance, but with other products. So for example, like what you just described, residual value. So instead of focusing so much on the borrower's credit, can you look at the the assets value? And that's how people think about houses, right? In the houses, you're you're lending to this random consumer, this 40-year-old John Smith, and you're giving him you know a million dollars to buy a house. He's not good credit, probably, right? You're giving him a million dollars.

27:50 It's a lot of money for a z amount of income. But you're okay because you know the household's value. So is there a way we can atta attack it from the asset? So that's why we do we do the residual value insurance. ~ the other aspect is just like, okay, there's some aspect of the off-taker creditworthiness. ~ and that's on a individual. Maybe you you know you know on the on a market level it's hard to in insure it, you can't insure the entire market.

28:13 But a individual off taker, that individual end user of the compute, if they disappear overnight, can you insure that risk? Just their their default rather than an entire market level default. So That I believe you can insure ~ is there an issue of the operator, the the neo cloud or the the project operator, them screwing up the operations, there's an SLA breach, ~ and the the customer walks away from the project because of SLA breach. That I think you can also insure.

28:41 ~ that that kind of performance obligation. Or delivery timelines. I think you can also insure delivery times. We're thinking about doing something there where you know a lot of these off takers don't want to give a down payment because they don't think you're gonna get the racks installed on time. So if we can guarantee the racks will be installed by in two months and if it's not installed by two months Interesting. we will pay the the the contract value for the the the the time that goes over that for up to like let's say like if it's a month late we'll pay one month of off take back to the off taker so they can take that money and put it somewhere else at least use it on on demand or somewhere else.

29:17 So can we do products like that where we can insure the delivery risk, the delivery timelines. But end of the day, the truth is, really what keeps people up at night isn't the asset value. What keeps people up at night is credit risk. Hundred percent. They're scared of an AI bubble, right? So we can't insure that. And I I would say that's something that's uninsurable. It's just it's it's a giant monster. Like it's just it's too much. But can we tackle some parts of it?

29:43 Where we probably can't insure the entire monster, but we can handle one foot, one arm, like some part of it. ~ so that that's how that that's how we think about it. Yeah. I would I would like to take an the the the evil take here, right? Like if AI has a bubble and then basically what you're underwriting is still the industry beta. Like if the industry beta goes away, then the residual value will be highly affected.

30:08 The out the I mean the credibility of the off taker doesn't really matter that much. Like do you take that into account ~ when you are doing your synthetic backstop? analysis or these product the the the the credit products that you put that you that you are providing right now. The insurance product you're providing. Yeah, there's a lot of consideration going into that. Like, I'll be honest, so like a lot of my capacity providers, so like the reinsurance and the insurance companies I end up working with when we structure these products, that's what's on their mind.

30:36 They're like, what if blah blah blah blah blah. What if blah blah like what if ~ it becomes illegal to use AI? I've had that. Like someone told Yes. me, what if AI becomes illegal? Like there's a lot of what ifs that I have to answer and explain how we think through it. At the end of the day, it comes down to ~ working with the right people I would say. Like we don't want to insure anything that moves.

30:59 We want to insure the right counterparties. ~ we don't want to work with everybody. We want to work with the right people. And that helps mitigate some stuff. we also just don't write super long policies. I'm not gonna sell you a 10 year policy on the residual values. So we try to cap our exposure that way in terms of duration. And the other way, right, is We don't promise you the world. Like I've had people who ask me for an eighty percent residual, in like three years time. I'm I'm not gonna do that.

31:30 Like I just won't like underwrite that. Like I'll I'll underwrite something a lot more conservative. So that's another way. We're just conservative. ~ and I would hope the industry stays that way. I hope the industry stays conservative. Don't give money to everybody, give money to the right people, ~ with the right terms. Yeah. So let's go for this prime instead of subprime. ~ Yeah. and then ~ can you explain and expand a little bit about the off take and interruption product that you just have?

32:01 ~ I think this is very important, especially ~ know, we have seen recently the dynamics between the buyer and the seller on the compute both the market side are very different. For example, my early days, we've seen some neo clouds that want to have long term contract and then off take say no, no, no, no. Let's do short time, right? That's right before last year. And this year, of the neo clouds suddenly say, hey, hey, I don't want a super long-term contract.

32:26 I want to have like two or three years max. And then ~ I can do whatever I want for the inference for more like higher value there. So for these ~ interesting contracts, it's coming to at the right time. Because let's say if there's only one year contract with the off take with the off taker, as a lender on our side, we think it's definitely quite risky because you know Most of the time these asset cannot recoup recoup most of the asset value within a year. It takes like two or max three years to get the money back.

32:57 So that actually comes quite a quite a good product in our eyes ~ to some of these contracts that we're seeing lately. Is that the reason why you offer this right now? What are the rationale and how you think about and whether it's popular among your clients that are buying it? Yeah, so I'll I'll be honest, so it's a very new product for us. We've been ~ working on it for the I'd say the past month, and it's arguably now more popular than our residual value.

33:24 ~ I have more people who ask us about the off-take interruption. ~ the reason it came about was because you know when I talk to these lenders and I talk to these financing parties, what they always tell me is, you know, Bernie, the residual value is great. But we don't want to even get there. Like the idea of us having to repossess, recover the equipment, sell it, that's a nightmare. I don't even want to think about that.

33:49 Can we just somehow insure the offtake will always happen? Because as long as the offtake happens, the project is fine. So that that that's that's been a thing that you know people brought up to me many times where they don't want to to be honest, they they don't want to call it asset back financing. They just wanna give a loan and not even worry about the asset. They'll take it as collateral, but really the offtake contract is the main Yeah, that's a value is so, yeah.

34:15 So Yeah, so so that that that was like a driving, you know, discussion point I had in the past year and I was like, okay, is there a way where we could insure some of the offtake continuity? ~ and and Mm-hmm. and kinda somehow lock that in. Well, it's kind of scary to do that because then you run back to like, what if there's a bubble? What if there's default? So the way we structure it is we we can give like this ninety day window of coverage. It's not infinite, it's not for three years, it's for ninety days.

34:46 In the 90 days, we will actively work together on getting this stuff released, find a new offtake. So if the off take leaves because of a default, if the offtake leaves because they're unhappy with the SLAs, they don't think you're matching the SLAs, if they just don't renew because they they found somewhere else cheaper or they're just running out of money, you know, whatever reason, we can cover that interruption and build a product that kind of protects the lender's debt service.

35:13 And yeah, the the e the essence of that product is for a lender you just wanna be sh safe on every deal. But for us, we actually think you can release you can find new off take. We don't think it's too impossible. We actually think it's very possible that hey, you know, if this off taker walks, you'll probably find a new off taker and we can help you find that new off taker. It's like it's not impossible.

35:36 So because of that, we we th we think it's a safer product to do. And it it really just trades off that uncertainty for the lender and gives them peace of mind. And for us, we don't think the risk is too exaggerated. There is some risk there, but we don't think it's too exaggerated. We think we can mitigate it. We think we can release in those scenarios. That's kind of a quite this bit this is actually very innovative, very, you know, like 90 days of window where you can allow people to find even you help them together to find a new off taker.

36:10 so that's that's that's very interesting. ~ and do you have a requirement on, let's say, the contract that you can do this on, whether it's one year, two years, three years contract, or even on demand, you can do this on demand, definitely not like one or two years ~ duration like that. ~ I I would probably have to to understand more about the transaction. But yeah, we as long as I would say w I'll take nine months is the minimum.

36:37 I I probably under nine months I'd already feel like ~ it's it's unnecessary risk. ~ like you you you have to just be okay with the idea of releasing. You're not really if you if you're underwriting a nine month contract then you you know that the offtake will end. So I'd rather do stuff that's a bit longer than that. But yeah, we we we we'll we'll look at it. and you know, at the end of the day w th this policy by the way is like a one year policy. ~ it'll be like a one year bond and we just won't renew for year two if we think this the deal is too messy.

37:11 Right. So like there is like an incentive on your end to keep the deal healthy, keep the tr transaction healthy, because we won't renew the bond for s the year two if if th there's too many things scaring us. ~ so there is Makes sense. some like effort on both ends to make sure things are smooth, but even if there are hiccups, at least for the lender, debt is on time every month. ~ really this product w one way to also think about this product is in a way it's like a artificial debt service reserve account.

37:43 It's like an interest reserve account, Mm-hmm. Mm-hmm. right? And instead of pulling from the reserve account only, you can now also pull on this bond. Or you can stack it with the reserve account. So if you have a three-month reserve account, you now also have 90 days from this bond, that's six months. Six months to figure it out when s when stuff hits the fan, ~ when something goes wrong, you have you have six months to figure it out.

38:00 wow. Just enough time. Yeah. Yeah. So can y if you could, what is it? Rough approximate pricing for that, for for something like Mm-hmm. this? Yeah, so what we're looking at, it depends on the and so we we're very aligned with the lender. And so we would price this around two hundred to four hundred bips on the loan amount. Well it f well then it depends on the LTV and what's the loan amount versus the revenue ~ contract with the off taker, blah blah blah.

38:30 So it it there's a lot of variables there because it depends on the off take contract, depends on your loan amount, th depends on like the LTV. But I I would say roughly ~ two hundred to four hundred bips the loan amount and we collect that payment monthly alongside the that service, alongside the loan or lease repayments. So it's aligned with your interests as ~ if you're a lender, we will collect our premium alongside you. we're not gonna get ahead of you, we will collect it alongside of you.

38:57 If we miss our premium, it means you also miss. your loan payments, but vice versa. Yeah. Yeah. And so with again alignment of interest, like if we're not getting paid, we're gonna we're gonna go find off take as soon as possible. That's that's very very great. We're on the same boat, you know? And who who are the who are the usual parties for these products? Like of course, lenders at a hundred percent going after insurance. other than lenders, do you see other more financial players in the market trying to, you know, get in these products or be the you know either the the buy side of the products ~ on on both ends?

39:33 Yeah, so what I would say is like the the larger transactions, they don't need credit backstops as much because they're ideally Mm-hmm. underwriting better credit. ~ so when you're working with really, really large enterprises, you could kind of have some confidence on their credit profiles. But if you're underwriting neo clouds that are two years old, that's when you you look at these products. ~ in terms of who we we spend a lot of time with, we spend a lot of time with private credit lenders.

40:03 ~ we spend a lot of time with ~ equipment finance companies, ~ or lessors, depends on how you call them. we spend a lot of time with ~ you know the neo clouds themselves, and a lot of neo clouds ask us for help with financing. So we'll come in, we'll take a look at their transaction and say, hey, ~ we can introduce you to A, B, C, D lenders, but you need to go fix this aspect of the project or we won't make the introduction. So so we spend a lot of that time with Neo Clouds too.

40:31 Occasionally, I've had a few different Neo Clouds look at the products for their equity investors as well. So ~ among the products, ~ residual value insurance is actually less ~ Highlighted for lenders, and it's actually been more popular with lett leasing companies, less source, equipment finance companies, and equity investors. Equity investors, high net worth individuals who are investing into these neo-cloud projects, also like the idea of residual value because it kind of locks in their future return profile on the project. ~ lenders are much more interested in the off-take interruption. Yeah.

41:09 Right. Got it. Got it. And then we we have all already seen the NVIDIA trying to give a backstop ~ to some of these new clouds or some of these assets there. And you was you just mentioned you're focusing a little bit more on the new type of new cloud, where you know they need to have that backstop or credit synthetic backstop in order to fulfill or enhance the credibility ~ alongside of the off taker. Are you like the ev the it the do you ever come across a question where people would say, you know, you're basically going to insure the asset that's less people wanted or less people, you know, more riskier in in nature?

41:49 And would that be a ~ you know a selection a selection bias or selection skill for you where you just subject yourself to these ~ risk a little bit more risky asset compared compared to other deals? Yeah, so I think that's the that's the fun of insurance. It's balancing, Yeah. you know, like what you you what we you know that's called adverse selection. It's balancing adverse selection and making money. 'Cause if if I went and chased all the big deals only Chances are they don't need insurance.

42:20 They don't want insurance. Or if they do want insurance, they want it at such high risk, like they want me to insure like it's 80% residual. I'm not gonna do that. So it's it's it's like for those big deals, they don't want insurance, they don't need insurance, they they're they don't perceive as much risk in the projects. It's in these smaller projects where there's more perceived risk, but arguably, I believe sometimes the risk is overblown. So they Mm-hmm.

42:46 might perceive that this project has a lot of risk. ~ and maybe three of them do, but one of them doesn't. And I'll go after the one that doesn't, because I believe that that one's better. So it's it's a ~ in the in the insurance world they call it gatekeeping. So it's like how do you not insure everybody? How do you say no to people? Which is actually very hard. Like it it feels bad for me when I have customers approach me and they ask me for the coverage and I say, No, I'm not gonna insure you. I Just too kind.

43:13 don't want your money. It feels bad. Like I don't want help you, I don't want your money, I'm not gonna insure you. Like it just feels bad to say that. But that's the truth of the business. You have to carefully pick risk ~ and choose which risks you want in your portfolio in your book. ~ so ~ that that's a delicate process and and you know the best underwriters the best insurers ~ are able to do that and the ones that blow up don't do that.

43:40 So like that's what we have to figure out and that's our challenge. That's what makes it make a difference. And who who actually who actually ~ reinsure you if you are the one insuring these new cloud companies? Because you know in the insurance world we have, yeah. Yeah, so so for these deals, for these transactions, typically I actually don't inju ~ issue the policy. Typically in these structures, I work with a you know regulated insurance carrier or insurance company and they issue the policy. They issue the policy and so you would receive a policy for a blank blank company and that company would be regulated in in that jurisdiction, ~ or if they're not regulated in that jurisdiction, they're they're allowed to do business there.

44:24 And they will usually have ~ a credit rating from AM best. So AM best is the main credit rating agency for ~ insurance companies. So it's equivalent to like SNP for insurance. So that the these guys carry the credit ratings. If I issued the the policy, which I theoretically could if I really wanted to, I could do that. It's not impossible to like I we have the infrastructure to do that, the the licenses, etc to figure that out.

44:49 it wouldn't be as valuable. Because if you go to your credit committee, if the lenders you know bring this to their credit committees, the credit committee is like, this policy issued by this young man named Bernie seems so reliable versus policy issued by this large legitimate insurance company with an AM best rating that's been around for 30 years. It's a difference, right? So typically the policy will be issued by a legitimate, you know, insurance company that we work with.

45:13 Behind that insurance company we may line up reinsurers. And that's that's our business, that's our problem to solve. Because sometimes these insurance companies aren't as comfortable with the risk. So to get them comfortable they have to syndicate some of the risk out to reinsurers behind them. So we'll line up reinsurance, ~ and work with these like regulated, offshore entities that have these balance sheets reserved for reinsurance and they will line up behind this ~ insurance company.

45:41 So that that that whole business is kind of like how the specialty insurance market works, how like Lloyd's and these different ~ yeah reinsurance markets work where usually you interact with one company but behind them they syndicate the risk out and and then And figure out how to split it. And so that's that ends up being my problem. Especially on the larger transactions. my God, it's a nightmare. Like ~ you know, I'm working on this ~ $500 million transaction.

46:08 It's huge, great, but it it also is a nightmare because I have to like juggle all these different ~ insurance counterparties to syndicate the risk. but yeah, that's that's a little bit about how we do it. Well, that's the standard's a a big deal coming. But are you are you think of moving like vertically to become like a policy insurer, ~ issue in issue policies yourself one day to have to do that like long long term history? 'Cause you know, nobody in the market, I think, ~ or in in the supply chain has better knowledge than you or yourself versus the other like, you know, your policy issuer there, right? So do you wanna be the one ultimately issuing these policies?

46:49 Yeah, so i you know, today the way we're set up We're not like just a simple broker. We we actually have ~ a lot of influence over the design of the program. And the the Mm-hmm. in the industry it's you're known as a program administrator. That's what they call it. They have a lot of terms. So as a program administrator, I have influence over the policy form, the language in the policy. I have influence over pricing.

47:15 I have influence over underwriting. So I actually get a lot of the experience that you just described. I I that track record, I'm already building it. So then that's my track record as a Is it my balance sheet? No. But it's my track record. do we want to get to a point where we actually start putting our balance sheet on deals? I don't know. Because that's really expensive. And you know, as a sm younger Yeah. company, that you know, would put up a lot of capital doing that.

47:42 But it's something that's on the radar for us. And one way we might do it is when we're RDing like these new products, ~ instead of raising reinsurance for a new product that's not not tested yet, we might just ~ park money and It's called ~ posting collateral. We might post collateral into like a trust account. So you so if we're insuring a million dollars of risk, we'd have to park a million dollars ~ just to test new products without ~ getting outside capacity. We just do it with our own balance sheet.

48:11 That that way we can test products quickly. so we might consider doing that. ~ but in terms of like going chasing big deals and then putting like a hundred million dollars as collateral in an account, it's not the best usage of company funds, in my opinion. And that's the job of reinsurance. Like that's why reinsurance exists. Exactly. Yeah. So you can borrow their balance sheets. So ~ that at least that's how we think about it right now.

48:34 Exactly. I totally agree. I mean, even similar to us when we do the financing, we do work with those big financiers out there where ~ you know, they have a great balance sheet. We have the knowledge and know-how and also the expertise, and we just help underwrite, help them do the deals there. So leveraging, you know, the resources, that's what we call. ~ Mm-hmm. final question before we leave, is AI in a bubble? There's nobody that I think, you know, that's the best question to you from your angle there.

49:02 ~ this this is spicy. so I Ha ha. think I think I so I'm I'm I'm AI pilled. I'll start there. I'll say I'm AI pilled. I use Claude Code every day. I use AI tools every day. I think AI is super beneficial. A little scary, but very good. ~ and very very effective for work. Now Is the the market sustainable? Are the the is the excitement and the amount of money flowing around sustainable?

49:33 I don't know. And I I I I think not. I think we're gonna see a correction. I don't know when the correction will be. I don't think it's gonna happen in the next year, but I think it might happen in the next five years, where we'll see a huge correction and it'll be depressing. ~ it will be bad. and The reason for that, the reason I think it's it we're we're gonna see that, is because I'm seeing a lot of unqualified operators and just counterparties getting access to financing when they really should not.

50:08 I'll be honest. Like I I meet some people, I'm like, Wow. This guy probably should not be running a, you know, and and raising thirty million dollars to buy GPUs. ~ that doesn't really make sense. Or like you see a lot of these AI startups where like, wow, they don't have a business model. Like their business model is spending VC dollars as quickly as possible. So there's some stuff that's just unsustainable. so I think there will be a correction.

50:36 But it doesn't mean I don't believe in AI and I don't believe these GPUs ~ are worthless. I think these GPUs hold a lot of value and I think AI is very beneficial. But I do think there is a little bit too much excitement around, and I I hope people underwrite conservatively. I think there's some some overly aggressive underwriting right now where you're just it's just bad credit. so I that that that makes me a little concerned.

50:59 and I just hope these startups you know, 'cause it it frustrates me 'cause like I meet a lot of startups and they'll tell me, like, Bernie, ~ why would we do an off day contract when we can just rent it on demand? Because we're wake so much more money on demand. Yeah yeah. Yeah. Increasingly more. Yeah, increasingly more. And it's like, yeah, you will But there's a lot more risk too. And so I I hope the startups who think that way, there's not too many of them. There's there can be some of them.

51:29 That the probably like the the as a economy we can sustain some of that. But if there's too many people who think on demand and not on ~ off take contracts, it will we will be more fragile. Like the economy b will be more fragile. So that's that's my view of it at least. Right, makes sense. Makes sense. Well, as you said, you know, we can be the gatekeeper of the industry and we should be, ~ because we are like, you know, helping mitigating and and doing the financing here, ~ which is the water gate for for everything ~ right now, the constraint of the market. So thanks, Bernie. I think that's all we covered today.

52:06 Very, very thoughtful and insightful, ins ~ insightful serving on how instant products being made, what the market is going to. If people want to learn more about American compute Where should they look for it? Yeah, I mean so Please please go to our website. That's amcompute.com. ~ You know, we have lots of resources there. So like a lot of the stuff I mentioned, the data, methodology, we don't publish all of it. Well, we publish a good amount.

52:30 And so if you go to our website, you can read reports, research. Hopefully they're useful. ~ And you're you're very welcome to book a call with me or my team. And we're we're happy to share a little bit about what we're seeing. You know, we can do everything from just basic appraisals of equipment, you know, technical due diligence. That's actually surprisingly popular, but a lot of people just don't want to pay for insurance and a multi-million dollar policy, but they want maybe a report just to understand some of the risks in the deal.

52:56 So we're happy to provide those. ~ And yeah, and we can do these credit backstops and these more complicated insurance structures if necessary. ~ Regardless, we just want to be helpful, so feel free to reach out. Right. Thanks, Bernie. Anyone that wants protection, go find Bernie there. All right. All right. Let's conclude the day. Thank you. Thanks, Kony Right. Bye.