GPU-Backed Debt and the Case for the End of SaaS Credit

Transcript

Gabe (00:01) So we've seen the tokenization and really the financialization of markets like real estate and these entities like Fannie Mae that have derived from these financialized markets. However, where is the case of this with AI and what does this look like and what is the debt market and the financialization of debt in AI and compute look like? Well, I'm here today with David Choi, the co-founder and CEO of USDAI, which is working on tokenized debt in relation to compute and data centers. So David, can you take us through, you first started working on financing and underwriting stuff, artworks like Warhols and these other major art and now moving into the AI and compute space. Can you take us through that journey a little bit and what you're seeing in terms of the compute markets as of late?

David Choi (01:01) For sure. I guess in terms of my personal background, kind of like coming out of college, my first job was in art financing, ⁓ where we were looking at ways to find indirect ways of valuations that weren't just based on very limited data sets, where you would have like maybe three or four comparable trades in the last year, which is not enough to be like what they call a mark to market, which is when that's a previous valuation. And they didn't have cash flows. ⁓ with these assets. So we tried using these Monte Carlo simulations on the sale leaseback of the artwork where you have the right to buy back your own artwork as a way to estimate the pricing. And so you had to assume your own volatility index, which we had to create ourselves. But it was just kind of teaching me in that early educational process of my career where you can't rely on just simple data points. You have to use time to your advantage, pricing. and just other data sets that you can infer. And then afterwards, I worked at Deutsche, my first job. Again, in terms of esoteric assets, my first deal was actually doing a securitization deal on timeshares, which is single mortgage divided up. Two weeks is the timeshare, so 26 times multiplied by 300 rooms in a hotel. And then that's where the real money came from in timeshare, not from the sales themselves. So it's just like interesting to see how prevalent this was when it comes to asset-backed financing. Just like most things in the world just don't have a price, but you can definitely discover one based on a market price. And yeah, that's kind of like, you know, worked in crypto ever since for the last couple of years.

Gabe (02:47) guys.

David Choi (02:55) ⁓ doing a little bit of MEV, doing a lot of investments. And really seeing how DeFi has developed mostly as what I call a money markets, ⁓ which might be a different definition to what most people view it. But I always viewed that most of DeFi was money lending against money. ⁓ And I feel like it still has the incredible tools to create financial systems for things that aren't money, which I would redefine as the foil of ⁓ being capital markets. where you have things you're giving loans against capital, things that are productive. And I think people also divide it as like, there's loans and then there's margin. Most of DeFi is just margin, so it's not actually lending. So I think we're kind of seeing the evolution and this is why you see rates so low in DeFi, because you're just providing margin. You're not actually providing productive, scalable, expansionary capital that is what credit really is for, to build the economy out.

Gabe (03:52) So you have this deep background in underwriting essentially and then kind of moved up the stack to MEV and then to DeFi and now in this specific field of the compute. ⁓ You mentioned a little bit about art not having a lot of these data points. Can you one, of take us through the landscape of the debt instrument in this space for compute? And then also, is there enough data points for ⁓ this industry in the first place?

David Choi (04:25) Yeah, there's some asset like GPU backed loans that people try to experiment with, but they're just really underdeveloped. Go on Bloomberg, you click a single CoreWeave bond where they say it's the ABS, but for GPUs. ⁓ Well, first off, it's against a single counterparty. Not that, you're underwriting CoreWeave, you're not actually underwriting. ⁓ And then CoreWeave is like a counterparty, but you're not actually underwriting a basket because that's how get to ⁓ a number, right? You have 10,000 different loans and then that's when you can get to enough. Like when I give you a loan personally, I'm underwriting your ability and my trust in you versus the loan to 10,000 different unsecured individuals peer-to-peer loans. That's when you get to a default rate and you get to a number. And that's how you use law of large numbers in your advantage. So that doesn't really exist for GPUs because that requires 10,000 different loans being put together. That's when you can actually start having data points and numbers and start creating metrics that are derivative from the underlying asset. ⁓ That doesn't really exist in this space because ⁓ it's just really difficult to structure. ⁓ This exists for airplanes. This exists for equipment like tractors. This exists for automobiles and houses ⁓ where you have a very deep secondary market because you have the data points. have a lot of numbers. You have lot of ⁓ softer data points. ⁓ It just doesn't exist for GPUs. The way that GPUs have typically initially been financed is very similar to Bitcoin miners, which was convertible debt, where the company would take out that very, very diluted loans, which would get converted out of the money strike at multiples of valuations, which always gets struck because it's fast growing companies. And yeah, if you were a seed investor in some of these Neo clouds, you're already diluted 95 % because It's just a capital intensive business and you're selling equity for depreciating assets just to keep up with the times. It puts a lot of pressure on your equity. ⁓ And yeah, the financial tools available for ⁓ these assets have the industries is very limited. And you can't really do equipment finance without that. It's just not really a thing. It just doesn't make any sense in terms of, weighted average cost capital as a financial concept. You're really, really pushing the cost of equity and not really pushing cost of debt. So the tool sets that a lot of these companies have is limited. Like we were talking with this data center who was talking to a middle market investment bank and he got a four page presentation on why you can't get financing for its 200 mil GPU cluster. ⁓ Because he's like, yeah, you just can't get a fixed APR loan for this. It's not availability in the market. And the reason being is that. ⁓ to your point earlier about Fannie, ⁓ when you get a loan from a bank, the bank doesn't hold the loan. They resell it. ⁓ Otherwise they wouldn't have enough money. like, you know, kind of like fractional reserve banking. they don't be actually holds to debt because if you can't resell it, you're holding the loan. ⁓ This is how was ⁓ in the 1920s, 1910s, like before Fannie, like as a comparison of what the world existed without it.

Gabe (07:38) Mm-hmm.

David Choi (07:51) You used to get a five-year balloon loan against your house, 50 % LTV, ⁓ and every five years they would just roll it until the Great Depression happened and you would have to give up your house because they wouldn't want to roll it because they were holding it on the books, ⁓ short-dated loans. ⁓ So it was only with the creation of, I guess, the ability to package up the debt and make it tradable ⁓ where you have a forward flow to sell into.

Gabe (08:09) and Okay.

David Choi (08:21) That's when it became a deep capital market. That's why you can get 10 % down low interest rate loans in your house because of that credit extension. But that doesn't exist for GPUs because the time it takes to structure that debt is so long. Giving 10,000 loans and then making it tradable is an onerous process by banks, which takes years to do sometimes. And if you're talking about year-long structuring cycles for an asset that depreciates in the same period, because they're faster depreciating assets, it's not possible, structurally not feasible, ⁓ to create these instruments for these asset classes. ⁓ So the market's in this weird period where it needs this instrument. But the legacy traditional finance just doesn't move in the same speed. ⁓

Gabe (09:00) Mm-hmm.

David Choi (09:21) And that's why it's never been a product. And it's why it was never made for Bitcoin miners. Bitcoin miners don't have the same product either. It was never made for GPUs because GPUs, most of the GPU miners, or sorry, Neo clouds ⁓ were Bitcoin miners and they use the same methods, which is converted to that. ⁓ But not everyone wants to debut their investors by 95%, including their own ⁓ equity. So yeah, it's a limited tool set. This is why we push into private credit and just to like,

Gabe (09:38) Mm.

David Choi (09:51) I'm sure the audience knows, they see Blue Owl ⁓ and then they also see Oracle borrowing all this money from another legacy tech business where they're doing term loans because they're all trying to compensate for this lack of this core financing mechanism, asset-by-financing, ⁓ because it doesn't exist. They go into borrowing against Oracle. They go into Blue Owl, ⁓ which was definitely a big private credit firm, but can they finance trillions of dollars? We don't know. mean, how fast can they raise money from their LPs, right? ⁓ That's their replacement cycle. ⁓ But if it's tradable, ⁓ you can create the trillion dollar markets. But instead, we're just pushing it further and further out. ⁓ And this is why ⁓ you're seeing a big issue in the financing markets ⁓ in the AI space. ⁓ And installation is getting delayed. ⁓ Projects aren't getting financing. ⁓ And now you're seeing the scarcity increasing with chips because demand's increasing, but supply isn't.

Gabe (10:58) So basically there's this void, Where, like a gap. where the banks are not servicing and just the system is essentially inefficient. The financial system is inefficient in servicing this certain sector. We can talk about it in a second, but there's like the hyperscalers and then the Neo clouds and you guys are servicing these Neo clouds, which are the long tail. But if you can kind of ⁓ smooth over this bank issue, what exactly is the issue? Is it the risk? Is it the just the timeline to underwrite these these ⁓ assets and instruments? Is it an education issue? ⁓ Kind of. Yeah. And you kind of double click into like, what is the core the the core disconnect here?

David Choi (11:47) What blockchains do really well is that it's incredibly asynchronous. Like it's just a continuous, it's a perpetual, right? That's why we have perpetuals for ⁓ like hyperliquid, because it is perpetual. So you can just roll over the contracts again and again and to add infinitum. ⁓ And it's also great for stable coins because it just goes on forever. And I just want to preface it there because I think it's a digestible metaphor. ⁓ What we're creating is kind of like a depth perpetual. It's something where you don't have to be, ⁓ it's not something we issue every year. Like, hey, we're issuing this ABS that has this maturity date. We're issuing a single debt instrument, which is SUSDAI. ⁓ And whenever a loan gets created in our product, it gets added in to that debt instrument into Infinity. So you can always enter a loan and then you can obviously repay it at different times in the next guy. ⁓ This is how TradFi works. TradFi does everything synchronously. ⁓ If you, like I said, 10,000 loans, you have this like tenor of maturity date and that's why you have so many. ⁓ If you go into the book, you have so many of these bonds. ⁓ It's because they all have a maturity date. ⁓ And that just doesn't work for ⁓ like 24 months refresh cycles on different chips where you have your, if people are saying like, it's going to take two years. I'm like, you're already like two generations out. Like you're. You're not at the hoppers or the block wells. You're not the Rubens and maybe the Feynman's. ⁓ So it's a timeline of securitization. It just doesn't fit into the same speeds and cycles as AI or tech or hardware refresh cycles. ⁓ It's just too slow. no matter how, even if you get to the most efficient speeds, which is probably when you have a lot of data, a ton of data, like real states amount of data, which is which take like 100 years to develop, ⁓ then maybe it might work, but not right now. Right now, it's a brand new asset class. ⁓ And there's definitely data. It's not necessarily the extent that you'll see unless every single American owns a GPU, which might be possible in the future. ⁓ It's not that infeasible to envision. ⁓ But ⁓ enterprise GPU, not personal PC one. ⁓ But ⁓ that's when... you might be able to create that structure. ⁓ for now, you need to create something that is a lot more accommodating to the speed at which you can structure the debt, which is what blockchains are perfect for, really. And that's why, know, companies like Figure kind of came to the blockchain to solve a very, very similar issue ⁓ for Helocs, where they can speed up the process rapidly by having asynchronous systems. ⁓ And that's why DeFi is great. So for us, ⁓ the moment of loans originates immediately priced and added to the SUSDA pool, ⁓ no matter what maturity that they enter or exit from, because we're just seeing their monthly repayments at different times. ⁓ And that's why it's pretty nice. It's kind of like, if I explain to more Tri-Fi people, it's kind of like, I would say, infinite, ⁓ you don't require a warehouse facility ⁓ for more Tri-Fi folks. ⁓ It's kind of like represented in the tokenized form itself, which is SUSDEI.

Gabe (15:14) Okay, so this perpetual approach suits the maturity of the asset class, essentially. yeah. And so how do you prevent against the risk of like, quote unquote, kicking the can on these assets?

David Choi (15:22) the short-daiting this event, yeah. Yeah, so we. in terms of the underwriting work or in terms of ⁓ the repayment schedule, I guess.

Gabe (15:46) Yeah, I would say maybe the under, maybe both actually. Yeah, maybe if you can kind of talk, or how you were thinking about both of these without things going bad or stale, cetera.

David Choi (15:50) OK. ⁓ Yeah, we actually have a much easier job than many of the curators in this space for DeFi protocol. I guess I can describe it in two ways. One, to address those respectively, one would be permutations of underwriting is very, very limited for us. So in terms of how much work we have to do as a governance protocol, as an USDA protocol, and by the way, to clarify, the CEO of Permian Labs, the developer behind them, ⁓ not the protocol itself. But ⁓ the permutations of work is very limited. There's maybe eight chips, of which the eight chips is very clear parameterization that we can just follow. ⁓ And and then secondly, ⁓ it's extremely high cash flowing assets. One of the highest cash flowing assets in the world in of the scale. So that makes it really easy to generate liquidity ⁓ as a result of that. ⁓

Gabe (16:35) Mm-hmm.

David Choi (16:52) So in terms of the permutations that we look at, there's very few limited number of chips. So we don't underwrite companies ⁓ and we don't underwrite loans to small companies. ⁓ We're not a SMB like lending business. We're very much a GPU backed and GPU like asset backed and turning the ⁓ assets into bearer assets on chain and then giving loans against them. ⁓ Much like Aave doesn't care about your FICO score. ⁓ Justin Sun, if you wanted to, you could get a loan from Goldman. But he chooses Aave because he could just put up his collateral and get cash against it really fast. ⁓ In our case, we don't care if they're all birds or whoever they are, because I actually don't think there's ⁓ as much value in the enterprise value of the companies versus the value of the chips, ⁓ because that's what we're undermining. And there's only a few chips, like I said. There's like B200s, B300s, GB 200s, GB 300s, ⁓ maybe 5090s and H200s. And whenever a new one comes in, we just get rid of the other ones that are older, just because it's just a refresh cycles. ⁓ But there's not that much variables as a result, versus like, hey, I'm looking at this point with this LTV based on the volume metrics based on where it's listed. There's a lot more risk based on this define integration. And I'm like, And you have to do that across 30 different assets. It's tough. And it's a hard job. Versus eight assets, which is just a chip to sell. So in terms of like kicking the can, it's actually there's not much kicking to do because it's very, very easy to not easy. Sorry. It's very, very straightforward. What we're really looking at and just being specialized, special in these ships. And because we're so aggressive in terms of the debt service coverage ratio, which is like what we really measure, like what is the most they could pay because kind of like Fannie, you kind of want to be pessimistic, impatient and aggressive to have the deepest capital markets. The more that you are impatient, the deeper people will trade it because they know it's ⁓ programmatic. Fannie definitely wants to be a program if it wanted to be just like a straight up like machine. ⁓ As for like ⁓ liquidity, ⁓ you have so much cash flow January from this that the repayment of debt is incredibly fast. And I'll give you the real numbers of our protocol. We look at the debt service coverage ratio, is these are mostly homogeneous assets. So it's not like this one is that different from the other one. We just look at how much they pay every month, which ends up being 3 % of the principal and 1 % interest. you multiply that by 12 months a year, that's about 48 % and 50 % of the loan repaid. And that's kind of what we look at. And that's how you generate liquidity. And you're not actually pushing the next greater fold because you just know like, I can just redeem my money if I'm less than 8 % of the pool taking two months. And with a 20 % buffer plus a 50 % repayment, the 70 % of the loan book kind of being pushed back into cash in a single year. That's pretty good. It's actually very good. It might not be the 100 % capital efficient, but I think DeFi users prefer a little bit of a buffer in terms of their liquidity preferences. But at 70 % of the loan book be turned into cash in a year is pretty damn good, I think. And it meets the impatience of the money market participants in DeFi.

Gabe (20:16) Mm-hmm. So your answer to that question about how do the investors necessarily become comfortable, right? Is these things are cash generating machines and basically they can take on this and model out essentially what they should receive, et cetera.

David Choi (20:56) As long as we sequence it correctly and organize it economically and make it really, really rules-based, think people get pretty comfortable. I myself have lost so much money on the Grayscale premium trade ⁓ to the point where I got in podcasts because one of my threads went viral. ⁓ I know what it's like to have your money not be redeemable by primary redemption mechanisms and where people assume secondary and liquidity mechanisms. And this is also true for Goldfinch. This was true for like usual money. This was true for like ANZIN. I mean, I can go down the list of every single failed RWA experimentation and almost 100 % of the time, it's because you can't get any money back. I was like, how did you design this thinking like just putting an AMM pool on UNICEF is to solve it? I'm like, no, no, you need to make generate cash. I'm not going to, I can dabble in STRC discussions because there's some yield generated, but if you can't redeem it for the principal, there's always going to be a risk. ⁓ For us, we make it very clear how exactly you get your money back. And that's how we organize a lot of our tech around. We call it QED, which is inspired by Flashbots' MEV boost. Which is pretty much every block you generate some liquidity or a reward and then you bid for it. What we do is that every repayment, we turn into an auction. It's like a tender offer that happens every month for 4 % of the entire NAV. ⁓ Yeah, and then as long as you're programmatic around it, think liquidity organizes around it, and it becomes ⁓ anti-fragile for these kind of use cases.

Gabe (22:30) Yeah. Yeah, just to touch on that RWA piece, I think in a past interview you said some of these RWA firms are like, quote unquote, toxic waste to the ecosystem. If maybe if you can like dive in and if you feel inclined, share some names and maybe why you kind of feel that way about these these protocols.

David Choi (22:40) Yes. I've just been in this space for so long. I know what happens with adverse selection. People think that we're adverse selection and I'm like, well, we're not underwriting ⁓ these companies, right? They're not like companies asking us for money. We're literally taking their underlying collateral and giving a loan against it. So unless you think one of our eight assets are ⁓ adverse selected, you can make a claim, but they think we're giving unsecured loans, very different, right? Versus asset backbones. ⁓ I guess what I don't like is people taking Dino yields, things that are like ancient, like real estate loans and bringing it on chain. I'm like, good. So you're taking the most capital efficient systems in financial history. And you think it's going to be better to give mortgages on chain. I'm like, no, that's that you just adding more fees and making the R and R inefficient. Just like adding more fees like, and the interest is lower. but you're taking the same amount of risk. Doesn't make any sense. I think what is the And that's why I think it's toxic waste. think it's just, I mean, I'll say this a bit with like a lot of these like tokenized private credit stuff that was coming on. I was like, do you really want to buy this stuff? And they're like, no, it's a good brand name. I'm not going to say which ones, but I'm like, the brand name doesn't mean you're underlining the underlying assets. You're literally taking something that's illiquid and you're pressing it nav par. Anyway, it just... They don't actually originate on chain. You see A16Z complain about this too. They're like, guys, you're just tokenizing securities. You're just adding fees on top of fees on top of fees and then the same underlying asset. If you don't originate on chain, and that's what happens in our protocol, you're actually originating. The fees are generated on chain and the token holders own the fees. ⁓ And we own the vertical stack of, not we, sorry, the protocol owns a vertical stack of origination, structuring, and distribution, which is the three layers of the financial process. So yeah, think it's, I think it's toxic waste. And this might pivot into like the next part of what I think is happening into credit in and of itself is changing. And we're at this perfect juncture. I think SaaS credit is, is dead. I think SaaS credit for what it was in the Zerp era, where we were giving unsecured loans against what seemed to be great consistent cash flowing ⁓ corporate loans because they're just generating great subscriptions year over year. ⁓ That era is like shutting down. This is what private credit is going against such wide compression is because those companies are going into, their stocks are crashing, right? Because you're just one codex prompt away from replacing the entire business. ⁓ Thing is AI companies aren't that much better either, right?

Gabe (25:36) Mm-hmm.

David Choi (25:55) You can't really give a loan to a cash flow company because this is going to get computed away. But what you can give loans to is not against the company unsecured, ⁓ but rather against secured asset backed loans against what they do have that is valuable, that generates entire business proposition, which ends up being 90 % of their costs to compute. ⁓ So what you're seeing is the end of this chapter of SaaS ⁓ cash flow based under unsecured corporate loans. and as being completely replaced, not by new companies, but by assets and just by GPU loans. So GPU loans isn't a niche is what I'm saying. It might be the replacement of all of credit that we know today. It's just, it's a super sector like replacement. It's a superstructure replacement. And I think people kind of are missing that point. And kind of having a hard time understanding this, but I'm like, and

Gabe (26:40) Mm.

David Choi (26:55) I'll say this too. This is also why I think this is the second largest business in AI. ⁓ Number one business is Nvidia because they create the chips. ⁓ The third largest business is OpenAI because that's the LLM. ⁓ But they only make 20 to 40 billion a year. Nvidia makes 500 billion next five quarters. Wall Street makes 150 to 200 because you're paying interest on the chips. ⁓ Because all the chips are definitely so to me, it's the second largest business. And that's only going to get bigger and bigger. And that interest is very, very high for Wall Street. And I don't think that's toxic waste. think that's the this is the birth of a new sector. And this is what we should be tokenizing, not Dino yields that are fees on fees on fees of tokenized securities, but really originated on chain. Anyway, sorry to go on. Yeah.

Gabe (27:48) So your thesis is a lot of these RWA companies on chain, We're going after the companies which were bad assets and kind of like the new category opportunity is you guys can not go after the companies which are you maybe the data centers or the providers but more so the actual asset which are the the GPUs and you're seeing this as the second largest asset class. So you have the compute which is Nvidia. You have the the debt structure around the compute and then the models. Is that a good articulation of kind of your and how you're seeing the landscape play out.

David Choi (28:30) Yeah. And I feel like this is what DeFi has always been great at, which is bearer asset lending. And this is kind of what we're congruent to. Like it is like GPUs as the isolated collateral away from the companies themselves. And I think this is something that DeFi people are also pretty like comfortable with because it is a low trust environment. There is a cost of underwriting. And I think a lot of people kind of misunderstood that in the early days of RWAs. They tried to push it off by just saying, We're just going to do it off chain. ⁓ V1 of RWAs was just like, let's underwrite on chain, and we're just going to do it ourselves. This is ⁓ what you saw with the first wave of RWAs. ⁓ I hate naming them, because I do like a lot of the people. But they were trying to underwrite themselves on chain. Didn't work. It just takes too much work. There's misalignment. Now, I guess the current iteration is that, yeah, we'll just use like a bank, know, like a top tier guy. And I'm like, this still still doesn't address the issue ⁓ versus doing what we did well ⁓ in DeFi, which was, Hey, there's a specific set of assets we could underwrite based on it being asset backed and then underwriting the asset itself, which is a lot more scalable. It's kind of like these five boxes that you like update the parameters for every quarter. It's isn't that different from Fanny, right? Like you have like three buy boxes. If it fits into it, then it works. Otherwise, we're not going to accept it. And it has to fit into this, otherwise we're not going to take it. That's what a governance protocol does. And I think it did it very well. yeah, and I do think this is the second largest industry in AI specifically, not assets, but in AI in general, ⁓ is the financing of chips. ⁓ I don't think anybody's trying to disrupt it. And I know for sure they aren't because ⁓ we chat with like Nvidia and they're like, there's not many startups trying to disrupt financing. This is not something they think of a startup like a disrupt. They think of private credit funds or our banks trying to finance these large deals. But a startup is to do it. ⁓ It's a very different proposition. And to me, a stablecoin kind of is a bank, right? ⁓ It's just taking, there's a checking and savings account. In the case of us, it's a USCI and SUSCA. We're not a bank, but it kind of follows the same concepts. ⁓ And we're... the deals generated from the checkings can subsidize ⁓ the loans that we're giving on the savings account. ⁓ And it can be competitive with these counterparties. The only difference between us and many of the other counterparts that they're introducing their customers to is we're not trying to be vultures and trying to ink every covenant out of the counterparties. ⁓ We're just trying to scale the AOM and define this asset class, ⁓ which is very different. ⁓ Anyway, yeah, think it's also what DeFi could be great for in reestablishing itself as the center of this new wave. ⁓ I think people forget that NVIDIA stock went from like. $33 to like $10 over the course of a few months because ETH went to ETH staking. ⁓ When Nvidia was 100 % beta to crypto, it's time to bring it back, know? Like ⁓ crypto used to be like the center stage of Nvidia's fate. ⁓ And now I think DeFi is in a great position to be center stage of AI ⁓ and we're trying to be that proxy.

Gabe (32:00) in So you mentioned in video a bit, and I know you guys work directly with them for ⁓ basically onboarding clients and they potentially would send you leads and things like that. What are they saying internally about, you know, this long tail, the Neo clouds ⁓ debt, you know, sector of the market? There's the hyperscalers and then there's this longer tail where like on the hyperscaler side, there's you know, there's billions and billions of dollars being moved around and you're seeing these huge But yeah, curious, what is NVIDIA saying and how are they looking at this? Do they have any ideas? Do they kind of align with your thesis and your framing? Are they seeing things a little bit different? Yeah, curious to kind of hear like the inside take on that.

David Choi (32:47) Yeah, I'll say this too. It's not like an exclusive thing that they're working with us on. It's just that Nvidia works with a lot of partners, a lot of funds and a lot of financing counterparties. So I don't want to imply anything exclusive there. They work with their ecosystem. You can go back to their numbers. Obviously at the $500 billion, like metric I said before, about 20%, 30 % of that is for emerging neoclouds.

Gabe (32:57) Yeah.

David Choi (33:14) know, sovereigns that are just starting to build up their supply all the way to like ⁓ non-U.S., European or Asian counterparties that are going to be the next coal weaves. ⁓ But they just don't have access to the American credit markets, which is by far the best in the world. If you want a loan in the U.S., it's way easier than getting a loan in Europe or getting a loan from a non-Chinese bank in Asia. ⁓ It's just like, credit is really easy in the U.S. But it still hard for a of these counterparties to get a loan. So they know it's a big issue. There's very, very little competition for the USEF protocol because credit just doesn't exist in a lot of places in the world as easy as this, especially for a growth sector that seems like it's high risk in different parts of the world. yeah, mean, 20 to 30 % of the revenues may soon be completely stopped because of this massive private credit retreat. People were depending on private credit for smaller deals. ⁓ But as you see, most of the private credit in the world is in full retreat right now. ⁓ This is a great place for DeFi to step in and tell that void. it's definitely, ⁓ if I was looking at what's stopping the deals that's happening right now, and I guess I'm going to kind of segue into that as like, what is ⁓ the difference in terms of bottlenecks between the different between the revenue stack, ⁓ the 80 % to the 70 % of the hyperscalers, ⁓ they're doing gigawatt build-outs, right? ⁓ They're massive. And that's what powers the limiting factors for them. Funny thing is really, they can always get more money because they're massive ⁓ American pillar like corporations like Amazon. But finding gigawatt is near impossible. That's like amount of energy that Denver uses, right? Like that's how big these properties are. And they're like, They are hyper-object scale build outs. If you were to build like 15 Las Vegas spheres, that's about the same amount of cost. It's like $50 billion for a gigawatt. It's fucking massive. ⁓ Whereas if you want to find like two or three megawatts, maybe 10, you can find that in the US or in Europe or in Asia, sometimes in Africa or Latam, definitely at Latam and near like a lot of the major dams. But you can't get financing. They have the opposite issue. There are sites all over the world that they can get. But who's going to finance a sub 300 mil GP financing deal where there's no equity upside and it's just a yield deal, an infor deal, ⁓ and they just have to organize everything together? ⁓ And they're like emerging operators. They'll have the chips eventually in the off take, but it's hard to get the off. It's like a big chicken and egg problem because they didn't, they might have the brand names. ⁓

Gabe (36:07) Yeah.

David Choi (36:10) But they need to finance it. That's like the thing that just stops a lot of deals. So that's the segmentation of bottlenecks. If you're big, you have a power issue. If you're emerging, you have a financing issue. ⁓ And yeah, that's how we've seen the market develop at least.

Gabe (36:30) and you're gonna be serving and serving currently this emerging market and the financing issue and you're seeing that this is gonna be the second largest. So on those points too, when going back to the asset, why specifically the Nvidia chips? And if I'm not mistaken, you guys are solely focused on those. Why are you not using any other providers in their grave?

David Choi (36:56) well, they're 90 % of the market. So that's kind of like the big reason. It's like, we do have like the most data from them because they are the most traded. We work with a lot of the ITADs, which is something called the IT asset disposition. The only thing that we've kind of like real, ⁓ created a big network for, ⁓ they're the guys who originally were doing server asset dispositions and with very little residual value. Chips have like extremely high residual value. ⁓ but, ⁓ they, ⁓ they, we, talk with them all the time. ⁓ And they don't really see ⁓ these other chips being traded. They're not really used. They're not really put up into, a rental marketplace. then people rent them out. They call it spec, or just per hour rentals, by weeks instead of by years. That's not really developed. So they don't really have a base price to defend the value of an AMD chip versus Nvidia. That's actually just pure. pure gold, right? It's just cash. can resell that instantly because you know that you can actually make money on it the next day. So not enough data, no off-take market, ITADs aren't trading it, less than 10 % of the market, and Nvidia chips are just big enough. Eventually one day, but yeah, I have not seen Google TPUs trade in the market once.

Gabe (38:20) And who are you seeing as the runner-ups? Because I know at one point, NVIDIA is working with Grok or the equity loan system with Grok. yeah, who are the emerging players or technology ⁓ that you think is going to maybe hold up a fist or be potentially like debt that you see or assets that you see that you could support?

David Choi (38:44) ⁓ I would say like AMD and Google are the ones that have been maybe brought up. Typically what they do is like they do these guarantees of minimum hourly rentals to get financings. That's what we've seen in the market at least. So like, hey, you get like this $2 an hour guaranteed if the contract fails. And that's a guarantee by like say AMD. So you're really just underwriting AMD. You're not even underwriting the value of the chip. I mean, we'll do that deal. If we get a guarantee from AMD, yeah, I'll then put it into the USDA protocol and get liquidity on it. I'm sure governance will approve. ⁓ But otherwise, yeah, it's still very, we're still very far. There's a lot of new companies kind of coming out trying to disrupt Nvidia. ⁓ But I think Nvidia would just buy them just like ROK, like you said. If a new competitor came out that's really better than Nvidia. They're gonna price it 95 cents to the dollar to Nvidia anyways. So it's still gonna be like pretty credit worthy because they're gonna be a max capitalist. But yeah, I think Nvidia is probably just gonna buy their way out. They have so much cash and I'm sure these ship companies wouldn't mind that either.

Gabe (40:04) Yeah, yeah. And where do you see is like the opportunity? So like there's the broader thesis where, you know, the credit markets, there's this opportunity to be restructured, especially right now when things are hurting. But where's like the next biggest opportunity that you're seeing that maybe people aren't accustomed to just because you're uniquely suited in like the financialization piece and on the GPU side?

David Choi (40:32) I think a lot of people talk about agent to commerce as if it's like a ball in a mitten. Like, yeah, you can pay your agent in this way. And that's agent to commerce. And I'm like, okay, that's sick. It's just like a plaid. Or they view it as like a credit card, or they think they'll make money out of interchange here and there. But what I think people often forget is that what is 90 % of all expenses in all of AI? It's compute. Pretty much, people are astonished by the salaries by OpenAI or ⁓ by these other companies. But I'm like, have you seen the compute though? It's so much bigger than the salaries or the talent acquisitions. ⁓ Even that OpenAI. investment by Microsoft. Most of that was in Azure credits. It wasn't in cash, the famous one years back. Just because it's so expensive to run all these models. This is why you're getting kneecapped on Cloud right now, because it's a compute issue. They have a lot of smart people and they hire the best, but it's nowhere close to the amount of compute that they need and are spending and are subsidizing. The next stable coin that's going to be created is going to be related to AI. I call the next great one, Tether. was originally obviously with the perps winning circle with DeFi being the great pair. The next one's gonna be related to AI without a doubt. And if it's related to AI, it's gonna be related to compute. Our thesis is, agenda commerce is thinking about it incorrectly. Like it has to connect to compute because that is where the money will be spent. Our theory is that if you can lower the number one cost that is controllable in compute being the debt. you may be able to create the next stablecoin, which is the best way to subsidize for why compute costs are so high, because the debt related to is expensive. ⁓ So if you can expand the credit correctly, I think there's another business model after you build the loan book, which is creating the settlement layer for all of rental contracts, all for the debt contracts, all for all of compute itself, ⁓ which I think a lot of people kind of don't see as a business model. is what we see as something we want to do in the future.

Gabe (42:58) Got it, got it. And are you guys seeing yourselves as a stable coin, especially like with the Genius Act and some of those other regulations? Like how are you navigating that for these instruments?

David Choi (43:13) This is why we work with PayPal and that's why our backing is on PYUSD for USDA, which is in no relation to any GP loans at all. ⁓ Or I mean, it's not exposed to the risk, but the yield generated from that is then ⁓ used to help subsidize ⁓ our counterparties in terms of their compute costs to some degree, if they are using that currency to settle in it. ⁓ PayPal I think is I mean, they're, they're FinTech, but they're also a great group of lawyers, probably number one in Silicon Valley. And they know what they're doing there. ⁓ But I, what I think with the Genius Act is interesting. ⁓ I think a lot of people focus on the Genius Act and clarity, but I think what they miss, at least for us, is the one big beautiful bill, which is like the introduction, introduction to bonus depreciations for our borrowers. That's our favorite bill out of all three of those. ⁓ which makes it very interesting to, I ⁓ mean, this could be the one of, this is in some ways ⁓ a tax efficient, a tax ⁓ optimizing ⁓ stable coin to some degree as well. ⁓ But I'll give you a use case of what we did recently ⁓ of like how we used USDAI, which is just one-to-one to PYUST by the way. ⁓ to lower the interest rate of one of the loans that we did. So one of the loans that we did, we required three months of debt service reserve account, which is just like, you if they miss a payment or two, can always draw from this account to cover up. It's kind of like, just like a rainy day fund. And we forced them to hold it in, we didn't force them. We like give them the option to hold it in

Gabe (45:00) Yeah.

David Choi (45:09) ⁓ USDAI, but, if they did, could lower the interest rate from 12 % to like 11. And they always say yes. and this is a company that has never even, ⁓ never even used a wall before, right? But they didn't have a treasury management system. They didn't, they didn't, they didn't have some guy to manage the treasury like second by second against, ⁓ all the cash they're holding. ⁓ so like, yeah, I mean, we'll just hold in USDAI and, ⁓ we would take that 11%. ⁓ So that's just like a clean use case of how we were able to use USDEI and SUSDEI to effectively lower the interest rate of compute. But yeah, that was just an interesting use case there.

Gabe (45:53) Interesting interesting. Yeah, recently we had ⁓ Chunda from from Paxos Labs actually and Paxos is labeling or is is supporting the the PY USDs and then you guys are working with PY USD so our PayPal USD rather and Like even just on our show specifically. We're seeing like this composability of Services come come together ⁓ Literally within a couple episodes. I'm curious for you. This insurance piece has come up is basically what you're what you're talking about now ⁓ Are you seeing people building this insurance product? Is this something that you guys think you need to take on? ⁓ Yeah, how are you looking at this this this insurance piece and almost covering, you know the loan paybacks and etc

David Choi (46:45) Yeah, the insurance fees was created. It was like in collaboration with Barker. It didn't exist in the world before we kind of did the first POCs and made it feasible and work with Barker, which did a back-to-back policy with Mean Agree just to get a place. But that was just one of the things we were going through with the obvious proposal. Initially, we were rejected because of some risks in the way we did the tranches risk with Filo, which is first loss. ⁓ We realized that they covered the entirety, not just 5%, but the entirety of the loan book. ⁓ So we worked with the Barker team to create something that ⁓ would meet the standards of ⁓ more institutional depositors. ⁓ And I guess what the common risk was, like, if you were a depositor and if you didn't understand how depreciation worked for chips, the obsolescence risk, ⁓ you're going to price it at a much higher BIPs than ⁓ what a professional would, right? Be like, it's probably like another 500 bips of risk per, ⁓ I think Tom Rees did analysis of what the real rate should be. And he just threw numbers because he just guessing. ⁓ But unless you actually price the risk, which is what we did with ⁓ Barker and Munich, ⁓ it's like we had to price the risk for them. It ended up being like around 75 to 150 bips. ⁓ Because given our aggressive ⁓ loan amortization schedules, they priced it pretty low ⁓ because they were looking at the data and the risks associated with them. ⁓ And now if I was doing my personal money, I probably wouldn't even take it because I know the amortization schedule that I'm doing is extremely aggressive. I'll definitely take catastrophic risk, but in terms of value insurance, yeah, like, I don't know if I'd take it, ⁓ I have like, should, USA protocol should because ⁓ selling a yield product, That's trying to show scalability. ⁓ So I think it was a great ⁓ innovation that we did in turning GPUs into a NASA class.

Gabe (48:50) Nice. And just before we wrap up here, ⁓ at one point you were talking about internet capital markets and internet money markets. And if I'm not mistaken, I think you're not a fan of internet capital markets specifically. Can you kind of maybe clarify that and give your take on these two different framings of ⁓ eventually this class?

David Choi (49:16) I am a fan of Internet Capital Markets. I'm not a fan of the people that ⁓ are defining it, are trying to put it for like, being coins. yeah, like things that are generated on the Internet. To me, like Internet Capital, like the word capital means that has to be productive in nature. Like things that are like actually created from, I guess, Internet and can be used and more effectively like ⁓ be productive, which for me is like cash generating. ⁓ And that's when you can actually have true ICM. ⁓ But right now, most of DeFi is ⁓ money market like. I mean, there's this entire era of like, say, Eigen layer restaking. then people are like, it's going to be productive. I'm like, and it wasn't. Nobody actually used restaking for anything. It's just money. It's just wrapped money against wrapped money, getting loans against money. ⁓ It's just unproductive assets. It was just margin. ⁓ So I am hopeful that it will come about. It's just like, ⁓ What are you actually giving loans to? ⁓ Is it making money? If it's making money, ⁓ I guess instead of calling it internet capital markets, they should just call it fixed income. ⁓ Like why does a business take fixed income? Well, it's because it's generating revenue and it doesn't want the volatility of floating rate. So wants to borrow at a fixed rate because it's a tractor and it has a farm and it makes this much revenue a year. And it's very confident that it can scale. ⁓ I think right now people are comfortable floating rate because nothing's productive. ⁓ It's not a capital market. It's the using floating rate because it's just arbitraging inefficiencies across market. So I'm hopeful for the future. ⁓ I guess I just didn't like how people are using it because they didn't understand the etymology of the words they were using.

Gabe (51:07) Yeah, and that kind of goes back to your just original thesis in general is, you know, these need to be productive assets and assets specifically that are generating revenue that can be underwritten. Do you think that only exists off chain or do you think that there's also opportunities potentially on chain?

David Choi (51:28) Yeah, I think there are opportunities on chain. It's just asset selection is pretty important. And I'm trying to think of what another good, interesting market would be. I'm just so biased because I like my project so much. So I always say GPUs. I think what you're with hyperliquid, like what you saw with pre-markets is like technically an asset class because it's creating something that is not really seen in traditional markets. If it's all something that existing financial systems can't create, that's just a great wedge into like building the other stuff. ⁓ And I think ⁓ that's when you can really build out the markets around it. But just by taking assets that have better counterparts in traditional markets ⁓ usually means that the product shouldn't exist. It's like, yeah, just a deeper market off chain. ⁓ Why bring it on chain for like 5 % efficiency? ⁓ So it's usually when you can really, really have a zero to one moment. ⁓ But I know it's kind of a cop-out answer. It's hard for me to create analysis outside of the product that I'm ⁓ Yeah, so OK. I guess I'll give that question a pass. I don't have a good answer for that. Yeah.

Gabe (52:55) Fair enough, fair enough. Well, yeah, David, thanks for coming on and appreciate talking about all this stuff. And yeah, hope to have you back on soon.

David Choi (53:04) Thanks, Gabe. Thanks for the fun questions.

Gabe (53:06) Yeah, likewise. See ya.