We hear about IPOs being oversubscribed all the time. What does that actually mean?
It means that the underwriters and bankers are doing their job in terms of marketing. It's complete nonsense. Oversubscription doesn't mean anything because it's a non-binding commitment from institutional investors saying, in a perfect world, here's how much stock I would buy. I know I won't get that, so I'll name a giant number and I know I'll just get a fraction of that. And so bankers then turn around like that and say, oh, look, there's so much demand for this IPO. It's 5 or 6 times oversubscribed. It's complete nonsense. If the peak of the Chinese bubble this summer, some of the IPOs were 8000 times oversubscribed and then fell by 50%. It means nothing. It's unbelievable to me that the media parrots it because you are literally carrying water for underwriters. You should get commission. I tweeted about that. I'm going to go ask for commission.
Hello and welcome to another episode of the Odd Lots Podcast. I'm Joe Weisenthal and I'm Tracy Alloway. Tracy, we're recording this, November 18th, 9:34 a.m. I would say arguably we're right smack dab in the middle of 2 or 3 of the most important days in the AI industry so far, for sure. So we've obviously had the Nvidia scandal that broke last week. God, it feels like a lifetime ago. Yeah, I know we had Nvidia earnings recently and on Thursday, tomorrow, hopefully when we release this episode, we'll be getting the Anthropic IPO. That's right. So as far as we know, Anthropic is going to begin trading tomorrow. We'll see if it happens. But yeah, lots of stuff between the Nvidia earnings, the Anthropic IPO. Prior to all of this, the market was weakening already. So even setting aside the stock, not to mention the midterms as well, and the midterms, the other tell was prior to all this, I mean, like what people call the AI trade peaked in June. We've mentioned CoreWeave credit spreads numerous times on the podcast, etc. So there's already fragility. There is one chart though that the bulls point to, Michael Dell being the latest, he tweeted this year, but they all do: GPU rental rates, which you would think is like a very clean measure of demand for AI. They continue to go up. And so you could point to all this market volatility and all this stuff and questions about ROI, etc. GPU rental rates still have been on the rise.
Yeah, that's right. And even after the midterms we saw a little bit of a spike. So the argument, the bull case argument is that demand for GPUs is still very strong.
Yeah. And if anything, if data center construction is going to slow down, then maybe existing compute goes up even more. However. However, a big question with that chart has to be like, what is it actually showing? And does it incorporate other things like the cost of actually securing GPUs versus just how much the GPUs are worth?
Totally. Because we haven't seen evidence in the productivity data that like companies by and large are meaningfully getting productivity gains out of AI. Right? And that's what you'd expect to see if there's going to be ongoing help for GPUs.
So you know it would be funny. Yeah, sorry. Just imagine if we got a big productivity dip in the fourth quarter because of the release of Grand Theft Auto 6, and that knocked over the AI trade. That would be, that would be ironic. I don't think we're going to see that, but it probably won't be bigger. But that's the other big thing that is happening today is the release of Grand Theft Auto. We're having a multi-part episode today to discuss this historic moment, but right now, first up, we really do have the perfect guest, someone who understands both the sort of GPU financing market, but also credit markets more generally. Unfortunately, I think there's a lot of people in tech who like, just like don't care about credit markets. They're all here to learn about credit for the first time.
Yeah, there's a big divide. So a lot of the people in Silicon Valley, they view AI as just this pure technological story. Whereas on Wall Street, and I would argue everywhere else, it's a financing story.
It's a little bit like the shale boom, isn't it? Yes. So this is my preferred analogy for what we're seeing right now is the 2015 sort of break in high yield. But we can get into that with our guest. All right. Well, I'm excited to say we have the perfect guest because we're going to be speaking with someone whose career has really straddled both the tech world and the credit world. We're going to be speaking with Stanley Lee, chief product officer at Liquid Compute. Stanley, thank you so much for coming in studio on Odd Lots. Why don't you tell us, what is Liquid Compute real quickly? And just like, how did you get here, your background where, as I said, you straddle both worlds?
Yeah, no, thanks for having me on. Definitely a topical time to be talking about all of this. So Liquid Compute, as a startup, we're focused on building a lot of financial infrastructure in the compute space. So a marketplace for people to buy and sell capacity. Obviously, this can be used to relight capacity if there eventually might be defaults. We maintain a rental index that tracks the prices of GPU rentals to work with a lot of the lenders that we work with and financing for potential hedges and ways to lay off residual value risk, which nobody really felt the need to a few months ago, and everybody is starting to feel the need to do that now. And then we work with financing, you know, for a lot of these folks that are buyers and sellers. I think one of the biggest things that hit me, and I'll get into my background in a second, was, you know, I didn't really see the GPU marketplace as just a marketplace of buyers and sellers. It truly is a credit and a financing ecosystem. If you look at a lot of buyers and sellers of compute at the below investment grade level, you are looking at the buyers being potential creditworthy off-takers for high yield and leveraged finance lenders to underwrite, and these operators are potential borrowers. They're either borrowers now for this cluster, or if they do a really good job, they'll be borrowers for their next cluster. So I think it's a financing ecosystem. My background is in credit. I spent five years at Barclays trading high yield bonds and CDS, and then a few years on the buy side, most recently at Point72 trading high yield and distressed credit, moved over earlier this year. But it's definitely, I would say, compute. As much as I love the high yield market, compute is definitely one of the most interesting credit markets that I've seen so far, for sure.
Do you think people in the AI industry, do they understand the financing aspect of all of this? Because it feels like 2026 is kind of the year where they're having to wake up to it. And again, we're talking a lot about the hyperscalers, like big cash-rich companies that have never had to borrow that much in the market before. And now we're seeing them take on loans and bonds, issue new equity and things like that. Are people waking up to the reality of the financing aspect of all of this?
Yeah, I think absolutely. I mean, at the top of the market, we as a sort of country and society have created this dual-track economy where AI capex sustained so much of GDP growth in this country. You know, you guys have talked about that in the past. And it's trickling down to the middle market. A lot of it has been funded by venture equity, and private credit has been financing a lot of these mid-market GPU loans. And I think a lot of people that are in tech and in Silicon Valley definitely did not really care that much about credit. Credit is very boring. It's negatively convex. And it was not something that, you know, people had to care about that were building innovative solutions. They're definitely focused on it right now.
What are you talking about when you see mid-market lending to buy GPUs? Give us a sort of like vanilla example of how that works, who the player would be, what they're buying it for, what type of lender they're looking for, and how that transaction occurs.
Yeah. So, you know, obviously there's all sorts of transactions and it's very hard for something to be so vanilla for a market that did not exist 12 to 18 months ago. Sure. But I would say about as vanilla as it gets, we see these multi-hundred-million-dollar, three-year fully amortizing loans that go to a mid-market neocloud operator, whether it's a startup or a neocloud that has a few billion dollars market cap.
Without naming names of your clients, what is a mid-market neocloud operator that someone could Google?
On the smaller side, a lot of these are private converted crypto and bitcoin mining folks. Some of them might just be quite literally a couple of people. And I think, you know, we can get into why when there's a shakeout in middle-market neoclouds and a lot of distress, a lot of those operators are going to have a hard time. On the larger side, you know, if you were to just screen neoclouds and look for anybody with a few billion dollars market cap or below, that's what you could think about. And over the summer when a lot of these loans were being written, they generally, about as vanilla as it gets, would be a 70% LTV, three-year fully amortizing loan. And at the time, these were pricing, you know, hold on, 70% LTV, 70% loan to value. So you buy $1,000 worth of compute, you're putting up $300 worth of equity and you're borrowing $700. Now, not only is it $300 worth of equity, a lot of these mid-market neoclouds are not even putting in their own equity. They're using customer prepayments that they're charging the large inference providers for. And that's why there's been so much consternation in Silicon Valley about these prepayments getting higher. And that's only going to keep going from here. I mean, just look at what CoreWeave did last week, which public markets rewarded them for, which they should. CoreWeave did two things. They required higher prepayments from their customers to push more of the burden of the upfront cost onto them, and they went back up market. Their current delayed long-term loan that's being marketed is back up market to off-takers, whereas the last one was a combination of Anthropic and Jane Street, which is more high-WB type operators. This is trickling down to the middle markets. Lenders in the private credit space are doing the exact same thing. They're asking for lower LTV structures and higher upfront payments. And they're saying, okay, we don't want to finance this company that maybe is worth $500 million or $1 billion. Let's keep it to the list of the 5 or 6 insurance companies that are worth $20 billion. And this is why, you know, spot pricing and on-demand pricing and short-term pricing is remaining so high because it's so indicative not just of the underlying demand, but of the buyer's cost of capital. A lot of the buyers of spot, on-demand, and short-term pricing wish that they could buy term. They cannot.
So for example, term, this means they wish they could buy compute on a longer scale.
So they wish that they could put 30% down at $4.30 per GPU hour for B300s, and they cannot. And maybe they could, you know, six weeks ago. But increasingly their neoclouds can't sell them that because those neoclouds are unable to take that off-taker to a lender and get financed. So a lot of these people, you know, the spot and the short-term market is just as much of a function, I think it's even more of a function right now, of cost of capital than underlying demand dynamics.
So this was going to be my next question, which is like when you think about how a loan is priced, like my understanding is that most of the price is based off of like the creditworthiness of the tenants, like the off-takers. Right. And then if the off-takers have to get continuous financing from presumably venture capital in order to actually pay what they owe to the data center operator, like that seems like a pretty big risk, right? At a time when maybe venture capital is starting to get more nervous about the overall ecosystem.
Yeah. So 100%. And I would say two things. And these are both shaking out in what's going on in the financing in the neocloud market. The first thing that I would say before getting into the second half of the question is you're absolutely right that at the start of this market, all anybody cared about was who the off-takers. Yeah, everybody viewed these loans as off-taker credit wrappers. And CoreWeave and the bigger neoclouds, where we didn't have to worry about execution risk, were the ones borrowing, as we have hundreds of smaller neoclouds. Now the operator is being underwritten more and more because take-or-pay agreements that these off-takers sign are only as good as if the guy has to pay, and the guy only has to pay if the neocloud does their job, you know, maintains uptime, meets SLA, etc. So now if you're in the private credit world, a lot of the lenders that I talk to, they see 5 to 10 different loans a week with the same few off-takers. And so what they're saying to me, you know, I had a very large private credit guy just last week say, all right, Stanley, I don't want to see any more loans where it's the neocloud's first or second deployment. Only show me experienced operators. So six months ago, it was 100% off-taker profile. Now it's definitely more split. And then onto the back half of your question. Absolutely. It's super risky because at the top of the market, CoreWeave is able to ask for higher prepayments from its customers because those customers have cheap access to capital. They can continue to issue debt, as it's been quite well documented. The hyperscalers continue to do this. Middle-market companies, a lot of these large inference folks, they've already been raising at a blistering pace, which is why a lot of these VCs are so over-deployed. The cost of venture equity getting higher is a little bit tough now where I think the gap is bridged, and I think this is what's going to truly catalyze the consolidation of neoclouds, is lenders are going to go back to these neoclouds and say, okay, you know what, I will do your loan. I'll do it in the mid-high teens, but I want a 50% LTV structure. So those neoclouds can accomplish that by two things. They can either go to their off-takers and ask for even higher prepayments, which they're starting to do. The problem is, those off-takers don't have unlimited coffers. This is venture equity, or those neoclouds can put in their own equity. So some of the neoclouds that I'm talking to are already raising more and more holdco-level equity so that they can put in equity to lower the LTV of these structures. This is where the smallest neoclouds, where it might be, you know, a couple guys who are converted bitcoin miners, where they're using customer prepayments to bridge the 30% equity, it's going to be hard if they don't have. The best way that I could say it is access to debt is going to equal access to equity in the middle market.
So you said something interesting, which is that if you look at a company like Anthropic, its compute bills used to be mostly out of financing, but as a share of revenue it's come down. So they have a lot of revenue to come in and it makes it a little bit more sustainable. Let's look at this chart. This is a chart of the Silicon Data We're Moving to B200 index. So fairly contemporary chip. It continues to rise. It's up to $7.10 an hour. As recently as August it was in the fives. Tell me, when you look at this chart, what are you seeing?
I mean, I definitely am seeing a combination of things. Number one, there is demand, like there's demand there. It's been quite well documented how much VCs have continued to deploy this year. That being said, VCs deployed less in Q3 than Q1 by about 25%. So it goes to your point about potential stress in that part of the market over time. Also, a lot of that liquidity is supposed to come back with these large IPOs over time. So part of it is demand. A lot of it is the cost of capital. I think a huge part of this is the cost of capital.
But just to be clear, when you say this is demand, and of course, on some level, of course it's demand, this demand is coming from big VC equity checks to companies who then use that equity to buy compute rather than, in the case of Anthropic, revenue being used to re-buy compute. This seems like a very big difference.
It's a huge difference. And I think this is where the vulnerability in these residual value assumptions are. I think if you just take a step back, and there are 30 different variables that go into residual values, whether it's new chips coming out, supply coming online, remarketing capacity, it really is right now a demand-side function of rental rates, as you're seeing. And the demand side is a function of two things: the willingness and the ability of the buyers to pay. Right now, we have not had to worry about the willingness of these buyers to pay, but the ability of these buyers to pay is from venture equity. So if more and more people can't go the Anthropic route of converting equity-funded compute purchases into revenue-funded compute purchases over time, then this is going to be an issue. And I think this is also, you know, this market is then, it's like other commodities, like oil, where a vast minority of the amount of compute that's procured and traded is done in the spot market. But these are the prices that we use as our reference rates. So I think it's incredibly, incredibly vulnerable, for sure.
This is where I start to get kind of nervous. And I hesitate to bring this up, but I start thinking about correlations during the subprime crisis, right? Where like everyone thought that defaults were going to be relatively isolated from each other, but no one understood that so much of the economy, so much of the financing trade was like built on the same housing exposure. And that ended up all being extra, triply linked. So if I think about venture capital, like a lot of that is coming from enthusiasm for the AI trade, and then it's getting plowed right back into the AI trade, and you can kind of see the link and the risks there. I wanted to ask though about, I guess, repossession. If an operator actually fails, which, you know, could be a realistic probability in the coming year or so, what happens? Like, do they actually have the right? If an operator fails, can they kick out the tenant? And then like, does the lender get access to the underlying collateral? And that gets like resold out.
Yeah. So this is a really important question because, you know, in a lot of the work that we do with a lot of our private credit partners, we're trying to standardize as many of these sort of middle-market loans as possible. The most important thing that I would tell anybody who's lending in the space is make sure that you have the right to do that. And also, these compute contracts need to be transferable and assignable. So some of them don't, unless it's like specifically spelled out in the covenants.
Correct. I mean, there's no, I wouldn't say there's no, but there's very limited standardization.
Okay. Increasingly lenders are focused on this. And also increasingly, the compute buyers, who are the insurance companies, are focused on getting non-disturbance agreements with their existing lenders because their paranoia is, okay, well, if my operator fails and I have this contract that I signed a year ago where I'm paying $3.80 for a GPU hour, if the current going rate is $4.30, the lender is not going to honor my contract. They're just going to reset it to someone else. Now, the actual remarketing and relighting is not as complicated as a lot of other commodities because you don't have to ship the GPUs anywhere. You just transfer SSH keys. This can be done, and we can do this in like a matter of minutes, for example. Obviously for larger scale projects it might take a bit longer. But yeah, I mean, I think this entire ecosystem has been built on the idea of two things that did not exist 6 to 12 months ago. Number one, that recovery in the compute asset class is relighting the capacity. Relighting the capacity is a lot harder to do operationally. You need the right legal structure in place from the term sheets than it is from an engineering perspective. And also the idea that through the construction of financial markets, derivatives, potential hedges, there are ways to lay off some of this residual value risk. But I think this is why it reminds me so much of, there's no perfect analogy, but I think aircraft finance is a very interesting analogy because when you look at pre and post financial crisis with aircrafts, there were large prepayments that were required. So, you know, the Boeings, the Airbuses of the world, post-2008, everybody thought, okay, this is not very financeable. Well, what made the asset class financeable was the construction of financial infrastructure around this illiquid real asset market, regular appraisals. So I think compute, you know, what we're focused on building is exactly that financial infrastructure. And I genuinely believe that Silicon Valley should embrace this credit crunch. It obviously is frustrating and it sucks right now. But if you just look at all of these other historical buildouts, nothing is so linear. We believe that AI is so extraordinarily revolutionary, and that is why capital markets have been so extraordinarily generous to the AI buildout over the last 12 to 18 months. There will be corrections and steps back. What's happening right now is it's happening in the neocloud market, going back to the GPU market.
When people talk about this boom, they reach for the telecom buildout as an example. But another potential analogy from that same time that doesn't get as much attention, a friend pointed this out to me, was a company like Sun Microsystems, which sold, I'm looking at a thing from Waverly Advisors. Sun Microsystems was the darling of the internet boom. The company provided servers that powered dot-com companies. Every company from Pets.com to the globe was buying Sun equipment. So here is an example of VC capital being used to flatter the boom in a specific piece of equipment. Talk to us a little bit more about like how correlated the entire trade is to essentially the VC environment, the huge checks that they're raising, the huge checks that they're writing, the speed of zero to a billion that we see, or 10 billion that we see, some of these startups, and just this voracious demand for compute out of VC money.
Yeah. I think this is incredibly, incredibly correlated. And I think that what it really is coming down to are two things. Number one, how much can we infiltrate AI productivity into the overall economy? And number two, it's so far very little so far. Yes. And number two is the evidence that more and more companies can generate revenue to buy compute and get off of equity. And I think the idea is because of what you just said about right now, there's a lot of skepticism. It's very healthy and it's very natural for credit investors to take a step back, especially when this is happening at the public market level. I think if you were sort of in the VC lens, you should think to yourself, okay, well, Wall Street is doing what Wall Street needs to do. You know, they're managing other people's money. They need to properly price this risk. We should be focused on investing in the financial infrastructure so that if AI is as revolutionary as we all truly believe it will be, when the evidence of that starts to come in, the financial infrastructure to finance this is there. So right now it is entirely, I wouldn't say entirely, but it is truly dependent on proving to the world that all of this equity-funded compute is going to generate revenue to buy that compute in the future.
I was just remembering the Sun Microsystems, the famous like statement from the CEO. He told Bloomberg, like, you don't need any transparency on our sales numbers, do you? He really said, yeah, something like that. There were signs. Yeah, there were signs in retrospect.
So the other thing I want to ask about is, you know, you mentioned the need for further financialization, I guess, of this market. And a lot of people have complained about the lack of a hedge, and we've spoken to a bunch of companies who are in the process of trying to build something that could hedge compute exposure in one way or another. But like, I also, I don't mean to keep bringing up 2008 because I really don't think it's like a perfect parallel. But I remember in 2008, the thing that kind of made the subprime market boom was the invention of CDS, the hedge. And so that, like, it kind of worries me a little bit. But talk to us about why people want the hedge in the market and how it would actually impact the market's growth.
Yeah, I mean, I think there is a very natural desire for this hedge because credit, at the end of the day, credit is a negatively convex product. You're sort of writing puts at the end of the day, you need to make sure you're compensated for that. So this is why these spot players are so hard to finance with credit because the value accrual of spot being higher goes to the equity holder. Any kind of hedge that acts as a floor on recovery is compelling. It's basically artificially, synthetically creating the same thing as an Nvidia backstop. That's what it is. You have to, you obviously have to trust the index and minimize basis risk. But even when CDS started, you know, there was a massive basis risk. And there still is basis risk between bonds and CDS. I think CDS is a really great analogy because we definitely are one of these companies that's also trying to work on hedges, given we were closer with all of these lenders than a lot of these other companies. And, you know, I do get a lot of demand from private credit for these hedges. And what it looks like will be a strip of put options on our index, usually the B300 index. And it makes a lot of sense if you have been convinced that relighting the capacity rather than selling the capacity is what recovery will look like, that is the recovery that you should hedge. So insurance on GPUs that are insuring scrap value is not quite as useful. I would say that, you know, I've been talking to these lenders for the last six months and nobody really was focused on these hedges. I would say we have at least one or two, two to three inquiries from large private credit firms a day for the last two weeks inquiring about indicative pricing on B300 puts. And but I think it makes sense and I think a lot of it is, you know, any way that you can flow a recovery that is direct value accrual to the creditor, which is why the Nvidia backstops were obviously so important for so long.
So great, let's talk about that. We don't really know yet how Nvidia is going to be financing neoclouds and so forth going into future financing demand of its own chips. But certainly a lot of questions, and let's say the new Nvidia management, whoever it is, really takes a more conservative posture than Jensen and his family, really. What does that do? How do you see that playing out?
Yeah. This is, you know, a $1 trillion question for the ecosystem right now. Obviously nobody knows. And obviously I can't predict what they will do. I will say that if financial conditions get more conservative in the market, which they already are, and Nvidia simultaneously gets more and more conservative, that will be a major stressor on the ecosystem. It will be a major stressor on the ecosystem because Nvidia has been very relevant in helping to finance both sides of the equation. Nvidia has backstopped and invested in many of the neocloud equity raisers. They have also gone to the credit side and provided backstops to the creditors. Now these backstops, because GPU residual values are so high, they haven't been hit on anything yet. But I think that's a possibility. I genuinely do believe, though, that if you are Nvidia right now, now is the time to stabilize the middle market. And that doesn't mean putting your own balance sheet down. What you should do is be incentivized to work with the people that are seeing the mid-market neocloud distress. Some of them will fail, and that is totally okay. And it's healthy. What you don't want is a total freefall in the ecosystem, no control on the roulette and remarketing capacity. And you want this consolidation to happen in a very structured and healthy way. The consolidation will happen either way. My concern would be if Nvidia pulls back entirely and the middle market is kind of a total free-for-all and it's left for dead. And the reason that this is my big concern is because what has made this asset class increasingly financeable to private credit over the last 12 months, it's that graph that you showed me. It would be B200 rental rates going up. It's the idea that these chips continue to earn more and more. If you have this massive diaspora of neoclouds defaulting left and right, and all of this remarketing capacity, and I have already had larger neoclouds call me and say, if you see distressed purchase orders from smaller guys, I want to buy it. I want to buy it on the cheap. The more data points that you get, that might impact the direction of this graph going up and to the right. All it takes is one Wall Street Journal article about GPU residual values falling, even if a lot of it is distressed capacity people are happy to buy, and private credit might say, okay, whoa, whoa, whoa, let's take a step back here. Let's reevaluate our priors. What do we think of residual values?
Do they actually, how often would they have to reprice based on like falling GPU collateral values?
It's a lot less falling GPU collateral values. And it's a lot more these mid-market neoclouds not being able to get financing.
Okay, okay. Well, as long as they're reassessing their priors, they should be okay. What? So the other thing again, hopefully this is coming out the same day that Anthropic starts trading, another view from cloud land would be like, it just doesn't matter. There's going to be so much demand from Anthropic and probably OpenAI that if a bunch of these like smaller inference companies dry up, it doesn't matter. There's just an infinite pool of demand from the really big AI labs. What is the role of them in this? And can that be counted on to put a floor under GPU prices wherever they are? Well, chip prices, whatever they are.
Yeah. I mean, there will be demand, there will be a lot of residual demand, but there will be demand from hyperscalers too, who are also spending a lot of their own capex for their own buildouts. But I don't think the demand will dwarf the negative sentiment. And at the end of the day, Anthropic is being rated, as you know, a high double-B company. This will affect their cost of capital. So even if Anthropic is going to gobble up a lot of this compute, the cost at which they will do it is dependent on sentiment in the sort of middle market.
Can you just talk a little bit more about, again, going back to the B100 chart or the H100 chart, how much does this chart, is it affected by this phenomenon that you're talking about? Where, or any of these charts, where the people buying compute regularly, maybe on the spot market, can't lock in long-term agreements? The significance of that to what looks like sustained demand.
Yeah, I think it's drastically affected because if you look at these charts, most of them, if they're worth anything, will exclude hyperscaler pricing. You're looking at a lot of neocloud pricing and where transactions are absolutely happening. And it's quite interesting because if the chart is going up and to the right because credit is getting harder, that is not really a healthy and sustainable environment. If there's a breaking point and people are starting to run out of money, that's where the floor can drop out from below. The spot market is pretty thin. You know, a good question that I got from, you know, one of the large private credit firms is how many bids do you think really need to get hit for the floor to fall out in these graphs? And the answer is probably a lot less than people think.
Yeah, this is what I wanted to ask. So talk to us about the standard, I guess, covenants in some of these deals. So are there like trip points where things become problematic and, you know, the borrower has to start, I guess, renegotiating, blown, or like returning stuff to the lenders?
Yeah. In the middle market, it's a lot less about covenants. Well, the covenants do matter. The most important are your lien on the hardware, the ability to access this, an agreement with the, tripartite agreement with the data center as well, to be able to relight the contract. But if there's a default, there's a default.
And the default might come from like an LTV ratio or something like that, like, oh, I mean, we're already seeing LTV come down. Okay.
Nothing, nothing. We were seeing 80% LTV loans get priced over the summer. I think post-Labor Day it was more 70%. And now people are asking for 50 to 60%. The other thing that I would say is, and this is where a lot of the more punitive covenants are coming in, a lot of these mid-market neoclouds are coming to credit for emergency sort of rescue and bridge financing. And these are extremely, if you're a lender, you want to be protected. And these are, first of all, I think they're extremely interesting investments, but they're extremely tight covenants. And if there is not a takeout in 3 to 6 months, and that is relying on future takeout capital from private credit anyway, lenders will, you know, foreclose and seize the equipment.
Last question for me, I think, and this is sort of big picture, how different, I've been asking this of various friends, how different does the vibes in San Francisco feel today than they did, say, two weeks ago? It's pre-IPO. I think this is a different world. Like, how different worlds?
To us, it's a different world. And I think it would be a different world regardless of the Jensen situation. The Jensen situation just exacerbates that, for sure. But it's a different world, you know, I talk to a lot of folks who are asking me, why are my neoclouds asking for more and more prepay? And if you, for example, you know, I'll just give you an example of how much people value the cash on their balance sheet and their equity. You know, we were, I was showing a deal to someone where you could rent, you know, three years' worth of B300s at $4.40 for a GPU hour, 30% down, or I'll show it to you, zero prepay, high fives, mid-high fives. They will choose the zero prepay, mid-high fives. And it's not that they're asking themselves, they just care so much about their cost of capital. But that spread, I think, is holding on the capital.
Yes, 100%, is hanging on there.
And the other thing, the other thing about that is if you're one of these companies, how do you grow faster? You procure more compute. How do you procure more compute now and then go back to the VCs and raise even more money? You get more compute contracts signed. So the more of these lower prepay contracts that you can sign, even if it's higher TCV and you're signing way more take-or-pay agreements than the cash on your balance sheet, the more you can then go to the VCs because your revenue is going up and to the right because of all this compute that you've been purchasing. So it goes back to your point about VC equity being the true variable to us.
I was thinking so much of this seems to be like a cultural misunderstanding between Silicon Valley, which cares about charts going straight up and to the right, and the credit market, which cares about convexity and charts that are like sloping and going downwards.
Yeah. No, I could not agree more. And I think as people, and you opened this by asking about how people in Silicon Valley, how much they know or care about credit. For me, as someone who was in high yield credit for years, this is like the Super Bowl. Nobody cared about how you look at it for the longest time. Everybody for you. Everybody cares right now. But no, I mean, I think you're exactly right. There's this mismatch and there's this frustration of this idea of credit markets and the lack of financing are killing startups and innovation in this country. And I think if you're sort of in Silicon Valley, you should be thinking to yourself, okay, capital markets have been so generous to us. It's very natural for them to maybe take a step back. We should focus on proving that equity-funded compute can be revenue-funded compute, and we should also be investing in the financial infrastructure to make this asset class financeable. Because like I said earlier, if it truly is as revolutionary as we all think it is, and I do believe it is, the proof will be in the pudding. These companies will perform. And when they're performing and we've underwritten stronger neocloud operators, the financial infrastructure that you should be investing in now will be there waiting to finance the rest of the buildout just fine.
Final thing for me, but the big story, everyone, you know, there have been a million chances of like Nvidia central bank of AI, and I'm sure some of that is true. But the key thing I'm taking away from this conversation is it's really about VCs and the scale of those checks and the scale of the LP checks that come into VCs and how ultimately and the speed with which, I don't know, your cursors, etc., going from $0 companies to a billion companies and how much that's just pushing cash up into these GPU prices.
Yeah. And I think, you know, if you're sitting there in the VC world, the way to get around this is these companies need to produce and make the economy more productive, like you said, and converting this into productivity that's driven by revenue and signing more enterprises as your customers rather than, yeah, just more AI app companies. That's how you get around that.
All right, Stanley, thank you so much for coming on the podcast. Fantastic. Yeah. Great. So, Tracy, we just wrapped up our conversation on the GPU question specifically, but of course, there's lots of questions that people have about the math all around and all of these still very high valuations, particularly for a lot of the private, soon-to-be-public, depending on the timing, could very likely to be at least one of them public today as we're recording. I mean, the expectation is that Anthropic gets to 1.4 trillion. Yeah. Which is this. Once you get into trillions, it's like numbers that I can't even conceptualize. This is like 1.0 to about 2 trillion, etc. But anyway, we're recording this November 18th, but we expect that this comes out, that it will be Anthropic trading day. And the question is for many of these companies, just like, do these numbers, do they make any sense and do they add up? And, you know, you've seen these things like the owners of the data centers have to collect like 10% of U.S. GDP by the year 2030 in revenue for something like, are there any shot at this? Anyway, I'm very excited to say we really do have the perfect guest, someone who has been writing about this and talking about this and investing in this world, investing in the tech world for a very long time. Someone who himself has some serious claim himself away. All that's right. So we've known for a very long time. So when we saw some questions about the math on this, we're going...
to be speaking, of course, to Paul Kedrosky, partner at SK Ventures, a fellow at the Oxford Mann Institute of Quantitative Finance and AI. Paul, thank you so much. Let's actually start kind of like on the conversation we just had. We see these numbers for like the 180, 108, 200 index. They're all going up and to the right, Nvidia. And she said on its conference call last night, and I thought this was very interesting, or sorry, I shouldn't say, Tuesday night. Nvidia actually said on its conference call, well, with token prices going down, people are more excited to spend more on GPUs. I don't know if it makes strictly. And he said, what's your take when you see these numbers going up that seems bullish? Why is it like that? Seems like great, there's demand for AI. I'd like more residual value. What do they, what do you see when you see these charts?
Same thing. I just think people, and this is the story, has been the story all along on this stuff, is that people confuse number goes up and to the right with the viability of the business, your ability to pay off the underlying financing. Right? I mean, these are two very separate issues, and people conflate them all the time. That if you say, you know, the likelihood of being able to generate a revenue not at the levels required, given hurdle rates, margins and everything else with respect to this, the CapEx build out, people immediately assume you're saying something that you're not, which is you're saying that, by the way, AI doesn't work and you're all idiots. Now, of course, that's not the problem, right? I mean, the issue isn't that it actually, that AI doesn't work, but the underlying, the math is really difficult, as you alluded to off the top. And so you're forced to kind of seize on all these proxy measures. Right. And one of the proxy measures people like the most is, hey, look, the rental prices of GPUs and data centers as proxied by these various indices that we have out there are not only just, they're not only are falling, they're actually going up, but there's a host of reasons why that could be happening. And so it's really this for me is like the drunk guy looking for his keys in a field, you know, because it's like, well, I'd rather look under the lamp post because it's easier to find it there. But yeah, you lost it in the field. This is an availability bias problem.
Okay. So which part of the math doesn't exactly work? And by the way, I broadly agree with this. Like you can have a groundbreaking technology that actually isn't economic at its current valuations. But let's drill in what exactly is like the arithmetic that doesn't work.
Yeah. So I mean, it's pretty straightforward. I mean, there's been a bunch of different estimates that have gone at, I've done at Columbia, some folks at the Columbia Graduate School of Business did it. There's actually a, I think it's a Bain study floating around, but they all come, we all come back to the same approach, which is to say, take a number for CapEx over the next, say, by 2030, 2032, in terms of the total spending on building out, this is the sort of data center moment, and then in turn say, what kind of hurdle, right, does it have to have? What kind of gross margins and what kind of margins do we need to have on that? And then say, okay, fine, flip that around. This is all basic, you know, MBA 101 finance, and say, okay, what does that imply as a, as a piece of project finance in terms of the revenue required to justify the spend? It wouldn't be any different if I was building a factory. It just so happens we're, you know, out to, you know, $10 trillion as opposed to building like a small auto plant or something like that. And so the numbers you come to are really dramatic. You're looking at numbers, depending on what you want to play with in terms of margins and hurdles, something like 3 to 6 trillion in terms of annual revenues by, say, the 2032 time frame, if you're one of those, you know, hopeful optimists who think it's required that you actually earn a reasonable rate of return. So those are the kinds of numbers you're looking at. And then you try to put that in some kind of context. And Tracy, I think it was Tracy indicated that that's at the top there. Like, well, that's, you know, 18% of global white collar work, or that's larger than the global software industry. We can all play this game. But the reality is it's just a very large number and it's not obvious where it's going to come from. And so that's where this discussion starts and often ends.
It's the total addressable market of like everything, is everything.
Yeah, that's right. I often joke, by the way, that if you're going to play the TAM game, this total available market game, and Goldman was the one who kind of set off the parade of doing this, you might as well just say, let's reduce humans down to basic elements. I mean, why stop at like the TAM for white collar labor? We might as well just reduce people down to atomic components and say, okay, here's what the raw value of humans are, because this is a game. Because what does analysis of humanity like our carbon component? I don't know why people don't go all the way with this. And so the problem, once you start making this argument, and you guys spend a lot of time in commodities, is once you reprice the corner of a commodity, but pick copper as an example, if you reprice 15% of the copper marketplace, I have a news flash for you. You haven't actually repriced 15% of the copper market. You've repriced the whole market because that's how commodities work. So the underlying assumption is that white collar labor is in some sense a TAM. And we're going to reprice a significant fraction of it. What you're really saying is I'm going to reprice all of white collar work. Now you can argue whether or not that's even possible. But the point is, the TAMs that people work from make absolutely no sense. If you accept the first principle, which is that I'm going to take a significant fraction of that TAM, well, now you've just repriced the market, right? It's no longer that large. So now you need to own more of it and so on. So the math breaks down in these ways.
Let's talk about how the revenue could get there. First of all, speaking of copper, it's actually getting clobbered lately. It's down like 5% today. And if you overlay a chart of like copper futures or miners, it looks like the Nasdaq because it's the commodity. If you overlay with the chart of semiconductors.
Yeah, that's right. Yeah. Yeah, it's like the chart that's of the commodity. That's the proxy.
We, we did an episode with Alexander Altman in Berkeley, sort of trying to talk through the math. And his basic view was this, that like, even if companies really adopt AI and mash, so they get some sort of ROI that's limited how much they're going to spend. And also, like, it's still not clear that any big company has actually found positive ROI from using AI, except maybe in their engineering department. He just says like it's not enough, especially if you don't get consumers paying for AI too. He says like to actually get the numbers where they would need to be, you need a consumer showing up. But for now, it looks like most consumers are pretty happy with like Gemini's summaries and Google search, and that that's just not going to get there the way like a paid service where like a family is like, we need to be spending $200 a month on ChatGPT Pro. Do you agree with that?
Oh, totally. I mean, that's no, which is never going to happen. So the whole story has to be the enterprise story, right? So you, I don't even believe that there's any way to get to these kinds of numbers without an enterprise story. And the enterprise story, which is really the Anthropic story, right? That's the whole bull case for Anthropic, is that it is the, the enterprise player. The only way to get there is to assume deep penetration of the enterprise with Anthropic and mass substitution, because there's only two ways you can generate the revenue. You're either substituting for other things, or you're complementing it and making them so much more productive that the whole, the whole pie grows. Right. And what we're seeing in the early days, and you mentioned this earlier, is that the only place where you're seeing that substitution productivity effect is in coding, in software engineering. And I think this is incredibly important. And it's one of the tenets of, not the Christopher Nolan tenet, but the other one, one of the tenets of venture capital is that your early customers look nothing like your later customers. If you project a business based on the early adopters of your product, you're making a terrible mistake. So in the case of large language models, the early adopters, the most enthusiastic buyers who were obsessed with being at the frontier of performance, were willing to pay any price, were token maxing their brains out and vibe coding their way to 10,000 fitness apps.
We're coding engineers. We're coders.
Yeah, and Joe, Bruno look nothing like that, right? They don't. We're not at the frontier. Things are good enough. I talk to coders who are baffled that people use Muses. They're like, don't you realize that's a horrible model? People don't care. It got them a restaurant reservation. This is the Muse tells me you have two emails that tell you you're going to be there, and it's supposed to be somewhere at 3 p.m. tomorrow. You might want to cry. That's a totally fine use case. I do not need Fable 5.2 to tell me I have two emails that are like to tell me that I have two reservations conflicting with each other. Exactly. And so. And it gets worse though, because software is really unusual in another way, which is that it has a what's called a tight gradient descent, which means that the feedback loop, as you let it work in its own undirected way, at whether it's RSI, recursive self-improvement, or it's something more mundane, like just improving some old financial code, you can let it loose and do that because there's ground truth, right? The software works or it doesn't. And so it can, it feeds back to itself and it can improve. Now try doing that with most of white collar work. I defy you to give a PowerPoint presentation to a model, let it loose for 12 hours and say I bet what it comes back with is better. It won't be. Not because models suck, but because there's no ground truth. Who knows what better means in that context? So if you let models loop recursively on things that don't have ground truth, you get exploding errors, you don't get better product. And this is a deep problem as you transition from coders to normal users, which is, which is exactly what we're facing right now.
Well, this is something that I've been thinking about even in the context of coding, because people were very excited about AI and its ability to code early on. But now, with all the outages that we've seen this week, I mean, like, you know, what is that productivity if all you're doing is like creating new bugs that have to be continuously fixed?
Probably not. I mean, in the number, yes. And the numbers are staggering right now, if you've looked at it. But like the CVEs, in terms of critical vulnerabilities in the first, in the first three weeks of October, we identified, I'm going to go to take the numbers. It's like 8% of the critical software vulnerabilities ever found, ever found were discovered in the first three weeks of October. 50% of all the critical software errors, the bugs ever found, were found in the first half of this year. Yeah, and it's just accelerating it. So this is a deep problem that this is a fantastic bonanza for the cyber security guys, for Palo Alto, for CrowdStrike and lots of others. But it's not particularly good for anyone else in terms of producing productivity. And then we get into this problem of, of, you know, the last few days where it's felt like I maybe should just unplug everything. I'm not even sure anymore whether I'm like, my, my toaster has got something going on inside. So this is, this is the problem is we're seeing again where the productivity is, and the productivity is with people who are able to use these tools really quickly to find vulnerabilities in other systems. And this is a deep problem.
And just on that point, to take it one step further, because, you know, this is finance after all, who actually benefit from that, from that other than like ransomware vendors?
Well, who benefits from that? The reinsurers. I mean, it's, if you talk to reinsurers, they can basically name their price now. I mean, it's amazing to talk to these guys because they, you can't price it properly, the risk, because even Anthropic said this in the S-1 was you can't price the risk. It may be larger than the global market for insurance, I think was in there, risks.