РАСШИФРОВКА 🔗
00:00:02
Speaker 1: Bloomberg Audio Studios. Podcasts. Radio.
00:00:06
Speaker 2: News.
00:00:18
Speaker 3: Hello and welcome to another episode of the Odd Lots podcast.
00:00:22
Speaker 1: I'm Joe Weisenthal.
00:00:23
Speaker 4: And I'm Tracy Allaway.
00:00:25
Speaker 1: Tracy, we're recording this.
00:00:26
Speaker 5: Yeah.
00:00:27
Speaker 1: November 18th, 9.34 a.m.
00:00:30
Speaker 3: I would say arguably we're right smack dab in the middle of two or three of the most important days in like the AI industry so far.
00:00:39
Speaker 1: For sure.
00:00:40
Speaker 6: So we've obviously had the NVIDIA scandal that broke last week.
00:00:44
Speaker 4: God, it feels like a lifetime ago already. We had NVIDIA.
00:00:47
Speaker 6: Earnings recently and on Thursday, tomorrow, hopefully when we release this episode, we'll be getting the Anthropic IPO.
00:00:54
Speaker 2: That's right.
00:00:55
Speaker 3: So as far as we know, Anthropic is going to begin trading tomorrow. We'll see if it happens. Lots of stuff between the NVIDIA earnings, the Anthropic IPO. Prior to all this, the market was weakening already.
00:01:07
Speaker 1: So even setting aside the scandals.
00:01:09
Speaker 4: I forgot to mention the midterms as well.
00:01:10
Speaker 1: And the midterms. That was the other tail risk.
00:01:13
Speaker 3: 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 Delving, the latest, he tweeted this chart, but they all do. GPU rental rates, which you would think is like a very clean measure of demand for AI.
00:01:40
Speaker 1: They continue to go up.
00:01:42
Speaker 3: And so you can put to all this market volatility and all this stuff and questions about ROI, etc., GPU rental rates still have been on the rise.
00:01:50
Speaker 4: Yeah, that's right.
00:01:51
Speaker 6: 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. And if anything, if data center construction is going to slow down, then maybe existing compute goes.
00:02:04
Speaker 4: Up even more.
00:02:05
Speaker 2: However.
00:02:06
Speaker 6: 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 these GPUs versus just how much the GPUs are worth?
00:02:20
Speaker 3: Totally, because we haven't seen evidence in the productivity data. that like companies by and large are meaningfully getting productivity gains out of.
00:02:28
Speaker 2: AI, right?
00:02:29
Speaker 3: And that's what you'd expect to see if there's going to be ongoing health for GPUs.
00:02:33
Speaker 6: You know, it would be funny 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.
00:02:43
Speaker 1: That would be ironic.
00:02:44
Speaker 4: I don't think we're going to see that.
00:02:45
Speaker 1: Probably won't be big enough.
00:02:46
Speaker 6: But that's the other big thing that is happening today is the release of Grand Theft Auto.
00:02:50
Speaker 2: That's right.
00:02:50
Speaker 1: Anyway, I'm very excited to say.
00:02:52
Speaker 3: We have 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've all had to learn about credit for the first time.
00:03:12
Speaker 4: Yeah, there's a big divide.
00:03:14
Speaker 6: 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.
00:03:23
Speaker 1: It's a little bit like the shale boom, isn't it?
00:03:25
Speaker 5: Yes.
00:03:25
Speaker 6: So this is my preferred analogy for what we're seeing right now is the 2015.
00:03:32
Speaker 4: Sort of break in high yield. But we can get into that with our guest.
00:03:34
Speaker 6: All right.
00:03:35
Speaker 3: Well, I'm really excited to say we have the perfect guest because we're going to be speaking with someone who His career has really straddled both the tech world and the credit world.
00:03:43
Speaker 1: We're going to be speaking with Stanley Lee.
00:03:44
Speaker 3: Chief Product Officer at Liquid Compute. Stanley, thank you so much for coming in the studio on Oddlocks.
00:03:50
Speaker 1: What do you tell us?
00:03:50
Speaker 3: What is Liquid Compute? And just like, how did you get here? Your background, where, as I said, you straddle both worlds.
00:03:56
Speaker 5: Yeah, no, thanks for having me on. So Liquid Compute is 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 in 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 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 I didn't really see the GPU marketplace as a just 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
Load full transcript
credit worthy 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 0.72 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.
00:05:18
Speaker 4: For sure.
00:05:19
Speaker 6: 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?
00:05:45
Speaker 5: 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 sustains 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 people had to care about that we're building innovative solutions. They're definitely focused on it right now.
00:06:19
Speaker 3: You say 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.
00:06:32
Speaker 1: That transaction occurs.
00:06:33
Speaker 5: 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. 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 dollar market cap.
00:06:56
Speaker 3: Without naming names of your clients, what is a mid-market neocloud operator that someone could Google?
00:07:01
Speaker 5: 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 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 gonna have a hard time. On the larger side, if you were to just screen NeoClouds and look for anybody with a few billion dollar market cap or below, that's who you could think about. And over the summer, when a lot of these loans were being written, about as vanilla as it gets would be a 70% LTV, three-year, fully amortizing loan. And at the time, these were pricing at 12%.
00:07:38
Speaker 3: So hold up, 70% LTV, 70% loan to value. So you buy X for $ 1, 000 worth of compute. You're putting up $ 300 worth of equity and you're borrowing $ 700.
00:07:47
Speaker 5: 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 because and it'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 upmarket. Their current delay draw term loan that's being marketed is back upmarket to IG offtakers, whereas the last one was a combination of Anthropic, HRT, and Jane Street, which is more high BBB type operators. This is trickling down to the middle market's 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 a billion dollars. Let's keep it to the list of the five or six inference companies that are worth $ 20 billion. And this is why 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- By.
00:09:08
Speaker 3: Term, this means they wish they could buy compute on a longer scale.
00:09:12
Speaker 5: 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 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, 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.
00:09:40
Speaker 6: When you think about how a loan is priced, my understanding is that most of the price is based off of the credit worthiness 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.
00:10:00
Speaker 4: pay what they owe to.
00:10:01
Speaker 6: The data center operator, that seems like a pretty big risk at a time when maybe venture capital is starting to get more nervous about the overall ecosystem. 100%.
00:10:11
Speaker 5: And I would say two things, and these are both shaking out in what's going on in the financing and 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-taker is. 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. taker 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, maintains uptime, matches SLAs, et cetera. So now, if you're in the private credit world, a lot of the lenders that I talk to, they see five 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 Neo Cloud'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 IG debt as it's been quite well documented the hyperscalers continue to do. These 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 offtakers and ask for an even higher prepayment, which they're starting to do. The problem is those offtakers 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 of guys who are converted Bitcoin miners, where they're using customer prepayments to bridge the 30% equity. It's going to be hard. The best way that I could say it is access to debt is going to equal access to equity in the middle market.
00:12:59
Speaker 3: 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. It makes it a little bit more sustainable. This is a chart of the Silicon Data B200 index. So fairly contemporary chip. It continues to rise.
00:13:20
Speaker 1: It's up to $ 7.
00:13:20
Speaker 2: 10 an hour.
00:13:20
Speaker 3: As recently as August, it was in the 5s. Tell me when you look at this chart, what are you seeing?
00:13:28
Speaker 5: It's a combination of things. There is demand. There's demand. 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.
00:13:52
Speaker 3: 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 rebuy compute.
00:14:13
Speaker 1: This seems like a very big difference.
00:14:15
Speaker 5: 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, remarketed 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, this market is thin. It's like other commodities like oil where a vast minority of the amount of compute that's procured and traded is done so 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.
00:15:15
Speaker 6: 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 they ended up all being extricably 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 risk 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.
00:16:00
Speaker 4: What happens?
00:16:01
Speaker 6: Like, do they actually have their 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?
00:16:12
Speaker 5: 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.
00:16:30
Speaker 4: So some of them don't, unless it's like specifically spelled out in the covenants?
00:16:33
Speaker 5: Correct. I mean, there's no, I wouldn't say there's no, but there's very limited standardization. Increasingly, lenders are focused on this. And also increasingly, the compute buyers who are the inference 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 $ 380 per GPU hour. If the current going rate is $ 430, the lender is not going to honor my contract. They're just going to re-let it to someone else. Now, the actual remarketing and re-letting 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 six to 12 months ago. Number one, that recovery in the compute asset class is re-letting the capacity. Re-letting 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 to the Boeings, the Airbuses of the world. Post 2008, everybody thought, okay, this is not very financeable. What made the asset class financeable was the construction of financial infrastructure around this, a liquid relet market, regular appraisals. So I think compute, 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 build-outs, 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 build-out 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.
00:18:43
Speaker 3: It's interesting going back to the GPU market. When people talk about this boom, they reach for the telecom build out as an example. But another potential like 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, is that 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 VCs, VC capital being used to flatter the.
00:19:22
Speaker 1: Boom in a specific piece of equipment.
00:19:24
Speaker 3: 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 one 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.
00:19:46
Speaker 5: 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- 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. 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.
00:20:56
Speaker 6: I was just remembering the Sun Microsystems, the famous statement from the CEO.
00:21:02
Speaker 4: He told Bloomberg, you don't need any transparency on our sales numbers.
00:21:05
Speaker 2: Did he really say that?
00:21:06
Speaker 4: Yeah, something like that.
00:21:07
Speaker 1: There were signs.
00:21:08
Speaker 4: Yeah, there were signs in retrospect.
00:21:10
Speaker 6: 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 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 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.
00:21:54
Speaker 5: Yeah. I mean, I think there is a very natural desire for this hedge because 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 obviously have to trust the index and minimize basis risk. But even when CDS started, 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 work closer with all of these lenders than a lot of these other companies. And 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 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. But I think it makes sense. And I think a lot of it is any way that you can floor recovery, that is direct value accrual to the creditor, which is why the NVIDIA backstops were obviously so important for so long.
00:23:36
Speaker 1: So great. Let's talk about that.
00:23:38
Speaker 3: 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 has a sort of takes a more conservative posture than Jensen and his family, really.
00:23:59
Speaker 1: What does that do? How do you see that playing out?
00:24:01
Speaker 5: Yeah, this is, you know, a trillion dollar 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 Neolab equity raises. 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 free fall in the ecosystem no control on the re-let and re-marketed 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 of 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 remarketed 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?
00:26:12
Speaker 6: How often would they have to reprice based on falling GPU collateral values?
00:26:17
Speaker 5: 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.
00:26:25
Speaker 3: Another view from Cloudland 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 smaller inference companies dry up, it doesn't matter. There's just an infinite pool of demand from the.
00:26:42
Speaker 1: Really big AI labs. What is the role of them in this?
00:26:46
Speaker 3: And can that be counted on to put a floor under GPU prices wherever they are?
00:26:51
Speaker 1: Chip prices wherever they are?
00:26:52
Speaker 5: 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 build-outs. But I don't think the demand will dwarf the negative sentiment. And at the end of the day, Anthropic is being rated as 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.
00:27:16
Speaker 3: 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? Or any of these charts where the people buy and compute regularly, maybe on the spot market? can't lock in long-term agreements and the significance of that to what looks like sustained demand.
00:27:41
Speaker 5: 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 actually 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. A good question that I got from 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.
00:28:17
Speaker 4: Yeah, this is.
00:28:19
Speaker 6: 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 the borrower has to start renegotiating the loan or like returning stuff to the lenders?
00:28:33
Speaker 5: Yeah, in the middle market, it's a lot less about covenants. Well, the covenants that matter the most are your lean on the hardware, the ability to access this, a tripartite agreement with the data center as well to be able to re-let the contract. But if there's a default, there's a default. And the default might come from- And.
00:28:47
Speaker 4: What about like an LTV ratio or something like that?
00:28:49
Speaker 5: Oh, I mean, we're already seeing LTVs come down. 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 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 three to six months, and that is relying on future takeout capital from private credit anyway, lenders will foreclose and seize the equipment.
00:29:31
Speaker 3: How different, I've been asking this to various friends, how different does the vibes in San Francisco field today than they did, say, two weeks ago.
00:29:40
Speaker 1: It's pre and post Jensen.
00:29:41
Speaker 2: I think.
00:29:42
Speaker 1: Is this a different world?
00:29:43
Speaker 5: It's a different world.
00:29:44
Speaker 1: Yeah, talk to us.
00:29:45
Speaker 5: It's a different world, and I think it would be a different world regardless of the Jensen situation. The Jensen situation just exacerbates it, for sure. But it's a different world. I talk to a lot of folks who are asking me, why are my neoclouds asking for more and more prepay? I'll just give you an example of how much people value the cash on their balance sheet and their equity. I was showing a deal to someone where you could rent three years worth of B300s at $ 4. 40 per 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 arbing themselves. They just care so much about their cost of capital. But that spread, I think.
00:30:26
Speaker 1: Is going to continue to widen.
00:30:28
Speaker 5: And holding on to capital. Yes, 100%. And having it on their real issue. And 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 and 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 to your point about VC equity being the true variable to watch.
00:30:58
Speaker 6: I was thinking so much of this seems to be like a cultural misunderstanding between Silicon Valley, which cares about charts going straight up into the right and the credit market, which cares about convexity and charts that are like sloping and going downwards.
00:31:11
Speaker 5: Yeah, no, I could not agree more. And I think as people and you open this by asking about. how people in Silicon Valley, how much do 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 high-yield credit for the longest time.
00:31:25
Speaker 1: Happy for you.
00:31:26
Speaker 5: 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, we're 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 AI 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 build out.
00:32:15
Speaker 3: Just final thing for me, but the big story, everyone, you know, there've been a million charts of like NVIDIA, central bank of AI. And I'm sure some of that's 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 the speed with which, I don't know, your cursors, et cetera, are going from $ 0 companies to deca centibillion companies and how much that's just pushing cash up into these GPU prices.
00:32:47
Speaker 5: Yeah. And I think 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 just more AI app companies. That's how you get around that.
00:33:06
Speaker 1: All right, Stanley, thank you so much for coming on the podcast. That was fantastic.
00:33:09
Speaker 5: Yeah, great.
00:33:26
Speaker 3: So, Tracy, we just wrapped up our conversation on the GPU question specifically. But, of course, there are lots of questions.
00:33:34
Speaker 1: That people have about the math all around.
00:33:37
Speaker 3: 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.
00:33:48
Speaker 6: As we're recording, I mean, the expectation is that Anthropic gets to $ 1.
00:33:51
Speaker 2: 4 trillion.
00:33:51
Speaker 6: Yeah.
00:33:51
Speaker 2: Yeah.
00:33:53
Speaker 6: Which is, once you get into trillions, it's like numbers that I can't even conceptualize.
00:33:59
Speaker 1: And it's like one point, like, you know, one point they were talking about two trillion, et cetera. But anyway, we're.
00:34:04
Speaker 3: 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.
00:34:13
Speaker 1: Like, do these numbers, do they make any sense?
00:34:16
Speaker 6: Do they add up?
00:34:16
Speaker 1: And do they add up?
00:34:17
Speaker 3: 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 and revenue for something like that? Are there any shot at this? 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.
00:34:36
Speaker 1: The tech world for a very, very long time.
00:34:38
Speaker 2: Too long.
00:34:40
Speaker 3: Someone who himself has some serious questions.
00:34:42
Speaker 4: He gave himself away already.
00:34:43
Speaker 2: Yeah, that's right.
00:34:44
Speaker 3: Someone we've known for a very long time. Someone who himself has some questions about the math on this. We're going to be speaking, of course, to Paul Kudrowski, partner at SK Ventures. a fellow at the Oxford Mann Institute of Quantitative Finance and AI.
00:34:56
Speaker 1: Paul, thank you so much.
00:34:58
Speaker 3: Let's actually start kind of like on the conversation we just had. We see these numbers for like B100, H100, H200 index. They're all going up and to the right. NVIDIA actually said on its conference call, well, with token prices going down, people are more excited to spend more on GPUs.
00:35:17
Speaker 1: I don't know if it makes strictly any sense. What's your take when you see these numbers going up? That seems bullish.
00:35:23
Speaker 2: Why is it?
00:35:24
Speaker 3: Like, that seems like great. There's demand for AI, like more residual value. What do you see when you see these charts?
00:35:31
Speaker 2: 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 and 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... the likelihood of being able to generate a revenue not at the levels required given hurdle rates, margins, and everything else with respect to 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's actually that AI does work, but the underlying 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 in data centers as proxied by these various indices that we have out there are not only just, they're not only not 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 lamppost because it's easier to find it there. But yeah, you lost it in the field. This is an availability bias problem.
00:36:43
Speaker 6: Okay, so which part of the math doesn't exactly work? And by the way, I broadly agree with this. You can have a groundbreaking technology that actually isn't economic at its current valuations, but let's drill in. What exactly is, I guess, the arithmetic that doesn't work?
00:36:59
Speaker 2: Yeah, so I mean, it's pretty straightforward. I mean, there's been a bunch of different estimates that have gone at it. I've done it. Columbia, some folks at the Columbia Graduate School of Business did it. There's actually, I think it's a Bain study floating around. But 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 sort of data center moment. And then in turn say, what kind of hurdle rate does it have to hit? 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 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 out to $ 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 timeframe. If you're one of those 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. They're 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.
00:38:27
Speaker 4: It's the total addressable market of, like, everything.
00:38:30
Speaker 1: It's everything.
00:38:31
Speaker 2: That's right, right, 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 human 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.
00:38:54
Speaker 1: A teardown analysis on humanity.
00:38:57
Speaker 4: Like a carbon component.
00:38:58
Speaker 2: Yeah, a teardown analysis. 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, 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 for 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. 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.
00:39:49
Speaker 3: Let's talk about how the revenue could get there. First of all, speaking of copper, it's been 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.
00:40:03
Speaker 6: If you overlay a chart of semiconductors, it's like perfect.
00:40:06
Speaker 1: Yeah, it's like the commodity that's a proxy.
00:40:09
Speaker 3: We did an episode with Alexander Altman at Berkeley, sort of trying to talk through the math. And his basic view was this, that even if companies really adopt AI and math, so they get some sort of ROI, that's limited how much they're going to spend. And also, it's still not clear that any big company has actually found positive ROI from using AI, except maybe in their engineering department.
00:40:35
Speaker 1: He just says it's not enough, especially.
00:40:38
Speaker 3: 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 consumers shelling out. But for now, it looks like most consumers, like, are pretty happy with, like, Gemini 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.
00:41:04
Speaker 1: Do you agree with that?
00:41:05
Speaker 2: Oh, totally. I mean, which is never going to happen. So the whole story has to be the enterprise story, right? So 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.
00:41:17
Speaker 4: Story, right?
00:41:17
Speaker 2: That's the whole bull case for anthropic is that it is 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 we're already doing, or we're complementing it and making them so much more productive that the whole pie grows, right? And what we're seeing in the early days, and you mentioned this earlier, is that the only place we were seeing that substitution productivity effect is in coding. in software engineering. And I think this is incredibly important. And what'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, We're token maxing their brains out and vibe coding their way to 10,000 fitness apps. We're coding engineers. We're coders. Yeah.
00:42:21
Speaker 5: And Joe.
00:42:22
Speaker 2: Normals look nothing like that, right? We're not at the frontier. Things are good enough. I talk to coders who are baffled that people use Meta's Muse. They're like, don't you realize that's a horrible model? People don't care. It got them a restaurant reservation. So this is the deep problem.
00:42:36
Speaker 3: The Meta's Muse tells me you have two emails that tell you you're going to be there. It's supposed to be somewhere at 3 p.m.
00:42:42
Speaker 1: Tomorrow. That's a totally fine use case.
00:42:44
Speaker 3: 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.
00:42:52
Speaker 2: Exactly. And so it gets worse, though, because software is really unusual in another way, which
is that it has what's called a tight gradient descent. which means that the feedback loop, as you let it work in its own undirected way, 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 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 that 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 exactly what we're facing right now.
00:43:49
Speaker 6: 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, is that productivity if all you're doing is like creating new bugs that have to be continuously fixed?
00:44:09
Speaker 4: Probably not.
00:44:10
Speaker 2: I mean, in the numbers, yes. And the numbers are staggering, right? I don't know if you've looked at it, but like the CVs in terms of critical vulnerabilities, In the first three weeks of October, we identified, I'm going to go think of 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 bugs ever found were found in the first half of this year. And it's just accelerating. And so this is a fantastic bonanza for the cybersecurity 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, you know, the last few days where it's felt like I maybe should just unplug everything. I'm not even sure anymore whether, like, my toaster's got something going on inside. So 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 benefits from that other than like ransomware vendors? Well, who benefits from that are reinsurers. I mean, if you talk to reinsurers, they can basically name their price now. I mean, it's amazing to talk to these guys because you can't price it properly. The risk, because even Anthropic said this in the S1, was you can't price the risk. It may be larger
than the global market for insurance, I think, was in their risk section.
00:45:37
Speaker 4: Oh, yeah, that's right.
00:45:38
Speaker 2: Yeah. Right, right, right. So if you talk to reinsurers like Swiss Re, Munich Re, Chaplin, All of these guys will tell you the same thing as they rub their hands gleefully is we can set any price we want right now. It's a tremendous interest boom time for reinsurers. These are the people who are benefiting, not productivity.
00:45:53
Speaker 6: Yeah, I find the insurance angles on AI so fascinating and we're going to do more on them for sure.
00:45:58
Speaker 3: I was thinking, you know, last night there was another outage. And it was some story and it was like, this wasn't a cyber attack. This was race to like patch a bug. And I was like, well, what's the difference? That's actually great for the cyber attackers. They don't even have to try. They can get the shutdowns of the critical systems without lifting a finger. And so functionally, it sort of ends up being the same thing.
00:46:21
Speaker 2: The funny thing was, not so funny, it's actually horrible, but the intersection of this stuff with NVIDIA is really interesting. Like this morning specifically, there was an NVIDIA, a piece of ransomware, like, I wouldn't say provider isn't the right word, but let's just say a ransomware outfit claims that they have access.
00:46:36
Speaker 1: Ransomware as a service.
00:46:38
Speaker 2: Yeah, yeah, yeah. RAS. They claim they have access to NVIDIA customer data, including shipment data. So, you know, these things are now intersecting, right? Because this is one of the big stories with respect to transshipment and NVIDIA. Now, all of a sudden, all of these worlds are coming together.
00:47:09
Speaker 3: We believe that as of the time this recording is coming out, that in a few hours, Anthropic will begin trading. There are some existing publicly traded AI companies, but they're all in China. Their talks have done terrible. Now, the difference is people would say, look, these aren't frontier companies. They don't have a ton of compute actually to serve inference. So that's not a biz. Like they just don't have that many data centers. Yet there's lagging edge. It's true commoditized intelligence at the lagging edge. And so they would say, well, the bull case for Anthropic is it's different. And the bull case is that they can sell a lot of compute at positive unit margins and they're at the frontier. And that they can monetize the frontier to an extent that even a model that's six months behind the frontier can't. And that therefore they're fundamentally different as a business. Give us your read on Anthropic as a business right now.
00:48:02
Speaker 2: So my argument all the way along on Anthropic has been the sooner they stop aggressively training models to be at the frontier and focus largely on inference, the faster the unit economics get better. And I think, you know, I joked for a while that The first frontier model
company to stop training models is the winner. And I was doing a presentation the other day on this, and I was talking about it and showing that in the last 18 months or 24 months, Anthropic has actually shipped more vertical products than updates to its models.
00:48:27
Speaker 3: Yeah.
00:48:28
Speaker 2: Which has largely been missed by people, which is really interesting, right? Because that speaks to exactly what I'm describing. They're not backing away entirely from training models, but they're moving really quickly up market and saying, We need to be, so they're launching in legal, they're launching in finance, they're launching in healthcare. So they've launched more vertical apps than they have updates to their models. At the same time, you know, the Epic Capabilities Index and others shows this kind of flatlining. So that's the future, right? But then the deep tension becomes, now I'm competing directly with the people who also give me distribution, right? We saw this with Anthropic Launching Office products. Well, they sell almost 48% through Microsoft Marketplace. And through marketplaces like Microsoft. And so now you're competing with your investors. And so now you get into this problem of, well, why am I even distributing your products at all? Because you're actually competing directly with me. And so then, of course, you're forced to take the next step. step, which Anthropic actually hinted at in the S1, and which goes to your question about the future unit economics. They said, we're going to have to move away from being so reliant on cloud rentals and distribution through these channels. They didn't say why, but we know why. It's because they're competing with their own suppliers. So they're going to have to be directly responsible for more data center construction. So that's kind of– the S1 was really sort of an anthropic 1.0 and they're already telling you that because we're launching more apps and we're moving away from pure reliance on frontier models that we have to move up market, which has direct implications for the CapEx build-out and for what their future obligations look like, which look more like $ 800 billion than the commitments that they had, which were purely lease-based.
00:50:03
Speaker 6: Yeah.
00:50:03
Speaker 3: We hear about IPOs being oversubscribed all the time. What does that actually mean?
00:50:07
Speaker 2: It means that 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, leak that, and say, oh, look, there's so much demand for this IPO, it's five or six times oversubscribed. It's complete nonsense. At the peak of the Chinese bubble— This summer, some of the IPOs were 8,000 times oversubscribed and then fell by 50%. It means nothing. It's unbelievable to me that the media parents it because you are literally carrying water for underwriters. You should get commission.
00:50:45
Speaker 1: I tweeted about that. I'm going to go ask for commission. I shouldn't have done that.
00:50:48
Speaker 4: It's the same with allocations in the bond market.
00:50:50
Speaker 2: Yeah.
00:50:51
Speaker 6: Can I just broaden this out a bit more? So with the AI unwind that we've seen recently, or like some of the weakness in the AI trade, one of the arguments you hear from the bulls is, oh, this is, A healthy market correction, the old cliche. And we're starting to see some differentiation between neoclouds, for instance. I was looking at pricing of some data center loans, and 2025 data center loans are seriously underperforming 2026. And the argument there is that, well, underwriting started getting better in 2026. So people are starting to differentiate within the market. What's your response to that particular argument? And also, actually, I should just add the other argument is with the midterms and a potential data center construction slowdown, that compute could become even more valuable. So people who already have it might be in a better place than they were, you know, three or four weeks ago.
00:51:44
Speaker 2: That last one in particular, I really struggle with. This is one of the most cynical arguments out there that good news is good news and bad news is good news. There's like a meme about a large hairy investor who I won't name, who's sort of standing their way at the apocalypse and saying, you know what, this is also good for us. And this is the problem, right? That once you start playing this game of finding reasons why both tight supply and wildly abundant supply are good for you, what you have there is theology. You don't actually have an argument anymore. You just have a belief and all data supports it. And so this is the problem with that sort of, That sort of position. And I think you have to engage directly with the problem, which is, Tracy, as you were saying, an increasing fraction of the pre-2026 loan book out there is trading at at least mildly impaired levels. And that's happening just as Congress says that there's a sort of a know-your-customer initiative working its way through, which is already showing up. There was a story this morning, I think it was on Bloomberg. in loan T's and C's with respect to you need to know who your customer is on this stuff. And it's not just loans that are being issued in the U.S., but loans in Southeast Asia, which is really unusual. We're essentially seeing U.S. know your customer T's and C's showing up in loans originating outside the U.S. Well, think that through. What's the impact of that? I don't want any part of those loans because I can't really know in a cloud world my customer in any complete sense. So what's that going to mean? That's going to mean that
loans become harder to get, more expensive, and spreads go out. And that's obviously really bad for neocloud providers.
00:53:09
Speaker 3: Can you talk a little bit more about VC world? Now, I know you're SK Ventures, you're more seed stage, so this doesn't apply to you. But one of the things that we've seen in VC world is just huge checks, huge valuations, huge growing. A company like Cursor, for example- The company's been around like three years total, something like that. They've raised, maybe like they've raised close to $ 5 billion, and then they sold to SpaceX at $ 60 billion, et cetera. These are just unbelievable numbers in a short period of time for startups. How much of, and again, maybe this goes to the GPU chart, how much is just they're like so much, like everyone's writing big checks. They're writing big checks to the LPs. They're writing big checks to the portfolio companies. The portfolio companies are writing big checks to compute providers, et cetera. And that it's just like cramming down big checks all the way down, flattering that top line numbers across the board.
00:54:14
Speaker 2: No, no, it's totally true. And ramp data shows it, right? If you look at the ramp data, one of the more popular sources of data on spend, the largest spenders on a per capita basis are uniformly venture-backed companies. Something like 80% of spend in terms of the leaders, in terms of per-employee spending, are all venture-backed. So you're seeing this phenomenon in action as that venture money in turn becomes spend that in turn goes directly spending on frontier models, which are obviously the most expensive and highest margin models on an inference token basis at those companies. And so, but this is a game though, right? Because you're seeing exactly what happens when it doesn't pay out. Cursor is a sort of a larger scale example, but there are myriad other ones where these sorts of adventures in token maxing aren't turning into products that anyone cares about. And there's great data from the National Bureau of Economic Research on this showing this dramatic spike in the number of apps being produced but a complete collapse in the number of ratings per app. So you have this disconnect, right? We have, in a sense, slop showing up everywhere, but some of it's got incredible capital backing behind it, which probably is going to have to be largely written off because there's no value being created. And that's the deep problem.
00:55:22
Speaker 3: All this investing, Tracy, it's all going great, except for the fact that there's no productivity gains, no value, and a lot of cyber attacks.
00:55:31
Speaker 1: But other than that, it's all going great.
00:55:33
Speaker 4: Everything's fine.
00:55:34
Speaker 6: Actually, this reminds me. So one of the reasons CapEx expenditure on AI has gotten so huge is because it's all couched in these existential terms.
00:55:42
Speaker 2: Oh, yeah.
00:55:43
Speaker 6: So, you know, the hyperscalers have to spend
loads of money because whoever gets to the frontier first wins everything.
00:55:50
Speaker 2: The U.S.
00:55:51
Speaker 6: Government has to support this industry because we're in a life or death race with China technologically.
00:55:58
Speaker 2: Uh-huh.
00:55:59
Speaker 6: And so you have this like ever upwards impulse on expenditure. I'm thinking how to ask this question, but can you maybe clarify where you see the Trump administration in terms of its support for AI now? Because on the one hand, with the NVIDIA scandal, it seems like they might have been turning a blind eye to China sales, but post the midterms and with all the headlines. We now have China hawks in the ascendancy.
00:56:24
Speaker 4: What does this.
00:56:25
Speaker 6: Actually do for CapEx, for government sanctioned CapEx expenditure, which I know you've been writing a lot about?
00:56:32
Speaker 2: Yeah, so I've argued that Essentially, we're running a private sector stimulus program aided and abetted by the federal government, right? Among the largest stimulus programs in the history of the country. And it's being done by the private sector as, I suppose, goaded and aided and abetted by policy. So you can look at policy in a host of different ways. There's been a series of executive orders, over a dozen now, that have all in one way or the other made it much easier to build data centers, to connect to the grid. Or, for example, to work around EPA rules with respect to sulfur dioxide emissions. I mean, there's a host of things that have come from the federal government, all of which uniformly push this whole moment further. And I used to think that it was all somewhat, I don't know, like almost naive that they didn't really realize what was going on. But subsequent reports in the context of this NVIDIA episode have made it really clear that The White House not only knows it, that they're terrified of ending it, but they've realized that one of the biggest things pushing capital markets forward and building people's 401ks, as the White House likes to talk about, has been AI CapEx, which is something that I thought for a long time, they didn't actually know it and were doing it by accident. No, it's totally on purpose. And so this is the deep tension is they actually do know and they are running by proxy this private sector stimulus program and it's having all of these consequences, but in turn, you know, something like 50% of the gains in the last four years have been Mag7 related, blah, blah, blah, blah, blah. So, you know, this is the deep tension in all of this. And so now, post-midterms, all of that begins to change, right? I mean, you can pick your favorite data point. Mine is that If Texas were a sovereign, which I'm sure we could end the sentence right there.
00:58:08
Speaker 4: Some people would like that.
00:58:09
Speaker 1: Many Texans would like that.
00:58:10
Speaker 2: Yeah, let's just stop right there. They would actually be the third largest builder in terms
of spending of data centers in the world behind the U.S. without Texas. Then you have China. Then you have Texas.
00:58:22
Speaker 1: Wow.
00:58:22
Speaker 2: It's staggering, right? And yet, so this is the third largest pseudo-sovereign worldwide in terms of AI data center construction. But we have a go slow governor-elect coming in that even slowing it down is material in terms of the impact. So this is just one example, but it's across the board. But just that's the most striking one to me was that that change goes directly to the heart of this de facto policy the administration is running that's kind of this private sector stimulus program. It's policy-aided both at the state level but at the federal level. So it all changes from here.
00:58:54
Speaker 6: There was a really great Deutsche Bank note out earlier this week where they talked about how a lot of the compute spending basically gets recycled into U.S. stocks in the same way that petrodollars used to go into U.S.
00:59:08
Speaker 2: Treasuries. Well, global trading. semiconductors is actually now larger than oil and gas.
00:59:12
Speaker 5: It's crazy.
00:59:13
Speaker 2: People don't realize the scale of all of this to the point that that recycling is now fundamental to what's happening in capital markets. Yeah, totally.
00:59:21
Speaker 3: You know, you obviously kept your head on your shoulder throughout this cycle, which is admirable. But I feel like in the wake of the Jensen scandal, the phrase that I keep coming back to is like, the cold light of morning. You know, you party all night and you have fun and stuff like that.
00:59:37
Speaker 4: You literally party all night.
00:59:39
Speaker 1: Yeah, and then you wake up one morning and it's like it's 7 a.m.
00:59:42
Speaker 3: And you're like, oh, you're putting your hands over your eyes like, oh, my last night was terrible.
00:59:47
Speaker 1: That's what it kind of feels like.
00:59:49
Speaker 3: And I'm curious from like the VCs that you talk to, et cetera, is this cold light of morning vibe a real thing?
00:59:57
Speaker 2: No, they're like jilted boyfriends. They all want their girlfriend to come back.
01:00:01
Speaker 1: Bring it back.
01:00:02
Speaker 2: Let's go back to last night. Yeah, yeah, yeah. I was joking to someone the other day that they should start sending out like Indigo Girls mixtapes. Maybe it'll all work out for them because it's just so obvious that people have a really emotional reaction to what's happening. And there's all kinds of great work on this in the finance literature about superstar CEOs and the consequences and how they earn outsized pay packages. But in the venture industry, it's, you know, Steve Jobs in black sweaters and it's never stopped being that way. It just keeps looping round and round. And so they see it through that lens. And it's uniform the case that, you know, if Jensen would just come back, everything will be fine. And, you know, and we'll also send him a mixtape with Indigo Girls.
01:00:41
Speaker 1: Yeah.
01:00:42
Speaker 6: You know, I've been reading some of the message boards on Reddit, like the NVIDIA subreddit, and the.
01:00:47
Speaker 1: Cope is high. Yeah, they really do. They really do.
01:00:51
Speaker 6: Actually, on that note, I would love to get your take on this. So we hear a lot of people describe NVIDIA as the central bank of AI in one way or another.
01:01:00
Speaker 4: Because you're so.
01:01:01
Speaker 6: Good at going through the technical details in a very compelling way, can you just go through, like, What exactly does that mean? And then what is your outlook for what happens now to all that NVIDIA financing?
01:01:13
Speaker 2: Yeah, so the way to think of it, and I've been guilty of using that metaphor forever, and the way to think about it is that in a sense, NVIDIA gets to print money. And the way it gets to print money is by... using its own balance sheet to create money on the other side, much like the Federal Reserve does. And it allows the banks, in each case these companies to whom it extends credit, to essentially create money from nothing, often on no payment terms, who in turn spin around and buy things back from NVIDIA, often on delayed payment terms. And that continues to recycle because that attracts venture capital and that attracts more building of neoclouds. And so it goes round and round and round, not just because it's one for one, because it's highly levered. It's a fractional banking system. Circular financing is inherently a fractional banking system, very much like what we see in the shadow banking system in the United States and then in private credit 18 months ago. And now with a single entity doing it, the fractional reserve banking system of 2026 was Nvidia. It was no longer anybody else. And so what they were doing was essentially this, was creating money through the circularity of these deposits that in turn generated all of this incremental spend. And so what we know about fractional banking systems is when they begin to go in reverse, they don't go in reverse linearly either. They go in reverse geometrically because all of that pulls back to the core. And you saw some of that alluded to last night, or two nights ago, depending on our temporal loop here, in the NVIDIA conference call when Colette, the new acting CEO, was talking about how the impact on neoclouds. And she said two things that I thought were really important. One was that we're still huge supporters of neocloud. And by the way, we're going to be much more discerning. As soon as you start becoming more discerning about neocloud, that leaves you with basically no neoclouds. So, you know, being more discerning about neoclouds leaves no neoclouds. And so this is the problem. And that's the fractional banking system going in reverse.
01:03:00
Speaker 3: There was also a line in there. And our guest from this weekend, Stacey Rasgan, was on that call. There was a line in the call.
01:03:06
Speaker 4: He
asked the question.
01:03:07
Speaker 3: He asked the question, which was like something like NVIDIA had said either in the statement or in the prepared text about like, well, you know, the cash doesn't go to the cloud until they like really break ground and get working. Is this a new policy or is this an old policy that you're telling us now?
01:03:22
Speaker 5: And they're like, no, no, no, no.
01:03:23
Speaker 1: We've always been doing it this way. But we really want to let you know now.
01:03:27
Speaker 3: That, like, it's really important for us to communicate with you that we don't just give them the cash before we can take it.
01:03:32
Speaker 1: I got one last question. You sort of alluded to this.
01:03:35
Speaker 5: There is this.
01:03:36
Speaker 3: Idea out there that anthropic and maybe open AI will not always be in the token selling game and that the future perhaps for these companies will be to use primo intelligence to enter into all these other kind of verticals. Health is an obvious category where it may be, law, etc., And that that's where the money will be made.
01:03:56
Speaker 1: As an investor, and as.
01:03:58
Speaker 3: Someone who speaks to investors, what is the thinking now about buying companies?
01:04:04
Speaker 1: You know, there's this meme.
01:04:06
Speaker 3: Gavin Baker said this on a podcast. He said, well, Dario thinks there's only going to be one company left in the world, and it'll be Anthropic.
01:04:13
Speaker 1: And it'll be Anthropic and the U.S. government.
01:04:15
Speaker 3: And I'm just curious— That seems far-fetched, but it does not seem far-fetched that Anthropic is going to get into more verticals directly. So I'm curious from your perspective how that affects your thinking when you're writing checks, this possibility that you might just be competing with products from the major frontier labs.
01:04:33
Speaker 2: No, it's 100% the problem right now is that you– it's like– The analogy I make is it's like the dark days of Microsoft, whenever Microsoft, if it saw you selling something on its platform and thought that seems to be generating decent revenues, it would build one itself and then integrate it into the operating system. It famously threatened to shut off Netscape's oxygen supply because it was going to launch Internet Explorer back in the day. This is that same phenomenon at a vastly larger scale. The problem, of course, is that... Anthropic is only as good as the data it gets as it enters these verticals. And what's happened is it's much like In Dr. Strangelove, don't tell people about the bomb before you build the bomb because the problem is now they've told people that we want to be in your vertical. And so as I talk to people more and more, there is this new reticence to share all of this data with Anthropic or even to partner in as many ways because of this fear that they're climbing up the boat like in Jaws or whatever and they're going to eat everything as they go up it. They can't
do this whole clock. They have to have everybody's data.
01:05:36
Speaker 1: A lot of anxiety there.
01:05:37
Speaker 3: Paul Kudrowski, so great to catch up with you again. We'll be watching today. It's Wednesday, but on the day this comes out, we'll be watching. We're finally going to have one out there. Really appreciate you coming on Outlaw.
01:05:51
Speaker 2: Cool, guys. Talk to you. Have a good one.
01:06:04
Speaker 3: That was fun. It's such an exciting week. There's so much news. I'm really excited to see where Anthropic trades.
01:06:10
Speaker 6: Yeah, I mean, well, I guess by the time this comes out, we might have some clues.
01:06:15
Speaker 4: But I do think the credit crunch in AI....
01:06:20
Speaker 6: I mean, it is a credit crunch. That's what we're seeing is probably the biggest story out here. If you think of AI as a financing story, then great, Anthropic is raising some equity. But there's a whole bunch of other players out there who are finding it more difficult to raise capital.
01:06:35
Speaker 3: I find this metric of revenue to compute to be very interesting. Look, a lot of these companies are very far from turning capital. profits, let alone gap profits or anything like those. But in the meantime, one metric that we should watch for both the industry as an aggregate versus or any company is how much of their compute is being purchased from equity versus how much is coming from revenue. And we know that the better, you know, the more successful companies like the anthropics of the world are plowing a lot of their revenue back into compute. But that's healthier than if they were, you know, needing to tap the equity market in aggressive ways just to get access to chip.
01:07:19
Speaker 2: Yeah.
01:07:19
Speaker 6: But at some point you need someone to pay for all the compute that they're providing.
01:07:24
Speaker 3: And you need to not, as Paul said, you also somehow need to get off the treadmill where in addition to all your money is going to training for the next model.
01:07:35
Speaker 1: Otherwise, where's the cash flow?
01:07:36
Speaker 6: That's the real problem. I think like there's no out. in terms of development, right? Especially for the frontier hyperscalers, especially.
01:07:44
Speaker 3: And this was the conversation that we had last week with Tian Yang earlier this week, which is like it's a race with no finish line. It's a Sisyphean marathon.
01:07:54
Speaker 1: All right.
01:07:55
Speaker 4: Shall we leave it there?
01:07:55
Speaker 5: Let's leave it there.
01:07:56
Speaker 4: This has been another episode of the All Thoughts Podcast. I'm Tracy Allaway. You can follow me at Tracy Allaway.
01:08:02
Speaker 1: And I'm Joe Wiesenthal. You can follow me at The Stalwart.
01:08:05
Speaker 3: Follow both of our guests, Stanley Lee. He's at StanleyLee21 and Paul Kudrowski at P.
01:08:11
Speaker 1: Kudrowski.
01:08:12
Speaker 3: Follow our producers, Carmen Rodriguez at Carmen Arman, Dashiell Bennett at Dashbot, Cale Brooks at Cale Brooks, and Kevin Lozano at Kevin Lloyd Lozano. And of course, you can chat about all of these topics 24-7 in our Discord, discord.gg slash oddlots. And for more OddLots content, read our newsletter, which you can find at Bloomberg.com slash OddLots.
01:08:34
Speaker 6: And if you enjoyed this conversation, if you like it when we dig into AI financing, then please leave us a positive review on your favorite podcast platform. If you're watching on video, then please like or leave a comment or better yet, subscribe.
01:08:47
Speaker 3: And if you're a Bloomberg subscriber, you can listen to all of our episodes absolutely ad-free. Just go on over the Bloomberg channel on Apple Podcasts and follow the instructions there.
01:08:57
Speaker 4: Thanks for joining.
01:09:21
Speaker 5: Thank you.