Speaker 1a lot of mainstream financial media outlets have concluded that maybe the AI build-out is approaching its end. So if you were to judge this or assess this currently, are we early or are we
Speaker 2late? So quickly, before we get to today's show, this episode is brought to you by Pith. Pith Network is the fastest growing financial data company today. Trusted by Fidelity Investments, the US Department of Commerce, Revolut, Kalshi, Jane Street, Coinbase and many more. They provide real-time financial market data across over 3,000 equities, commodities, crypto, FX, rates and again a lot more. They're the first and largest provider of 24-7 financial indices and offer the widest array of market data for the lowest cost. So go to pith.network to take advantage of their free trial. Join me, Raoul Pal, as I go on a journey of discovery through the macro, crypto and exponential age landscapes. In The Journeyman, I talk to the smartest people in the world so we can all become smarter together.
Speaker 3Hey everyone, hope you're all well out there. It's great to see you again. It is time for another quarterly dose of Exponentialist meets RV Pro. It's time for another conversation between me and Andreas. There's been so much going on. We're going to dig into a whole tonne of things. So let's get started. We've got five or six big topics to talk about and I just can't wait to get into this. It's always a great conversation. Andreas, how are you?
Speaker 1I'm good. My portfolio suffers a little bit, to be honest, these days, but let's see whether we get a rebound into the end of the year. How are you feeling about things right now, David? Because it kind of feels like everyone is in Camp Doom again, related to AI, right?
Speaker 3It really does. Yeah. And if I'm smiling when you're saying that your portfolio is not feeling so great, it's only because I'm not feeling so great. It's only because mine is not feeling so great either. And it's nice to know I'm not alone. Yeah, it's a strange old time. People are very much in Camp Doom. There's a lot of nerves around. I mean, we're going to get into all of that. There's a lot of weirdness around and I'm super keen to just, I mean, of course, we're all reading on the platform how you're thinking it through, but to dig even a layer deeper on how you're thinking all this through. Because yeah, it's a weird old time right now when it comes to technology and marketing.
Speaker 1Hmm. So David, you've obviously been watching technology cycles for decades. And we're sending live during a week where I guess a trillion has come off memory makers. We've seen Anthropix prospectus leaked with a, I think it was an 81 billion loss. Some of it was non-cash, actually a big part of it. But anyway, it made for very, very gloomy headlines. It seems like a lot of mainstream financial media outlets have concluded that maybe the AI build out is approaching its end.
Speaker 3Yeah. I mean, and this is typical of the way the media approaches this story and the way markets to a certain extent approach this kind of story. A somewhat hysterical and very sort of short-term approach, if I may say, where it's either sort of the only thing that's happening and it's going to happen to the next degree or it's over and it's dead and it was always just hype and it was a fad. I've written a lot about the two ways of thinking through this big story. There's the story that markets can tell. There's the story that's in markets. That's one story. And then there's the story that is the story we talk a lot about inside The Exponentialist of this epic civilizational shift to a new kind of economy, a new kind of civilization, intelligence is abundant. We're very, very early in that second story. And that's the fundamental point here. In terms of the first story, the story that markets can tell, the AI build-out that is priced. Look, yeah, I think it's true. I mean, it gets obviously incredibly complex, but it does feel that we're moving to a new stage of that story. I mean, what happened last week was a lot about kind of stressed financials, you know, core weave, these neo-cloud builders borrowing huge amounts of money from banks and bond investors to fund new data centers, to sell to Meta, sell compute to Meta and open AI and Microsoft and all this stuff. You saw what happened with core weave last week. Like it's bonds fell to sort of distressed levels. It fell 15% in like a day. The market is more and more and more nervous about the borrowing that is fueling the build-out. And so it feels to me that we're moving to a new stage in the story that's legible to markets, where it's not going to be the case that just everything to do with AI all just goes up at once. And you have to kind of position yourself for the next stage, which perhaps is a rotation into the users of intelligence like pharma and robotics and other use cases. That's my kind of first draft thinking on kind of where we are in the cycle. I don't think we're anywhere near like ultimate bubble levels or, you know, the bubble is about to burst. I can't see how you make sense of that. But the market is incredibly tetchy. It overreacts to the slightest bit of news, like it did with this open AI revenue news, even though it turned out that it was mainly, you know, mainly the product of accounting processes, you know, open AI had said, or people had talked about 70 billion, then open AI said 50 billion. It turned out that this was mainly a matter of how you do the counting, but the market didn't care. And that helped spark what happened last week. So it's so tetchy. It's so nervous. It overreacts to tiny bits of news. And there could be a, there could be an event. There could be a correction of some kind. If you get another piece of news that, that triggers a cascade and we're moving to a new stage where you have to position for that new stage. But I don't think this is a bubble. And I don't think the bubble is about to burst. You know, Howard Marks said something brilliant. I think it was last week. If you, you know, if you, if everyone's talking about, are we in a bubble, then we're not in a bubble because one of the hallmarks of a bubble is absolute disregard of risk. And people aren't there. So I don't think even the sort of market sentiment supports the idea that we're in a bubble, but yeah, I mean, super keen to hear. How you are, does that align with what you're thinking and how you're thinking all this through?
Speaker 1I think it's a tremendous quote from, from Howard Marks on, on the bubble stuff. I've also said that we have a bubble in people calling it bubbles. Yeah, that's, I think that's more accurate than, than us being in a bubble. But having said that, David, I think you're onto something when you refer to the funding stress that we've seen among the neoclouds. First and foremost, as I've always said in my time in financial markets, you cannot argue with the bond market. The bond market is basically always right. Because if you're currently faced with high financing costs, it's your job to convince the bond market that they should provide you with a Obviously, a huge competition for capital, also given that we have massive fiscal deficits in the US and elsewhere. But there is also a need for the new clouds and, more importantly, maybe the max sevens to come out with a compelling message to bond markets now. Why do we do all of this CapEx? The first glimpse of this, we basically got that from Meta yesterday. I think a lot of people had anticipated that they would scale down CapEx. They did not. They upped the range for CapEx. This is also what I'll expect ahead of the reporting season from Alphabet and Amazon and Microsoft next week. I think the executives are still very upbeat. On the prospects of building more. The question is whether they come up with a sufficiently compelling message to calm down funding markets, because that's very needed now. Even though the income side of the build out is still doing well, I mean, we still see rising prices of everything related to compute capacity.
Speaker 3Right, GPU prices are still, GPU rent is still rising. So, there's no indication that demand is, that there's oversupply.
Speaker 1No, exactly. And I'd actually like to show a chart, which I think is a mind blowing chart. When. We discussed this whole AI build-out in the context of the dot-com, which is a hot topic, to say the least. On page 20 in our slide deck, you'll get the slide deck after this to all of you watching. I've just made a simple comparison between the dot-com and what we see right now. So on the right-hand side of the chart here, you have the price of bandwidth between year 2000 and year 2003, and we got a 90% drawdown in the price of bandwidth. We could obviously get there with the price of compute, but we're nowhere near anything that resembles the dot-com. And more importantly, in my opinion, we have the utilization on the left-hand scale. The utilization was very, very low of the whole fiber build-out to begin with during the dot-com bubble, which is by no means the case now. Even Anthropic explicitly state in their… IPO prospectus that they struggle to get the capacity they need. They've gotten no's from several of the max sevens because they cannot deliver, and they say that it's the biggest issue for their business right now. So I've been very vocal that I think the issue for Anthropic since June is basically that they've been capacity constrained, and that's why OpenAI has taken a lot of markets here since. So I don't see any signs that we're… We're approaching sort of an abundant data center capacity yet. And I would need to see more clear signs of that to conclude that the build-out has gone too far. But I don't know whether that rhymes with the observations that you're making, David.
Speaker 3I think it does, yeah. I mean, you know, I can't get past the fact that GPU rents are still rising. So you don't have… You don't… That's the opposite of what you would expect, obviously, if there was oversupply right now. And exactly… As you say, I think the dot-com analogy that everyone is so fond of, you know, the primary thing for me is they built tons of those fiber cables, and tons of them sat unused for years, which, again, is the opposite. Like, no one… You know, people can't get their hands on GPUs, so the price is going up and up and up still. So that analogy doesn't work. And, I mean, I think we'll talk a little bit more if we get the time about historical analogies later. This is the era. This is also a bubble. And historical analogies, what is the right analogy for what is happening now, which is, you know, as essentially a historian of technology, it's a question I love. My concern is, you know, if I try to be pessimistic, do you get… People are worried that the weakest borrowers, like CoreWeave and the NeoClouds, are too stressed, that they're not going to be able to repay. And do you, you know… Could you imagine… Could you imagine a scenario where you get kind of some… You know, you get a piece of news or an event that triggers that worry again, and then you get kind of contagion where people start to worry about the hyperscalers, and you get a serious correction. Is that a… My question to you, I suppose, is, is that a viable scenario? I suppose it's one that plays on my mind. I'm not deep enough in markets to think through properly, I don't think, whether that is a viable scenario. And the final thing I would add is… Yeah. This is another… This is a big second leg of the question, but… And then what happens if, on top of those kind of worries, rates rise again? You know, we just had an inflation print that was higher than people thought it would be. Surely, then, the chance of another hike is higher now. What does that do to this whole situation, given that what fuelled last week is concern about debt and the financial…
Speaker 1I think this is a clear risk, and I think it is a risk scenario that becomes particularly worrying if the earnings-slash-revenue cycle rolls over amidst this. So I'd like to show a chart on page four, which is one of the oldest earnings indicators among Wall Street strategists. So you have… You have the exports from Taiwan and Korea in light blue, year-over-year change, and then you have the earnings in the U.S. equity market 12 months forward in dark blue. And the business cycle and the earnings cycle is very correlated to the exports from some of the open economies in Asia. I think that makes sense in many ways, since both Korea and Taiwan are very dependent on the overall growth momentum of the world. And since they have small open economies, they are very dependent on the overall growth momentum of the world. economies, very export-oriented, they obviously feel the trends earlier than others. What we see still from Taiwan and Korea is an extreme acceleration in the exports, obviously very semiconductor-based. We're basically talking breakout high export growth across the board. My best guess, given that this is a correlation that has held true for more than a couple of decades, is that we're still far from the earnings peak. We also saw how SpaceX delivered a tremendous earnings report yesterday. I expect the hyposcalers to deliver on the earnings side of the equation next week. That should alleviate some of the worst concerns around their balance sheet. As long as they make money, who cares about the funding side? To put it a little bit blunt, right? If they fail to deliver on expectations. Yeah. then run for the hills. I think it's as simple as that.
Speaker 3That's the trigger, yeah. If the hyposcalers' revenue starts to look dodgy, that's what could trigger that contagion that I'm talking about out of neoclouds like Corweave and then into the hyposcalers, and then you would feel like markets really would panic massively and everything would be in trouble. You'd see the kind of thing. my sort of bellwether kind of scenario is, what triggers a massive drop in NVIDIA, and something like that maybe could. Yeah. Right? But what you're showing with this chart is we would expect hyposcaler revenue to. it's nowhere near the top yet, so it's unlikely that that is going to happen, and that again supports our general vibe here that we're not near the end of this cycle, we're getting into the complex middle of it. It's not as simple as it was at the beginning. Yeah. And there's kind of rotations to position yourself for inside this broader story, but we're not near the end of the story yet, I think is what we believe.
Speaker 1But at the same time, I think it's absolutely fair to acknowledge the risk from funding markets now. I think that's something more fundamental than the earlier AI scares that we've seen, right? The deep seek moment and such stories, they never really had a fundamental layer to them, in my opinion. But this one does, because the bond market is basically screaming right now, and you should always take notice of that. So it's a big ask that I have for the executives from Alphabet and Microsoft next week, that they need to calm down the funding markets because they're.
Speaker 3They need to calm down. Yeah. But obviously- And another hike just makes all this more acute, does it?
Speaker 1Yes. Absolutely. Yeah. Of course, we could. you know, if we get some benign inflation, we're pretty sure that's what's going to happen. Yeah. If we get some benign inflation reports and stuff like that, we could get some external help for this story, right? But it doesn't really seem like it's around the corner, right? So they'll have to say something clever next week on everything funding-related. Because as said, I don't think they'll scale down their capex yet. That's not my impression, but David, I'd like to ask you to unpack this thesis that we should focus a little bit more on the users of AI. I know that you consult with many companies, you spare with executives on the AI implementation cycle and all of that. So when you look at the implementation of AI among users, are we early or late? That's maybe another question than discussing the whole build-up, right?
Speaker 3Yeah. It is. It is another question. Yeah. It's a great question. I mean, if we're talking about the implementation of AI into the enterprise, we're really early. And this taps into kind of themes that we can dig into. It takes much longer than almost everyone thinks for a general purpose technology like AI to collide with millions of real people. real people in the workplace and for those real people to find ways to make that useful and And so I think that's a really good question. unfortunately, the type of people you often hear talking about this are people with very little direct experience of what that looks like. And so their visions of it are incredibly naive. And the first set of visions we heard were from the AI overlords themselves, who are a weird bunch of people who are very young and have only ever worked in startups. That's number one. And number two, overridingly, they just want to hype their technology and say how powerful it is and then it's going to sweep everything away and be amazing so you know the the kind of forecasts you you heard from them two years ago three years ago up until fairly recently you know there's not going to be any knowledge work jobs left in three years like everyone's going to go all entry-level knowledge work will be gone in 24 months um even sam altman himself has said much more recently those kinds of forecasts i talked about two years ago they're clearly not playing out and they you know anyone who knows anything about working in an office right with a with a whole team of colleagues knows that they were never going to play out that way it takes years for these kinds of things to play through you know for jane in accounts to figure out how to use ai make that useful for steve over there to also cooperate with that and bob and you know alex and all this stuff so we're super early in in enterprise use of AI and we're super early in enterprise use of AI and we're super early in enterprise use of of AI that's going to take time and that's why i in talking about historical analogies that's why i think that you have to go all the way back to electrification probably to find the real analogy for this because that was a general purpose technology that took a long time to build out and there was a big productivity lag but obviously the the the ultimate historic result on productivity was like immense and i think that's why i think that's why i think that's why i think it's immense immense you know intelligence is going to be a lot like that it's much more like that than it is like fiber cables which are just one thing um so yeah that that you know it helps you think through the analogies too when you when you think about how intelligence is is colliding
Speaker 1with workplaces david speaking of these crazy forecasts from the ai executives we haven't had the chance the two of us to discuss uh darryl amadei's block called pace the frontier um it's probably more than a month old now right but he he basically said that there is a risk within 6 to 12 months of uh ai swarms taking over the internet and you know there is even a risk that ai could erase the human race and whatever he wrote in that block right i personally found that to be almost lobbyism for uh regulation around the whole ai space right but how did you read that block was it more of a pitch for regulation or something that he actually found to be a true risk um it is
Speaker 3a pitch for it is a pitch for regulation you know i mean and we have to remember that both things can be true you know some yeah a set a set of statements can be true but that doesn't mean that dario or sam are going to say those statements just because they're true they will say things that are in their interest to say so the reason they're saying these things is because it's in their interest to say them and they want to kind of co-design a regulatory framework that bakes in the lead they have that makes it hard for for insurgents to come along and kind of you know eat their lunch and all the rest of it um at the end of the day i think it's the same time uh what they're saying has some truth to it i think you know i mean dario's trying to build a case in that essay so he probably overstates the case somewhat he does overstate the case somewhat i think um but it's not insane to worry about swarms of intelligent agents getting out into the internet and like causing havoc um i mean i said i think i talked about this chat a couple of months ago with rao you know this i you know obviously the the really eye-catching scenario is going to turn us all into paperclips end of humanity you know global nuclear war you know it's no it's like it's unfathomably unlikely but can swarms of intelligent agents get out on online and sort of mess things up like yes i think they can like they've been pretty close to doing that already um and should we have some some lines around that and could it get a bit you know serious or they're genuine like you know risks to human life like yeah i think there are and we should we should draw some lines around that but of course you know everything that dario and sam says is about upcoming ipos and and positioning themselves against one another and against the insurgents and all the rest
Speaker 1of it yeah and speaking of regulation of of the whole ai space a couple of days ago i think the first chinese mainstream large language model was excluded from the um u.s civilian government systems not really a surprise is it david that we're going to see stuff like that being implemented as well
Speaker 3no i i think it was it always kind of was on the cards it always kind of had to happen and this is just the start of the u.s and china building kind of rival um rival ai stacks you know i mean both want end-to-end control of the entire stack they want sovereign ai they want control of their own intelligence because geopolit you know the world we're heading towards geopolitically that is just of enormous importance so i mean i think what this is this these new kind of rules say that sort of federal agencies shouldn't use deep seek right um and yeah i mean how i i can't see how you can i can't see how it could be otherwise like you can't have the risk of like sensitive data going where it shouldn't go um and you want to you want to promote homegrown ai so david ahead of this
Speaker 1expo meets rv pro edition we we basically chatted back and forth on how to view the chinese ai supply chain versus the u.s ai supply chain i actually think that's a very very interesting investment topic as well right um so i i looked a bit at the market shares and maybe we can show page 11 for the audience since it is actually a very mixed bag of goodies when you look at the chinese market share of relevant hardware to begin with we can discuss so far in a second but if you look at hardware china is actually a majority uh marketplace in optical transceivers uh this is by the way uh to the best of of my understanding an area that is also under discussion within the u.s administration we've seen a couple of reuters stories on it whether to ban imports of of those exact um optical transceivers from china but if you look at for example high bandwidth memory they basically have no footprint at all as of now china they do have a footprint within the memory space um albeit a neglectable one so in d-ram space they have a little less than 10 percent uh market shares and in the nand uh storage space they have a little less than 20 percent so the relevant names here if you look at at the chinese equities to begin with in d-ram space it's obviously cmtx um and nand space it's yamze i think it's pronounced um so they have a couple of pretty sizable players within the ai hardware build out but if you really want to bet on the the ai supply chain being de-sinoed if that's a word i mean if you want to bet on the u.s administration blocking all chinese hardware for example to the u.s you'd really need to bet on the fiber optics companies in the u.s because they will get a big advantage from such an import um blockade with with china being the biggest player right now in that market uh while it's less relevant to the in for example the memory space but how do you view this david i mean it it does this really kind of compare to you know the whole race to get to space and stuff like that where we we simply have two superpowers competing against each other which ultimately means that we'll have two completely separate ai supply
Speaker 3chains here yeah it's i think that at the moment it would appear that that's the direction it will take i can't see how china and the u.s would want their their ai systems kind of interwoven with one another and dependent upon one another that just feels tremendously risky so yeah the space race in that way is is another good historical analogy i think you'll see two competing build downs and two races you know towards that i mean you see china now putting restrictions on set on some on sale of some of its own technologies or export of some of its own technologies including like the the frontier you know models the weights associated with those like it doesn't want its own frontier models being exported around the world um so yeah i think i think you're going to see two two separate build outs so is there an opportunity in sort of finding the kind of nvidia you know the who who's going to do the frontier ai chips in china who's going to do that kind of work in china and trying to bet on them you know i don't i know so little really about and i want to talk on this on on on the names to focus on in china so this is absolutely fascinating to me
Speaker 2so a quick break in your regular if you're serious about your future grab my free report called prepare for 2030 i think you've got five years to make as much money as possible and this guide will help you navigate what's coming the link is in the description download
Speaker 1it now it's it's definitely a um a topic that i'll try to exploit further in in coming weeks as well especially since we got this deep seek news uh this week because it's probably the first of many of such announcement i i would have to agree with that yeah and
Speaker 3you see i mean you see more and more u.s um companies or the evidence is more and more u.s companies pushing their enterprise spending into deep seek because it's cheaper you know it's cheaper and it's good enough so deep seek you know the deep seek is a genuine competitive threat to open ai and anthropic and how they're going to deal with that all of these questions are live you "and david
Speaker 1speaking of the quality of models um open ai has seen pretty decent momentum since the astro model was released uh it was by the way great marketing campaign they had around that whole release in my opinion it really made the rounds everywhere what's your take i mean i know you've been a fan of of fabled but do you find astra to to create more value for you what's your take on the recent developments on the software side of things yeah
Speaker 3it's been really interesting to watch because i think as you've said um on the platform before i mean everyone around christmas which feels like an eon ago now when it was only really i don't know excuse me it was only like what 10 months ago it's really nothing but it feels so long ago the momentum was so strong with anthropic and i think it was opus 4.6 at that time i mean i remember kind of a lighting on 4.6 over christmas and just being blown away like it just it felt like such a step change it was so nuanced the writing was so good uh and the the momentum was really with anthropic and then they were releasing all these like design tools and plug-ins and the markets kept moving in response to those i remember you talking about that they had all the momentum we've really felt the momentum shift to open ai or in my my sense is the momentum has shift to open ai astra is a great model um but the the core point here and this taps back to um you know i mean in terms of the benchmarks that fable versus astra like it's a toss-up like i think astra is like supposed to be um better at computer use you know fable is better at like long chain reasoning and it's better at like long coding projects it's a toss-up in terms of the benchmarks but the point really and this taps back to what i said about deep seeking it being good enough is that for the ordinary user now for not in fact for 99 percent of users um these models are much of a muchness like very very few users are touching the edge of these models you know what i mean are reaching the capability limits of these models you're talking about models that can solve like insanely difficult mathematics problems that hardly any of us have any understanding of you know we're not reaching the limits of of these models in our in our daily use enterprise is not reaching the limits of them in its daily use because enterprise is like a ton of consultants at accenture using it for their research and all this kind of stuff that's not reaching the edge of these models that's why they're switching out to deep seek um so and this taps back to something i talk about a lot which is yeah essentially you know intelligence good enough intelligence is bad being commoditized the kind of intelligence that 99 of people use is being is becoming a commodity there's a huge choice of models and that's a genuine including free ones that are open source and that to me is the deepest structural challenge for open ai and for anthropic is what they produce the the fundamental product they produce intelligence is being commoditized and so why does anyone you know you get a race to the bottom on prices and why does anyone choose them in particular or stay with them and the answer is going to have to be for them in what i call like ai magic not the intelligence itself but features and context that makes that ai useful in your day-to-day life and magical to use and that's why you see openai doing dot you know the agent that's why you see anthropic doing all kinds of fun things and it pays such attention to its ux you know that's why meta muse is a you know the new agent from meta is so exciting to people because just raw intelligence is now a commodity pretty much so what kind of magical experience are you going to serve me that i think is the quest open ai and anthropic have to go on to win
Speaker 1yeah and i in my opinion astra and everything related to the uh also the campaign around astra was a big step forward uh on ux for open ai uh which is a very very important um point you made there anthropic basically won that ux battle uh last year and and during the first half of this year uh but open ai is is um is fighting back uh yeah they're
Speaker 3fighting back because they realized that that's that's where this race has to go you know um i mean you know i mean look at a tiny signal of this right fable is the frontier model but i almost overwhelmingly use opus 5.5 you know because i don't need the extra intelligence that fable is bringing you know unless unless i might use fable if i'm asking it to kind of like do some long two-day project to build me some kind of platform or dashboard you know which i've done a few times but in just in day-to-day research i'm not getting anywhere near using the extra that fable brings me and actually fable is a bit verbose and convoluted in its writing so like it's so intelligent it's hard to understand sometimes basically so i've gone i've had to go back you know down to opus 5.5 which works really well for me that's where we're at now and open ai realized that yeah they need to produce magical experiences they need reasons to lock you in as a consumer to get you to stay that's what dot is all about that's hardware devices are all about they're about giving the models like context so it's they're listening to you all the time learning about your life seeing your physical environment and then it will have years of context on you so then you're not going to leave go to another model because this model knows all about you and it's known about you for years that's what they're
Speaker 1trying to do and maybe david this is the exact reason why dario amadei samuel and elon musk can all agree that they would love to see some regulation because yeah regulation of ai would in turn become some sort of a soft flaw under the token pricing in my opinion that's at least the practical outcome of fierce regulation and that may be needed given what we're seeing from uh open source competitors and so on and so forth because that the only worrying trend for me right now as an ai investor especially since i'm very heavily in many of these hardware names is that we see the average price of a token falling it doesn't worry me fundamentally speaking from a macro perspective but it worries me as an ai investor because if the token pricing converges towards zero how on earth are we going to pay for all of the chips there's something there's a missing link there unless you actually ensure that there is a soft flaw under the token pricing it it's not worrying yet i mean we we got a glimpse of of anthropics numbers yesterday that wall street journal league from their ipo perspectives i actually think we can bring up the numbers on the screen here from page eight in the slide deck um if we set aside that massive uh net loss from a fair value change of their conferred a convertible note which is basically just giving those early investors more shares when the ipo is done i i'd actually argue that they've managed to outpace their compute costs quite clearly over the past 12 months which is comforting i'd say um if you look at the revenue as a whole you're going to see that that's going to be a little bit different than what we've seen over the past 12 months which is comforting i'd say um if you look at the revenue as a whole if you look at the revenue as a whole that's going to be a little bit different than what we've seen over the past 12 months um if you look at the revenue as a whole that's going to be a little bit different than what we've seen over the past 12 months um if you look at the revenue as a whole that's going to be a little bit different um if you look at the revenue side it's up from 1.1 billion during um h1 in 25 to 16 billion roughly in h126 in h126 in h126 and yeah we're basically talking about 15x and yeah we're basically talking about 15x and yeah we're basically talking about 15x something like that while compute costs are something like that while compute costs are something like that while compute costs are up four to five x i think that up four to five x i think that up four to five x i think that that's that's decent i mean and much that's that's decent i mean and much that's that's decent i mean and much better than feared in many ways so sure better than feared in many ways so sure better than feared in many ways so sure the token pricing has the token pricing has the token pricing has been on a slippery slope but they're been on a slippery slope but they're been on a slippery slope but they're still able to make money on their still able to make money on their still able to make money on their compute it seems
Speaker 3compute it seems compute it seems yes exactly exactly but i do i do know yes exactly exactly but i do i do know yes exactly exactly but i do i do know what you mean about that concern you what you mean about that concern you what you mean about that concern you know especially for open ai and know especially for open ai and know especially for open ai and anthropic anthropic anthropic there's just an underlying concern here for them that the ultimate value of this whole ai story doesn't in the end go to the makers of the intelligence it will go to the people who have massive distribution and can use that intelligence to serve those people magical experiences and then probably like sell advertising against those experiences or whatever it is you know like it like meta you know meta can reach billions of people with facebook and whatsapp and instagram and now it has this muse agent that does all these kinds of useful things for you and you can talk to it in whatsapp and all the rest of it where does that leave open ai and anthropic this there will still be a business to be had from being at the very frontier there will still be customers for the very best models and the very best intelligence but it's not a business of the size that people are perhaps reckoning on now if they believe that open ai and anthropic are going to be like the meta and the alphabet or the google you know what i mean of the future because what if meta and google are the meta and google of the future doing all this wonderful a stuff ai stuff for people where does that leave open ai and anthropic you know because as you say the cost of just a token just raw intelligence when you have deep seek and a bunch of chinese open source models doing it for free and it's good enough for almost everyone that the cost of a token falls to zero so on they're on this mad race to lock people in with magical experiences of their own to get like billions of users and lock them in um and look they're doing incredibly well like they they're you know they they have i think open ai has got billions of users right but it's that's a very difficult path to tread and they're still trying to figure it out and open ai you know talked about this big device with johnny ive and everything which is clearly part of the kind of magical experience but where is the device yeah we've got to figure out where is the device where is the device we still have not heard anything about it
Speaker 1a practical example from my own life and my own business setup i'm a mostly a client of anthropic and all of the skills all of the automated setups that we've now created across by um my business units it would take me some time to move all of that to another setup uh especially since i find claude to be the most you know user-friendly to to to work uh with am i anywhere near my pain point um price-wise when i look at my bill from anthropic no not really i think um so i think they've managed via the ux part of it to create some stickiness already um so i think they've managed via the ux part of it to create some stickiness already um so i think it is doable to make a business case even though i perfectly agree with you that it doesn't necessarily matter from a business perspective to be at the very frontier um the question is how sticky they can make this product before it's
Speaker 3too late yeah exactly exactly how sticky can they make it before it's too late is exactly the question and yeah look i agree i mean it would be a pain for me to move out of anthropic you know i've got all these projects set up and all this context and i can just jump in and interact with it and you know i think it's it's really important for me to know what i'm talking about and it knows my different work streams and you know it knows you know the essay i wrote six months ago so yeah they've already done you know anthropic have already done a really good job of building some lock-in um but you can feel people like meta coming along right and i can i can also feel that sometimes now i mean these are tiny anecdotes but like if i'm in the google window and i just have a general ai type of query i'll now just type it into google right and i wasn't doing just to just to use their ai mode and i wasn't doing that a year ago because i wanted to use anthropic or open or you know or chat gpt because it was better and now it's like yeah the google one's fine for this kind of question so just tiny little things like that it's going to be very interesting it's going to be very interesting i mean are you when when anthropic and open ai ipo are you immediate an immediate buyer of that
Speaker 1probably into what's saying no to both um i've had the view for a while that open ai is better equipped to deal with the compute capacity part of the equation i think anthropic will have to ramp up their purchases um i don't know whether you have the same experience but claude is on average substantially slower than its competitors um and i think that's a good thing i think that's a good thing i think that's a good um yeah because of a lack of data set of capacity for inference um so i i think they have to do something about that again i'm not close to my pain point in terms of i mean i could just do something else while it's working right but um it still annoys me a little bit that it's slower than its peers um and it is substantially slower i've tested like equal tasks many times and it is just slower um so we really have you know a a setup where the two major players that are going to ipo they've taken completely different routes in terms of their compute capacity and their you know forward commitments to those but it seems like it's open ai that has the upper hand on that question right now so i'd lean towards open ai just for that um just because of
Speaker 3that for now yeah yeah yeah and you know in the meantime obviously they're doing incredible things i mean look at what open ai did with the 700 you know maths results that it just dropped like a nuclear bomb on maths um a few weeks ago i mean you know they're doing incredible things at the frontier of intelligence um so yeah i mean huge kudos to to what's been achieved inside those companies it is absolutely incredible yeah
Speaker 1david i just wanted to show a chart on the rental prices of the gpus because we've touched upon it a couple of times uh during the show already on page 15 we have the live data uh on b200s the black wells and also the older h100s so both the old and the new gpus are still um going up in in terms of rental prices which is a huge surprise i'd say um and one thing that i find extra interesting here is that you know i've i've been through basically all of the big investment banks research offerings on on ai uh i've asked for the assumptions built into their discounted cash flow models uh when they uh value the um the bigger tech companies and i've not found one investment banks not assuming falling gpu rental prices over time they all assume that which is probably also the right assumption so i'm not sure if that's true or not but i think it's true i think it's true and i think it's true and i think it's true and i think it's true and i think it's true and i think it's true say 10 years down the road but as long as they tick up i'd argue that the market is a little bit too worried about the financing side of the equation since the you know the asset side of the equation is also going up if if we had falling gpu prices and rising interest rates that would be a very very very nasty cocktail for a neo cloud for example but still a neo cloud could lock in compute capacity forward with one of the bigger players at 40 50 higher prices than they could first of jam if we look at the current spot price developments that is absolutely bizarre in many ways that we're up 40 50 in rental prices on gpus just since first of jam so i yeah that is that is crazy yes statistic yeah and and if you compare to the interest rate by well the bond market right david interest rates are probably up 40 50 for many of these companies over the same time span and they do not fund the whole build up with debt so i'd actually argue net net they're better off than first of jam if you look at income versus costs because of the income side doing as well or even better than the cost side and that's probably the part that is neglected ahead of the um max 7 reporting season next week because as long as this is the case if your income side is doing better than your cost side even though there is clearly stress around your funding side isn't the logical conclusion to go ahead and do more capex i'm not sure i mean that would be my best guess yeah i
Speaker 3mean it's it's it's hard to walk you with that isn't it and it's just it's it's just you know people systematically underestimate the amount of compute and the amount of intelligence that we're going to use or that there is a use for you know because the market and the mainstream media just has a very limited imagination in terms of in terms of the uses of intelligence and they see as this incredibly new exciting sort of rarefied thing with a very sort of narrow set of uses but it's not that i'm going to say that it's not that i'm going to say that it's not that i'm going to say that it's not that i'm going to say that it's not going to be like that it's just it's going to be abundant it will flow through everything and there will be all kinds of uses of it i mean look again look at meta music news you know and this taps to a point that rao makes a lot which is very important you know the biggest user of ai is really yet to get started or is perhaps just getting started now it's other ai agents you know ai agents are going to make lots of use of ai themselves um so that whole story is yet to play out you know and that will probably be the vast majority of use of compute eventually so there's why there's a huge road left ahead in terms of the amount of compute in use and the amount of things we that intelligence is doing in the world around us you know think about when there's all these autonomous vehicles driving around and all these humanoid robots and other next generation robots the the compute is is immense so yeah this idea that this compute won't be used or we're building too much at the moment is just is is kind of crazy really but that's not to say that the market at some point won't get so scared you know and sentiment won't turn so hard that there'll be some kind of big correction that's scenario that plays on my mind you know i'm very confident in the long-term secular trend of where this goes technologically but how market sentiment sentiment changes and it feels so on a pinhead right now is much harder
Speaker 1to call yeah i've had a couple of very uh interesting and very practical um almost mind-blowing experience experiences with my token usage just over the past couple of months i have a bookkeeping system and i have an accounting system both of them are you know i've had a bookkeeping system and i've had a bookkeeping system both of them are you know i've had a bookkeeping system and i have an accounting system both of them are you know i've had a bookkeeping system and i have an accounting system both of them are you know i've had a bookkeeping system both of them are you know i have to have them you know i have to have them you know i have to have them due to the law in denmark because i need due to the law in denmark because i need due to the law in denmark because i need to be able to have a i need to show to be able to have a i need to show to be able to have a i need to show all of my receipts digitally to the um all of my receipts digitally to the um all of my receipts digitally to the um tax authorities and all of that so it is tax authorities and all of that so it is tax authorities and all of that so it is basically all everything is digitized and basically all everything is digitized and basically all everything is digitized and in um in accounting and bookkeeping in in accounting and bookkeeping in denmark now denmark now denmark now not until not until not until was it six or seven weeks ago was it six or seven weeks ago was it six or seven weeks ago um um um it couldn't talk it couldn't talk it couldn't talk to claude but now they've opened up to claude but now they've opened up to claude but now they've opened up claude for these bookkeeping and accounting claude for these bookkeeping and accounting claude for these bookkeeping and accounting systems which basically means that systems which basically means that systems which basically means that my token usage has gone up my token usage has gone up my token usage has gone up materially just because of that because materially just because of that because materially just because of that because it can now it can now it can now connect to other workflows that i connect to other workflows that i connect to other workflows that i couldn't connect to before couldn't connect to before couldn't connect to before and these are just two examples very and these are just two examples very and these are just two examples very recent examples of very practical recent examples of very practical recent examples of very practical obstacles that will be solved with obstacles that will be solved with obstacles that will be solved with a connection into cloud or connection a connection into cloud or connection a connection into cloud or connection into open ai into open ai into open ai that's also stickiness by the way that's also stickiness by the way that's also stickiness by the way because as soon as you connect that to because as soon as you connect that to because as soon as you connect that to your cloud your cloud your cloud um um um and again a practical example why they and again a practical example why they and again a practical example why they haven't opened up for a deep seat haven't opened up for a deep seat haven't opened up for a deep seat connector right so and i um connector right so and i um connector right so and i um i i think that is a very smart move again i i think that is a very smart move again i i think that is a very smart move again by by anthropic and open ai that by by anthropic and open ai that by by anthropic and open ai that they're working to towards getting sort they're working to towards getting sort they're working to towards getting sort of these interconnections with all sorts of these interconnections with all sorts of these interconnections with all sorts of software systems across the globe of software systems across the globe of software systems across the globe and all of that and all of that will also lead to an increased token will also lead to an increased token will also lead to an increased token usage without you really thinking about usage without you really thinking about usage without you really thinking about it because it's just it just happens it because it's just it just happens
Speaker 3it because it's just it just happens right right exactly and that's just one right right exactly and that's just one right right exactly and that's just one you know you know you know one tiny example that you say has one tiny example that you say has one tiny example that you say has increased your token usage materially increased your token usage materially increased your token usage materially i mean we're so early in that story i mean we're so early in that story i mean we're so early in that story you know think about when um you know think about when um you know think about when um claude or whatever it is you use really claude or whatever it is you use really claude or whatever it is you use really connects like meaningfully to your connects like meaningfully to your connects like meaningfully to your phone and understands everything that's phone and understands everything that's phone and understands everything that's going on in your phone going on in your phone going on in your phone right and then understand and it's also right and then understand and it's also right and then understand and it's also then location based so it understands then location based so it understands then location based so it understands where you were yesterday and where you were yesterday and where you were yesterday and where you were the day before that you where you were the day before that you where you were the day before that you know then think about when know then think about when know then think about when you're when claude or whatever you use you're when claude or whatever you use you're when claude or whatever you use like connects meaningfully to like your like connects meaningfully to like your like connects meaningfully to like your house house house you know and understands your physical you know and understands your physical you know and understands your physical environment and what's going on in your environment and what's going on in your environment and what's going on in your house and what the season is and your house and what the season is and your house and what the season is and your heating and all all these things i mean heating and all all these things i mean heating and all all these things i mean again like again like again like token usage just it's very very easy to token usage just it's very very easy to token usage just it's very very easy to imagine those kinds of happenings where imagine those kinds of happenings where imagine those kinds of happenings where that just exponentially increase that just exponentially increase that just exponentially increase token usage for the individual ordinary token usage for the individual ordinary token usage for the individual ordinary user user user and all kinds of connected devices and and all kinds of connected devices and and all kinds of connected devices and then like i say all kinds of robotics then like i say all kinds of robotics then like i say all kinds of robotics uh uh uh yeah you know we're so so early in that yeah you know we're so so early in that yeah you know we're so so early in that story story story but we're moving to a new stage in the but we're moving to a new stage in the but we're moving to a new stage in the kind of market story we had the initial chapter we're moving we had the initial chapter we're moving we had the initial chapter we're moving to something of a new chapter now
Speaker 1to something of a new chapter now to something of a new chapter now yeah i i i agree with you that uh yeah i i i agree with you that uh yeah i i i agree with you that uh it won't be as straightforward to just it won't be as straightforward to just it won't be as straightforward to just buy ai hardware and lean back as it was buy ai hardware and lean back as it was buy ai hardware and lean back as it was a year ago um yeah i i i have to agree a year ago um yeah i i i have to agree a year ago um yeah i i i have to agree with that
Speaker 3with that with that let's see do you think do you think then let's see do you think do you think then about rotating into like other about rotating into like other about rotating into like other well like how do you think the well like how do you think the well like how do you think the implications of that through implications of that through implications of that through you know do you think about then you know do you think about then you know do you think about then rotating into like the users of rotating into like the users of rotating into like the users of intelligence like robotics and pharma i intelligence like robotics and pharma i intelligence like robotics and pharma i mean maybe robotics is still too early mean maybe robotics is still too early mean maybe robotics is still too early right to rotate into right to rotate into right to rotate
Speaker 1into at the moment yeah at the moment yeah at the moment yeah i i hold a few robotics names but i i've i i hold a few robotics names but i i've kept my robotic sleeve fairly small this kept my robotic sleeve fairly small this kept my robotic sleeve fairly small this year year year i'd actually like to show the chart on i'd actually like to show the chart on i'd actually like to show the chart on page three page three page three uh to give you sort of an overview of uh to give you sort of an overview of uh to give you sort of an overview of how of how i think about the timeline the current bottleneck is very like the current bottleneck is very like the current bottleneck is very like memory gpu driven right i i think memory gpu driven right i i think memory gpu driven right i i think there's still a power bottleneck ahead there's still a power bottleneck ahead there's still a power bottleneck ahead of us um of us um of us um especially given that you know we've just especially given that you know we've just especially given that you know we've just discussed this discussed this discussed this massive demand for compute and massive demand for compute and massive demand for compute and we know that we know that we know that the power bottleneck is one that will be the power bottleneck is one that will be the power bottleneck is one that will be very very relevant over the coming years very very relevant over the coming years very very relevant over the coming years because a lot of data some capacity is because a lot of data some capacity is because a lot of data some capacity is simply not online because of a lack of simply not online because of a lack of simply not online because of a lack of power power power um so i'm focusing on power as the next um so i'm focusing on power as the next um so i'm focusing on power as the next bottleneck um in in my bottleneck um in in my bottleneck um in in my in my portfolio and then i agree with you in my portfolio and then i agree with you in my portfolio and then i agree with you that uh the green that uh the green that uh the green uh one here the i call it the deployment uh one here the i call it the deployment uh one here the i call it the deployment but you know everything from robotics but you know everything from robotics but you know everything from robotics to physical ai to inference at scale i to physical ai to inference at scale i to physical ai to inference at scale i mean those chip makers that design chips mean those chip makers that design chips mean those chip makers that design chips that are that are that are ready made for inference ready made for inference ready made for inference uh i think they'll do well into 2027 uh i think they'll do well into 2027 uh i think they'll do well into 2027 2028 that's kind of 2028 that's kind of 2028 that's kind of the next phase uh so i think we're on the next phase uh so i think we're on the next phase uh so i think we're on the same page here the same page here the same page here uh i you know i've written 27 28 i i uh i you know i've written 27 28 i i uh i you know i've written 27 28 i i personally find a little early to go personally find a little early to go personally find a little early to go all in on that but um all in on that but um all in on that but um rotating towards it yeah sure that's rotating towards it yeah sure that's rotating towards it yeah sure that's starting to make sense yeah
Speaker 3starting to make sense yeah starting to make sense yeah yeah yeah i mean an energy is something i yeah yeah i mean an energy is something i yeah yeah i mean an energy is something i mean energy is a whole mean energy is a whole mean energy is a whole world that we haven't talked much about world that we haven't talked much about world that we haven't talked much about like you say right right now the like you say right right now the like you say right right now the fundamental bottleneck is energy fundamental bottleneck is energy fundamental bottleneck is energy getting enough energy into these data getting enough energy into these data getting enough energy into these data centers is just a huge challenge so it centers is just a huge challenge so it centers is just a huge challenge so it makes sense that that is your current makes sense that that is your current makes sense that that is your current your current bottleneck and of course your current bottleneck and of course your current bottleneck and of course that sort of you know i mean inside the that sort of you know i mean inside the that sort of you know i mean inside the exponentialist and a lot in gmi we talk exponentialist and a lot in gmi we talk exponentialist and a lot in gmi we talk about about about the way this whole story excuse me is the way this whole story excuse me is the way this whole story excuse me is is in the end a story to is in the end a story to is in the end a story to make intelligence more energy efficient make intelligence more energy efficient make intelligence more energy efficient you know get more intelligence per unit you know get more intelligence per unit you know get more intelligence per unit energy we talk a lot about that inside energy we talk a lot about that inside the exponentialist that is the the exponentialist that is the the exponentialist that is the fundamental direction of travel fundamental direction of travel fundamental direction of travel um and i'm trying to do more to think um and i'm trying to do more to think um and i'm trying to do more to think that through as a kind of actionable that through as a kind of actionable that through as a kind of actionable investment theme and right now that investment theme and right now that investment theme and right now that does it is like the power trade it's does it is like the power trade it's does it is like the power trade it's like how can we get energy into these like how can we get energy into these like how can we get energy into these data centers and then like what are the data centers and then like what are the data centers and then like what are the chips that make if inferencing and chips that make if inferencing and chips that make if inferencing and intelligence more efficient intelligence more efficient intelligence more efficient i think is a really really valuable i think is a really really valuable i think is a really really valuable thing to think about in terms of thing to think about in terms of thing to think about in terms of deploying yeah in terms of your deploying yeah in terms of your deploying yeah in terms of
Speaker 1your portfolio basically portfolio basically portfolio basically david let me ask you a final question david let me ask you a final question david let me ask you a final question uh we're soon approaching the top of the uh we're soon approaching the top of the uh we're soon approaching the top of the hour here hour here hour here i'll see you in a quarter's time again for this expo meets rb pro session until for this expo meets rb pro session until for this expo meets rb pro session until then what's the most important thing then what's the most important thing then what's the most important thing that you're watching in this whole ai that you're watching in this whole ai that you're watching in this
Speaker 3whole ai build out i think in terms of the build out in terms of the build out in terms of the build out as you said the kind of balance of as you said the kind of balance of as you said the kind of balance of revenues against debt revenues against debt revenues against debt and the hyperscaler revenue results and the hyperscaler revenue results and the hyperscaler revenue results um is really the big thing to watch i um is really the big thing to watch i um is really the big thing to watch i just that's the thing to watch in terms just that's the thing to watch in terms just that's the thing to watch in terms of what turns market sentiment or at of what turns market sentiment or at of what turns market sentiment or at least least least to my mind and then and then you know to my mind and then and then you know to my mind and then and then you know just in a very simple way just in a very simple way just in a very simple way what are people's expectations and what what are people's expectations and what what are people's expectations and what is the reality around rates is the reality around rates is the reality around rates um because those feel to me the things um because those feel to me the things um because those feel to me the things that could that could that could just turn market nerves just turn market nerves just turn market nerves into some kind of big market event into some kind of big market event into some kind of big market event uh yeah does that make sense i mean uh yeah does that make sense i mean uh yeah does that make sense i mean you're you know speaking to a deep you're you know speaking to a deep you're you know speaking to a deep person who's so deep in markets does person who's so deep in markets does person who's so deep in markets does that make sense and what how are you that make sense and what how are you that make sense and what how are you what what are you watching
Speaker 1what what are you watching what what are you watching i you know i'm i you know i'm i you know i'm i'm so focused on these hyposcalers next i'm so focused on these hyposcalers next i'm so focused on these hyposcalers next week and their quarterly results week and their quarterly results week and their quarterly results uh and i guess for ones i typically uh and i guess for ones i typically uh and i guess for ones i typically don't watch earnings calls but i have don't watch earnings calls but i have don't watch earnings calls but i have to hear because to hear because to hear because it's so important that they manage to it's so important that they manage to it's so important that they manage to convince the market that capex is still convince the market that capex is still convince the market that capex is still a good idea a good idea a good idea because that's the feedback i get from because that's the feedback i get from because that's the feedback i get from every single portfolio manager every every single portfolio manager every every single portfolio manager every single sell site strategist i talked to single sell site strategist i talked to single sell site strategist i talked to too much capex too much capex too much too much capex too much capex too much too much capex too much capex too much capex that's the message everywhere capex that's the message everywhere capex that's the message everywhere and i don't think the executives share and i don't think the executives share and i don't think the executives share that view so they simply have to that view so they simply have to that view so they simply have to convince the market why that it's a good convince the market why that it's a good convince the market why that it's a good idea and i i sincerely hope that they idea and i i sincerely hope that they idea and i i sincerely hope that they will because i think the capex is a good idea as well i think the idea as well i think the idea as well i think the results are great results are great results are great um so that is the main task for them um so that is the main task for them um so that is the main task for them next week next week next week and hopefully they'll succeed because and hopefully they'll succeed because and hopefully they'll succeed because that is basically the way i've that is basically the way i've that is basically the way i've positioned my portfolio On top of that, David, which is a story for a whole other show, a deal with Iran would calm down things a lot.
Speaker 3I was just going to say, yeah, does it turn out that a lot of this is about watching what happens on Iran and its knock-on effect on energy and then on inflation, right? A lot of it hinges on that.
Speaker 1Yeah, it really does. But we can do 60 minutes on that.
Speaker 3Yeah, we should do another 60 minutes on the whole broader geopolitical question, which I mean, you know, and to wrap up saying, you know, you guys talk about that so often on Pro and I love listening in and I love listening on how you're thinking and reading on the platform, how you're thinking through in such detail. The, you know, the collision of markets and this build out is absolutely fascinating to me. So your work to me is so valuable. You know, I spend a lot of time thinking through the broader technology. Story, yes, the market implications, but also the sociological implications, the implications for us as human beings. So that kind of deep focus you have on the markets is just is just really valuable. And I love these conversations.
Speaker 1Likewise, David, it's been a pleasure to host this session. The Exponentialist meets RV Pro with you once again, and I look forward to see you in a quarter's time again.
Speaker 3Yeah, likewise, you know, we'll put it in. The Diary for for three months from now. But I reckon if some if some kind of red alert hits, if something big happens, we'll just whip out the webcams and do it and do do an emergency session and let people know what we're thinking. So let's keep our eyes. I mean, of course, we've both got our eyes very closely on this. If something big happens, we'll be back sooner. But yeah, Andreas, great to talk. And thanks for tuning in, everyone.
Speaker 2This episode is brought to you by Pith. Trusted by Fidelity Investments, the U.S. Department of Commerce, Revolut, Kalshi, Jane Street, Coinbase and many more. They provide real time financial market data across over 3000 equities, commodities, crypto, FX rates and again, a lot more. You obviously enjoyed the episode because you're here with me at the end. But listen, don't forget to go to realvision.com forward slash join and grab a free membership. It's an incredible community packed with alpha, great investment ideas and the research that you need to help you unfuck your future. So get started now. Go to realvision.com forward slash join.
PodcastTranscript AI