Ai The Biggest Capital Misallocation In History Market Talk With George Noble

TITLE: AI: The Biggest Capital Misallocation in History | Market Talk with George Noble CHANNEL: George Noble DATE: URL: https://youtu.be/ro7AhBylxz8

---TRANSCRIPT---

Welcome everyone. I’ve been waiting with great anticipation for this particular podcast. If we we’ve got this is going to be the definitive podcast on AI for this year. I suspect it’ll be seen by millions. We have two of the leading experts on all things AI. Gary Marcus who really needs a little introduction. Um he’s been for four or five years now calling out the problems with artificial intelligence. a very successful entrepreneur and one of the leading critics of what’s going on with AI. Welcome, Gary. Good to see you. Thanks for being here.

And we have my old friend Julian Garren who we’ve known, oh my god, I can’t remember if it’s 25 years, 30 years, but one of the smartest cookies around. We did a fantastic uh podcast back two, three months ago. Um Julian had is is seen his more than his share of uh economic cycles and in particular has called out the folly of the current AI boom and and how it could upend the economy. So Julian, good to see you again.

Yeah, good to see you George.

And then my friend, my partner in crime, Jack, nobody special. He’s very special. Um he’s forgotten more about AI than I’ll ever know. So he’s the brains behind this podcast. I’m just the ring leader. So Jack, good to see you, my friend. And um I know it’s you and I talked about this podcast with great anticipation. So I’m going to hand the ball to you, my friend, and uh have at it.

Thanks, George. And I wanted to go right to Gary with this one because Gary, I have been covering the AI bubble since before anybody even was willing to call it a bubble. You know, way back in summer of 2023, I first picked up on the shady accounting practices that were driving this bubble. And since then, I’ve gotten a lot of criticism. Mainly, I get a lot of, “Okay, boomer, you just don’t get the tech. You You just don’t understand how good the technology is, which number one, Gen X, thank you very much.” Uh, number two, okay, maybe I don’t get the tech because I clearly don’t get it. I I’ve used large language models. I’ve used image generators. They’re neat. All right. They’re cool, I’ll even say, but I don’t see trillions of dollars in value in them. So, maybe I really don’t get the tech. And you know, you’re one of the founding fathers of this industry, one of the original researchers. And so, I don’t know, are these are these Twitter bots right that I’m arguing with? Maybe I don’t get the tech and it really is going to change everything and we’re in for this world of white collar unemployment, or are there inherent limitations to this technology, and it really isn’t that impressive?

Well, that’s a lot of questions all at once. Um, let’s start with this. Large language models are driving most of what people are talking about. That’s just one way to build AI. There going to be other ways to build AI. If you look at the history of AI, there’s been like, I don’t know, half a dozen fads that people don’t even remember anymore. Like a lot of the people who would call you a boomer have never heard of expert systems. And there was a time when expert systems were on the top of the world. They had high valuations, those companies. And like I said, these guys haven’t even heard of them. Um, so large language models I think are technology that’ll stick around for a while, not forever. They have some serious liabilities. Um, and let’s start maybe with those liabilities. One is they’re wildly inefficient. They use massive amounts of energy, massive amounts of chips. Uh, they need the entire internet to train on and that’s not really enough. So they have to supplement that with synthetic data. It’s madeup data. um they uh you know just cost a massive amount of money to do anything. You compare them to the human brain runs on 20 watts and you’re like this just cannot be the answer. So the first is the inefficiencies. The second is the lack of reliability. That’s something that I think I can rightfully say I am OG on because in 2001 I pointed out that these kinds of systems is really ancestors the current systems would hallucinate. And what I have heard from the people who run around calling you boomer for 25 years is oh we’re going to fix that. Give us a little more data. We’re going to fix that. But for 25 years I’ve been right. They haven’t fixed it. You know even the latest models that just dropped last week still hallucinate and they still have other kinds of reliability problems. So something else I pointed out in 2001 is if you take this path unless you do something fundamentally different, you are going to have problems with reasoning. And they still have problems with reasoning. And I cannot tell you how many times I heard, oh, scale, scale, scale, add more compute. We’re going to solve this. Now, what I’ve been saying is you need to have what we call neurosymbolic AI. And that’s a mix of the stuff that’s getting hyped all the time and old-fashioned AI that’s not getting hyped at all, but is actually pretty useful. And probably the best real advance lately is claude code that actually does the thing that I keep telling them to do. They run around saying Gary’s always wrong, but the field is actually doing exactly what I said a number of years ago we need to do many many years ago and subsequently.

Um so we are making a little bit of progress by not doctrinarily sticking to scale scale scale. But the thing is that all the money is still going into scale scale scale. All the hype is around that. All the VC pitches. If it turns out that the right answer to AI is not like kind of the world’s biggest supercomputer trained on as much data as possible, then all that money may be a waste. Like we may well find more efficient approaches to AI, more efficient in multiple kinds of ways. Needing less data, needing less memory, needing less energy and so forth. And we may be sitting there with $2 trillion of infrastructure that people don’t really need. I mean, just think about it. The brain runs on 20 watts and we’re building out terowatts after terowatts. like that just cannot be the right long-term solution. There’s a there is a world in which it’s the right short-term solution and it is the right short-term solution for a few problems like coding, but the idea that this is going to last is crazy. And then I mean probably we’ll get into things like well what about depreciations? You put all this money into these chips, maybe they turn out to be the wrong chips, maybe somebody makes better chips. Like I I also have been pounding away on the economics since 2023, August 2023. my my case, I wasn’t early on the circular financing stuff that you were, but I was making the argument back then that like if the stuff is unreliable, how are you going to make that much money on it? And that was when people were putting in hundred billion dollars. And I thought that that was kind of a wacky idea. Now they’re putting in, you know, trillions of dollars literally and like we can talk about this like where is the revenue? Like my favorite example of this right now is maybe this stuff can actually work for coding by using neuros symbolic AI, not just scaling. We could talk more about that, but for now it doesn’t matter. Um that whole industry, the software industry is only $570 billion is the number I saw. And they’ve got to make trillions every year. So even if they captured the entire software development industry, which is ludicrous, and we could talk about why it’s ludicrous. um even if they capital the whole thing like they’re nowhere near to the revenue they need now that they just you know they’re playing double or nothing over and over and over again you know cuz they’re never making money. It’s never profitable. So they’re like well let’s put in twice as much money maybe then we’ll be profitable. But the thing is if you keep putting in twice as much and you know sooner or later now you need to make like 16 times or 32 times the revenue and it just starts to seem like really really crazy. The only thing I can say is if you really had reliable AI that really worked solidly, that might actually be worth trillions of dollars, but we’re just not anywhere near close to that. And maybe we can dive in a little bit more like why I think on a technical side we’re not close to this thing of AGI, artificial general intelligence that people imagine.

I can tell you the the reliability thing really jumps out at me. I mean, my my background, I’m not a financial guy by background. Uh my degree is in mechanical engineering and I worked in nuclear power for 15 years. So when I look at the hallucination rates, which I’ve seen estimates around 20% of what it tells you are just factually incorrect, I mean that kind of an error rate is terrifying to a nuke. Um we don’t tolerate 1% failure rates on our equipment. And so you know, anything that is like life or mission critical in any kind of industry just can’t tolerate a failure rate like that. The consequences are too high. The only real applications are ones where the failure rate isn’t a serious problem. If you’re talking about a nuke, there’s just no way you can use this stuff. And this is just insane. Like maybe you can use it to write a report or something like that, but you can’t use it for your your critical computation. The same in medicine. I mean, like people keep talking about using this stuff in medicine, but there’s study after study that shows that there’s actually, you know, fairly serious problems with using it. The reliability is just insane. Do you remember the Pentium FD bug like 20 years ago? like the the the the young kids don’t know this, but there was um Intel made a a a CPU back when those were, you know, the thing. And it had a mistake in arithmetic, but it didn’t make this mistake all the time. In fact, anytime you did integer arithmetic, which is kind of the easier version, it made no mistakes. But somebody kind of proved that like I think it was once in a billion operations or once in a trillion operations on floating point, it would make a slight mistake. And it was a huge scandal. cost Intel $500 million. It was in the news all the time. You know, this was like making a mistake once in a trillion. People were like or once in a billion, whatever. People, oh my god, I can’t believe that, you know, Intel did this. And now suddenly we’re in this environment where like people have forgot about accuracy and forgot about reliability. And that’s fine if you’re like brainstorming, hey, you know, write me an advertising jingle that that you know rhymes with the lines of in 12 Days of Christmas or whatever. And they’re like really excited about that. But if you have something where you really need to get it right, this doesn’t work. And so, you know, one thing we probably want to talk about is study after study has shown that the productivity that people were expecting is not there. Like just a minute ago before we logged on, there was a new Bloomberg piece. I didn’t read it yet, but maybe I can grab you the the headline. Um, uh, the technology worked, the value didn’t arrive. Bane concluded the report. There had been so many reports like that. It was there was one by MIT first MIT Nandanda that said you know only 5% of the companies were getting return on investment and there’s a guy named Rob Woodlin did a nice you know takeown of that study and he said the sample size is small and yada yada and he’s probably right about those points but there have been now like a dozen studies all pointing to the same thing. So the first study was, you know, imperfect, which is what happens in science, right? Somebody does a study, somebody improves on it, whatever. Um, the fact that the first study was not perfect didn’t mean it wasn’t, you know, at least sort of directionally correct. It was directionally correct. Now we have at least, you know, 10 studies or something like that. U McKenzie, Bane, etc. Everybody is saying the return on investment isn’t there. And then people will say well maybe the newest models and you know for coding maybe finally but in general the return on investment isn’t there and it’s because of the reliability right so there’s now a great term called work slop right so you know what AI slop is like you know make garbage that kind of like is sort of correct but not really like you want a picture and doesn’t quite do what you want so so um these people I’m blanking on their names came up with this term work slop and the idea is like you ask it to write a report and it looks good on first inspection because these systems what they really do is they mimic the style of the language that they’re trained on. And so they’ll get the style of the report you need, but if you look closely, there are problems and it’s actually a pain to look closely. You’re actually like an employee that you don’t have to watch like every single thing that they do and you’ll find that there’s hallucinated citations and maybe some of the numbers are made up and whatever. And then that actually costs you time. And so the net is you don’t get that much productivity out of at least in most cases. I’m not saying there aren’t any use cases, but that comes back to the economics. Like I if you’re spending trillions of dollars, it really has to work almost everywhere most of the time to be worth that kind of money.

So Julian,

yeah,

what Gary just said, right, scale, scale, scale has been the cry for three years now. and they have scaled to the tune of arguably in the trillions now that they’ve spent. Certainly the market cap of these stocks is several trillion each. And then when I hear a headline like the value didn’t arrive, talk to us about what that means because this is not just an oh well we tried better luck next time scenario is it? I mean just how how many eggs are in this basket right now?

Oh. Oh, it’s it’s it’s huge. And it it’s if this goes wrong, which I’m almost completely sure that it will go wrong, um then then we’re not just facing a sort of temporary one-year decline in GDP. We’re potentially facing a major kind of upset um in the capital markets um which could have major implications for multiple years. But I guess to to get at the real heart of this um kind of following on from kind of Gary’s technical work, my my kind of golden rule for for large language model AI and I distinguish that completely from kind of a narrow AI uh looking at a closed system with with very clear rules that can be measured and optimized. That that’s can be very useful and has been useful for 50 years. What we’re looking at is these general generalized large language model AIs. And my golden rule is that you can’t use a large language model um AI to create an app, a product or a service that’s going to be commercial across the supply chain. So if you if you imagine a sort of a healthy ecosystem for property, you’d have the bank making money while it’s lending to the to the property developer or the house builder, the house builder making money selling the house to the um to the um landlord, the landlord making money renting the house to the tenant and the tenant happy that they’re paying a third of their salary on an apartment because it’s much nicer than living in the park and so on. And you can see how that that’s kind of sustainable, but that’s definitely not what’s happening um in the large language model AI ecosystem. You’re seeing major losses taking place at the data centers. The data centers by making those losses are effectively subsidizing the frontier model generators. Um they’re then making significant losses. Broadly speaking, go AI is losing about two bucks for every one buck of revenue it gets. So its costs are three times its actual its actual revenue. Uh they’re effectively subsidizing the front-end app developers like kind of um Perplexity on the search side or Replet or or kind of lovable on the u or cursor on the coding side. Um and they’re then making losses and effectively subsidizing the end consumer.

I think the reason that’s important is because that’s clearly unsustainable. um you need to have continued funding if you’re going to manage to continue doing that until you find an app that’s mass market profitable. And if you don’t get that continued funding or if you start to worry about that continued funding, you’re then going to have to start charging people further up the chain. You’re going to have to stop your subsidies. And that’s why it’s kind of fortuitous that we’re we’re recording today on June the 1st because this is the day that Microsoft starts charging not a flat fee for its use of um of GitHub copilot but actually starts according to usage starts charging according to usage and that’s when people have to start thinking am I making money using this and that’s when the trouble really starts and so I think this is going to cause a huge amount of churn and huge amount of trouble and is probably the beginning of the end of the process. Um,

can I

Yeah,

I’ll just add one thing. I I thought that was a beautiful metaphor or it’s not even a metaphor, but uh the the comparison you made with with the the chain all the way down to the rental. Um, and I think that the um change that you’re talking about has already started last week in particular. So um a lot of what sustained the most recent excitement was this idea of token maxing and it was tied to agents. So we we went from things like chat GPT to agents that would do not just answer a query but do things on your behalf which was a terrible idea because they’re not reliable enough and you don’t you don’t want to give your credit card number to a system that’s not reliable. But we’ll leave that part aside for a second. um they require many many tokens in order you know a lot of use tokens basically just a measure of usage for these things right and so some people got very excited they’re like the more my employees are using AI the better and this term came out of token maxing like and like Amazon briefly had a leaderboard they were like rewarding people for using the most tokens all the employees were getting the message use AI as much as possible I posted about a month ago on Twitter um token maxing is stupid. Prove me wrong. Nobody ever did because it is in fact stupid, right? Because what it is is it’s saying I’m burning money without regard to whether it’s useful, right? So of course employees are to game that and like they can write little macros and burn as many tokens as you want. And then what we heard last week was the beginning of the end of that in multiple ways. So, um, uh, uh, Madison Mills, I think is her last name, um, reported in, um, Axios that somebody had, uh, wasted $500 million. One customer had wasted $500 million in a month. We had, um, the Uber uh, COO, I think, saying, “We’re spending all these money, but I’m not really sure that it’s working.” And they had already spent their whole token budget for a year in the first four months. And there were several other studies um or or anecdotal reports and so forth like that. And so like I I actually put like an RIP for token maxing a couple weeks ago seeing this coming um on Twitter and I think it really did die last week. like the whole mood changed and you the most promising thing that has happened for the AI industry lately is that Anthropic is about to report a profitable quarter or at least that’s the rumor which would be amazing but you have to look at the details of that one detail is they did that in the height of this token maxing mania where people had these unlimited budgets but now they realize they can’t do that the customers realize they can’t do that which goes to Julian’s um long explanation um long and correct explanation and so that’s going to stop and so the next quarter is not going to be like that and then oh by the way also Elon gave them what’s looks to be like a couple billion dollar subsidy and the profit was like $550 million and so you take away the subsidy the onetime subsidy from Elon and the token maxing and it was not you know like that quarter was a total anomaly it’s probably not um going to be sustainable and going to be back in the scenario that Julian describes where there’s kind of subsidies all the way up the chain there’s not actually being mining made by anybody but Nvidia that can’t be sustained.

Yeah. If I could elaborate a little bit on on what both of those guys just just bought up and you know you took us to XAI and Anthropic and the the invisible subsidy there. Um Michael Bur had a fantastic Substack post over the last couple of days about a name I’ve never heard before, Valor. And I’ve never heard the name, but I knew this exact scenario was coming. It’s exactly what I’ve been warning about. And in this post, he talked about this special purpose vehicle which a chapter right out of the Enron playbook. They create a shell company and that shell company buys the chips. And in this case, it was Apollo, XAI, now SpaceX and Nvidia got together and created this company Valor. They bought 5.6 billion worth of Nvidia’s chips. All right. Now, right off the top, 1.9 billion of that came from Nvidia themselves. There’s Nvidia buying their own product and reporting it as revenue, right? Like like the Girl Scout whose dad bought all the cookies and then she’s the best Girl Scout because dad bought all the freaking cookies. Apparently, that’s a trillion dollar business proposition now. Uh but the remaining three and change billion dollars came from private credit firm Apollo and they shopped this around and they ended up selling this debt to Athen. And Athen is an insurance company that is owned by Apollo. So, a sol, you know, XAI bought these chips, but didn’t buy the chips. Nvidia bought them. They’re all located in XAI’s data center. XAI doesn’t report them on their books. They don’t report the debt on their books. It’s offbalance sheet debt. Who’s actually holding that debt? Grandma’s annuity and and with the life insurance company. And so, if the depreciation bomb does go off as Bur suspects, who’s going to take the loss? It won’t be Nvidia. It won’t be SpaceX. it’ll be grandma’s annuity that that eats that loss.

I mean, that actually reminds me something else we should touch on briefly, which are these insane rules to change the index funds to accommodate um the IPOs. Um you guys will say it more more carefully than I did, but um the way in which SNP is about to change its rules such that basically everybody is going to be obligated through their retirement funds to have a piece of SpaceX, which is completely overvalued, like just absurdly overvalued. So, so Gary, let me interrupt you. Uh, I’ve known Julian a long time. I’ve never seen the facial looks. It’s like, Gary, you should consider doing standup. Both of you guys. It’s like this stuff is like so insane. We have to let Julia It’s Look, you look

funny thing happened to me on the way to a recession.

I know. Julia, you look like the cat ate the canary. Multiple canaries. You want to weigh in a little bit what Gary was saying in particular the financial angle what Jack was talking about this this BS with the wi-i with with the VIE and in and shades of Enron. I mean the fact that this might be technically legal. Jim Chainos is has is has crafted the the saying legal fraud. All right. problem with the accounting system is, you know, it tries to capture the essence of what’s going on, but it’s not perfect. And the bad actors have a big incentive to always try to frontr run and anticipate and and and view the accounting rules as an obstacle course. And so in a case like this, the accounting system does not fully portray what’s actually going on. So Julian, you want to have a whack at that?

Yeah. No, sure. And I I’ll broaden it out. in in kind of one of the big questions we’ve got is well why has this taken so long to start to unravel and and its incentives. Um if if you’re running a kind of one of the big four kind of hyperscalers and earning 50 to 100 million bucks a year um it’s a life-changing event getting to keep your job for an extra year. So if you can extend your kind of job by getting involved in the next higher round of open AI or anthropic having previously invested then you can boost your income statement on paper. Um and at the same time if you get them to agree to take compute off you you can increase your revenue and so it starts to look very attractive until you realize that there’s no returns in the ecosystem to actually support it. And and so now that the the the kind of hyperscalers have gone from being capital light cash generating machines of four years ago to being capital heavy cash-free machines today. They’re now starting to rely on the favor of strangers. Um because it’s now no longer down to them and how much cash they want to burn. It’s now down to the debt markets. and the debt markets are going to have their say and they’re going to start to question how valuable this stuff is and we already got the clear signs of that when the Oracle kind of announcement which initially last September caused their shares to explode. Um and then as people started to work out just how vulnerable they were to uh failure of Open AI to kind of pump hundreds of billions of dollars into kind of uh into data centers over the next four years. um the CDS started to blow out. Um so my my kind of thing is well you know there are a lot of these stories. We heard stories like this during the dotcom crisis. Um we stories about overordering of equipment. We’re hearing those again as well. Um the question is how how big is the problem? And um so what so what I do to try and identify that is that um we look at something called a wixle spread. If if you kind of I’ve been in economics for about 40 years. It always it never ceases to amaze me how few people at the sort of top levels of government or industry or the professions understand the kind of economic foundation of civilization and progress because it’s not technology. What what it is is is working hard and skillfully thrifting, creating a surplus, and then building um a piece of capital, whether it’s a house or a hospital or a office or a shop or an electrical grid um that makes return. And by making return, you create a foundation for your own future. But what you also do as that’s repeated millions of times across the economy is you create opportunity for people other people to work with that capital um use their own knowledge and their own skills uh and generate income. That’s that’s the basis of economic kind of society. Now this is not what’s happening with AI. So the question is how how big is the problem? And what you can use is something called a wixle spread. Nutwixel was a Swedish economist from about a hundred years ago and he said that the neutral rate of interest was a couple of points above the kind of potential nominal GDP growth and if you set rates at that level um then a company who knows their own business who doesn’t think they can make a return borrowing at that rate uh they kind of stand pat um they run for cash and they shrink. company that does think they can make a return, they borrow and expand and that way capital gets allocated as well as possibly can be allocated. Um, if you set rates too low and and kind of the Wix bread was at its biggest discount ever during the pandemic at minus 12%, we’d never seen anything like it. You create this huge incentive to borrow and buy assets um and build assets. Um and the problem is those assets get bid to perfection. So later in the cycle when rates rise, growth slows, the assets start to fall in value, the debt and equity behind them begins to get impaired and you get a downturn. So how much of this stuff has there been? Well, I estimate that the misallocated capital in the US economy, this is the whole economy, not just AI, um is now twothirds of GDP. It’s now 23 times bigger than it was on the eve of the dotcom crisis. Now, this isn’t just AI. It’s anything that’s been used with used or anything people have used cheap capital to buy. So, it includes crypto and it includes private equity and it includes private credit for instance, especially the 2020 and 2021 crashes. Um, but that’s a real problem because that capital isn’t making return. it doesn’t fund itself, so it’s not supporting growth. Uh, and if it starts to fall in price and the debt and equity behind it starts to get impaired, then we’re in for a really serious downturn. One that I think is going to take extraordinary measures from the Fed and the Treasury to try and guess get get us out of.

What’s the scenario that you see for how this falls apart? I mean, I think we’re agreed on this call that what we’re seeing now is not sustainable. I can give you my own scenario. Um, but I’m curious about yours.

Okay. Um, look, look, I think the the key thing is is kind of identifying where the if if the system the ecosystem can’t make returns on its own. Um, and we can talk a bit more about the details behind that. Um, if it can’t make money on its own, then it needs funding. um if it needs funding then it’s kind of basically got to go to the debt and equity markets to get it. So I think the first kind of cracks um are basically when the when the debt markets start to question whether they want to fund any part of that ecosystem. So I think we’ve we’ve clearly got the first cracks in that um kind of the the kind of data center um debt and Oracle’s debt is not trading particularly well. Um and that’s your first sign that funding is going to get more difficult. Now the obviously one of the other major funders is is the banks. Um kind of some of our clients are bank credit kind of departments. Um and they they’ve been telling me that while they’re sort of in some cases forced to make some of these loans because the bank wants a relationship with the client so it can make money elsewhere in the bank, um they are also capable of pushing back to some degree. and they’ve told me that they’re dancing near the door in terms of their willingness to do this. They’re clearly getting more cold feet to a greater degree. We’ve heard kind of specific examples of Deutsche and others who are beginning to step away. Um now that’s that’s the funding side. Um clearly the IPO season is going to be kind of critical here because if we start to see the key kind of um documents from say OpenAI and Anthropic um that start to disclose their real profitability. I think that’s going to then start to cause some real problems because when you look at OpenAI it’s a real struggle to see how they’ve got a road to profitability. And I think that the OpenAI’s abandonment of sorrow which in the press was reported to be costing 2.4 billion but where their revenues couldn’t have been more than say 500 million. That’s a that’s another example of of what they’re having to do to clean up shop ahead of the IPO. But that was meant to be one of their banner kind of products. Um, and I guess the final one is is that as we start to see kind of charging that’s more realistic um, kind of for customers, we’re going to see what businesses can actually survive with real costs. And my kind of analogy for that is that that if you kind of offered me a a chauffeer driven Rolls-Royce for 10 bucks a day, well, I’ I’d find I’d use that. I’d find that pretty useful. I could think of a few businesses I could um I could uh I could start kind of using it, you know, free Rolls or cheap Rolls-Royces. But if you then suddenly started charging me what it actually cost with a bit of profit, then all of those businesses would fold. And that’s the reckoning that we’re now coming up with in the past week as you’ve called for the death of token maxing. And as we see kind of Microsoft and and increasingly others step away from kind of set pricing to pricing according to usage, companies are having to go out and find out is this actually making us money and they’re finding it very difficult to prove that it is.

I’ll just add one thing. I’m more more or less in agreement with all of what you said and learned a bunch from it. Um, I’ve always seen Open AI as the weak link in the chain. I’ve been arguing since late 2023 that they would eventually be seen as the wei work of AI. And I think now they have multiple problems. They have a management problem. Anyone in the right mind would no longer trust Alman. They have um uh we could talk about that if you want, but I I I assert it and I believe it to be correct. um they have a competition problem which is they had a lead and they squandered it and now both uh Anthropic and Google are caught up or close or you know um and so they don’t really have a lead. They don’t have any kind of mode whatsoever. They’re less capital efficient than Anthropic appears to be and they want to IPO at the same time as Anthropic and as you say they’re burning a lot of money. At some point people don’t want to write checks for them anymore. Like if someone has to choose between the IPO of Anthropic and Open AI, which are theoretically going to be valued at about the same price, it’s very hard to make a case for choosing Open AI uh over Anthropic. And I think that that’s going to impair their ability to raise the cash that they need. And and so my imagination has always been that they would fall first and that that would have huge ripple effects throughout the industry both psychological because they were the poster child for all of this and economic because you have all all of these commitments that they’ve made that they may not actually be able to honor and so that’s going to you know lead to at least in my scenario that I’m imagining lead to problems even for Nvidia and for you know lots of other folks.

Yes, I’m I’m I’m with that. And I I guess the one thing I’d also add is that the front-end apps have even less of a moat because because those guys are using kind of um open AIS and and Anthropic’s kind of models as the base of their business to create a front end whether it’s search or coding or what have you. But Anthropic and OpenAI can simply kind of copy their front end, subsidize themselves while charging properly for the the apps and those guys are set to get into trouble as well. But yes, I I I agree with you. Out of the frontier model, guys, Open AI is definitely definitely the weak link. And I think that I think this attempted IPO is going to be difficult. And I know there’s history between Musk and Open AI. Um but the fact that he’s chosen to subsidize anthropic rather than them just makes an even bigger chasm emerge. Um and in general it’s it’s tough it’s tough to keep those modes going. You can as you as you know you can use synthetic data to recreate.

This has been one of my points for quite a while is that on the technical side there is no mode because everybody is using the same underlying AI technology which is the large language model. It didn’t have to be that way. There have been many approaches taken in the history of AI over the last um 50 years, but everybody converged on this one thing. And not only did they converge on this one thing, but they converge on doing it in the same way, which was they all converge on the scaling hypothesis. And they all converge on the same data source, which is basically the internet. And so they and then they always have the same results, which is like they never solve hallucinations. They get a little bit better on these benchmarks and so forth. So it’s like they’re all conducting the same very expensive experiment and with exactly the same technology which means nobody has a technical mode you know worth you know even a couple of months there was just a study in fact and then I I’ll stop by um I think it was meter but could could one of these benchmark um organizations found that the open models are only four months behind the closed models and so like that’s another problem you can’t have a sustainable business model if everybody’s basically playing the same game with the same recipe. It only takes 4 months to catch up. Like there’s just no way.

Could we just I just want to jump in here because time is flying so quickly. No conversation.

You can add a little extra time. This is really

Yeah, that’s great. That’s great. Let’s go down this rabbit hole. Um Jack, I think you described uh SpaceX as an AI company with a rocket thing attached to it. All right. I think we’re all in agreement here that this is like sheer madness. I mean, it’s complete insanity. and and you look at the way they’ve they’ve walked the valuation up the last couple years with these private market transactions. Um so Julian, Gary, whoever wants to have a have at it first speak I mean the rocket thing is easy enough to value. Starink’s easy enough to value but um XI looks like a complete croc. All right. So, could you just explain um how troubled um XAI is and and just what a what a travesty the the SpaceX IPO is. I don’t know who wants to have a shot at that first. Go for it.

I mean, I can take one piece of it, I guess, whatever order. Um one piece of it is that XAI was not very successful as a so-called frontier model AI company. So, they make Grock, which gets used by people on Twitter, but is just not really competitive. Elon poured in a lot of money. He made a lot of his usual hypy statements, like I think he said at one point, I think there’s like a 10% chance this is going to be AGI this next release or whatever. And it never amounted to anything. I mean, nobody I think in the industry takes them seriously as, you know, a leading model provider. Everybody thinks they’re in the game, but they’re a follower. They’re not a leader. And this is despite, you know, some pretty impressive things he’s done on the infrastructure side. You know, I think Elon does understand infrastructure pretty well. He understands certain aspects of scale pretty well. I don’t think he understands AI that well. And on the technical side, I don’t think that they were particularly innovative. And so the net result is like they’re just not a major player. And in fact, the transaction they just made with Anthropic should be alarm bells for everybody because basically what they said is we can’t win this technical game. So, we’re going to dump all of our chips onto the market. Like, I it was insane. They they basically they they made a subsidized deal to Anthropic to get rid of 220,000 H100s that they didn’t really need that badly. Like, that should have been a sign to everybody that like, how can we be building up so much supply when he’s dumping some? It’s like, if I had cars and I sold 220,000 of used cars of this model, I wouldn’t think that was good for the car vendor. Um, and so they they’re trying to reposition themselves as a cloud provider basically because they couldn’t really win on AI and they’ll be a decent cloud provider, but it’s not clear how much money there really is in that. And the whole thing is then predicated on a fantasy that we’re going to build um these, you know, data centers in the sky. I had a friend who who wrote me a very funny text message um friend who’s very sharp on engineering. He’s like, “Any place on earth would be better. Even the bottom of the sea, you know, all these places are going to be better for all kinds of different reasons.” Like, um, he’s like, “But, you know, or you could just go to Montana and it would be a lot better than any of those choices.” Like, it just it’s not it’s not sane. Um, and then also in the IPO, and then I’ll I’ll turn it over. Um, my friend read it more carefully than I did, are crazy things like um point-to-point rockets, so you can go like from New York to Tokyo. Nobody’s going to go from New York to Tokyo on a rocket. That just like this is just fantasy stuff.

Can I uh add to what Gary just said there? And I I love the data centers and space criticism. I I mock that regularly on my show. Have you have Have you ever done something and thought to yourself, “Wow, that would be easier in orbit.” No. No. Nobody does that. Um

I I have a friend who shot a movie and he needed zero gravity stuff. So I’ve got one friend who actually has an answer for that.

Okay. Well, there’s the vomit comet and even that it’s, you know, NASA’s 747 that simulates zero g. There’s cheaper ways to get zero g or close. Jo Julian, have a crack at it.

Okay. Well, I’ I’d sort of comment on the broader issue, which is that when we saw the dotcom kind of bubble get into trouble. It was for a combination of the kind of reasons that we’re thinking there may be difficulties, i.e. funding of dotcoms and the difficulty of those dotcoms to raise revenue that then led to a shutdown in orders for routers and switches from Cisco for instance. But then that then fed through the entire ecosystem. But part of the reason that there was trouble was because there were a lot of IPOs in 2000 and these three together are multiples of what we saw uh combined in 2000. And in addition to that, um, this is an environment where with a macro hat on, liquidity, broader macro liquidity is set to deteriorate. So, we’ve had this front-loaded boost with the Fed doing adding to bank reserves with uh the banks increasing their money to non-bank financial institutions, including kind of data center funding and private equity. um and with a bit from the Treasury General account getting drawn down, all of that’s set to turn around and start reversing. And so if you got kind of a move from FA um feast for liquidity towards famine by the end of the year and then you add this big call on everybody’s cash um with these hypervaluated businesses on top of that, you’re you’re really running a risk that the bottom’s going to fall out. not just of those businesses but of the whole market.

You know, within a few days of that XAI anthropic deal getting announced where, you know, XAI says we have this whole data center we have no use for because Grock kind of sucks. So, we’re going to rent it all to Anthropic. Within a few days,

just in fairness, in in fairness, they they built multiple data centers. So, they’re leasing out the first one, but you know, on depreciation side, it’s only a 2-year-old data center. They they’ve kept their newest ones, but they they basically dumped their 2-year-old data center. the hoppers, the H100s predominantly. And within a few days of that announcement, Meta comes out with a a statement that kind of got brushed over in the media. And for the first time, Meta admitted just maybe we’re not efficiently allocating our capital, which really Meta after the metaverse, you think. Uh but what they said after that, they said, “If we did overbuild, that’s no problem. We’ll just do a cloud offering. We’ll just we’ll just rent them out like XAI just did.” Well, who’s Meta going to rent it out to? The same people Nebus rent out to and Cororeweave and Iron, OpenAI and Anthropic. At the end of the day, they are like the GPU renter of last resort for this whole industry. And those two businesses, as you guys are pointing out, are loss burning cash infernos that are hopelessly dependent on an endless influx of investor capital to continue making those rentals. What happens if even one of them slows down the rate at which they’re gobbling up compute? Now all of a sudden there’s nobody to rent this to. You know, it’s like the guy who bought a 100 houses and says, “No problem. I’ll just Airbnb it all.” Well, what happens if people stop traveling? You’re stuck with assets and the carry cost and the debt with nothing to do there. So, you know, it’s not just XAI that’s doing this. Meta’s talking about doing it now, too. And Meta doesn’t even have a commercial AI product yet. They have, you know,

I mean, just a year ago, people were like stealing H100s off a truck, and now people are like, “Yeah, I got some. What do you need?” I mean, J Jack uh uh uh coined the um brilliant line a couple months ago talking about the neoclouds and and maybe each gentleman could speak on the neoclouds. Jack, I think he called them what was it? We works with GPUs.

Yeah, I I didn’t want to I didn’t want to step on the uh open AI was the work of AI space.

No, we we works with GPUs. I mean, and what happens business model, these open-ended growth things, it’s just it’s unbelievable that this is happening. I mean, is there any sustainable business model?

I got a funer who would love that proposition. His name is Masa. I think you should look him up.

Oh, that’s been a little he had a little real estate venture, didn’t he? Too.

Yeah. And and by the way, Julian, maybe you want to talk a little bit about SoftBank’s role in this whole thing because they are either driving the car or in the passenger seat of pretty much every one of these transactions.

Well, yeah, and they’ve they’ve been huge. Again, the question is how how far that can they continue going? And I think kind of the the story was that um when uh when kind of um uh son had to sell his Nvidia shares in order to fund his commitments to to Open AI, he was crying publicly in the boardroom. Um I I really question given he had to borrow substantially against his shares and had to sell other shares. I really wonder whether he’s going to double down on that. I think it’s again it’s going to be difficult and

there was some reporting last week that people inside a soft bank are really really concerned and and like basically being told they can’t talk about this but they’re they’re very worried about the level of commitment there to open AI.

I Yeah, I would be I I would be as well. Can can I just sort of um kind of slightly kind of turn the conversation a bit because we we’ve been talking about AI not being commercial or large language model AI not being commercial and and

not commercial enough. There’s some commercial

but but the whole system is definitely not commercial. And I guess the way I’ve been thinking about it and I was wondering if you could expound on this Gary is that there’s kind of four reasons it’s not commercial. So the first reason is kind of because it uses Let me interrupt for one second and then give the four reasons. So what you really mean there is not commercially viable, right? So there are commercial applications. You’re using the word in way that I’m not used to. I just want to make sure. So like there are commercial applications. Some of them make money but the thing as a whole doesn’t make sense. It isn’t vital.

Yeah. The whole ecosystem isn’t making money like

Yeah. Okay. Okay. Go ahead. So give us your four reasons.

Yeah. So, so the four reasons number one um they’re built based on correlations but correlations with a probabilistic interface to convert those correlations back into an answer and those aren’t always accurate. In fact, can’t always be accurate. Second, the errors compound themselves because the way that they jam in prompts in with the sorry, jam in the beginning of the answer in with the prompt. And so if that beginning of the answer has an error in it, it makes it more likely there’ll be more errors.

That’s right. That’s particularly pernicious in agents, right? So agents,

you take multiple steps and the output of one is essentially the input to the next. So you make an error anywhere along the way and let’s say you’re taking 60 steps, then you’re screwed. And so agents are particularly prone to error for exactly that reason.

For that reason. And then the third reason is because we’ve hit a scaling wall as as you’ve discussed. And and I and my my thing is I I see sort of arguing ahead of time that there will be a scaling wall is is is right. Um but is but is theoretical if you like. But the proof in the pudding was was kind of um OpenAI spending 50 million on chatbt3 which came out in November of 2022, 500 million on chatbt4 which came out in March of 23 and was demonstrabably better and then spending 5 million on chatbt5 which was meant to come out in Q323 but didn’t do and then when um you know the information and other kind of um kind of uh journalists um talked to people inside OpenAI, they were saying it didn’t come out because they couldn’t it wasn’t sufficiently better to justify the 10 times training tag. Um and when it did eventually came out, it was it was kind of to wide kind of uh lack of interest. Um and so the training costs and also the fact that they’ve run out of of real data um mean that it’s extraordinarily expensive to train these things with only very incremental improvements in accuracy. Um and then the final issue is um is security um because they kind of jam in the the beginning of an answer in with the prompt and then carry on doing that. If you if you want to kind of code to say a Pac-Man game and it finds kind of Pac-Man games in its training set and puts those in as the answer to your kind of prompt, a bad actor can have a a prompt that they sneak in with that which effectively asks, you know, for all your bank details and bank passwords, etc. And that’s fundamentally very difficult to solve because you need a pristine coding set. But who’s got one of those? and you’d then have to build it. That’d be kind of tens of millions of coding hours. And you’d have to trust that your coders didn’t take shortcuts and borrow stuff from the web themselves. So those four reasons, each of them seems to be critical to the lack of commerciality of these businesses.

I basically agree with all of those. Um I I think that there’s a pretty good analysis of where things are. I could have some quibbles about the nature of that security problem and maybe broaden it out a little bit um to say that there’s a general problem which is that they don’t follow instructions well these large language models. So um some people call that the alignment problem. Some of that has to do with like AI safety and we haven’t talked about that. But some of it is you you can tell it something like don’t hallucinate and it will still hallucinate or don’t use copyrighted materials. It will still use copyrighted materials or don’t tell people how to use biological weapons and it will still tell tell them that and you can say write really secure code and people actually put stuff like that in their system prompts but it is no guarantee that that will actually happen. So like broader just than security though you’re quite right is that there are no guarantees from these systems for anything and that again poses problems in in you know critical domains and so yeah I think you’re right even any one of those four problems would be pretty serious and putting them all together is even more serious and then I guess one other asterisk which is that claude code is working as well as it is not I think because of scaling but because they’re adding all of these other gadgets to it so instead of using a pure LLM which is no longer having that much return. They finally went to symbolic AI and they have 500,000 lines of code and 50 tools that it uses and so forth. And so the actual innovation is not from the scaling that everybody’s spending the money on. And so like I don’t know what the metaphor would be like you know the the the dog is chasing its own tail or something like you know it’s just kind of lost right now.

You know Julian three of those four things that you just mentioned your reasons why it’s not viable. um the fact that they’re probabilistic language models, so there’s always going to be a certain error distribution. The errors compound plus the scaling wall. All three of those remind me of something Gary said almost at the beginning of the stream. You you mentioned briefly synthetic data, Gary. And I’ve been wanting to circle back to that because I think that is a huge landmine that this industry is already already.

Yeah, let’s talk about that. Yeah. Um let me insert one thing and then talk about the synthetic data which is Julian left out of his analysis the lack of moat which is also absolutely huge problem right it forces these things to be commodities and that just makes the economics really really hard especially when you keep scaling the models to be bigger. Okay so synthetic data um there you might contrast that with kind of organic data. Organic data is just everything you find on the web like some human actually said it or it’s a report from somewhere or something like that. Um, so basically they trained on all of that. They then started taking data not just from written data but like transcribing YouTube videos because they realized the more data they had the better they were doing. And they they got pretty desperate. They took all this copyrighted materials decided to ignore copyright law and so forth. And even that wasn’t enough. So what they started doing is making up data. Well, how do you do that? Well, in some cases it’s really easy. Like let’s say um these things suck at multiplication. you want to make them better at multiplication, you make up multiplication problems. You take a classical AI algorithm or not even an AI algorithm, just a classical algorithm, and you generate, you know, you make a times table. 3 * 4 is 12 and and so forth. And so you make up data. There’s nobody maybe on the internet bothering with all all of these examples, but you just have a program write them. And that’s what synthetic data is. It’s not a new idea. Um, people have been, for example, doing this in the driverless car industry for a long time, making up data. So they’ll they’ll run a driving simulator. For a while they were some people were actually using Grand Theft Auto 5. Um and so they make a driving simulator and they make more data. Then people made, you know, fancier versions than Grand Theft Auto that were, you know, more tuned um towards actual roads. So you you simulate what might happen. But the problem with that has always been that you can’t simulate everything. So in math you can actually kind of cover a lot because you kind of know what the domain is. Same thing in coding. in driving it’s been kind of marginal like we really have not solved the driverless car problem even with a lot of synthetic data and people have been working on that for a decade. Um there it’s because there’s scenarios nobody envisions. My favorite example was um someone used the Tesla summon uh feature at an airplane trade show. So they’re like hey look how cool my Tesla is. and it ran directly into a three and a half million dollar jet because there was no no none of the synthetic data had thought oh well we should put jets and runways into our training data right and what we have found with these recent so-called reasoning models which I believe depend heavily on the synthetic data is they work well in closed domains I think Julian maybe touched on that briefly before they work well in closed domains um you know Alph Go was like that it’s a closed domain the rules haven’t changed in a couple thousand years, you can play against yourself. That’s making synthetic data. That’s fine. But if you want to simulate what happens in the world at large, you can’t make a perfect simulation. You can’t anticipate everything. And so there are limits on synthetic data. It really does work pretty well in these closed domains. And it doesn’t work that well in kind of open-ended domains.

So if I could summarize the thing with the Tesla, it’s like here’s this thing in the road. You didn’t tell me not to drive into that specific thing. So, I’m just going to go ahead and drive into it.

That’s right. And it it it highlights the difference between how a human works and how a lot of, you know, these contemporary AI works. Um, there might be better AI at some point, but contemporary AI tends to be sort of like, this is kind of an oversimplification, but kind of like it looks up in a video library. What do I do in the closest video in this library? It’s not literally that, but it kind of gives you the gist. Whereas a human is like, huh, jet expensive, big, maybe I shouldn’t run into that. So the human solution is reasoning and the AI solution is basically like looking for similar things in some vast database. It’s not exactly that, but it’s roughly that. And that’s when you get into trouble. Another example of this, at least a couple years ago, were these river crossing problems. You have like a man and a goat and and um some cabbage they have to get across. and a guy named Colin Frasier came up with all these examples um that were similar to the existing examples but not identical to and somehow subtly different. One one of my favorite ones I think it was Collins but maybe it wasn’t was um a man and a woman have a boat and need to get to the other side. And one of these systems like Claude or something like that wrote this long-winded answer about like the man goes across and then he swims back and then he goes and gets a boat blah blah blah. And like I told this to my daughter who was maybe eight at the time and she’s like why don’t you just put the man and the woman in a boat and go across and the the you know a human can reason about like what is the situation on the ground whereas what the AI was doing was basically analogizing very loosely speaking to these existing problems without really understanding what a man is, what a woman is, what a boat is, what it means to get across the other side but just trying to find text that is similar. And then it got so desperate that I think it was anthropic actually had in its system prompt instructions about how to deal with river crossing problems because they couldn’t solve it organically. It was getting so embarrassing that they did special purpose, you know, tweaking to deal with the river crossing problems. Um, you know, what you really want is an AI system that understands abstract concepts like efficiency, person, vehicle, crossing, and so forth. And we don’t really have that. and we kind of approximate it. The approximations work really well until they kind of don’t and they break. And that’s why the stuff is still not really reliable.

Gary, what you’re talking about, it’s interesting, too, if I just think about markets and um quant trading strategies. A little bit of a rabbit hole here, but novelty, novelty screws up quant trading strategies. And what you’re telling me kind of reminds me that you can’t deal with novelty. So, for instance,

same thing. Imagine you’re trying to I’m sorry, we got to go back on Elon Musk. It’s a self-driving cars. All right. I was at my Wharton reunion a couple weeks ago and they had one of the major professors of mobility, whatever. And he went through this whole thing, but hey, after level four, and the only person on the face of the planet that still claimed he can do it with with level three is Elon Musk. There isn’t no single. All right. So, now what do they do? They set up these geoence things in Austin where, you know, it’s contained and blah blah blah blah blah. So, it’s like a photo op. have three cars running around. It’s in a little contained environment and there no crashes. Yeah, of course. But the minute you introduce novelty, the thing doesn’t know what to do.

That’s exactly right. I mean, it’s slightly oversimplified because there’s some novelty that’ll get right some of the time for some reason, but it does not really have a robust solution to novelty. You cannot ever trust that in the face of novelty it’ll work. Which is why really, as you’re saying, it’s kind of like a glorified monorail system right now. Right. monor rail system is like you know it runs around the airport in three places and it’s driverless like big big deal and what they’re doing in different ways both Whimo and Tesla I guess are trying to avoid novelty by constraining which roads you go to learning as much information about those roads as possible and then they’re still using remote operators even within those limits and so you know level five self-driving what that means for people who might not remember is you can type in any two locations like you can on Uber and your system will will take you there safely. Nobody has that in actual operation. The closest they have is limited roads where they know something about those roads and have you know probably plotted out to avoid certain trickiness and so forth. Nobody has what you know a normal driver would should be able to do normal human driver.

Um and sorry let me say one one other thing which is one of the ways I got into all of this is my dissertation was about how children understood novelty in inflecting verbs in the past tense of English. So like if I make up a verb what should the past tense be? Um so uh I did some of this work with Steve Pinker. He had this great example of um Yelton out Gorbachov and Gorbachov. And so you what’s the past tense of out Gorbachev? It’s going to be out Gorbachev. You know you add ed to it even though you’ve never seen it before. novelty. And what I discovered in my dissertation is that the neural networks that were popular then that are ancestors of the ones now are bad at novelty. And then in 1998, I made a kind of formal demonstration of how multi-layer perceptrons that are the ancestors of today’s systems could not deal with principled versions of novelty. I I made a distinction. I said they can generalize to nearby cases. I called that within training space. But if you go outside that training space, you can’t be counted on. I give very simple examples like um if you trained on um identity function, basically copying the input to the output for even numbers, the systems wouldn’t generalize to odd numbers. I said, look, this is a deep problem here with extrapolation. They can interpolate, but they can’t extrapolate. And there’s been a little bit of progress on that. There are certain ways in which LLM are better than the models that I wrote about in 1998, but fundamentally novelty remains the problem in in basically the same way as I described almost 30 years ago. And so when these guys say, “Hey, more data is going to solve it.” I’m like, “No, you need a principled solution to this problem that has been an Achilles heel for this technology for 30 years. Until you get there, it’s it’s just not going to work as well as you think.” And yeah, we found some commercial applications, maybe not commercial in Julian sense of making uh actual profits, but at least you know, making some revenue. Um, but that novelty problem has not been solved, which means no, you can’t put a chatbot in a car and expect it to drive safely. Like, you know, if it were really AGI, that would be trivial, right? True AGI, you should be able to just hook it up to your car and it’ll drive, right? I can, you know, I can do that with the average 16-year-old. Um, but they’re just not actually there. They have not solved that problem. Novelty, that is still the core of why this does not make the trillion dollar sense that they needed to make. End of rant.

Julian, um, could we try to tie this together a little bit? So, you know, it’s just extraordinary. We’ve been going a little bit over an hour and any one of these topics have been discussed are just like a sheer head scratcher, but the hits keep on coming and enjoy. I’m like you. I could hardly contain myself listening to Gary like are you kidding me? But as I say, you can’t make this stuff up. But on a more serious note, it’s not just, you know, AI investment first order effect capped on obvious. Well, you know, it’s accounted for 105% of GDP growth, blah blah blah. Yeah.

But when you start going up the chain and as you pointed out, maybe you could kind of weigh in a little bit, the wealth effect and the huge increase of uh market capitalization, what that’s done to consumer spending. And then once the you know what hits the fan, the fallout and what will come afterwards because we’ve got we’re going to have so many stranded financial assets, physical assets as well. Um you know it was lovely your rendition about how um you know Apollo they’re buying this stuff you know they’re it’s marked to model they put it in their insurance company it winds up in grandma’s portfolio. So like can you just sort of zoom out a little bit as our crypto friends would say not the first order effect like okay the AI stocks will go down fine but bigger picture I mean this has the potential to really make dot bomb look like child’s play.

Yeah. And it’s it’s so much bigger. So yeah, f first order is that that kind of uh a combination of the wealth effect and the kind of uh data center buildout effect is adding about 3 percentage points to nominal GDP. So if it just stops getting bigger, uh then we take 3% off nominal GDP growth. If it starts to reverse, you take three to six off then. Now that’s a recession in and of itself. But the next issue is all this misallocated capital this 2/3 of GDP that’s misallocated because once that starts to go once you start to see that impaired the problem is it’s not rate sensitive. Um these things didn’t stop getting invested in because rates went up in 2022. In fact they were invested in even more. what they’re sensitive to is is the impression about whether they’ll ever get to the point whether they’ll create a commercial ecosystem uh which we don’t think they will. Now in that environment you’ve got a major problem in terms of having to deal with the weight on your economy. um not just from the fact that the immediate contribution to growth is gone, the fact that you’ve put in a false capital stock, one that doesn’t support future growth. Um and in that environment, um it’s going to be very difficult to deal with. So my view is that in order to deal with the downturn uh we’re going to have to see extraordinary measures from not just the Fed but also from the kind of um fiscal authorities from the Treasury in order to try and kind of uh arrest the slowdown and turn it round because just cutting rates are they going to be able to do that? So my view six months ago was that we were going to reach a point where they were going to demand bailouts. They were going to say, “What about China? We can’t afford to lose to China.” Um,

in fact, OpenAI was already shopping loan guarantees for data centers. Um, their CFO, I think, said that at a conference. There was a lot of push back, but I was assuming that the play here was going to be some kind of bailout. Taxpayers um, we’re going to have to do it. But now the numbers are so large because the investments have kept going. I don’t know if like that’s even feasible anymore.

No, I I don’t think they’re going I don’t think they’re going to be able to bail them out. I think

we’re not talking about mortgage bonds here that the Fed can buy and sit on their books for 20 years, right? We’re talking about GPUs that are going to be worthless in a couple of years. So, this is not they’re not going to create some kind of troubled asset relief program and park all this stuff on the Fed’s books and act like that’s going to make the problem go away.

The magnitude just seems too high to me. Like, I don’t this is not my specialty, but my intuition is like because they keep doubling the debts and probably will continue until it falls apart. Like, it just doesn’t seem viable.

Yes, it’s and it’s it’s not viable, but also it wouldn’t work. Um because the only way it would work is if the thing they were buying actually fundamentally made a return. So yes, mortgages fundamentally will make a return if they’re at the right price. So if the Fed buys them cheaply, they should be able to make a return. They won’t be able to make a return on GPUs if people decide that they don’t want to train and in and use them for inference. And so so what they’ll have to do the fiscal for is to try and counteract the the kind of the results the impact on unemployment and the impact on falling growth and the impact on broader asset prices. Um so and it’s not going to be easy. And there’s another difficult thing which is because of the sort of move to higher deficits and kind of the move off Swift taking Russia off Swift so China moves away from treasuries and because Europe and Japan have bad demographics and they’re more spending more themselves so they have less money to buy treasuries. All of those factors um mean that there’s not going to be that much relief from interest rates flowing through to rate sensitive areas like kind of um housing and autos and so on and uh and office. And so once you have all that, my view is we’re going to move into a highly reflationary area. They’re going to have to pump a lot of money into the system just to try and stabilize it. They’ll probably do a bunch of fiscal as well. I think there’s a significant risk that they that Trump follows through on the Mirin doctrine um and starts to look to send capital offshore to try and weaken the dollar and improve the balance of payments. And in that environment kind of with again with my macro hat on, what tends to happen is the second derivative of US liquidity is offshore dollar liquidity. And when you start pumping offshore dollar liquidity into the system, um then you start to see it getting lent out again in dollars to commodity finance, trade finance to emerging market investment. So I think what’s going to in a sense the phoenix that are going to come out of these ashes um is going to be a major resources and emerging market boom

and and also in that scenario implicit in that and and Jack, you have to have a go at Julian. Um, you can’t own enough gold in that scenario, can you?

Yeah, I was I was about to flash my little shiny, my my silver that I keep handy for just such an occasion cuz when when I hear about them spinning up the printer again, which you know, look, they always use the playbook from the last crisis and they’ll keep doing it even though I think something changed in 2020. I think the uh Zerpenber, as I call it, zero interest rate policy and money printing, I think that stopped working in 2020. They’ll try it again, but I don’t think it’s going to work now that the inflation genie is out of the bottle when they cut rates and print money into inflation that’s already 3% pushing forward. Um, that’s just going to make the problem bigger. And you’re talking about a this misallocation of capital, Julian, that you described earlier in the show is the direct result of the grotesque loose fiscal and monetary policy in the years prior to this. I mean, you know, it’s it’s not a coincidence that we got the most grotesque misallocation of capital in history after the biggest monetary expansion in history. They’re this it’s the same thing. all these guys who think they’re in on the next big thing and that they’re these visionary tech leaders. No, you’re just a kid at the monopoly board that reached into the bank and grabbed a handful of 500s and this is the inevitable result.

Yeah. I mean, Julian, doesn’t it ultimately lead to I mean, if we’re solving for X and you think about, you know, second, third, fourth order effects that these financial assets tied to AI and and and financial assets tied to debt and and the US long and the whole thing that value may they’re going to try to maintain value in nominal terms, but certainly in real terms that capital is going to get destroyed.

Yes, I think you’re in a real risk. So, we’re already seeing the problem starting to emerge. Um, just last week, we hit kind of 19-year highs for 30-year Treasury yields and and kind of longer than that for for yields of similar longdated bonds in Europe. The and the reasons I cited, you know, the demographics in Japan and so on. Um, and the budget deficits are all part of that. But if you directly seek to debase, so you directly try and send money offshore, which is part of this mirin doctrine that that kind of the Trump administration has kind of signed up to, um, and you put fiscal in at the same time and the banks are still solvent, which they likely will be, so they can keep lending into this, then as uh um as Jack said, um, you’re going to see the inflation genie come back out of the bottle again. And in that environment, you risk turning the US into a sort of an emerging market play where its attempts to get itself out of trouble can get it into even more trouble in terms of putting its currency at risk. Um, and if we start to see that happen, um, that flips everything. That that’s when, you know, that’s when debt has much less value. Um that’s when you know anything long duration, anything that’s expecting returns a long way in the future that can get crashed uh at the expense of things that are here and now like commodities, resources, the emerging markets that will benefit from this. So So yes, I think the risks of going again with easing are going to be much higher than the risk of going with easing after 2008.

So So Julian, what would you tell your clients and what would you tell to the average viewer watching this thing? because we’re trying to help investors. All right, it’s clear we don’t we don’t want to own AI related assets. Okay, but what would you tell your average client institution? How should they prepare for this? What should they do?

Okay, so my my view is is to kind of look at this as if it’s a kind of corollery just bigger of of what happened coming out of the dotcom kind of crisis. So say say we’re in September, August, September of 2000. Um and it was right to start avoiding or shorting um kind of tech stocks. Let’s say avoid just to kind of keep so and then moving into value um kind of like kind of uh like resources like emerging markets. And what happened to begin with was the tech went down faster than everything else. So you made money on a relative basis and once they finally got enough liquidity into the system the value the the resources the emerging markets went up much faster than the tech. So I think you can play it on a relative basis now um that you want to shift a relative portfolio into those kind of value and resources and benefactors of an ultimate kind of Fed and Treasury bailout. Um but when you want but you want to be at the down to at the bottom if you want to play it absolute. So I’d be cautious now if you’re an absolute investor um overall and wait and gradually get into these things. Um and I would be playing the split if you’re a relative investor right now.

And a question for both of the two of you. So, one of Soros’s uh great lines was the way to make money in the market is to figure out what does the market believe that’s not true, but more importantly, try to pinpoint when the market will realize it’s been had. Like, it’s all well and good for Jack, Gary, you and I to be screaming about AI three years ago. Good luck with that, right?

Um, predictions are difficult, especially about the future as Yogi Bear once said. So, um, you know, something will happen. you know, the way things work, it’ll probably be something we haven’t even thought of. You know, as as as as our friend Louis Gab would say, you know, this is like a bug in search of a windshield. All right. What’s going to trigger it? I don’t know. All right. And maybe, as Gary said, it’s already started from last week. I don’t know. But if you had to think about, it’s questions for both. If you had to think about the timing on all this thing coming unstuck, I mean looking at the semiconductor stocks screaming and the neocloud screaming and okay, Nvidia’s lost its mojo. It doesn’t have a beta compared to these other things. But how would I mean clearly I want to stand aside and and maybe you know what we can’t figure pinpoint the timing. It’s just you know as my friend John Ro would say we’ll be the second to know. Let the market show itself. The market it will show itself. How would you think about timing the day new this whole thing?

I mean the timing is incredibly hard. Um the metaphor that I keep using is wy coyote at the end of the cliff in Bugs Bunny and yeah like he only falls when he looks down on the question. It’s really a psychological question right we’ve gone through the economics and we’ve gone through the technical AI side like it is clear that this is not commercially viable and it’s risky in multiple different ways. So open AI could not come up with enough cash. The companies could decide they don’t want any more tokens. Someone could come up with a better technology that doesn’t require quite as many GPUs. Like there all kinds of things that could be a kind of decisive blow. But ultimately it is about psychology. Like I always think about what if I was alive during the tulip craze. Like at some point like it still makes sense. You’re like all these people are insane but if I flip you know my tulip I can buy a house and so maybe I should do this. And like you know you don’t know exactly when that’s going to end. even when people are doing it and you see that it’s insane. And I I sort of feel like that’s the case now. I think that the death of token maxing last week is really bad, but how long does that take to ripple out? Is it a week, a month, a year, five year? Like it’s it’s hard to know.

Jo Julian, you and I were were having it, I believe, back at the time of do. So I I know for myself in in 99 my therapy bills went up enormously. They got they made it back and then some in 2000. But trying to make sense of the nonsensical just, you know, forget it. But Julian, how would you answer the question?

Oh, look, I I I think we’re just I think it’s a weight is beginning to build here. Um and um and at some point we’re going to have enough straws to break the camel’s back. I think as Gary said, the death of token maxing, I think the beginning of companies like Microsoft charging kind of not a flat fee but by usage. Uh I think the way that kind of um the uh the credit markets and the banks are starting to kind of price up CDOS’s for for the um data centers and Oracle etc. I think tightening liquidity and I think um the the risks around IPOing companies that are deeply lossmaking. I think all of those things can come together. And then I think that the final one that’s sort of doing the rounds over the last kind of couple of weeks is that now everyone’s facing the the sort of issues that they may be put onto usage rather than flat fee for you for using these models is that they now need to with internally within their businesses justify the return they’re making on their use of this stuff. And that’s that’s really difficult. And so I think that process going on internally within businesses which we won’t get told about. We’ll only see the results of it months later. But that process could be one of the straws that breaks the camel’s back. And I’ll add one other thing there we didn’t talk about which is a lot of what’s been driving the companies um like say Fortune 500 companies that are the customers of OpenAI Anthropic etc to use it is really FOMO. It’s like they don’t actually have a revenue case, but they’re doing because they’re afraid their competitor is going to do this and so forth. So, one interesting thing to look at is with token maxing losing its allure, will people change their view on the FOMO and be like, maybe I should let my, you know, competitor waste $500 million on this and maybe I don’t want to. Like, I I don’t know what that psychology will be like. But this is a moment where some of that psychology could change where people could be like, you know, I actually feel like I’m being had here. maybe I’ll let other people.

Um, so FOMO has driven it. The people might at some point say, you know, I don’t want to miss out, but also like this doesn’t make any sense for me anymore.

And there’s also a fear of coming to your senses, right? The big tech CEOs now realize they’re trapped in this game of capex chicken. They they know they’re stuck here, that they can’t keep burning this money forever, but if they stop, their shares are going to tank. So they they they don’t have an out here. Um, and they know it. Yeah, they’re starting to the things like renting out XI’s data center and Meta’s comments about renting. So, you know, how do you model when that point is when a room full of imbeciles comes to their senses? I I don’t know how to model that, but you do know at some point between

but that has happened historically, right? I mean, we’re not talking about something that never happens. It rarely happens, but it does sometimes happen. People are like, “Oh my god, why am I buying these tulips?” And then it changes, right?

Yeah. Let me just say one thing at this point. I remember clearly the summer of uh summer of 99. I was already a boomer. Summer of 99. I go to one of this was Adams Harkness Hill emerging growth technology conference in Boston at the Marriott. I’m sitting there as a room full of 500 people and the management one after getting up going on about do and blah blah blah. Small detail went down 90%. But but I digress. And I remember sitting there. I was sitting next to a a fellow who was a few years older than I. And I leaned over to Bill Crane. He was a salesman for one of the major firms. I said, “Bill, I don’t understand a word of what this guy’s talking about.” It was, you know, it was the the CEO of Yahoo or something like that. And he says he he leans back. He says to me, “George, don’t worry. Nobody else does either.” But the only difference was they’re all bulled up and they’re long. Okay? I was like, I wanted to shoot myself. Okay? It’s deja vu all over again. Julian, you’re you’re visibly shaky. You know what I’m talking about.

No, no, I’m with you. I’m with you. Yeah.

Oh my gosh. Jack, anything else you want to touch on here? We’ve been going almost an hour and a half.

Yeah. I’ll just leave it. My parting thought is uh you turn on these guys on CNBC and they all say earnings supports the rally and big tech earnings are solid. It’s all paper gains on illlquid shares. All right. And you’re going to see more of it. It that hasn’t hit its fever pitch. Google’s big earnings beat, half of that was them arbitrarily marking up their stake in Anthropic. The 30 billion of their 60 billion in earnings was just them increasing the value of their stake in Anthropic. They’re going to do it again after the IPO. Um, all while that’s happening, money in the bank is going down. As Julian pointed out, they’re now capital heavy cashlike companies. You can do all the accounting fraud in the world. You can sell to yourself. You can’t fake money in the bank. And all these companies are running out of it.

Yeah. Gentlemen, this has been extraordinary. Uh Julian, I want to thank you. Gary, I want to thank you. Um we’ll have to do this together, do this again in a couple months. um anyone who’s invested not just in these companies but in the market because I saw a statistic the other day and Julian since you co cover the markets pretty carefully and Jack so people point to the percentage of the S&Ps in tech narrowly defined but someone did a calculation of the percentage of S&P stocks or market cap that actually trade they’re correlated to the move in the AI trade and it was like 60% of the index so it’s become like one giant trade And so therefore, anyone who’s invested in in public markets, even if you don’t like tech stocks, you have to watch this. This is this is fantastic. I Gary, I can’t thank you enough. Julian, Jack, let’s do this again in a couple months. Thank you guys.

Yeah, this was great. I’d love to do it again.