← Big Technology Podcast
Loading summary
Transcript79 lines
- [00:00]
A
If we are in an AI bubble, what could an unraveling look like? Let's talk about it with investor and analyst Paul Kadrosky right after this. Welcome to Big Technology Podcast, a show for cool headed and nuanced conversation of the tech world and beyond. We have a great show for you today. We are going to tackle what I think is the strongest argument that all this AI investment is going to lead to, well, a collapse. Because our guest today, Paul Kadroski, thinks that we are in the midst of an AI bubble. He has been making the case far and wide and has not backed off, despite the fact that this technology has gotten much better over time. And so this will be a really fun discussion to ask. Basically, even if everything goes right, are there economics on the downside going to be so bad that it will still fall apart? So Paul, it's great to have you on the show. Welcome.
- [00:49]
B
Sure. Great to be here.
- [00:51]
A
All right, let's just start with the spending and the return necessary to, to make that investment pay off. Right. If we're in an AI bubble, as you argue, there's going to have to be some level of overspend and then an inability to make those returns materialize. So first off, can you just talk about the magnitude of spending going into the AI build out today and how that compares to maybe previous infrastructure build outs and the rest of our economy right now?
- [01:22]
B
Sure. I mean, there's a thousand ways to kind of put it in context for people. But one of the ways I try to do it is to compare it to, as you say, prior infrastructure build out. So you can go back to the 19th century and canals and railroads, or you come forward to the 19th, well, the late 19th century and early 20th and talk about electrification and rural electrification or the, or the interstates, World War II, rearmament, the fiber optic build out. These are all these moments in western economic history, in particular U.S. economic history, where we had this massive infrastructure investment that in some ways not, obviously the analogies are never perfect, but in some ways are analogous to what's happening today. So one way to think about the sizes of each of these moments is to think about their contribution to gdp. Or you can think about them in terms of their contribution to GDP growth, you can think about them in terms of their contribution to non residential fixed investment. There's lots of ways to back into this so you can kind of provide some context. And it doesn't really matter anymore which one of those you use. We're the winner. So we're now currently larger than everything except for, and this was an unfortunate analogy I made recently on a German interview. As I said, we're now larger than everything except for World War II rearmament, which doesn't play as well in Germany as it does everywhere else. But nevertheless, the point being that as a percentage of gdp, as a percentage of non residential fixed investment, as I've documented for the last year or so, in terms of its contribution to GDP growth in all of those metrics, we've now exceeded all of the largest capital expenditure paroxysms impulses in Western economic history. And again, you can say to yourself well so what? Or anything else, but that's sort of a separate question. So the point to start off with is this is a really, really unusual moment in terms of the scale of capital expenditure normalized against all of these other capex moments. And then we can get into whether or not any of those analogies matter or whether this time is different, the favorite sort of responses to these kinds of things or all sorts of other stuff. But the point is we've now reached that moment and it's in lots of other measures. It's now the largest tech, is now the largest piece of the high yield bond market. It's now the largest piece outside of financial services of the investment grade bond market. So in terms of new issuance tech companies themselves are now at a point where for the last two years people repeatedly told me that it really didn't matter because they were doing it out of cash flows. And so it would only become worrisome if this was becoming out of debt. Well, guess what, as of the second quarter of 2020, this is now more than 50% of the funding for data centers is external financing, which is obviously the term of art for off balance sheet and out of your own cash flows. And now of course the same people who were saying that a year ago were saying that that would matter, is now saying, well that's perfectly fine now by any of these metrics, gdp, non residential fixed investment, percentage of gdp, percentage of GDP growth off balance sheet financing. We're now at a point where this is a remarkable historical moment, full stop.
- [04:30]
A
Yeah. And a way that I like to talk about this is first of all, it's not only a bigger magnitude than these previous build outs, but it is a bigger magnitude contracted into a fewer number of years. So just to put a.
- [04:44]
B
That's a really important point. Yeah, yeah, yeah. And that's a really important point that you know, for the, for example, to put that in context, electrification took almost 30 years. The build out of the US railroad was a multi decade proposition. The interstates were a decadal proposition. Even the build out of the fiber optics big backbone was probably four and a half to six years, something like this. So this is a higher scale of spending happening at a much more rapid pace. And that matters in the context of capital markets because you don't have time to slow down and consider exactly what's happening and where are the returns going to come from. But that's a problem for another day. But just to put it in context, that's an appropriate context, right?
- [05:24]
A
And so if you have let's say a build out that goes over a couple decades or even five years, you have these, I think this is what you're talking about, you have these stop points where you put some, some investment in, you get some time to marinate in your projections and then you say should we put some more in? Right. And of course in many of these buildups that we talked about there were collapses. But what we're seeing now is this rush in to invest in the AI infrastructure build out without those natural stop points and the numbers are bigger. So we're looking, this year it's looking like big tech alone will put something like 700 billion in towards capex. This year. I think last year was something like 350 to 400 billion and next, next year is projected to be 1.5 trillion in build. Yes, now we're going to, you know you mentioned a lot of the different dynamics about like where this money's coming from and that's important. But let me just put this to you to begin with. With the level of, of investment that we have coming in to this type of build out, what is the return that's going to be necessary to justify these investments? So let's just take like 700 billion. What does what for? For even investors to like, I don't know, not go under or I guess a lot of this is big tech, but like what are the numbers that we need to be looking for for those numbers to be rational?
- [06:44]
B
So you have to turn it around and look at it from the standpoint of the providers of capital, so alternative uses of capital and what return I could get on the same capital in another context. So the way that I try to analogize this loosely, and this is very loose, is that data centers from the context of many capital providers are real estate. They're really just multi tenant apartment buildings. They' so happens there's no humans in the apartment building, there's just GPUs and so from the standpoint of providers of capital who look at these as project finance and then by that measure try to compare the returns they're getting on this to the returns they're getting from doing project finance. So think about it in the context of commercial real estate, a strip mall, a multi tenant apartment building or whatever else. So increasingly the providers of capital for these things look at it in that context and say, well, what's the yield in terms of I'm contributing $100 billion to some massive metapro? What's my reasonable cash flow expectation? Very much analogous to what I might expect from the cap rate on a multi tenant apartment building. And is this competitive on that basis? That's the short answer to your question is it's very much a market based return that's required. The scale of the money is irrelevant in some weird context because it's really all about what sort of return can I expect and how does that compare to comparable investments? Again, in this context, CRE is the most comparable investment from the standpoint of external capital providers. So they say to themselves, we're looking at cap rates around 6.8%. 6%. Is that reasonable? Well, that compares reasonably well to the following five projects, but not particularly well to this project. So what it's provided is a way of putting the returns from these things in context. So it's wrong to say for better or worse, we're going to be putting in a trillion, Therefore I need 100 trillion out of this. That's not the way investors are looking at this. And it will lead you down the wrong path if you take that approach. Because now you're forced to say, well, I'm going to have to estimate what percentage of some giant number I'm going to earn over the next five years. And that's where you get into these loony arguments from some of the sell side analysts where they'll say things like well the TAM, the total available market for human labor is like $12 trillion if I get 20% of the TAM. This is ridiculous. This is just completely seat of your pan speculative stuff. So it won't get you anywhere in terms of understanding the calculus that's driving people to provide the off balance sheet financing for these projects. So the right way to think about it for better or worse, is to analogize it to commercial real estate and ask yourself what kind of cap rates they could get on comparable projects. And that is really the answer. Now that leads you into a trap. But nevertheless, that's the answer in terms of thinking about what kinds of returns are required to justify continuing providing of capital.
- [09:28]
A
Okay, this is really important table setting here and I think this is sort of worth digging into a bit because the way that you're framing this is actually suggesting that the, we don't even need to hit the best case scenario. Right? So like a way that I've thought about it is almost everything needs to go perfectly in order to return on these investments because they're so big. But if you're saying that this is just like being, being invested in the matter in the manner of a typical real estate investment, then that perfection is actually not necessary. And my level of concern goes down here. So let's say, say, let's just take an example. I'm meta, I invested $100 billion in data centers. So let's say you're like looking at like, I don't know, you could help me with the math here. You want to get us like a 6% return or 20 return on your, on your investment. You might not, you might only need to get 120 billion back. You know, if you're going to compare this to a, to a real estate investment. And now that I'm thinking about it, I'm like, well Meta makes what, like 30, 40 billion a quarter? That might be eminently possible with the outlay. So where's the concern here?
- [10:37]
B
Well, the concern is that the nature of the investment is profoundly different from real estate. So what you're really entering into is a project that not only has current capital requirements, but has ongoing capital requirements. This isn't just now and then I'm going to have to replace a tenant's drywall. This is a project which will require wholesale replacement of most of the hardware and probably changes in the cooling system and probably changes in other aspects of these data centers continuously and probably, depending on the math, anywhere from a four to seven year period. So it's nothing like an apartment building in the sense that most of the capex occurs upfront and then it generates a recurring annuity cash flow from which I bask in and generates compelling returns back to my investors. This is much more like a utility, a non regulated utility who has continuing capital requirements which continually dilute the returns because you're having to raise more capital all the way down the path. And this will continue for the lifespan of the project. So what you end up is from the standpoint of an investor, you end up with a duration mismatch problem. Right. So I've got what looks like a long duration project, like an Apartment building. That's actually a short duration project in the sense that most of the underlying assets need to be turned over relatively frequently or at least upgraded. Now you can get into all kinds of traps, the Michael Burry thing with respect to like, well, what's the proper depreciation schedule for, you know, GPUs or whatever. But the point still stands that you not only have upfront capital requirements, but you have continuing capital requirements. So that's problem number one in terms of thinking or making the direct analogy to commercial real estate back to data centers. The second one is that you need to also think in terms of the nature of why this replacement happens. Some of the replacement happens because of the MTBF, the mean time between failure of GPUs, which varies depending on what the GPUs are being used for, and the generation of the GPUs. So we have some GPUs that are failing inside of modern data centers on an 18 month cycle, some that are failing on a much longer period. So we've got constant churn just from that standpoint. And then we have familial upgrades in terms of upgrading to new generations of GPUs that cause upgrades. And then we have the whole replacement cycle of maybe we won't have GPUs in some of the upcoming data centers. There'll be increasingly inference specific ASICs, and we're seeing lots of that going on. So that's problem number two. Problem number three is we're paying a fixed rate of return on a depreciating asset, not just to capital, but also in terms of the thing under the hood that's generating the cash flow. So what's generating the cash flow? Data centers can be thought of as factories. And the thing that they produce, the widget that they produce, is this thing we euphemistically call tokens. And these tokens are among the most rapidly depreciating assets we've ever seen in a modern economy. That they've continually been falling 70 to 80% year over year on a constant performance basis for at least the last four years. And there's no reason to expect that to change. So you've got at least three different problems here in terms of making that naive comparison to commercial real estate and saying, okay, everything's going to be fine, look, look, these guys are good for it. And we've got these long duration contracts, we have the depreciation of the data centers, we have the continuing capital requirements, and then we have this unprecedented problem of a hyper deflationary commodity at the core of the revenue generation engine. Of these so called data centers. None of those existed in the context of any other cycle in the past. Railroads weren't going through hyper deflationary cycles and neither rural electricity. So fiber certainly wasn't. Fiber was actually the reverse. It became more valuable over time. So all of this is really unusual and makes the naive analogy to commercial real estate that brings in those kinds of investors who have showed up in huge numbers because they see this analogy incredibly fraught and probably perilous for them.
- [14:23]
A
Okay, so there's a lot here and I don't have a dog in this fight, but I'm going to do my best to advance the counter arguments to your arguments here.
- [14:33]
B
Sure.
- [14:34]
A
And, and you tell me what, what you think of them. Okay. And maybe I can do, you know, two in one here. So the depreciation that you're talking about is because 50, I think you've said 50 of the cost of data centers is in the GPU and the GPU has a lifespan. You know, someone, some people say three years, right? This is the typical depreciation argument. You put, let's say a Nvidia H100 in there, three years later, it either fails like you said, or you have to replace it with a black whale or a Reuben, whatever it might be. So, so, and so therefore these, these expenses in the data centers aren't just like you invest $100 billion in a data center and you get to live off the land for 20 years. The investment is, you know, you have to continually, you know, feed that data center with more money in order to make it work.
- [15:28]
B
Which doesn't work in the context of the NPV calculations that underlie a typical real estate project. Obviously that's completely different from the kind of math that we use to justify a multi tenant apartment building, for example.
- [15:40]
A
And then, you know, further on what you mentioned is that the tokens, right, with the things that these GPUs produce, they are depreciating, they are getting cheaper. No one will argue with them that
- [15:51]
B
the, the counter, I'll just, just, I'll just say they're not really depreciating, they're actually staying the same value. They're just deflating. There is a difference.
- [15:59]
A
Okay, right, Deflating, right. The, what you used to pay for a token is, is much cheaper now than it was previously. Okay, so here's, here's what the counter argument would be wrapped up into. One, the counter argument would be, you know, as tokens have gotten cheaper, people have wanted more of them because the, the AI models that used to use them have become a more powerful and be more capable. So therefore, you know, even if those tokens are cheaper people just the demand for the outputs of these factories have grown by a magnitude sometimes, you know, 10, 20, 30x than they were previously. And as you do that, you know, as that demand has grown, you know, people are willing to use even the older chips at rates that would be higher than when they initially came out with the less powerful models. So I was speaking with Core Weave at the end of the year last year or beginning of the year this year, you know, around new year time, and they said they were actually renting out H1 hundreds for higher prices than they had previously. So all of what you said is true. The counter argument that they would make is yes, and their demand for the tokens is higher and the old hardware is working well beyond that typical three to five year estimate that people expected. So what is your thought when people say that?
- [17:21]
B
So there's a whole bunch of nested arguments in there. So let's take them on kind of one at a time. The lifespan of a GPU in terms of just looking at it from an MTBF standpoint, I mean time between failure standpoint depends very much on what it was used for in its adolescent years inside the data center. The analogy I often make is if you could buy a used car, two used cars, one of them, both have like 5,000 miles on them. One was driven in a 72 hour nonstop race across the country. The other one was driven, that was the only, that's where all the 5,000 miles came from. And the other one was driven to church on Sunday per year. Which car would you buy? Well, I think we would all buy the car that was driven to church on Sundays. I want nothing to do with the one that was raced in some kind of bubblegum rally across the country. So in the context of GPUs, what we have is a generation of GPUs that were largely used for very intensive training purposes. And so the failure rates of GPUs used so intensively for training purposes are much higher than in inference specific usage. So yes, there's no question that if a chip is used exclusively for inference, which is to say token completion in response to prompts, then the lifespan will, all else being equal, likely be longer. And if I have a chip that didn't fail during training, then can I repurpose it potentially to be used for inference? Sure, there's no reason. But in aggregate There is this problem. The failure rates of chips that were used for training is very different from the failure rates that were used for inference. So we have this kind of mixed population of chips inside of data centers with very different failure rates. And people have a tendency to conflate this and just pretend that it's all the same thing. And it's not. That's not very helpful. Because if you actually talk to people who are running data centers, they will say this is exactly what we're seeing, is we see much higher failure rates. So there is this sort of blended problem that you have to understand the nature of what the chips were actually used for. And that's only going to become more profound in future because increasingly I often joke that the frontier model company that will, the most valuable frontier model company in future will be the one that's stops pretending to train models and actually just moves on to harnesses and moves up the stack. Because what we're seeing increasingly, if you look at things like the Epic composite index and other things, is that while models are still improving, they're improving at a much slower rate. And I often do this kind of Pepsi Coke test where I'll put a couple of different models in front of people using some kind of a harness like opencode and ask them to tell the difference. And everyone thinks they can tell the difference. And the reality is no one can tell the difference. And so we're at this point of convergence that increasingly the thing that differentiates models outside of marketing is price. Which is one of the reasons why on tables like OpenRouter or whatever else, it's now dominated by Chinese models, we're rapidly seeing this move away from any kind of premium pricing in terms of the models themselves. And I'm trying to get to the point about this kind of Jevons paradox, which is really what you're pointing to, this idea that as models get cheaper, as tokens get cheaper, we use more of them. And this is a, this is a common idea and we've seen it repeatedly play out in different ways over the last 150 years. But I think this mostly speaks to the enumeracy of people that they don't understand what a compounding price decline of 80% means in terms of what you would have to see in terms of growth. On the other side, you have to see around 100 million fold growth over the next six years. In terms of tokens. Is it possible? Absolutely, it's possible. Is it likely? No, it's not likely, but it could happen. But let's not pretend that it's one of the most probable outcomes. To throw out this and say but Jevons paradox, but people will use more is to really dodge the core problem of the geometric decline in the price which will only continue and get faster now that we've got increasingly price based competition because of the convergence of models. So that problem not only doesn't go away, it gets even harder in future. And now you're competing with sovereigns who have state subsidized token prices as China is probably the canonical example. It just becomes increasingly difficult to make the kinds of returns that your investors expect given the comparable cap rates that they're comparing them to. This idea that but it will work itself out because prices will continue to decline and magically we'll just use enough is both historically naive. This argument gets made all the time, has been made repeatedly in prior tech bubbles and people wave their arms and say this and it almost never works that way. And it's worse this time because at the core is this deflating commodity called tokens that is being used to pay a fixed cap rate in terms of what the expectation is from investors who have fronted capital for these instruments. So is it possible? Sure. But think about some of the carnage that it's already creating. Alex Karp was complaining on CNBC the other day. I'm sure you saw it that these companies are increasingly marching up market and trying to eat other. The reason why they're marching up market is because they see this coming and they're looking for higher return places to be because they see the collapse in the fundamental commodity that they're selling. No different than gold miner deciding they need to start making jewelry. This is the same phenomenon playing out. So they're marching and so that's going to have collateral damage in terms of them being seen as fair and unbiased players. Which will then play into the likelihood of companies going down the path of sovereign data centers and doing token inference generation inside their own organizations as that becomes increasingly possible. So I think think there's no doubt that we'll see this continuing growth. But whether or not the growth will be large enough to compensate for what will essentially be an asymptotic collapse to zero in terms of the price of tokens is mathematically a very hard argument to make.
- [23:03]
A
Okay, so this is a great point to dig into as well. So Karp of course went on CNBC and talked about how you can't trust the anthropics, the open AIs of the world with your data because they'll take your data, then they'll build their products that are, you know, sort of compete with yours. Obviously they're competing with Palantir. Right. Because they're going to go and they have these, you know, Palantir has forward deployed engineers. Now OpenAI and Anthropic have forward deployed engineers and they have effectively the intelligence underlying a lot of what Palantir is doing. Right. So if they sort of go up market, you know, you can all of a sudden, and this has sort of always been the fear about these AI companies is their AI would be smart enough that when they see companies building on top of it, they would just go in and take their business. So to me, seeing Karp on CNBC, yes, he was, he was sounding concerned about what OpenAI and anthropic might do to, you know, quote unquote, your business. But he's also talking about what they were doing.
- [24:07]
B
Oh, there's no question to his business. No question.
- [24:09]
A
Yeah. And then just, just from a like, because we're talking about the economics of these build outs and whether these companies will be successful from a pure, like sort of ruthless business perspective, is this the way that they can actually make these investments pay off is they say, all right, well, we have the intelligence because that we will agree that that technology is good in some areas.
- [24:33]
B
No, it's good in lots of areas. And I think. Let me just jump in here for a second because this is a common misconception is that I actually, this is probably the most. Yeah, yeah, yeah, yeah. It's probably the most transformative technology of the last hundred years. And that's, you know, it's obviously a strong claim, but I genuinely believe that. But that's not the same thing as saying that therefore it's fundamentally justified by default justifies the investments being made on its behalf. These are two very different things. And as a matter of fact, the former is almost required for the latter to fail. Right. Because if it wasn't a good story, who the hell would show up with lots of capital?
- [25:06]
A
Absolutely right. And you know, on this show all the time we talk about how like we, we think that this technology is real. And you know, you gotta question the economics because of many of the things we're talking about. All right, but let's just go back to this argument. So they're coming at, let's say they're coming after Palantir, Anthropic is coming after figma. Right. And you know, sort of the list goes on. As, you know, there's always Been this like is, does the entire economy effectively become like a wrapper on top of these AI models? And if so, what's to prevent the AI companies from going out and building into verticals that have been successfully captured by other companies? And so even though it, you know, it's, it's ruthless, etc. It's going to make Alex Karp jump out of his seat and you know, various TV appearances. Is that the route to, you know, having the, the business OpenAI and anthropic, having the business business to pay back the investment? Because if you're able to do that and jump up market, you can potentially justify all this investment by creating these massive businesses.
- [26:11]
B
Yeah, I just think it's reversing the logic and I have no sympathy for carp whatsoever. Palantir can burn or not burn, it doesn't matter to me. But I think it's reversing the logic to say that if the idea is to say I need to find a way to justify the investment, therefore it's okay for frontier models to move upward upmarket, eat the economy and become oligopolies, then I guess that's okay. I think this is a cart before the horse problem because we're not trying to. I'm not in the business of justifying what they're doing by coming up with societally toxic mechanisms that therefore make it work. That's not at all appealing to me any more than it would have been because we heard these same arguments way back in the go go days of Microsoft. Long ago, whenever Microsoft first launched Windows and some of the early operating systems were coming out, one of the things that happened was people built applications on top of the operating system. Microsoft saw those applications were doing really well and gu what they did, they launched their own. Now, most of the ones they launched were garbage. Microsoft turns out for a long time wasn't particularly good at launching applications, but they got better at it and over time ate a host of different applications in spreadsheets, word processors, all over the place. They essentially removed the oxygen supply for all of those different markets and moved up market. So that's not unprecedented. So it's not surprising at all that we'd see these companies do that. The difference this time obviously is we have a generic, a general purpose technology that has much broader applicability. So in theory can do this across a host of other domains, which at the very least should be cautionary and can do it at a much faster rate. The only thing I will say in defense of, in a weird sort of defense of the frontier companies is. And it's the same thing I say with my venture capitalist hat on is we have companies or startups show up all the time that say I have this amazing technology that really has high alpha, should be bought by every hedge fund and so on. And I'm like, okay, fine, why are you telling me then? And they're like, well, what do you mean? And I said, if your technology is so good and you can generate a competitive alpha with it, don't be an idiot. Go out there, raise some capital and invest it directly. Don't tell other people. So the fact that they're telling other people about this alpha generating technology is by default a refutation because if it actually worked, they wouldn't tell me. So the same logic applies to the frontier companies. So if the frontier company's technology is so amazing that it can eat the entire economy, why don't they just ingest the economy and stop selling it to us? Why are they even bothering to sell tokens at all? Why not just move up market immediately? So what that tells you is in the same way that these companies launching hedge fund tools don't actually have things that can do that, the frontier model companies know perfectly well they can't do that. They know perfectly well they can't move all the way up market to generate those kinds of returns. And the proof is in their own behavior.
- [28:56]
A
That was basically Karp's argument, which is why are you selling tokens? If you can increase my sales by 2x, why don't you exactly that up?
- [29:05]
B
That's right. That's right. And it's compelling. It's a compelling argument. You could argue that, you know, it's because Dario at Anthropic is so darn ethical that he refuses to do that. And I guess it's possible. Unlikely.
- [29:17]
A
Yeah. No, I mean, so just to. All right, I mean let's. This is good to go back and forth and talk through the arguments here. The argument would be that, you know, basically this technology is so new and it's moving so fast that it's going to take time to figure these things out. And so you can't just like, you know, on day one that fable comes out or you have mythos in house, you know, go ingest the entire economy. You have to do this like step by step. And a case in point is a cloud design where like Anthropic has been watching the design. Of course, like Mike Krieger used to run product there was on the board of Figma. And that's led to this whole issue and made, like Dylan Field, like one of Anthropic's biggest critics, the, the figma CEO. And so instead of like going out and saying we're just going to do this wholesale, we, we'll do it step by step. We'll. We'll see what the technology is capable of, see how people are using, using it, see what other solutions are out there, and then go ahead and build. And it just goes to this whole concept of the, the, the deflation of, of tokens is like, you know, it seems to me that we're at this point where everybody agrees that these models underneath are commoditizing. And owning the model model is valuable only in the way, only in your ability to customize your own products, to have that like, deep sync between your products and what the. And the models you build that nobody else could have. And that's where this is going.
- [30:42]
B
Yeah, I think that, I think that's broadly true and I think, I don't necessarily agree with you that everyone believes these models are commoditizing. It still feels to me like we're in the, I don't know, iPhone version 4 era where people get all excited about a new release, release, and then everyone whines because they say, well, this didn't change the world and it's not AGI. That's kind of like the fourth iteration of the iPhone, where people want to believe that there's breakthroughs still coming. And there were huge breakthroughs in the early days of launching large language models, but now it's not just commoditizing. But I've, you know, some data that I often show people that shows a kind of convergence that's also happening. So it's not just that there's kind of a plateauing phenomenon going on. It's that the variance among models, the best practices across all of these different models has kind of has converged to a large degree, reducing the variance in terms of the composite performance of various models, which means that the opportunity cost for changing models is much lower. So if the opportunity cost is much lower, unless they can lock me in, there's a huge incentive for me to constantly arbitrage and play back and forth across them, which is hence the rise of Chinese models, why Deep SEQ is doing so well all of a sudden, and why Gwen is. And others, and why open code is emerging as a viable tool for many people. Because there's this sense that I'm not really locked in at all. And the convergence means that the model differ while There are so minimal, as I can't tell the difference in the kind of Pepsi Coke phenomenon, which again, to cut to the investment chase suggests that the competition then becomes much more about marketing expenditure, another form of costs and price. So both of those augur poorly in terms of the investment returns for this asset class.
- [32:14]
A
Correct. And so this is sort of, I think we're both seeing it in a similar way, which is that the economics is going to force these companies to go up market. That's where things get interesting.
- [32:26]
B
Yeah. And I think that's going to accelerate with these. Assuming the IPOs happen, that's going to accelerate with the IPOs, because public markets investors will look at the underlying economics of this fund of the commodity called tokens and say, so what else you got? Right? And say, what are we going to do next? What are we, what markets are you going to move into? And so that's going to increase the pressure to do this absorptive move up market. And then you get into this problem that, and this was my complaint early on about the SaaS. SaaS apocalypse earlier this year is there's a deep misunderstanding about why companies buy software. It's not because they think ServiceNow or Salesforce or whoever is somehow, you know, bold innovators that could not be replaced. No, it's because they have a problem. They don't want to build it themselves and they want someone to sew or shout at. That's it. That's why people buy SaaS software. It's not. And so whenever you start building it for yourself, this notion that companies are going to increasingly use these frontier tools to build things for themselves, themselves or vice versa, that the frontier companies are going to be sued and shouted at by everyone on earth for building vertical apps for them. This will rapidly be disabused because it is a terrible business. You do not want to be in that position of continually having to service people whose main utility for your product is having someone to shout at or sue, which is again, it's a gross exaggeration, but it's a misunderstanding of why verticalized software exists, exists and why those companies exist to service the peoples in those verticals and to just naively say the frontier models companies will blively race up market in, you know, in service of their new public investors is to misunderstand why those markets exist in the first place.
- [34:01]
A
Right. I mean, maybe they'll have to though. That's, that's the thing.
- [34:04]
B
No, no, they'll have to. But that's my point is that It'll just, it won't be easy and even more importantly, it'll probably be very painful and costly. And so be careful what you wish for, I think is where you get to on that one.
- [34:16]
A
Okay, let me, let me make one more of the labs arguments and then.
- [34:21]
B
Sure.
- [34:22]
A
I actually want to get into some more of the weaknesses that you see and that I see. Okay. The other, the other, you know, you kind of winked at this time. You know, the people that believe this time is different. This is not like super, a super technical argument. This is sort of like the general argument that you might hear here would be somebody saying, you know what, Paul? This is different. You have these labs who've built magical, you know, thinking computer machines. Yes, they're, they're investing a lot. But in tech, what you do is you build an asset, you find a way to scale it through computers in some way and you mark it up and people will buy it because it beats any other alternative. And what you've seen recently is like, even in the past, let's say six, seven months, the capabilities have scaled dramatically. You've been able to like now leave these computers alone and they can code on their own and, and do a decent job to the point where like, they're not just useful for engineering, they're useful for all types of work. And so over time, you know, that there will, despite the fact that so much has been invested, like we said, you know, maybe $2 trillion that are coming between this year and next that will be so economically useful that the business is going to have to work out. And this is sort of why people are rushing toward it. Your thoughts?
- [35:45]
B
Sure. But again, this is a classic logical fallacy of assuming what you're trying to prove, right. So you race ahead and say, it has to work out because I need it to work out. And I'll go more deeply into this whole question of, of the this time is different thing. The corollary to the this time is different thing is there's always something useful left after these moments, right? Then the idea that there's always some useful assets left after the fiber bubble, years later, we could use the fiber for things. Even though half of railroads were eventually abandoned because of overbuilding, railroads are still hugely valuable. That's all true, but it's kind of an on sequitur. Well, of course it's true. We didn't build it because it was useless. We built it because it was useful. The issue is what are the consequences of massive overbusters are building in terms of spiraling consequences in the broader economy. Increasingly, some of the largest purchasers of data center related data, insurance companies. We know what happens whenever insurance companies get in the middle of this stuff. We've seen it in the global financial crisis. We've seen it repeatedly. So the right question is not pat people on the head and say this is all going to work out. Because it's always worked out in the past one, while it's always worked out in the past, it's nearly taken out the global economy at least four times. Times. So that's worth noting. And the other issue is, and this is, I think, the more insidious one that people miss because you'll see people refer to, there's this woman named Carlotta Perez who wrote a book called Technological Revolutions and Financial Capital, which in a sense is the bible for many of the most, I don't know, bullish partisans pushing some of this stuff. And they'll say, well, this is what has to happen. It has to, we have to have this kind of huge moment of spending and waste and everything else, but then it works out. Here's the problem with that. Our argument in prior episodes, people didn't know that. That's a really important distinction. We've created this reflexivity where now we justify overspending on the basis of prior overspending having worked out well in prior episodes where that happened, people were not justifying the overspending by saying, say in rural electrification, you know, this may look bad, but it worked out in railroads. No, no, no, no, no, you don't get to play that game. We didn't have that. So now what's happened is it's become a hermetically sealed yield fly, almost a flywheel in a sense because we're justifying things on the basis of information that we, that we didn't have in prior episodes and using that to justify an even larger overbuild. And that's why the notion that this time is different. It is different, but it's different in a really dangerous way.
- [38:15]
A
Okay, so, so I, I, I hear that. I accept that. The argument that I was trying to like put forth on your plate here is, you know, is a little bit different. I think, I think like the argument that I'm trying to get you to respond to is the Paul, it is AGI man. Like this is, well, yes, this is, you know, so what is your response on that front?
- [38:37]
B
Feel the AGI. So, yes, so the question then that turns into this one of what would you pay for a call option on AGI? This is essentially the. The argument is you cannot possibly overspend because the value. It's like saying what would you spend for a call option on your mortality? Well, mathematically I should be willing to spend anything. Similarly a call option on AGI. There's no discountable net present value that is too large. So once you start down that path, once I accept that premise that this is fuel the AGI or feel the immortality, then again I'm into this trap of well, yeah, absolutely. But the problem is that we're now going along with this cultish idea idea that we both now agree that what you should be trying to approve, I should assume, and therefore we should be willing to spend anything. And it's simply that becomes a toxic board game. It's tennis without a net. Right. There's no way for us to have a reasonable conversation once the other side of the conversation is what would you be willing to pay for a call option on immortality? What are you willing to pay for a call option on AGI? So you don't need that argument. We should be able to make an argument and say that this technology is very powerful and very important and transformable and here's the way it's going to change things without having to have, you know, it's. It's like in classic sort of agnostic theory this idea of inserting God into every gap in an argument where you can't find a good argument. This is a God of the gaps argument and I'm inserting AGI because now that allows me to create an undiscountable call option that I can't price. Therefore I should be willing to spend anything. And I reject that.
- [40:14]
A
Don't you think that all the money that's going towards this AGI or AI buildout, the people writing the checks have been told the AGI argument and therefore, despite all of the economic weaknesses that you've pointed out in our discussion are, are basically writing that check for a call option on AGI.
- [40:40]
B
To a degree. Investors that I've talked to are very cynical. So they're perfectly happy to use that in front of their own LPs, but they don't believe that in house. They look at this very cynically and with very cold calculating eyes and compare it to other similar real estate projects. It's really compared to other sorts of project financing from hydroelectric dams to long lived capital intensive projects. That's the hurdle that it has to clear in terms of promoting it. Sure, we can call these AGI factories. I was talking to a regional economic development official in New Mexico recently who had a hyperscaler show up and tell them, don't you want to be part of AGI factories? I was like, what? This is the pitch you're being made because you're signing up for the future because now you can help us build the factories that dictate the future of AGI. All of these objections you might raise in terms of the kinds of tax advances that they want with respect to water and power and real estate and other things. Doesn't matter because think about the scale of the call option I'm offering you. And it's a get out of jail free card. And it's really, I think, unfortunately offensive. But nevertheless, it's more marketing than anything else. And when I talk to the largest investors who are putting capital into this, they'll use it with their own LPs, but they don't do it in partner meetings.
- [41:54]
A
Interesting. So they don't do it because they don't actually believe it.
- [41:57]
B
No, they don't believe it. Not that they don't believe. I'll put it differently, it's not that they believe it or don't believe it. They just couldn't be bothered caring because
- [42:05]
A
they think they can.
- [42:06]
B
It's non material.
- [42:07]
A
So then why are they. Okay, if you, you're speaking to these folks, you're, you're seeing the, the clear problems here. What is the justification that they make in their mind? Let's say they take everything that you say and they, they sort of, they give it credence. Right. The fact that, okay, we've talked about you have to replace the GPUs, token prices are going down to make the, to make this work, the demand would have to be like 100x what it is.
- [42:35]
B
Or more like a millionx.
- [42:37]
A
But okay, a million X. Okay, let's just say that a million X. Why are you seeing the checks? Because.
- [42:46]
B
Yeah, no, it's, it's a bit crazy, but yeah. So why are they still writing checks? But you might say the same thing. It's back to the Harold Prince line back at the. During the financial crisis. As long as the music is playing, I keep dancing. This is that right? This. As long as the music is playing, they're all going to keep dancing because there is absolutely no incentive as any of the largest capital providers on earth, from sovereigns down to private equity and private credit credit to walk away because you get pressure from ELPs saying why aren't you participating in this? And then even worse, as a sovereign, as a sovereign wealth fund, and I've been inside these folks, is that once you're managing hundreds of billions of dollars, you start looking at opportunities not in terms of their economic value, but in terms of check size. And you say, I need to write a check for, fill in the blank, $100 billion, because I do not want to write $101 billion checks. So this weird filter starts happening where you now these projects are like, look at my friend Saudi, Qatar, I have this project that's perfect for you. You want to write 50 $100 billion checks. Nowhere else on earth can you write it other than these giant data center campuses like the Meta project in Louisiana, or take your pick. And so once you become develop a check size filter, the world starts to twist on its axes. And that's why these projects become even more interesting, because there's just nothing else out there like them.
- [44:06]
A
So I will just respond by saying everything that you just said sounds crazy to me that that is the way people operate.
- [44:12]
B
Oh, I was inside, I'll tell you a funny story. I was inside of a $700 million venture fund at one point that turned down five terrific projects. So this is at a very small scale. So think about it now as a sovereign, because, because the entrepreneur we want, the entrepreneur wanted $4 million and the fund wanted to write a $20 million check and said, you know, I can't do a 4 million dollar check. And they walked away from four terrific projects. And I thought, that is absolutely, to use the technical term, batshit. And then I thought, I'll never see that again. And I've now seen it repeatedly inside of some of the largest funds on earth, looking at projects through the filter of can I write a large enough check? Because I have this much out, I have this much capital burning all in my pocket. So it doesn't mean that I'll just give it to any random, I'll direct checks to anybody for anything. But it does change the way that you filter the landscape of viable investments in a really and truly perverse way. So these point masses of capital worldwide are part of the issue, right?
- [45:06]
A
So to basically sum it up, we've Talked like for 40 minutes so far about the logic of investing in these, in these AI projects. And we've gone through all the logic and I've made all the arguments, you've made the counter arguments. And basically what you're saying it boils down to is this is just a group think and convenience thing, which is why all this much money is going this direction.
- [45:27]
B
There's A huge component of that. There's also this, and I've made this, I make this argument all the time that the largest bubbles in U.S. history usually had either to do with technology, loose credit, government policy, right? Some combination of these things. The US in particular is very good at ones that also include real estate. So we can add that to the mix. So technology, real estate, loose credit, government policy. This is the first moment in US history that sits at the intersection of all four of those. So it shouldn't be particularly surprising that we have people who live in each of those bubbles who feel as if they can justify what's happening on their own bases. So I have real estate investors who see this as a real estate project and they're like, look it, I do long duration, high capex projects all the time. Don't tell me what to do. I have technology people telling me this is the most important technology in history. People always tell us that these things aren't going to. And they always work out. It's the Andrew Siena argument. And then the other piece that's pushing this to a real cliff is that it's also seen as an existential battle with some of our competitors around, with erstwhile competitors around the world like China. So there's this government component where we must win. We must win because to not win is to somehow foreshadow some future decline. And so the notion, the idea of sitting at the intersection of those four forces is incredibly important because you end up with four very powerful justifications, just one of which is the large funds with point masses of capital who need to write large checks.
- [46:54]
A
So the checks will keep coming until the music stops. On the other side of this break, I want to talk about what would cause the music to stop and what happens when it stops playing. We'll be back right after this. Hi, everyone. Alex Kanchwitz here. I want to tell you about a documentary I've made with Gravity to explore the future of AI agent security. Security. To find out if we're truly ready for autonomous agents. I sat down with MIT Professor Ramesh Raskar, former White House CIO Teresa Payton, Michelin's group chief Data and AI Officer Ambika Rajagopal, and Sharon Guy, a former executive at Alibaba. They each offer unique insights into this evolving landscape. We conclude with Rory Blundell, CEO of Gravity, to discuss the path forward. Forward. With Gravity leading the way. Join us on this journey. You can watch the full documentary at the link in the show Notes. This episode is brought to you by DeepL When I sat down with DeepL's founder Jarek Kutliovsky on YouTube recently, we got into the case for specialized AI DBEL voices. What it looks like when the stakes are real time Combat conversation. And honestly, it's something I wish I'd had for my own cross border interviews. Turning a language barrier into a non issue DeepL voice delivers live translation in over 40 languages for virtual meetings and in person conversations, helping people speak in their preferred language without losing flow or nuance. Whether you're meeting with a customer, negotiating with a supplier, or collaborating with global colleagues, it keeps pace with you in real time, easily handling the technical terms accurately, acronyms and product names specific to your business. So what you actually mean never gets lost in translation. And for the builders listening, Deep Bell's Voice API lets you embed real time speech, transcription and translation directly into your products. So go check it out for yourself. You can try DeepL Voice for free at DeepL.com tryvoice that's DeepL.com tryvoice Today's executives are more threatened, more exposed, and more vulnerable than ever before before, corporations spend billions on workplace security. But what happens when a threat finds your executives outside the office? 70% of attacks on executives happen at home or away from the office, and Ironwall understands a terrifying reality. If someone has a grievance against your company, the first place they turn to is Google. It takes them about five minutes to find one of your executive's home addresses online. And if their personal information is sitting on the open web, they're far too easy to find. The team at Ironwall knows this better than anyone. Anyone. They've protected some of the most targeted executives and individuals on the planet for almost two decades. Protect your people with continuous personal data removal, proactive prevention tools, and emergency support. So when someone goes looking for your executives, Ironwall ensures they hit a dead end. Go to ironwall.com bigtechnology Fill in the quick form and request your free risk assessment. The team will show you just how exposed your executives are and how to lock it down before threat reaches their front door. That's ironwall.com bigtechnology stop online threats before they become real world attacks. This episode is brought to you by AvePoint. Everyone's racing to roll out AI right now. Co pilots, chatbots, agents doing real work. But here's the part nobody loves talking about. All that AI runs on your data, and most teams have no single way to see it, secure it, and prove it's under control. That's exactly what AvePoint does for 25 years they've been the trusted layer beneath the world's most demanding data now extended across your entire AI estate, your data, your cloud, and the agents acting on your behalf. It's how more than 28,000 organizations deploy AI with confidence. So innovation scales without scaling risk. It's a single platform instead of a pile of tools bringing security, governance and risk resilience altogether. AvePoint the unifying trust layer for AI. Learn more@avpt.co Big Technology Podcast that's AVPT CO Big Technology Podcast and we're back here on Big Technology Podcast with investor and analyst Paul Kadroski. You can go and sign up for his great newsletter@paulkadroski.com all right, Paul, you know we talked a little bit on the other, on the first side of this, this break or before the break about the money will keep coming until the music stops. I mean, I imagine it would take something dramatic for the music to stop playing. What do you think could be the compelling event?
- [51:26]
B
So the argument I make is, you know, people fall into this trap of saying it's going to be this or it's going to be that. I think it's actually over determined in a statistical sense, meaning that there are so many different ways it can stop that the only thing you can say is that it's going to stop because it could stop because of a macro event that changes the hurdle rate that external capital providers are looking for. If I'm suddenly looking for high single digits and not six and a half anymore, well then all of a sudden data center projects with their deflating underlying token compricing looks much less competitive. So that changes things dramatically given that more than half of data center projects now, or half of the capital for data center projects now are external financing. So that changes things dramatically. So the providers of capital pulling back is an obvious source. Then obviously the post IPO phenomenon of having these companies having to generate competitive returns on the back of a deflating commodity and then moving up market and discovering the returns aren't there as they move up market and they continue to spend aggressively on capex. Investors become unhappy about it very, very quickly, as we know from hanging around this stuff for a long time. So it wouldn't take very much to have people feel like this is, this is a much less compelling investment opportunity than I felt like because they're having to immediately abandon the thing that I thought they were selling and now they're having to move up market and chase applications. I'm not that excited about that anymore. So there's a host of these different pieces. Another one obviously is, and we're seeing rumblings of this already is that as this becomes increasingly the state versus state existential battle that you could see export controls instituted. So we can't use Chinese models, Chinese companies can't use US models. We start balkanizing the market. The balkanized market looks much slower and smaller. Well, I don't know what I'm willing to pay for that. That changes things, government involvement. So we're talking already about anthropic or I guess it was OpenAI having potentially a 5% US share in it. What am I willing, how do I feel about that as an investor? Do I want the US as a co investor in my company? What does that change what multiple should be willing to pay? You can go down all of these paths and if these things are truly capex intensive, much like the railroads are, much like utilities, as I've argued, and we see this already in Microsoft and some of the other hyperscalers, then there's a re rating required. I'm not willing to pay a 30 times price earnings multiple in a company that essentially has a utility class capex usage and sort of an asymptotic decline back towards a more utility like hurdle. Right. So there's so many ways this can break and the way it doesn't break break is if it's actually a call option on AGI.
- [54:00]
A
Right. That seems like the only way because like I'm looking at the, I made a list of like the different arguments that you can make for like the fact that anything these big AI labs are going to sell will. The price will inevitably come down, it won't support the investment. We've covered a number of them. But like the open source models out there can make, you know, the, the proprietary model costs come down. You've talked about this in the past. There are these small models, okay, so you have small models out there that are doing as good of a job in some areas of as the big models and they can bring the cost down. Then another thing I wrote down to Zuck, right, Mark Zuckerberg probably sees it to his advantage to like not have OpenAI and anthropic dominate this next paradigm. And he's already trying. First he started with open source, now he's starting with his new proprietary model, but the cost is like 25%. Then you add in the fact that all these super apps that are coming out, which is sort of like the prayer for these companies, companies on the product side which we just both I think Agreed. Is going to be more important. Well, you're going to have a super app from open AI, you'll have a super app from Anthropic, you're going to have a super app from, you know, who knows what. And all of a sudden you're just like, why am I, you know, paying all this money to use the super app if I could just use a different one for cheaper?
- [55:11]
B
Well, and, and even, and even more. No, no, for sure. And even more fundamentally, these super apps or whatever you want to call, I think of them as harnesses, right? They sit on top and kind of orchestrate what models are doing. So harnesses like you know, cloud code, like Codex, which I think is being renamed. But anyways, Codex, like open code or whatever, all of these increasingly are like. The analogy I use is it's kind of, you've got a bunch of bratty kids and the models are kind of bratty kids and the harnesses are kind of like really, really high functioning nannies. And so they take the bratty kid and they make them actually do useful stuff. Right. So much of the improvement we've seen in the last 18, 18 months has really been about the imposition of harnesses, effective nannies sitting on top of bratty kids and not about the actual structural improvements in the models themselves. And that's a sort of a huge misunderstanding. But it's a reflective of where we're going in the future that investors increasingly are going to look at this stuff and say, well, why am I continuing, why are you continuing to spend a billion dollars on a huge training run for a model that may be out in 18, 18 months? Because let's not kid ourselves, like GPT 5.6 was not a massive training run. This was a relatively modest enhancement on an existing foundational model that was pre trained, pre trained probably two years ago now. And most of the gains we're seeing are harnesses and post training, things like what are called reinforcement learning with human feedback, RLHF, all of these other tools that are now coming in after the fact. Once investors look under the hood and see more and more of this, they'll be questioning why are we spending so much on pre training? Why are you doing billion dollar training runs anymore? If most of the gains and models are coming from post training in RLHF and some of these other tools or even like quantization and whatever else, there's going to be immense pressure for these, for the companies to cut back on that spending, which will have huge knock on consequences across the Hyperscalers and across the board, because that's the food system, the ecosystem they live in, that's the food system they rely on. So it's another way, way that this can potentially fail is when the realization strikes that a lot of this or increasing fraction of this training expenditure could be, could be wiped out with almost zero consequence for the future utility of what we, of what we will get from these models, given the increasing reliance on harnesses.
- [57:28]
A
Okay, and so then, you know, the obvious follow up here is, well, what happens when the music stops playing? You know, if it does.
- [57:36]
B
Yeah.
- [57:36]
A
So then assuming you're right, because we're talking about like again a compressed, massive, massive investment cycle with like many companies betting, you could say their future on it. All this like off balance sheet financing, which we really didn't get into today, but like it's not being financed in traditional ways, you know, we're seeing the return or the highlight of credit default swaps again. So talk a little bit about what happens if this goes under.
- [58:04]
B
Well, so it's in many ways analogous to what happened during the global financial crisis is that you find out how it is metastasized across the economy because there's an increasing fraction of the institutional investor population that mathematically, because this is such a large fraction as a percent in percentage terms of issuance in both high yield and prime, both high yield and investment debt over the last six to 12 months and it's a growing fraction of it going forward mathematically they must be holders of this, this stuff. So you're in it whether you like it or not. And so people are realizing that they're in it even by holding an S&P 500 index fund because of the concentration of hyperscaler and AI related names, which is something like 40 odd percent now. So even by trying to diversify, you're still in it. But it's even more insidious than that because it's sitting inside of what euphemistically might think of as higher grade investment bonds that are increasingly being up by hyperscaler related debt. And then that in turn is rapidly it shows up at a place like Pimco first and then they'll flip it and it ends up inside of some insurance company. It'll be spreading across European banks. It'll be in all the same places that we were startled to find US real estate and CMBS debt and then subsequently credit default swaps and credit CDS squares back in the 2007, 2008, 2009, nine period. So we're Gonna. It's literally following the same playbook, just with a different asset class.
- [59:41]
A
That is scary. So how are you playing it? I mean, you are an investor. Are you like shorting certain things or what is your plan here?
- [59:48]
B
So I very much so. My day job in large part is in venture capital. And so for the most part, we just don't invest in it. We just, it's, it's not obvious. It's not obvious how to invest around AI, because one of the worst things you can do as a venture capital capitalist is get into a marathon where there's a thousand participants, they're all at the start line, they're all well funded and well trained. And it's like, oh my God, I'm going to have to outlast all these people to get to the finish. And so you really have to pick your spots and try to stay away from these sectors where people are concentrating capital and doing it in a way that leads to much poorer returns. So for the most part, we were very active in a host of different areas, but not AI, which is perverse because it's not because we don't believe in AI, it's because we believe it's structurally a terrible place to be as an investor. And on a more personal level, in terms of assets, I just very loathe to haven't committed new capital to any sort of broad index class passive categories in over two years for that reason, because whether I like it or not, prior commission commitments now amount to a much larger commitment to this asset than I would like already. So I'm already over invested in this stuff just by the fact of having a pulse and having some assets in the market. And so you have to be very careful about not having it grow in a really, in a way that you wouldn't otherwise have noticed. So sort of personally and professionally, I take a different approach, but they're all kind of mirrors of the each of
- [61:18]
A
the other right now. This is not an investment advice podcast, I have to say that. But you're not like, for someone with such conviction that this is, you know, all going to come down. You're not taking any like, personal big short position where you could benefit. I mean, you even have a ton. You think this goes down in like a year and a half?
- [61:37]
B
Sure, sure, sure. Yeah, yeah, yeah. So it's not my job, but I mean, you know, and I get more pleasure from sleeping well at night is the honest answer. But I will say so. I'm, I'm, I'm working with two very large Hedge funds who put on some fairly complicated positions that I've been working with for close to a year now. So I am indirectly all biases on the table very closely associated with a couple of very large trades related to this stuff. And, you know, if it works out, it works out for me and them.
- [62:01]
A
I see. Class you2 more before we go.
- [62:04]
B
Yeah, sure.
- [62:04]
A
Or do you have to run? Okay, I just want to ask you about, about China. So China has a very different approach here, right? It is effectively. It is the. You could tell me if I'm wrong here. From my understanding, the government is basically backing a lot of this. So if everything goes up, it's just like government allocation, you know, wasted. Which like is that's, you know, happens around the globe on a Wednesday, right. It's not like, you know, their whole system collapses. So do you think that they're insulated much more than the US System, which relies much more on the private investment in data centers and AI training, You know, given like, let's say this all starts to collapse, China will still have the technology. Technology. And you know, that's ultimately. Right. I think we agree. What, you know, I mean, what matters in the end will be like, actually, I don't know, I don't want to put words in your mouth, but like, effectively what we're left with is pretty important. So China might have no economic collapse and, you know, a technology inside of this thing that they can keep investing in it and, and feel good about to a degree.
- [63:15]
B
The problem that China has, and I've been doing a lot of work on China lately, is that much like what happened with battery manufacturing, much like what happened with solar, that there is this huge incentive across the country to impress the central government by building these things locally. And so the Chinese premier has been recently cautioning the provincial governors stop building so many data centers because this is now the new thing. It was battery plants, it was solar cell manufacturing, and for a while 40 years ago, it was hydroelectric dams. So China has a history of under, of consumers underspending and regional and central governments overspending and which led to, you know, massive investment in real estate and ghost cities. I expect you'll see the same phenomenon, albeit with a little less social consequences once all of this turns out to be. It'll be much like what happened with their overbuilding in apartment buildings and residential and industrial space over the last decade,
- [64:07]
A
which they, I mean, they definitely shook a little bit, but they didn't crumble from it. Which is instructive.
- [64:12]
B
Which is instructive. And I think it'll look a little like that. And that's in large part because this is an economy not as reliant on consumer spending as Western economies in the US in particular.
- [64:21]
A
Yeah. So Paul, can I summarize what your position is, which is basically, we have a technology here that is commoditizing, that is effectively any move beyond just selling, sort of like pure intelligence is not going to be easy. And alongside that we have a data center build out that is getting a ton of money based on this prom on the promise that it will pay off. But ultimately we will be much more expensive than people anticipate. And that is going to lead. Those two factors combined will lead to an inevitable collapse.
- [64:58]
B
Yeah. And the only piece I would add to that is that structurally one of the reasons why the data centers will become an even more fraught business, even with all of the other pieces working out, is that this technology came to market faster than any technology in modern history and reached a billion users faster than anything. So under the hood there are vast inefficiencies. The technology industry is very good at wiping away vast inefficiencies, whether it's through the launching of new silicon or improving in software compilers or anything else. So that's all a long way of saying that we should expect token deflation to continue and even accelerate in future because of how quickly this stuff came to market and how many opportunities there are to drive efficiencies. So that creates just incredible pressure on the underlying economics.
- [65:43]
A
Okay, last one for you. What does AI look like after all this? Like, is there, you know, even though that there's a problem, there will be problems ahead. In your view, there will be a winner, the technology will continue to advance, do you think? I mean, you know, I, I know that investing based off of a call option on AGI might be unwise, but do you think that there's a chance that that is where this technology goes? What does the future look like in your perspective?
- [66:08]
B
Not this generation of technologies. There's a deep structural problem with large language models that they can't, they can't easily update the model weights in real time. So in that sense, these are not dynamic systems. And you know, Yann Lecun and others, the former research researcher at Meta, who's now off doing his own world model thing, there's lots of people who will say the same thing. And so I doubt that this is the path, but I do think it's incredibly valuable technology that in a sense will disappear in that it will become like electricity. It's a utility. It will become underlying a host of other things that go on all the time. And I will no more know who provides my tokens than I do from which hydroelectric dam the power came from. From that's powering my MacBook right now.
- [66:49]
A
So the OpenAI and Anthropics of the world, their future is like a power dam.
- [66:56]
B
I don't know where they are or who they are, but I guess they exist and they'll earn utility like rates of return.
- [67:04]
A
Okay, Paul, thank you so much. Really appreciate your time today.
- [67:08]
B
Yeah, sure. No problem.
- [67:10]
A
All right, great. Well, folks, do sign up for Paul newsletter. It's@paulkadroski.com this has been great. I hope we can do this again. Thank you again to Paul, and we'll see you next time on Big Technology Podcast.