Media greetums. I'm at Zetron and this is Better Offline. Today. We are joined by the mighty economist Paul Keadrowski. Paul, thank you for joining me.
Hey, how's it going.
It's going gray. Everyone's deeply upset because this week and the last week everyone has been saying, huh does AI have a return on investment? And it's I've really been enjoying it because it's like watching the dinosaurs look up and see the meteor. They're just like, what do you what do you mean? What do you mean this costs money?
Have you?
I don't know if you've seen the GitHub co pilot stuff.
Yeah, I actually put it a thing on it yesterday.
Oh sorry, Paul, terribly rude to me. That was you. You know me, I've got all sorts of crap on, so I haven't read it yet. But I'm excited to talk about this. I'm really excited. It's really and not only that. I mean there's I'll sort of triangulate with three different things that touched on different aspects of this at the same time. One was, obviously they get up copiled study, which we can get into as deeply as
you want. There was also a piece that came out in part from the Peterson Institute for International Economics.
Yesterday the day before. Jack Cook at Clark at Anthropics set it around who is obviously one of the co founders there, and it's and it's called where is the where is AI? And GDP statistics. And then of course there was the tobacco which I saw anonymous someone that had anonymously disclosed that they had spent almost a half five hundred million dollars because they had uncapped token expenses, and we'll discovered they've sort of blown their credit cards.
So anyways, yes, there's a bunch of flat as well.
Well, let's stop with the GitHub things. So for the uninitiated, get up Copilot, a coding tool from Microsoft. A couple of weeks ago I broke the story, of course that they were moving their users from a premium request model to a token based model. So think of it like
this with listeners. If you every time you use the cab service, you could just say drive me from the Upper West Side to Red Hook and that would just that would be one drive, and you get a certain amount of drives a month, and then suddenly the beginning of June they turned to you and say, yeah, you've got to pay by the mile, and you suddenly realize you've been taking ninety five mile trips. You've been asking to drive from New Jersey to Maryland, which I realized
is further than ninety five miles. Not a geographer, all right, But nevertheless, on the GitHub co pilot subreddit, people have just been posting, what the fuck what do you mean? My whole balance is gone in three prompts? What do you mean by that?
Uh huh, it's yeah, And this is part of the problem, right has this been? You can get all ekon of wonky about this stuff about the merits of metered pricing on a per token basis versus lump sun pricing, but in a sense you can think of this was the early pricing and token in terms of how tokens were metered out, had two really important characteristics. One is they were grotesquely subsidized. You weren't actually seeing the real all in cost with respect to the loaded cost of actually
providing you with those tokens. And then as a kind of don't pay a cent event up there with cost co it was being bundled, so you we had a second layer of masking with respect to what these what tokens were actually costing. And so once it becomes unsubsidized and unbundled, then you see, you know, your ass is dangling in the breeze of real token pricing.
I think it's funny as well, because for years people have been saying to be that's not happening. They're not subsidizing it. It's different. It's just it's like the the cost Go model, for example. People are like, oh, yeah, well they're making money other ways. It's like, no, they're not. They're just selling. In Microsoft's case, they were like, we're going to sell you one one thousand dollars for thirty nine dollars, right, what do you think? You think that's good? Do you like that?
It's lovely, It's a lovely come on. It brings people in. It's like the hot dogs and Costco, except Costco has other things on which they make a b outload of money.
Except the hot dogs cost like seven thousand dollars a packet. It's just I think it's quite deceitful personally. I think it's because these on one hand, we can make fun of these people. I will continue to do so. It's funny, but when you look at them, it is also quite depressing because they were intentionally misled. Like these people had no idea. It's not like these subsidize subscriptions, like, hey, if you use this many tokens while paying for them,
it would cost this much. Until they made the change. Microsoft release the calculator. It allows you to see that, but only once they'd announced it. So you have millions, I would say, the vast majority of people that interact with AI who have no idea what it costs, literally none.
Right and may and which which is made worse by some of the early over excitement, especially among large corporations that made the mistake of creating leader boards and right, So this is we got into this token maxing phenomenon. If so, if you're inside of which is obviously the idea that the more tokens you use, the better you do in your job review, because look at you, you're all AI. The problem, of course, is that this is a little bit like the Saudi's handing out humviies to
everyone in America. Wow, this is awesome. I love having a hum vy. And then you have to fill it up and so for a little while it was like, we were subsidizing these grotesquely uh profligate users of tokens, just like profligate users of gasoline. And then all of a sudden, the bill comes to you and you say, wait a minute, this thing's a pig. It's a lot of gas. I don't want to drive it for growth
through anywhere. And the exact same phenomena, it's true with respect to being again exposed to the having your ass hanging in the breeze of real token prices.
Well, the other thing is as well, is I'll just put our newsletter about this. You look at how these people are freaking out, and you also realize they have no idea what AI cost. It's not just like, wow, this is a lot of money. It's they're not even thinking in terms of cost. It's not like they know, I don't know. They're refactoring something, they don't know how much that got. They don't know how much anything costs.
So it's not like they can smoothly transition to token based billing because they don't they don't know, they have no idea.
And this is this is the deep this is the deep structural problem because they were brought in through the side door of bundled pricing, and now that's that's becoming unbundled, and of course that also has a is reflected in
what we're told. And I think the Wall Street Journal others have written about this, and it'd be interesting we see the final s wyans for some of the upcoming I p os that there is this attempt to try and even mask it in the financial filings where you get into this phenomenon what we used to call earnings
before bad stuff. And so what they're trying to do is hide the costs of training the models and saying that's not actually an operating cost, that's a capital cost, and we shouldn't have to show that as a function of what actually the margins are on producing tokens. And that is of course a cheat, right because if that's true, then you should be able to capitalize these things and
expense them over a long period of time. And we know full well that these are actually operating costs, because they tell us that every eighteen months we're launching a new major model. These things are not capitalized. These are operating expenses that should be treated accordingly with respect to
the actual cost of token production. So there's a multifaceted game going on here, both in terms of how it's being presented to users, but also in terms of how they're trying to sell it in the context of the upcoming s ones for the anthropic and open a IPO pilots.
Well, what's really funny as well about the idea of capitalizing training costs is they're never going away because it's not just free training shoving the stuff in the models. They have to constantly tweak them because that's.
Right and so it's just right, which from a class you know, from a classic my years ago accounting, whenever you have a regular and predictable cost that you have to expense, you have to incur to continue operating your business, that is no longer a capital cost, that's an expensable
item that should be expressed. Such so you get into, as I said, like back in the dark days of dot com and even the telecom boom, you get into this problem of earnings before bad stuff, where they want to they want to exclude all of the things that
make the numbers look bad. And then of course on the other side, you have this run rate problem where we continually hear about what the run rates are at these companies, and the window with respect to the run rate could be the last fifteen minutes for all you know, right, a run rate is just you extrapolate whatever is most convenient for you. So it's a problem on both sides.
Well yeah, actually that's that's what I love talking about run right. Everyone who listens to the show knows I'm a real run right pig because because like I've reached out to both open ai and Anthropic and said, hey, how do you find this number? And they will not respond. They will not They're very unfair to me, very nasty. They will not respond, probably because from what the information has reported. I don't know how open ai does it.
But Anthropic not only includes the amounts of money that Amazon and Google make in their revenues that's like when they resell them a but they also they do thirteen times the last month's APIs spend and twelve times the current days subscribers. So it's just there's so many ways. Also, so token spend, so just organizational token spend. That's not a recurring costs, that's just you can kind of I guess think, well, maybe people are spending this today, but
that person who spent half a billion dollars. That company spent half a billion dollars on AI. Right, that's not happening again, that person is that person is You're not going to get one half a billion dollar miss the bean every single month? Is someone just goofully? I also, genuinely I know the report of Madison Mills. She's respective report, she's very she's she's good, she is well sourced. It's just like I hope Anthropic didn't include that five hundred million in their annualized Yeah.
I'm looking forward to its showing up in a public company filing because it almost inevitably will this is going to be somebody can time item right.
And and then you think that that will though.
Oh absolutely, I mean half a billion, it's material for almost anyone. So I my guess is it's going to show up somewhere. It'll be really interesting to see. And my guess is at that scale, it's a it's a public company. So my guess is we will see that, we will know where that actually happened, and so it's
gonna be it's gonna be very entertaining. But that this is the deep structural problem, and it gets worse of course, because once you unbundled token pricing, and then you're looking at the actual year over year decline in quality adjusted token pricing, in token pricing, and you see that the inherent deflationary curve underneath the hood. Now, let's connect that to how all of these data centers that are producing these deflating tokens are being being constructed and increasing fractions.
Can elaborate what you mean by the deflating token, I'm not sure I understand.
So over the last since twenty twenty two, on an annualized basis, on a performance basis, so ignoring what this continual jump to the frontier, if you imagine sort of on a comparable token basis, per cross models, across the period, token prices have fallen anywhere from seventy to ninety percent year over year consistently back to twenty twenty one.
Right, but they're burning more tokens in the process.
So but let's put that aside for one second. Think about it the alternative the other way around. So if I'm now, my business is now I'm unbundling and I'm selling tokens, and that's the way customers. You're telling my customers to think about it. So now they're looking out and they start to see what happening with token prices, and if I go back one generation, maybe those prices are cheaper. Now we have this classic financial problem of
what's called a duration mismatch. Right, So I have debt funding the data centers that's ten to fifteen years duration and longer, which is predicated on fixed payments but being made on the basis of tokens where you're telling the customer to control your costs. You may want to look back in time and use an older model. So I'm paying for a fixed cost with a deflating commodity right this WEEKNTE from the over and over and over again.
These duration mismatches, especially duration mismatches that are built on top of debt and a deflating commodity, are absolutely atomic with respect to causing a blow up in people's obligations with respect to these these kinds of duration mismatch problems. So there's a there's a deep structural issue that this will expose, and people haven't quite realized it yet.
So you're saying that as the token costs get cheaper and everyone's being encouraged to use this less or more thoughtfully, that's happening. But they're building the data centers as if number will only ever go up and they'll only ever use.
That's exactly right, and so you've got this wonderful ye again the term of art. You got a duration mismatch on top of a deflating commodity that can only end very, very badly, and it was masked because for a while you weren't exposed directly to that. You were just paying
a straight up subscription, almost like Amazon Prime. And of course that doesn't work because Amazon primes costs across the border declining, whereas costs are increasing at the frontier declining in the in the in the back catalog, if you will have tokens.
And that's the thing with like an Amazon Prime for example. Yes, they have found like Amazon or not like Amazon people. I know many listeners don't love them. Then I agree, But it's like Amazon Prime. They fixed those costs by building their own logistics network and they found ways to they they had I don't know, ways to make that cheaper. No one has that in AI, No one like it's just we have three four years in and that one's like, oh, we'll do a six. No, we won't. That didn't work.
Like we have. We're like two or three generations of tranium inferentia TPUs. Still not profitable, still not profitable, right, we would know.
Since we would We wouldn't know. But I think and and and of course the problem is that if you look at I was just looking at some data yesterday with respect to how small language models are increasingly closing the gap with large language models, which is causing training cycles on large language models to have to accelerate, become
more expensive, throw more compute, added more reinforcement learning. The costs are not are particularly not not not declining, they're actually increasing sharply at the frontier because they're essentially being chased like you know, like the rabbits and like wiley coyote and the road runner, and so they're being chased into this very costly corner as a result. And that's
a classic commoditization problem. If you go back to the nineteenth I don't know, late nineteenth century, a very similar thing happened in railroads as people were racing desperately to try and find a way to build a corner and control themselves so that they could compete with all of these other upstart railroads, and of course all that really happened was Capex exploded, margins went to shit, and multiple railroads failed, and we led to the crash of what
eighteen seventy three, eighteen ninety three and arguably was a cause in the Great Depression. So you know, so you're playing out this exact same game because you're sitting in this high Capex world that's increasingly funded by debt and built on top of this duration mismatch, with token prices being now exposed and raw in front of people.
And the other thing is as well, is people are just literally in my piece today, people make this point about, oh, well, it's like the dot com bubble in the we will we will simply just we'll reuse these things in the future, like we'll just pick these up. And it's really I hear this from very smart people, are people who are not like beguiled by AI industry. But it's like, okay, let's talk about what happened to the dot com bubble.
So when it exploded, you had those some microsystems, the Ultra whatever it was, I forget forty three fifty grand a server, but that one server could run an entire company. You could run everything on it, databases, messy CRMs they had like did they have on prem lotus notes? Anyway, you had those things. But you could run that, and you could probably run that in a garage. You might need to use the washing machines plug, but you could
do it if you and that. Yeah, those things were fifty grand, so you probably get them at what twenty thirty large? Yeah, okay, great, what happens from the AI bubble burst? You can't just plug in an AI GPU a B two hundred GPUs but fifty grand. It requires about I think side look this up for a recently. It's like twelve hundred and fifteen hundred watts for a Sun Microsystem server, about the same for single B two hundred, which will require bespoke cooling, a server, hardware, RAM, all
of this other stuff. And then you'll find out that you can't do jackshit with the single GPU.
Yeah, you're gonna be a huge source of disappointment once you power it up. Your neighborhood lights all blink out and you can't still can't do anything. So yeah, I think that's exactly right. But that's but that's I call this I was just I got into this with someone recently who was making a similar I call it faith based argumentation. It's this kind of you know, quasi religious orthodoxy that requires you to believe the following five things.
You know, they always create more jobs than they destroy. We can always reuse assets after the fact. And I'm one of the things I always point to people is that almost half of US railroad lines built during the Boon years in the late nineteenth century were eventually abandoned. And did they find reuse. They absolutely did. It only took one hundred years, and now they're mountain biking trails. You know, let's let's wait around for that.
Railroads. They were railroads that didn't require electrification, like.
That's right, actually, so they were much more stable as assets, right, They didn't have the problems that GPUs and data centers do. Were not just the huge power and cooling requirements, but also they hearent the trajectory of the underlying technology where it changes quickly enough that you know, is a twenty year old you know, blackwell of any use to anyone other than as a paper weight. Of course, the answer is probably not. Whereas a railroad pretty heavy, they are
extremely heavy. I actually was messing around with one recently, and so, yeah, and so this is the problem. And again it's this sort of naive argumentation, not to mention that, you know, the old Canes line that it may be, it may be great in the line run, but in the long run we're also all dad. So it really
depends on your time horizon. And I find it honest, honestly, in the face of the kinds of consequential changes in the US economy, I had a very glib style of argumentation, where you're essentially patting people on the head and saying, you know, don't worry your pretty little head. This will all work out because it always has. And they're arguing from a data set sample size of five, which we wouldn't launch, we wouldn't lunch a drug on that basis.
Also, I think it helps them rationalize bad behavior because if you say, okay, it worked, the one that actually upsets me is, well, the dot com bubble worked out. It's like, yeah, the stock market lost eighty percent of its value, hundreds of pauses, of people lost their jobs, people lost everything in some cases, and at the end, it's like, okay, that was also completely different but you're being quite glib about the first part. But it's also yeah, it's okay that people burn a lot of money for
basically no reason. This is also allows you to not think about bad stuff. It allows you to This is.
The andrey Sonian argument that there's no point in introspection. It's just a really bad idea, right. I think about these things that the law sort and stuff out. But I also think there's a deeper issue. And we may have talked about this before, but the idea a lot of people treat as an article of faith. Carlota Perez's book Technological Revolutions and Financial Capital, and one of the things that they quote take away from that, which I'm not convinced they do. I think they only look at
the pictures. But anyways, one of the things they take away from her book and her work and other people's work with respect to these violent technological revolutions is the idea that it really doesn't matter because it always works out. Here's the difference, though in past episodes, we didn't tell ourselves that. So there is an element of reflexivity going on here, because once you know the plot and you act as if the plot is somehow f equals MA,
it's a law of physics. Then the whole game changes because now you're acting as if it doesn't matter what I do, because you think it doesn't matter what you do, because as you've got this idea in your head as an article of faith that it always works out. That wasn't true. Historically, no one in building out the railroads, rural electrification, the fiber bubble, no one was telling themselves in the time, this always works out. That was not part of the that was not part of the playbook.
The idea that we now tell ourselves these things is such a deep structural change in terms of the way this stuff happens that it amazes me that no one understands it.
Well. I think it's just it's the rationalizing, and it's also it gives you a way of avoiding thinking about true structural issues.
The right I'm sucking, I always call it. It's really kind of nice. It gives you comforuit.
It allows you to be like, well, Google isn't stupid for raising eighty billion dollars in equity sales. Yeh, Google, the largest companies in the world couldn't just destroy their companies by wasting all their money. It's like, yeah, go and type something into Google. Search anything into Google, and tell me if this looks like a company running a good business or a good product. Well, just a company throwing shit at the wall and being like, this works, right, you know.
And we have we have so many examples of companies that were lauded during the run up of prior episodes for being really understanding the way the world works and being a real path breaker and so on, baiting to whether it was the global financial crisis and the banks at that time, and my friend Jim Kramer's unfortunately timed comments about bear Stearns and all that kind of stuff.
Because people are so backwards looking and so extrapolative with respect of the way they look at things, they just can't see the discontinuity, the obvious discontinuities ahead, and so they extrapolate and extrapolate, and then they eventually, you know, extrapolate their way right off, right off a cliff. And I see a lot of that in this going around. And I don't know if you saw it, but there was a paper came out as a good example of this. There was a paper came out yesterday and this goes
to the heart of the token pricing problem. And it came out I think it was on SSR and or Amberg or yeah, National Bureaueconomic Research, and so the paper basically was about how there's been, as you and I both know, there's been this explosion in the number of GitHub commits and repositories or repositories and commits within them, and it's up something like two hundred percent. Of it's just over the last eighteen months, largely driven by harnesses
and everything and all of these coding tools. And of course then they looked at the other side of it was this is this profligate use of tokens, what does it led to? Because producing more stuff that shows up on GitHub is just their need immediate variable, and nobody in the real economy cares other than maybe Microsoft, and even they probably wish there was probably a little less
activity on GitHub. And so they showed that this was just they used iOS apps, Android apps and one other category anyways, and they showed that the number of reviews per app has declined sharply as the number of repositories
and commits has gone up. So essentially what we're seeing and This is the thing that I think is really important is these can be very effective, if wildly subsidized, productivity tools for coders, but the end economic result is mostly the production of sort of slop everything, slap apps, slap content, slop And so you're you're you're flooding and commoditizing these these markets that are becoming both saturated and
declining margins. And this is an incredibly important distinction that just because it's helping you produce more stuff, it doesn't mean that in the broader economy its ability to absorb it as increase, nor does it care. And that's what
this paper shows. And I think the idea that we're doing all of this work and what's increasingly become expensive work using tokens to produce things and makes coders very happy having, you know, things running in agentic loops, but the broader economy, you know, doesn't give a fuck.
And that have By any chance, did you read semi analysis is AI doc output? I think, which is is one of the funniest things I have read in my life. So for the listeners, you'll have a link to this, but it's basically, yeah, AI is, so AI output will be real before it is measurable. We can capture token spend, we can capture jobs lost. But unless AI's output is sold at a visible price, only token spend is captured in GDP. But which they mean, we don't actually measure
whether something is good. We just measure whether something like we just it's actually so that this is the ship a teenager would say when lying about having a girlfriend, like this.
Is voodoo teen economics. Yeah, it really is. And uh and it and again it goes to that that that National Buer of Economic Research paper, it's exactly the same thing I was mentioning at the at the top, there is this tremendous and I can. I'll send you the link if you haven't seen it. And it's called where is AI and GDP statistics Filling the measurement gap? Yeah, yeah, yeah, a couple of days or a couple of days ago. And they argue that essentially AI quality adjusted air output
is up more than two thousand percent per year. They with the estimates of like, you know, two hundred and fifty three hundred billion dollars on top of the and then but they could then essentially come to the conclusion that this is all true as long as you accept our readition of redefinition of GDP, and of course.
Ah, right, right, okay, if.
You allow me to redefine GDP, I could present you with some tremendous numbers, and the entire paper is absolutely fascinating as an example of what's what's often called motivated reasoning. I need to believe this. Therefore I construct an argument to allow me to continue to believe it. And the way I get there is by redefining a variable that's
already very squishy in the first place. Let's not pretend that you know, measuring GDP is much easier than, like, I don't know, measuring muons in a cloud chamber or something. It's still very hard, and we're about to You're trying to make it harder to justify something that's just not defensible.
And the AI dalk output one is great because they substitution dalk output is work that was previously done by humans and he's not done by AI. In our dark output monitor, we have identified roughly one and a half trillion dollars in tasks that current AI could substantially augment or automate, to which I say, why hasn't it done it right. This is the AI thing though, because specifically with AI, with other things, productivity is hard to measure.
It's hard to measure outputs with workers in knowledge work, especially like it's doable, but it's not like a linear path, except you're selling a tool that can theoretically do anything. If this did what they said it did, we would have gunfights in the stream. We would have the destruction of most knowledge work, and it would be happening a year ago. It would be half happening year ago, happening fully today. We would have the destruction of law firms.
We'd have the destruction of hyperscalers, because anyone would just be like, build me a Microsoft word, and it would build them a Microsoft word and they would use it and it would be functional, bug free, all of these things. They would be well. I mean, we've already seen a spike in litigation from pro se people representing themselves. But nevertheless, we would see law firms turning into two or three person shops that would beat the leading litigators because they would.
Have Oh absolutely, it would be very easy to see. I'll give you a.
Related example which made the rounds yesterday, and it kind of gets to the heart of this misunderstanding is there was someone who shall remain unnamed but has a popular newsletter and used to work at a certain venture fund put out something about put out the radiologists the radiologist paradox, which was the idea that back in twenty sixteen, Jeffrey Hinton computer sciences, a really pioneer in image models and deep neural networks, said in a talk that probably within
five years, if not ten years, a large language deep deep at the time neural networks learning models would be better than radiologists and there's really no reason to continue training them now. Of course, he said ten years on the outside, well it's now ten years later, and if you look at the data, we're continuing to produce more radiologists. And that analyst then put out a note yesterday and so did I think KO two or someone else and said, like, well, checkmate,
Jeffrey hinted, Look, we have a lot more radiologists. And of course this is a classic example of a profound misunderstanding of so many things it's once it's hard to keep track. One is that again it's not clear that being a selectively better than radiologists at certain things like identifying I don't know, prostate cancers or whatever else. If that's obviously that's not good enough. Radiologists do more than that.
But it also.
Misunderstands the nature of the employment market because radiologists, like most of medicine, has created a very comfortable little cartel for themselves. So even if there was gale force winds blowing at radiologists because of AI, the likelihood of you seeing it in such a short time, even if Hinton was right that they could in theory replace you know, a significant slice of what radiologists do, it's a misunderstanding
of the nature of the markets themselves. So it misunderstands both the technology and the nature of cartelized employment markets. Whenever you get these kinds of arguments, and yet it's used as an example of how the inexorable march of these things, you know, continues apace and it will always
be augmenting. And I just think there's so many sort of nested misunderstandings of how, how what pressures AI is having on employment markets and how we might see it where it might show up that then to take it up a level, so then do these calculations and say, oh, look, I can now come up with a defensible measure of how the augmenting function is working and then incorporate that ENERGDP. I kind of have to say bullshit, no, you can't because we're failing at the simple stuff.
We also just a very simple response is okay, let's say it can identify them better than radiologist right now? What like the radiologists, it's they don't just look at stuff like they are doctors. They've required like there's more to the process than just like yes or no. And also you are buying the experience. You're buying their experience and their connections and their ability to work within a hospital system.
And actually there's and there's tremendous papers on this treatment, oh absolutely showing how models do we do next?
Right?
Models in general in a medical context, and this is writ large applies to models the use in all complex environments. They tend to over triage trivial cases, meaning that if you come in with like a cut, like dude, this could be sepsis, Let's take you in and start doing tissue biopsies, and it's like, no, no, it's just a cot, leave me alone. And at the other end of extreme, a woman comes in with chess well with pain in her back, which sometimes is indicative of some kind of
cardiac event. They're like, yeah, it's probably just a strain. And so this idea of marching straight through and saying that the only thing that matters is the input data, and therefore I can use these things in these complex environments. We know these tendencies towards over triaging trivial cases and
under triag in critical ones. That's also true. Just as a side note, I gave a talk about this recently to the FED where I was showing how models do the exact same thing in financial markets, where they tend to become over aggressive when they should be conservative and vice versa, which leads to much more fragility and financial markets.
And yet you know, we march on. And this is the deep problem, is this kind of complete misunderstanding of the nature of how these things respond in these complex environments, and then the systemic consequences of doing it. Like, for example, you replace idiologists with something with a tendency to overtrioge, you guess what you're gonna get. Far more testing, much more testing getting done, which may or may not be
profitable for hospitals. But We'll have cascading consequences for people who have to have follow up biopsies because of things that looked like possibly malignant season turns out they weren't. And what we know from medicine is that for the most part, most things should be left alone.
Yeah, and again I keep coming back to the really simple thing, which is if these things were going to replace people, they would just do it. They wouldn't be everything wouldn't. I keep saying this, but everything wouldn't read like the riddler rod tip. They wouldn't just every every single AI jobs thing is like, well, it's AI affected careers that might be doing this in this time in this way. There was a CNBC headline last year it was like eleven percent of jobs can already be done
by AI. But when you looked, it was like, yeah, it was the labor simulates that we made, right, We didn't We just like we'll look at anything like it's the same.
Problem with the with the meter studies. Obviously, in terms of the duration, right, the duration of tasks met right right right those ones and were the duration of tasks where you can get to fifty percent likelihood of completion. And of course if that was a human using that as your as your as your benchmark. If that was a human, I would fire those guys, right. I mean, that's not a useful measurement in terms of how a human might think about a productive co worker. I don't
think about you just half the time you get shit wrong. Right, That would be kind of that would be something that would probably lead to review problems in the at the end of the quarter or a year. And so we do what's the line, Sam Harris's line, This is playing tennis without an ad right.
Yeah, it's it's just as and it's also just we treat these things like the fucking like gifted children. It's like, Wow, you could of the time do this maybe, And that is it's time for the New York Times to write it in hire article of it. We need an odd Lots episode that covers that fifty percent of the time this could do this. And it's just because you can't do the thing that every other obvious innovation is done. You can't do it where you just go, wow, this
does this. We could do this. Now, it's if this happens, and that is a load bearing if we might be able to possibly do this, we can't measure it in the way you measure other things, which is how we would otherwise distinguish whether something was good or not. So we made up a new thing, and wow, has it beaten the benchmarks were made up for it right.
And the problem, of course is is this is all becomes a bit facile and glib but everything else in terms of the arguments being made, But it has spillover consequences in the real world, which is the unfortunate thing is that let's follow the logic forward. If my job is I'm selling tokens, and tokens, I need to sell more tokens rather than less because I have to pay
the not on some fixed obligation. I'm going to construct more data centers and construct more larger data centers, and you end up with these massive mega projects like this controversial one that Kevin O'Leary has been lting right north of Salt Lake City. You know this that in its in the limit might be the size of Manhattan or larger. As people pointing out, this has consequences because the arrow of time only moves in one direction. I defy you to find an example of you know the old Talking
Head song where it. You know, this used to be a parking lot, now it's covered with flowers. The data
center is not going to reverse. Once you build these giant things in the real world with real consequences in terms of you know, sprawling out physically, but also sitting on top of water and power untangling, that becomes really, really difficult, as does, for example, having to spin up all of these new natural gas plants to power these things, because we're increasingly asking the hyperscalers come with their own
power behind the meat. Yeah probably, yeah, right right right when we're doing that at the worst possible time, because the combination of batteries and alternative sources ranging from wind and solar for example, are becoming much more effective and able to be more persistent with battery backup. And yet we're installing these co two intensive things with thirty and forty year life spans funded by debt that are almost
all likely to end up being stranded. Assets like the still be like the statues at Easter Island eventually, except natural gas plans.
Well, that's and this is what I've been saying. It goes back to the dot com thing. I was saying. It's not like an incomplete data center, which I think the vast majority I don't think any of these things get finished. The vast majority of them don't get fully powered like that.
Yeah, I think anything that's targeted over a giggle why doesn't get finished it fully agree.
And the funny thing is with that is people are like, yeah, the dot com bubble, when at burst people had the useful infrastructure that will cost just as much to finish in the future, except that you'll go to a credit for you go to a well probably not private credit in the end of this, but go to a bank like, yeah, I want to finish this data center. They will shoot you with a gun. Will they will. You will get headshotted by the bank manager for saying the words AI.
It's just right and it's going to be. I'll give you an even more. It's even more insidious than that. And I spend a lot of time talking trying to talk off the ledge, if you will, various regional economic
development people. I was just talking to some people in New Mexico about this, and the problem they have is, you know, they've been trying to land some large employer for twenty five years in these high unemployment regions, and so I'm entirely sympathetic to the problem that a data center hyperscaler shows up and says, listen, let me install this, give me the following, you know, giveaways with respect to taxes, and this will eventually, after construction, will have this many
jobs and so on, and you don't have to keep fighting for the Hundai battery factory or the FOD assembly plan or whatever else. It'll just be here spinning off tax revenues. And so what happens is a that looks like a pretty good bet because it's a fixed obligation in terms of what will be flowing back into your county for years to come.
And what do they do?
Then they start pre budgeting that and saying, okay, we'll start building new playgrounds, we'll start fixing the water supply, we'll be able to fund schools.
Great.
Now, okay, you've front loaded all of that stuff. What happens whenever the data center doesn't get finished, you're actually in a worse situation than you were previously. So it has real world consequences in terms of these annuity streams that are being dangled in front of people whose regions have suffered economically for decades, and that's going to be the story over the next twenty five years.
Yeah, it's going to be years of data center collapses. Even after the AI bubble bursts. In my opinion, there's going to just be years of this because you're already seeing a lot of this stuff is speculative. And even then, even if these things get turned on, as you said at the beginning, we are in an era where people are going to be trying to cut back on costs. But then there's the really basic answer, what do more data centers do? What do we get out of these
Because open ai has more compute than anyone. What are they doing? What's different? What's the difference? What does what? What? I keep hearing the term AI factory, and I'm like, what do you mean, what do you mean?
Oh, a factory full of geniuses? That's my favorite.
Oh the data center. Oh Jesus, a data center full of geniuses. I really dislike Dario ama Day. I hate how he sounds, I hate how he speaks, Just like, no, what are you fucking talking about? Because more data centers so far has not actually improved these products. It's not like there's not if you gave open ai another fifteen gigawats of data centers doesn't exist. But let's say they did.
Nothing Like, nothing is going to change about this. Yeah, I don't, and I don't think I But the other thing is as well, Hey, h is Vera Rubin gonna make AI profitable? Because if it isn't, this is probably the last generation. But that is I think at this point the thing.
I think very much so. And I think that's one of the other consequences here that's going on, and I think it's part of the one of the reasons why. And I don't know if you've noticed, but the Jensen has has gone from being very promotional the extra special vary to the third power promotional in terms of I saw today he was anointing Marvel as the next trillion
dollar company. And for me, this is really unprecedented. But it only works if you start thinking about it in terms of the ecosystem of buyers and sellers in the context of AI capex, and realizing that the more valuable all of these companies become, the more money is sort of flowing around this what we used to be called like a captive economy, and then it just recirculates amongst all the players as they become increasingly wealthy because their
stocks get bit up. And so this notion of having people suggesting that one of their sort of peer or quasi competitors should also be valued at a trillion dollars is really unprecedented, and it's only under You can only really understand it once you understand it that they are all essentially running printing presses in their basement, and the printing press is their stock, and they're hoping that the value of the printing press and the currency keeps going
up and that way they can circulate more script among them, which in turn turns into purchasing. And that's the that's the fundamental circularity at the core of all of this.
So as we wrap up, I wanted to get like, because I've already had emails and texts somehow, I don't know how they got my number. What do you think of this? Good? What does this Google thing mean? So Google doing that eighty billion dollar raise at the market, what does this tell you?
Well that they a couple of different things. One is that this is this is the equity raise.
Yes exactly, so ten billion from Berkshire and then some other like ten billion from Berkshire and then I think two different at the market sales.
Yeah, so, I mean, so this tells you that the appetite continues to be incredibly high for their pay for equity, which is surprising because for the most part the funding has been increasingly moving towards credit obviously right, and because the saturation of their cash flows with respect to having to sort of inoculate themselves against all of the other commitments they have, my favorite example being that Microsoft's a good example of this is that their stock based compensation
is so high that they have to which is obviously only handled through cash flows, that the way they inoculate themselves against it is they have to do stock buybacks. And once you start doing that, you've got a much larger commitment of cash, which forces you have to you
pay for hyperscalar data centers. You then have to start doing raises off balance sheet using SPVs and other kinds of funding vehicles so that they're able to do this is sort of surprising to me to a degree that there's still this much appetite for, you know, non credit equity financing of some of their future obligations, because it gives you no call on future cash flows. So what's in it for you as a provider of equity here? It's not clear.
Yeah, is it also assigned that the debt is running out? Like why would they do this instead of raising debt?
So there's no question about that as well. So that's the other side of this is that as of Q one twenty, what a year are we in? Twenty twenty six? I have to look around the room. That's bad. So as of six where the hyperscalers are now the largest issue of investment grade debt on in investment grade markets worldwide? They just passed the banks. So yes. The other answer to this question is is there is a capacity issue with respect to the further issuance of investment grade debt.
In a weird way, they would actually be better if they were issuing junk high yield because there's a higher appetite for high yield, but they just so happen to be currently anyways prime credits, so they're issuing investment grade and the appetite for that stuff is finite, which is why increasingly the marginal buyer for the most recent credit issuances from the hyperscalers is the usual suspects like European
Insurance finds Middle Eastern sovereign wealth. These are the people who famously tend to show up at the end of almost every bubble, and so here they are at the door again.
So yeah, you think, do you think that this is toward the end? I'm just looking for a hard.
No, no, no, no, I think very much that I think the blowoff top is the is these this year's three meg I p O s and and that kind of marks the the gonging of the bell with respect. To take the seriousness with respect, you have to take this inability of these companies to make money.
Paul. It's always such a pleasure to have you. Welcome. People find you.
Paulkadarski dot com is the best place.
Hell yeah, everyone, thank you so much for listening. I'm of course that zytron. You can catch me on this podcast Better Offline, Where's your head at? Subscribe to newsletter my principal form of income. I will be back with a monologue on Friday. Thank you all for listening, and goodbye. Thank you for listening to Better Offline. The editor and composer of the Better Offline theme song is Matasowski. You can check out more of his music and audio projects at Matasowski dot com, M A T T O S
O W s ki dot com. You can email me at easy at better offline dot com or visit better offline dot com to find more podcast links and of course my newsletter. I also really recommend you go to chat dot Where's youreaed dot at to visit the discord, and go to our slash Better Offline to check out our reddit. Thank you so much for listening.
Better Offline is a production of cool Zone Media. For more from cool Zone Media, visit our website cool Zonemedia dot com, or check us out on the iHeartRadio app, Apple Podcasts or wherever and you get your podcasts.
Sp