Plain English with Derek Thompson - Everybody Thinks AI Is a Bubble. What If They’re Wrong?

Episode Date: October 17, 2025

Two weeks ago, in one of our most popular podcasts of the year, the investor and author Paul Kedrosky explained why he thinks AI is a bubble. In the last few days, practically everybody seems to agree.I hate this. I don’t like feeling like my position is the same position as everybody else’s. Conventional wisdoms are often more conventional than wise, and I’ve started to wonder: Is there a bubble of people calling AI a bubble?Today’s guest says yes. Azeem Azhar is an investor and the author of the blog Exponential View. Like Paul, Azeem is a fantastic explainer and storyteller, and I’m satisfied that Plain English has now presented the strongest possible arguments for and against AI being a bubble. If you want to know where I land, you’ll just have to listen to the end of the show. If you have questions, observations, or ideas for future episodes, email us at [e-mail]. Host: Derek ThompsonGuest: Azeem AzharProducers: Devon Baroldi and Kaya McMullen Learn more about your ad choices. Visit podcastchoices.com/adchoices

Hey, it's Craig Horlbeck here to tell you that the NFL is back, whether you like it or not, and we are covering all the latest news, trades, rankings, and more on the Ringer Fantasy Football Show with my two co-hosts who are both named Danny. Check the Ringer Fantasy Football Show out on Spotify or on our new YouTube channel. Today, being wrong about artificial intelligence. Two weeks ago, in one of our most popular podcast of the year, the investor and author Paul Kedroski explained why he thinks artificial intelligence is a bubble.

And in the last few days, practically everybody I follow and read seems to agree. I hate this. I really hate this. I don't like feeling like my position on any issue is the same position that everybody else has.

It's not just me wanting to not belong to the team that everyone belongs to. It's my deep-seeded hunch that conventional wisdoms are often more conventional than wise. And I've recently started to wonder, is there a bubble of people calling AI a bubble? Today's guest says yes. Azeem Azar is an investor

and the author of the wonderful blog, Exponential View. And before I introduce Azeem, let's review why everybody seems so sure that AI is a bubble today. Go back to 1960s. The Apollo program allocated about $300 billion in inflation-adjusted dollars to get America to the moon between the 1960s.

in the early 1970s. The AI buildout requires companies to collectively fund a new Apollo program not every 10 years, but every 10 months. The hyperscalers, the big tech companies, and the frontier labs are collectively spending $3 to $400 billion every single year to bring artificial intelligence to life.

This is the most that any group of companies has ever spent to do just about anything. The hallmark of a financial bubble is tricky financing. Once you see companies going into debt or devising creative new financial vehicles to build something without a guaranteed return, you should start to be a little bit concerned. I think we're clearly already there. As Kedroski explained in our previous episode, these hyperscalers are creating what are sometimes called special purpose vehicles, SPVs, like black boxes into which they throw some money, and then some private capital firm throws some money,

and then that black box goes off and build a data center. The fact that they're already moving AI infrastructure off their books makes me a little bit itchy about the possibility that they don't want investors to see just how expensive this thing is to build. In September, the company Oracle's stock popped after it announced a future deal with OpenAI worth hundreds of billions of dollars. There was a catch. Oracle won't be able to finance that deal with its cash flow,

with the money it actually makes from its business. it'll have to take on a sensational amount of new debt. J.P. Morgan's Michael Sembalist, who has been on this show before and is absolutely fantastic, described the deal this way. Quote, Oracle's stock jumped by 25% after being promised $60 billion a year from OpenAI. An amount of money, OpenAI doesn't earn yet to provide cloud computing facilities that Oracle hasn't built yet, which will require 4.5 gigawatts of power, the equivalent, of 2.25 Hoover Damns, which America hasn't built yet.

End quote. The AI dream is built on a lot of yets and a lot of maybes. But Azeem says it's too soon to panic. Today we talk about the steel man case that AI is not a bubble yet. Like Paul, Azeem is a fantastic explainer and storyteller, and I'm satisfied that plain English has now presented the strongest possible case for and against AI being the bubble of the 2020s. And if you want to know where I stand on this $10 trillion question,

you'll just have to listen to the end of the show. I'm Derek Thompson. This is plain English. Azim Azar, welcome to the show. Thank you for having me, Derek. So I did this interview with Paul Kodroski on why artificial intelligence is one big, fat bubble. And it was one of the biggest episodes of the year.

And then I published a transcript of that article for my substack, and that was one of the most red newsletters I've ever published. So clearly this is a topic of massive interest and massive uncertainty. And I wanted to give somebody really smart a chance to essentially push back on everything, persuade me that I'm wrong. And your work was recommended more than anybody else's. So here you are to persuade me that I'm wrong, that Paul's wrong, hopefully in a way that is mostly polite.

But if it gets impolite, I'm fine with that too. I think maybe a useful place for us to start is with the very basics here. Azim, what is a bubble? Well, what is a bubble? I mean, that's the $10 trillion question. And I would say I'd probably agree with a lot of what Paul said to you. In my mind, a bubble needs to be defined very, very crisply. It's not just vibes.

So I say it has to have two components. Number one, there needs to be a significant market correction. it needs to be beyond the 20% bare market. It has to be 40, 50%, and one that sustains for a long time, the dot-com bubble market correction sustained for 15 years, the housing bubble for seven or eight years. But the second thing to distinguish this from just speculators getting too excited is that the productive capital investment that drove the bubble in the first place also has to decline significantly. And what we've seen in previous bubbles is it really needs to be 50% or, or more. So I'm looking for two tests, a decline in market valuations by at least 50% for a few

years, and productive capital for this technology dropping by about 50% again for a few years. Well, you've mentioned the housing bubble, the global financial crisis. You've mentioned the dot-com bubble. What feels similar this time around, and what feels different? Well, I lived through both of those, and I remember during the dot-com bubble, somebody broke into my office by climbing up the fire escape to pitch us their business to look for investment. I mean, so that is a moment that I hope will not be repeated. During the housing crisis, I had to get to know the face of Angelo Mazzillo, who was a CEO of Countrywide Financial and a subprime lender with a great tan. And for some, some reason, he was on my TV screen every single day. And for some reason, I still remember his tan. I don't remember that much visually about the global financial crisis or the housing crash, but I remember that man's tan, and it was spectacular. It was absolutely spectacular. And unfortunately, the reverse was true of his mortgages. So there are these, you see these

characters, you see these moments. The similarities with the dot com are that this is being driven. by venture capital, by Silicon Valley, by the promise of a new technology that will reinvent the world. We thought we might create a new nation in cyberspace back in 1999. But, you know, what's really different is that back then, nobody was really using these sites and these services. Outside of Yahoo and CNET and eBay, these things were empty. The line was from the Kevin Costner film, Field of Dreams, build it, and they will come. And, People build these things and nobody came. I mean, Pets.com spent $150 million to make $600,000 a month in sales. I mean, it was really, really confused. So I think that's where the similarity to some extent ends,

which is that what we are seeing with generative AI is that most of us are using chat GPT and many companies are and there's billions of dollars being spent. So that feels distinctly different to where we were back in 2000. And this is where it's confusing for me, because I have to admit that I use Chachbtee. I acknowledge that Claude is extraordinary, and some of these Gen AI tools are fantastic. And at the same time, I read headlines about thinking machines, the AI startup that's helmed by the former OpenAI executive, Mira Murati, which just raised the largest seed round in history, $2 billion in funding at a $10 billion valuation. That company has not released a product.

it has not described to investors what product they're even trying to build. There was a quote from one investor who met with Maradi, who by all accounts seems like an absolutely brilliant woman. And he said, quote, it was the most absurd pitch meeting. She was like, so we're doing an AI company with the best AI people, but we can't answer any questions, end quote. And so when I look at this, I have to admit, this does somewhat remind me of a dot com company that has vast promises for reshaping some aspect of retail, but very little underlying

distance that people can see, or maybe like some housing development that's built in the excerpts of God knows what, Phoenix. It's beautiful and perfectly financed, at least on the outside, and yet no one is moving into the houses. So to what extent do stories like this, at least give you like the tiniest little bit of a spidey sense that, like, wow, there is a lot of frothy expectation and funding going toward very expensive projects right now with no clear proof that they're going to actually throw off revenue. Well, I think thinking machines are a really great example. And I felt my spidey sense tingle as well. It's really just a hard thing to piece together, no business plan, no product, and that kind of valuation. But those things happen in the private

venture capital market from time to time, and they don't really see. spill over in the real world. I mean, three years ago, venture capital fundraising, prices investors were willing to pay was really, really crazy. That stopped. Prices normalized. They're getting expensive now. And, well, that does feel like it's heading towards that, that moment of bubbledom, if that's the right word. But the other side of that is, even these startups are growing like absolute topsy. We've just heard. that cursor, which makes a tool to automate coding, has reached a few hundred million dollars of revenue in this company's only three years old. I invested in an email tool that uses AI to help

you answer your emails. And I remember being really cross with the founders when I looked at their pitch where they said, we'll get to $10 million in revenues in the first year. And I was saying, that's just ridiculous. That's not going to happen. And this is the thing I don't like about your your pitch, they got to 17 million in revenue after about nine months and they're growing faster than ever before because customers want to pay for this. And we shouldn't forget that by the end of this year, chat GPT, that weird experiment that spilled into our lives three years ago, we'll be making about $10 billion annualized at the end of this year. And again, that's real money and that's money that's been made faster than Facebook got to that milestone or TikTok got to that

So, of course, you get these moments of exuberance, but they're also examples where real customers are spending real money on products that they really like. And here, let me argue with myself. I remember talking to the folks from Stripe about what they were seeing on their platform of AI companies versus non-A.I. companies. And they showed me really compelling data that shows that firms who self-describe as AI companies on Stripe are dominating revenue growth on the platform and surpassing the growth rate of any group in that platform's history.

So what's confusing to me, but also fascinating, is that it can simultaneously be true that revenue growth for AI companies is astounding and that the amount of spending on AI infrastructure is so much larger than the amount of revenue that we still might be looking at a bubble. Like we're investing in a $1 trillion economic market, as if it's a, say, $10 trillion economic market.

So that's where I think it's worth being very specific here. And fortunately, your analysis is very specific. You've said there is no one way to answer the question, is AI a bubble? In fact, there are five ways. There are five distinct tests. And what I want to do in the rest of our time together is go through each of those tests. So you define these tests as, and folks might not remember it as I say it now, but they'll remember it later. economic strain, industry strain, revenue growth, valuation heat, and funding quality.

No need to remember those terms now. We'll repeat them a lot through the show, but let's start with number one. Economic strain. This is basically asking the question, is investment in this sector large enough to bend the economy, the same way that railroads bent the economy or the dot-com bubble bent the economy? A third of recent GDP growth can be traced to data center construction right now. That's enormous.

Why isn't that big enough to make us afraid of a bubble in AI? Well, I've looked historically at 18 bubble instances back to the canals in the 1790s and where I could get the data. What I found was that the threshold where it gets really sticky is around 2% of GDP and really problematic at 3%, which is where, the railroads got to in the 1870s. And the reason seems to be, and it's particularly true in a big and complex economy like Americas, that it's a big and complex economy and it can absorb that sort of level. But I'd also add that it's actually quite a good thing at this moment,

where without the buildout of these data centres, the American economy might be heading towards recession. And the buildout involves things that America has stated, and authors like you, and your colleague have stated need to happen, which is getting back to building. When you build a data center, you pour concrete. I was speaking to the leadership team of a 140-year-old American engineering firm that built many of the famous old bridges that you've probably driven across. They're busy building data centers right now, and it's electricians and project managers and HVAC engineers who are there. And that feels to me like it's probably quite a good thing to be happening right now. So, so, you know, I would say that that feels at the economic level pretty

interesting, but let's also recognize there are some other consequences of it. So everywhere where the data centers are being built right now, we're seeing electricity prices rise. So that cost is falling on people who are not benefiting, frankly, from the large venture capital rounds in Silicon Valley. And we are starting to see communities push back. So just this week, a town in Wisconsin called Caledone, rejected a Microsoft data center, and they said, look, we don't like turning farmland into this data center. And I think that that is quite an interesting tension that's starting to emerge around where will the pressure point of this economic strain fall? Will it really be purely that it's just too much of the economy, or will there also be a political dynamic to it that

forces these companies to behave differently? There's two ways, I think, to tell the story. There's two ways, I think, to tell the story that you just told. One is that we used to not build things in the physical world, and now we're building things in the physical world. They're data centers, and that's good because building is good. That's one story. Another story you could tell, though, is that the data centers are pulling resources away from other things we have to build. Housing, residential investment has plummeted recently. Manufacturing employment is down. And one argument that Paul made is that AI and data center construction is like a black hole, a dying star, a death star that's pulling all of the planets around it toward it, all of the labor and the

resources and the capital, all being sucked into AI so that it can't be deployed to, say, build more houses. How do you feel about that interpretation that AI isn't, doesn't just represent new investment, it's crowding out investment that we might find more productive over the long run. Well, I mean, crowding out risk is real and the reason it will be happening is because the rates of return you can get to build a data center are just much higher than the rates of return you could build from building an apartment block. But these things can and often do level out. I mean, what you really want to see is a demand stimulus. I mean, if we take a look at America's energy infrastructure for 25, 30 years, Americans have underinvested in their electricity

grid, in their sources of power generation. And that's broken a very long-term relationship in our economies, that more energy is more welfare, it's more prosperity, it's more income, it's more innovation. And that there are rich companies putting in a kick to say, we need more power is probably going to be net net a good thing over the medium term. But of course, to get there, there is this difficult squeeze right now. And the beauty of building is that there's a lot of learning processes, practices that get developed that are transferable across into other types of building as and when the time comes. So I mean, I agree that there is this problem right now, this problem of focus, but the question is when we look out over three or five years, will we start to see

benefits emerging from it? So it turns not less to be a black hole and more to be an M-class star. Just to make a point really briefly, I heard you say that in the medium or long-term, this could be net good for energy generation. And I think that that qualification is very important, because in the next few months, quarters, years, it seems overwhelmingly likely to me that data center construction ends up raising electricity prices before it increases energy generation. But I acknowledge that both things might be true. I want to get that as something you just said a few minutes ago, which is you compared the buildout of artificial intelligence

to the railroads and broadband, the canals of the early 19th century. But there's a very important twist here, which is that GPUs, the chips that Nvidia and others sell, depreciate much faster than rail. You lay steel in the ground. that's steel in the ground that a train can flow over for decades. You dig the canals, that's a hole that will stay there and the water will bear steamboats for decades.

But GPUs age in dog years, I think, to use your joke. Nvidia is making high-quality chips right now that in two, three years, will not be the frontier. And so these companies, if they want the best kind of pre-training, will have to buy a new set of chips. I genuinely don't know how this is going to work. If these companies have to spend collectively half a trillion dollars every three years on the best GPUs, how are they ever going to make back this money?

Well, I think they'll make back the money if the revenues show up. So we can talk about revenues later in the discussion, but you've put your finger on one of the uglier parts of all of this, which is that of the $3, 400 billion dollars, will go into data centers this year, about 50 to 60 percent will be in those GPU. So we'll call it a couple hundred billion. And the companies themselves mostly keep them on their books for six years. So that's a long time, but it's not a canal or a railroad. And as we know, within three years, they have to move out of frontline service. So I think that there is a bit of an ugliness in the way

that this is accounted for. It's, you know, it is something that we should keep our eye on. Now, they may argue, well, the GPUs make money and we're paid back within two years, which is what it seems to be the case. And they still have some useful life in six. But it feels like it's a bit of a thin argument. And this is one area where I think we could ask for, hope for, but we won't get, you know, better rules around how GPUs should get depreciated, the kind of tax

incentives companies should get because they are effectively operational expenses and greater transparency. You know, somebody sent me an email about this and they said that in China, the accounting rules are such that computer chips are effectively depreciated over a three-year, maximum five-year period. So I think this is a, you know, this is one of those, you know, ugly pimples of this picture that we should just stare at and say this could turn into a problem. You mentioned revenues, and I think we should jump right there. Your second gauge is what you call industry strain. And that's basically an answer to the question.

Are industry revenues commensurate with capital expenditures, right? Are these companies making back the money that they are spending on the infrastructure to build AI? And this is where I'm particularly interested in your case, that this isn't bubble-licious, right? By your own calculation, data center CAPEX is roughly 300. $7,400 billion. AI revenue, as you've calculated, is closer to $60 billion. That's a 6x gap. That is an enormous gap.

How is that not worrisome? Well, I mean, it is worrying, but it's just not where other things were. I mean, in the case of the dot-com, we were an order of magnitude practically away from that in terms of the capital that had gone in and the revenues that were being generated. And the question is really how long can you sustain yourself at that point? The thing to note here is that we had no Gen AI revenues three years ago. And our estimates and the calculations that we've done are quite conservative. We've come in at about $60 billion for 2025. One of the big investment banks has suggested it's closer to $153 billion for 2025. In truth, it's probably somewhere in between there. So we're quite conservative. And I would add to that, Derek, that that $60 billion is just a rough measure to cover the capital expenditure. It doesn't involve margin. It doesn't involve profit. It doesn't involve a whole set of other things that you might need in order to justify this long term. But when you look at the

buildout of technologies during this phase, they are generally behind. on the revenue curve. It's not like building a hotel which you expect to have 80% occupancy the day you finally open it. So that's the kind of historical nature of all of this, that it's a technology, it's got to catch up, 16% coverage for now is okay if we were still there in a couple of years. I'd be a bit nervous, to be honest. And we're going to get to revenue growth in just a second, because the question that's being screamed here is, okay, well, it doesn't matter if revenues are low right now. If they're going to grow by 300% every year, eventually you're going to catch up. You just want to make sure that we pin this one point, because in your entire analysis,

it seems to me, this is the place where you think AI is most susceptible to, most vulnerable to accusations of being a bubble. You write that at the height of the U.S. railroad expansion in the 1870s, KAPX was running about two times ahead of revenue. In the late 1990s, the top of the telecom bubble, KAPX was running four times ahead of revenue. So a little bit more bubbly by that measure than the railroads. AI is 6x, AI cap-x spending is 6x revenue, right?

So I'm sorry to throw a bunch of numbers at listeners, but the bottom line is that by this measure at least, AI is three times more bubbly than the railroads and 50% more bubbly than the telecom buildout. By this measure at least, you would agree, I should make sure that we have this clearly on the record, you would agree this is worrying for now. Like the situation's dynamic, but this is a little bit worrying for now. it's worrying for now, but I will explain the way that I think about this, which is that, you know, imagine you're taking off from the runway in a plane I have to fly on Sunday. Halfway down the runway, you're at 80 miles an hour, and you're praying the pilot doesn't

pull back on the stick because that's going to be really, really messy. And so if we're 80 miles an hour, 90% down the way of the runway, we're going to hit the brick wall at the end of it, and that's going to be horrible. So that's why it feels, you know, I put this in the amber because we need to, in a sense, see what's happening with acceleration and see where that takes us. But this is the piece that really makes a difference. Ultimately, real revenues from real customers who are not being forced to pay, who understand what they're buying is a way of paying for any investment that you make.

And that, I think, is the biggest bet that's being taken. Let's go to the next gauge, because this, it really matters whether the plane, takes off. The question here is, is revenue rising or broadening fast enough to catch up to the infrastructure spending? If the answer is yes, not a bubble. You're just talking about an infrastructure buildout and revenue generation that is both growing basically faster than any general purpose technology in human history, so it's not going to be a bubble. The other explanation would be, or the other prediction would be, nope, the spending is growing way faster than revenue is growing, and that's why we're looking at an obvious bubble. Why don't you start us off on this?

point, what trajectory would the hypers need to keep operating income from plummeting in the next few years in a way that essentially caused a huge sell-off of the biggest tech companies? What kind of revenue growth would be necessary? And where might it come from? I mean, they'll need a lot. They will really, really need a lot. you know, we're talking about 100% a year for for a couple of years in and then perhaps lower

than that, after that, as the baseline rises. And that, you know, where's that revenue going to come from? It'll come from probably three areas. One will be customers, consumers and businesses paying for services and buying up or eating up all those tokens or, you know, paying an application that is doing something for them. Maybe it's managing their accounts. accounting or their logistics or their supply chain. The second thing where they'll start to get their revenue will be on return on advertising spend because AI can target ads far better and meta has given some evidence for this. And the final thing will be, can productivity benefits from AI actually reduce their costs or allow them to produce better products? And that works in two ways.

So in one way, it might be that you save money because you can ship a better product with the same number of developers. It might work another way, which is that in order to keep your product relevant in the market, you've now got to pay someone to provide AI services for it. And I think a lot of business services companies with software on the web are finding this at the moment, right? In order to say relevant, they have to provide an AI add-on, they're maybe charging $20 a month for it, but it's costing them $25 a month, but if they don't do it, they will lose customers. So that's the threefold way in which we get to this number. And I think it's also worth knowing that these are big numbers today already.

You know, the amount that goes on online advertising and buying software is well over a trillion dollars a year. And it's already growing at 14% a year. So I think part of the question is how much could AI increase that growth rate and how much would it displace existing companies who are not using AI? And so there is a journey. I think one of the big investment banks talks about a trillion dollars of revenue a year in 2030, which, you know, it's not a shoe in, but it's not completely pie in the sky. You know, it's somewhere in between. As long as we're talking about revenue growth, I think it's absolutely critical that we acknowledge that some folks think we're on the cusp of the last invention, right? AGI.

which if we invent, essentially, an intelligence in silicon that can do anything, surely one would expect that one of those many things that it does is earn a lot of revenue. So maybe give me a sense of, you know, if this technology really does take a leap in the next two to five years, and we build what some folks called autonomous agents, essentially AI white-collar workers that can do 99% of the job of any white-collar, human in the next few years. How would that change the picture here? Wow. Well, I mean, that would be really grim. It would be a little bit like, you know, feudal Russia, only the Tsar is even richer and the surfs are even poorer. So, I mean,

it's a, you know, when you paint it that way, you know, without there being really significant policy interventions, you'd essentially have a giant vacuum cleaner in San Francisco know, sucking up all of the money in the world through AI systems and even companies that are operating with their own products will be paying so much economic rent to the AI companies that they'll have the slimmest of margins, just like sharecroppers on a 17th century English farm. And I think that that direction absent any policy interventions, and if the technology worked out the way you described it, would probably be a direction. that we would end up traveling. Now, my bet is that the AGI vision won't play out the way people

think it will. I think it'll take much longer to get this technology in the economy. I think it'll just be harder to get that robot that can do anything and everything. And I also think that at some point, as we saw in Caledonia, Wisconsin, people would stand up and say, hey, we need to change this script. We need to push things in a different direction. I do think that one of the disconnects between folks like Paul Kodroski that are writing about AI being a bubble and folks like, let's say, Dario Amadei, the CEO or Sam Altman, folks at OpenAI and Anthropic that are building this technology and believe that they are on the cusp of building artificial general intelligence, a true superintelligence, is that,

you know, look, if you think that you're mere months from inventive, preventing God, like, what time do you have for concerns about a brief economic bubble? It's like, you think we're reaching, like, the end of, like, an end phase of human technological development. Like, you're not interested in the question of, like, oh, is CAPEX slightly outpacing revenue growth? It's like, no, like, behind this curtain that is mere inches from our fingers lies a technology that will absolutely change the world.

That's the only thing that I'm focused on. And so I do think that one of the things that maybe could get us in trouble, us as the economy in trouble, is that the folks building this technology have a faith in it that is psychologically, philosophically distinct than the people who are building or laying fiber optic cable. Even the folks who were building pets.com didn't think that they were ending human employment. People who were building the railroad certainly thought they were doing something quite grand, but they knew. exactly what they were doing. They were taking a carriage and putting it on rail and sending it across the country. Maybe before we go to the next gauge here, I do just want to pause and sit in the fact that one thing that could contribute to an economic bubble is the rather

unprecedented confidence of the people building this technology that the thing that they're building exists in a kind of biblical sense as the last thing we ever have to build. The technological rapture, as we might call it. It is quite odd to feel that sense of messianic belief that comes from some of the bosses and some of the people who work for them. But reality has a rude way of interrupting dreams. It just does. And things end up being messier than we might expect. And I think we could already see that with with generative AI today. So it is true that businesses have adopted it faster than previous technologies, faster than the internet and faster than the PC. It's also true that they're starting to see results pretty quickly and that many of them are struggling with it. But that doesn't mean that getting this technology, getting that imaginary all-powerful AI into our economy, into our hands across the whole of the US and beyond will be as simple as pushing an update to an iPhone.

And we've got quite a few steps before we get there. And what I start to see when I talk to business people, I was in Las Vegas a few weeks ago and I did a show of hands with 300 IT bosses. And, you know, half of them said, listen, we're not getting results now, but we should get results in a couple of years. And when you talk to them, it's really. challenging, practical stuff that is not far off trying to unblock a sink. Because companies are difficult. You don't just drop chat GPT and, you know, in Costco or Walmart and, you know, snap your fingers. It takes time. There's a lot to do. And I think that that has been part of the story of how rapidly technologies diffuse and deploy across economies. And so I take a view that these systems will get better,

they'll get better, quicker, that we have to get prepared for them to be really good. But reality is always more messy than the spreadsheet model. Your next and fourth gauge you title valuation heat. This is basically the question, are stocks overpriced? And my God, are stocks going absolutely bonanzas if they have anything to do with artificial intelligence? I mean, just a few statistics. In the last two years, roughly 60% of the SP500's growth has come from AI-related companies like Microsoft, Nvidia, and meta. And if you look at the reason for why some of these companies' valuations are soaring,

it often seems very circular. So, Nvidia, for example, will say we're investing $100 billion in OpenAI, but in return, Open AI agrees to buy billions of dollars from Nvidia. And then I think just last week, OpenAI signed a deal with Nvidia's rival, Advanced Microdevices, AMD, to buy tens of billions of dollars' worth of chips from them. But in exchange, Open AI is going to get like 10% of... of AMD. I mean, it's like, I will buy a billion dollars from you if you give me a billion dollars worth of your stock. Like, at some point, like this whole thing just seems like some

incredibly interconnected spider web evaluations. I guess I wonder how you see this as similar to or distinct from, say, the last bubble in in tech, the dot-com bubble. I think it's quite different to the dot-com bubble because while there was a lot of circular supplier sub-financing, that was provided as debt rather than equity, which has a technically a different categorisation. And the spending was often sham on the way back. And nobody was using those fibres. I mean, we didn't use those, start to use those fibers that were laid until 2012, 2013 when YouTube was really, really kicking off. So the differences here are that this has done in equity, a lot of these things are tranched, that means that they're dependent on people

reaching specific milestones, and businesses across America and across the world are demanding this technology like we've never seen before. Now, is it ugly? I'm actually looking on my screen at the moment at the spider's web map that you talked about, and it's really ugly, and you should look at that and say, this doesn't feel right. But let me give you one other example where perhaps we're familiar with it and it worked for a long time in vendor financing, which is in about 1919, General Motors set up something called GMAC.

And they said, dealers are finding it hard to get financed to set up dealerships. And Americans have got stable jobs but can't buy the car up front. So what we will do is we've got the best balance sheet, we've got the most cash, and we will finance both dealers and American homeowners. And GMAC still exists today, but within a decade, the loan book of GMAC was the equivalent of half a percent of US GDP. Wow.

So vendor financing doesn't always have to end badly. What it does do is it creates a new set of risks for the participants to behave badly. And then we need some transparency, right, to see what's going on between these, these relationships. How real are they? Are there things that are being hidden? We've got that experience from Enron. We've got that experience from WorldCom. And, you know, for me, this is exactly an area where we need the kind of inspection that journalists have been, have been doing. But right now when I look at it, I think, well, yeah, it could go bad, but it hasn't. And it seems like it's got a little while to go before it might. Is there any reason to worry that a vendor financing model that worked for a stable

economic sector like car buying in the middle of the 20th century might, if applied to a brand new sector like AI in the 21st century, present a kind of risk that we haven't seen before and don't know how to fully map onto historical analog? I think it's going to present all sorts of risks that we can't map and that we will be behind the curve as we were during the global financial crisis on bond rating agencies, not rating these buckets of terrible mortgages correctly. The thing that strikes me about a lot of these deals is that they are at this point still relatively transparent. There is a little bit of economic logic. It's not so much about NVIDIA saying, you're going to buy my chips from me and then I'm going to give you, you know,

money to do it. What's happening here is that Nvidia has the strongest balance sheet and it also has the monopoly on these chips. And it really wants to sell a lot. So in a way, this is almost a logical way of them doing this, particularly as Open AI, which is the one that needs most of these is such a young company. It's massively loss-making. And it doesn't have the kind of balance sheet that lends itself to buying these off the cap. Now, look, that is a very, very positive polyanish reading of what's going on. But it also is one that you might end up with if you were trying to think, how do we make sure we don't run out of fuel to meet our customer demand? I mean, I'm not sure what else you would do if demand is as hot as people claim it is.

Right. Your final gauge, gauge five. You call funding quality and the coming credit strain. By your calculation, about $3 trillion in global data center CAPEX is going to be spent in the next three years, $3 trillion. The big tech companies, which some call the hyperscalers, will be able to cover about half of that. And the bottom line is that the other half, the other $1.5 trillion has to come from somewhere else. Maybe that's somewhere else is private credit, securitized finance, new operators, maybe even governments. a lot of the data center spending, as Paul Kedroski explained

in our previous bubble episode, is happening off book. It's happening in special purpose vehicles where the hyperscalers are throwing some money into a box and private credits, throwing some money into that box. And the box is going off and buying a data center,

but it's off the books of the hyperscalers and that makes it easier to demonstrate that these companies are profitable, even though they're finding ways to finance the guts of artificial intelligence. I mean, once again, maybe this works out. It just sounds incredibly messy to me. How messy does it seem to you? Look, it's quite messy. The big companies who've had time to build profits and get cash are obviously

not needing to do this, but the younger ones, the call-weaves, which is a new hyperscaler or open AI, need to be able to fund this growth that they're seeing. And of course, they're going to find more and more spicy ways of doing it. At this stage, again, just thinking about where we are in the cycle, these things look exotic rather than poisonous at the moment. But the question I think will be what happens thereafter. And is there enough transparency? Can we see what's going on in these SPVs? Can we really see how far the GPUs have depreciated? Because if we can't, then that opacity allows all sorts of shenanigans to take place. And when you look at the way in which these booms turn to bubbles and bust,

and I did do some analysis on this, out of the 18 busts that I had enough data to really drill into, the funding quality was the trigger in nine of them. It was the stressor that broke in half of them. And sometimes it was out and out illegality. sometimes it was just that you couldn't see where the risk had gone, as we found with a global financial crisis. So this is definitely one of the gauges that we need to keep an eye on and hope that the capital markets will do that at this moment where there's probably no one else

to look to to keep everyone honest. Let's try to sum up here, because we've gone through so much. We've gone through the scale of the investment, the fact that investment is running way ahead where revenue is right now, the fact that you're hopeful that revenue will continue to grow at a feverish pace to catch up to the spending. That from a stock value standpoint, these companies are very, very healthily priced at the very least, but also they're building something that has enormous demand. And finally, that it's requiring new exotic financing structures, but you consider the structures to be more exotic than I think what you said toxic or poisonous. Tell me if you disagree with this incredibly simplistic, almost embarrassingly basic summary

of your analysis. If AI revenue grows 100% a year for the next 5 to 10 years, this is not a bubble. If AI revenue can't grow 100% a year for the next 2 to 3 years, we are looking at a situation where investors will reconsider whether these companies can afford to spend as much as they're spending in artificial intelligence, and we might see a drawback that is broadly defined as the popping of a bubble. To what extent is that missing something really, really significant here? Because it doesn't touch on financing quality, which you just said is often very key to this. But what does that summary miss?

I think that's a very, very reasonable summary. The gauge that matters most, is revenue growth. It is the signal that people want this. It's the signal that we can monetize it. And that's the one I'm trying to keep my eyes on and all the little ways that we can because, of course, these are mostly private companies. And if revenue growth doesn't show up, then, you know, we are in a pets.com or a telecom bubble moment where we've built infrastructure that might one day be useful that we can't fill today. And that's often the story of these, these infrastructure buildouts. Azim, I'm grateful that you walked me through your, your gauges.

The truth is, I'm personally still very concerned that AI is a bubble. And for me, the sticking points are probably two things. Number one, I just don't think that incoming revenue will arrive on time to offset these investments. It just seems too improbable that companies like OpenAI become a 20, 50, 100 billion dollar revenue gushers in the next few years, but something quite like that is necessary for them to justify current valuations. And then that leads me, I guess, to number two, which is that I don't think these valuations are going to make sense in three years. And when you add an investment

correction, right, the fact that these companies are going to have to pull back their spending, to an asset value correction, right, the fact that OpenAI is not worth, say, half a trillion dollars, well, that's a bubble, right? An investment correction and an asset value correction, that's a bubble. So just humor me here for the end. Let's say that in the next few years, your gauges move from the green to the yellow to the red, and you and me and Paul and everybody else in the world, everybody agrees that we are looking at a definitive bubble. What do you think a bust in AI looks like? In a funny way, we might be grateful for it.

Of course, there will be stock market prices going down, but what will have happened is that there will be a lot of GPU infrastructure that is out there, computing infrastructure that's out there, that companies, organizations with less money, could pick up at fire sale prices. And those assets will go to those smaller players who might have newer approaches. They may prefer open source. They may decide they don't want to chase after the machine guard. They may decide that pricing needs to be more sensible and there's more financial discipline. We might even see faster innovation alongside democratization with diverse teams living on this affordable infrastructure. So there is perhaps some light outside of the, you know, the doom that a bust might otherwise present us. I like how you added sort of a third

door for listeners and for myself to open. If door number one is AI is a bubble and it's a catastrophe, and door number two is AI isn't a bubble and it's fine, you've outlined door number three. AI might be a bubble, and that's okay too, that in the short run, we will have a price correction, but that price correction will, just as with the dot-com bubble, essentially create cheaper infrastructure for other companies to build something like the YouTube's and Netflix's of the future that took advantage of fiber optic cable that was laid in the 90s, early 2000s, and we will be grateful for this buildout, just as we are grateful for many bubbles that have happened in American history. I'm not telling listeners or you or myself what door to walk through, but it is interesting to consider all three doors.

Azimazar, thank you very much. It's been my pleasure. Thank you. Thank you for listening. Plain English is reported and hosted by me, Derek Thompson. This episode was produced by Devin Baraldi and Kaya McMullen. If you like what we're doing here, please rate and subscribe. New episodes drop every Tuesday and Friday.