AI’s Biggest Skeptic Sees a Bubble

Over half trillion dollars that has been put into AI, we're looking at maybe 60-70 billion dollars of revenue. So let's start with open AI. Just the cost of doing business is who are three times what they are making in revenue. Well you can't rely on it, it doesn't seem to be able to do anything. I disagree with that because there is a underlying reality here that hundreds of millions of people are using these tools. I get tremendous value. I'm immediately relieved that we're in a bubble. Do I have to believe that the tech sucks? Ed Zitren, welcome to the show. I've been excited for this conversation because lately everyone has been asking, are we in an AI bubble? This whole conversation is kind of broken containment. You've got mainstream economists, national pundits, even the CEOs of some of the AI companies are now unable to avoid the question. But you have been pushing this question farther and louder and longer than anyone else I know. You're kind of like the OG AI bubble guy. How did you find yourself in that situation? Why were you one of the first and now one of the most persistent? So we'll start in November 2023 when Sam Altman got fired from Open AI. Now there are other people like Gary Marcus was very early on the diminishing returns of large language models, for example, but I saw the reaction to Altman being fired and it was like a cult. All of the members of Open AI doing the Hunger Games thing. It was very strange. And so I thought, what the hell is this? This is weird because I don't know. I don't remember a cult like that for the iPhone or cloud computing. I don't remember the Amazon web serviceistas. So I went and looked into it. It's just generating text and images and I mean they don't seem that good. Why is everyone excited? And then I went one level further and I said, surely somebody will be talking about money. They were not all of these public companies, Google, Amazon, Neta, Microsoft. They were all yacking up the whole thing around AI. But not nobody wanted to say how much they were making. And indeed, no one wanted to discuss how much they were losing. The only thing I can consistently find is it's very expensive. And then in June 2024, Amira Ferrati over at the information, posted this great thing saying, Open AI could lose $5 billion in 2024. And at that point, I joked a fight. I was like, this is ridiculous. $5 billion in losses. That doesn't, that's an insane thing. And then the more I dug into everyone, public companies and private companies, the more I saw was just loss. I saw loss, loss, loss after loss. I didn't really see any business returns. Didn't see great revenue. I saw everybody losing hundreds of billions or billions of dollars. And then the actual products themselves didn't really seem to do anything that magical. I don't remember having to do aesthetics and do a 15-part course just to use an iPhone or a laptop. You mentioned Gary Marcus. There's this whole group of critics of contemporary AI industry who are really focused on making technological critiques. People like Marcus or even Jan Lekun who are out there saying large language models aren't all they're cracked up to be. This is a technological dead end. It's not going to get us to AGI. Then there's a group of people and I see you as one of the leaders there who are making a financial critique and saying the economics aren't there. This is an unsustainable economy. But how intertwined are those? In other words, for me to believe that we're in a bubble. Or can I just believe that it's an investment mania and the underlying tech is sound? So I think that that's a multifaceted answer there. So first of all, you can think the tech sucks without hating it. Large language models on their face are not a bad idea. They do somewhat interesting things in some cases. If they didn't cost 5 to 10 times their revenue, maybe if they didn't require stealing from everyone, maybe if they didn't require destroying our environment and polluting communities with gas turbine engine pollution. I don't know. Maybe if we removed all of these massive land blocks, maybe it would be good. I don't know. I think the technological side is something I've talked about a lot as well because it is fairly simple. These are probabilistic models, even at scale. Every product you use with AI is prompt engineering. They're not particularly precise. You can't rely on them to do the same thing one and once. Even if they do it 900 times, they're probably going to be doing him wrong at least. I don't know. Could be anywhere from 10 to 100 times to 300 times. You literally don't know because that's probability. Yeah. And just to say, anyone who's used these tools is familiar with the limitations too. Right. At the same time, you've got huge numbers of corporations spend bidding billions of dollars in enterprise licensing. You can always critique that and say, "Oh, they're just chasing the next great thing." They don't really understand what they're using. Just quick point. Where are you getting that number from? I think it's been reported in the outlets like the information that companies like Anthropic are largely B2B plays where there are companies or maybe workers at companies who are using them to do things like code. Does that not the case? See, that's the thing though. Right there is a myth. Anthropic does not break out their revenue based on cost. Nobody does. We don't have any idea because these companies do not break them down. Andthropic maybe a month ago said they had 300,000 business subscriptions. What the hell does that mean? Okay. So this could be individual users who happen to work at a company, maybe buying it themselves, maybe getting the small teams, etc. But there is a underlying reality here. The hundreds of millions of people are using these tools. Some of them are paying for them. What are we missing? Because I'll say just for myself, I get tremendous value out of ChatGPT and I'm using it more and more. So there is some underlying value there. But are you just saying there's no way to make a profit off of that? What value are you getting? I'm not actually trying to be argumentative. I'd like to discuss it because it is formative to the point I'm going to make. So you get different answers from different people. But for me, the most value that I get from it is things that are more complex or involved than a Google search, but maybe not quite at the threshold where it would be worth me spending time and money on getting professional advice. So for example, I got married recently. I needed to figure out whether I should do my taxes as joint or separate filing. It's actually really hard to get the major tax software to answer that question for you, like TurboTax. They're not built that way. But neither did I want to ask a professional to guide me through it. It's too expensive and complex. I didn't have the time. So I was able to get ChatGPT by providing them some of my basic figure, which was I already had on hand. So it was able to run a bunch of numbers that I could kind of sanity check. Google can't do that. TurboTax couldn't do that. I didn't want to pay H&R block to do it. What you're describing is search too. You were describing the growth of search. Had Google search grown to a more sophisticated platform versus being made worse so that they could get more advertising impressions, Google search would likely do this. I mean this with no harm or insult though. The idea of turning tax advice over to this is insane to me. That terrifies me. I would always recommend using a CPA. I think my account will kill me. The point is what you were describing is search. It is just the next generation of search. If you Googled should I 10 years ago? Should I file single? I've done this process myself by the way. I know how convoluted it is. Perhaps a page would have popped up from a CPA which had an SEO thing that explained it. You've read that and you'd go, "Oh right, I'm fine." Or I will do this. It is just an outgrowth of search. You can say that. But if you had found a good search result that answered your question, it would effectively be the same thing. I disagree with that because the Google search result will not have my numbers on hand. I will not be able to supply it with the kind of mathematical assumptions that it should be making about my income, my wife's income, student loans, the tax rates in Maryland and in the specific locale that I'm living. All the deductions that I plan on taking what it could do is offer me some generic advice. But then I'd have to wade through a bunch of different pages in order to decide which advice to take. It's one example, but I do think using chat GBT and LLMs has taught me a little bit actually about how limited search is. There are a lot of things that are just in this middle category where it doesn't quite make sense to ask a professional or to give you another example because I love you could take pot shots at this. I find I'm often in a situation where there's an information asymmetry and I'm talking to professional that has a huge informational advantage over me, like a doctor, right? And I'm walking into that situation fairly blind. If an LLM can kind of up level my amount of knowledge that I have of symptomology, potential courses of action, I can have a bit more of a balanced conversation with the doctor. Otherwise, I'm kind of at their mercy. And so this is just a better version of Google, as you're saying. I need to be clear about something. My best friend is a nuclear health and safety physicist and had a clear answer at one point. I must be clear that those subtle, those little, these, what you're describing here in both cases are situations of that could cause astronomical harm depending on how badly you rely on these things. While you may think that the information is correct. How do you know? Sure, you could learn some things that are useful for the doctor on a low end. Sure, you could be told file single or file joint fine. But if you messed that up, you could be audited. Oh, but that's also true. That's also true if I'm interpreting Google results, right? I mean, I ultimately am in verify where that information comes from. At least you can point to somewhere and say that an expert wrote this. At least you can say it's the Mayo Clinic. Oh, sure. At least you can say it's a CPA. With this, it is generating the answer each time. The reason I bring up the safety expert thing is is that Phil, one of the early test CDBIT chat GPT was to ask it how you would put out certain chemical fires multiple times. He was told things that would have exploded. My point is, is the if we're talking use cases, this is pretty mediocre. That's interesting. That's interesting. That's interesting. For the hundreds of billions of dollars, and then if we're talking safety wise, you're talking two different situations, doctor maybe not as much. Doctor, you might get in the doctor and say that's not true. Perhaps you'll have better questions. That could be useful. But there could be a scenario where you reading this much like your tax situation say, I don't need to go to the doctor. I'm good. AI mode on Google or chat GPT told me this. I just feel like and then you take a supplement. And then you took to put someone who I'm not in the habit of quoting. Joe Biden used to say, don't compare me to the Almighty, compare me to the alternative. The reality is, to me, I already have to make health decisions without complete trust in whatever expert I'm dealing with. Doctors have made mistakes all the time. So I'm already going to go into a doctor's appointment armed with what I can read online if I can. A tool like chat GPT can more efficiently do that search and aggregation function and then provide me with citations that I could look into. I mean, I don't want to spend all of our time on this use cases. But I do think it's illuminating. It goes back to the central question of does one's view of the technology necessarily lead to one's view of the financial prospects of this? And I kind of get the sense from you that because you're so downcast in the technology that has something to do with and it's maybe core to your predictions of the economics. Is that fair? I've used the technology a great deal. I am very well versed in how these work. I am very well versed in how the hardware works as well. My views on this are multifaceted and built on a great deal of reading. I think what you are describing is a better Google search, an outgrowth of that and nothing you're describing. However, you may feel about it, makes any of the financial side worth it or indeed makes up for the harms that this could cause. You can feel what like the technology's bad but have hope for the future. I don't know why you would. But I think the financial side just overrides any hope there is. And on top of that, the fact we're running out of high quality training data and we're hitting the walls of scaling laws in the training paradigm. Okay. These models aren't getting better. What we're seeing today is pretty much what they're always going to be like. And I'm not really sure what we have is worth the squeeze. Okay. Maybe we could dig into the finances a bit and there's a lot of different ways that we could do this. But maybe one way is you could pick one or two of the AI companies that you think is maybe most overvalued or kind of most unsustainable trajectory and just walk the audience through some of the key numbers or realities that point in that direction. Open AI, I just reported a story out based on documents I've viewed about open AI's inference spend and their revenue share with Microsoft. I'm going to focus on the inference spend because inference is the way that models generate outputs. You put something in inference happened output comes out. They spent 12.4 billion dollars on inference from the beginning of 2024 to the end of September 2025. Based on what I have seen, they may have only open AI may have only made maybe four point that's it's 2.27 billion through the first half of this year. Okay. And they're like 4.3 billion. Now I'm going to say at least because these are numbers inferred based on the because they pay 20% of their revenue to Microsoft. So these numbers are inferred from there. Nevertheless, we're talking company that just the cost of doing business is two or three times what they are making in revenue. And that's just the inference. That's before training, which I do not have the numbers on, but leaked suggest training is 6.7 billion dollars just for the first half of this year. This company is annihilating billions of dollars to make what? Two four five billion dollars of revenue. You've got the training part that you do with the model that kind of is the initial creation of the model. And that's a huge expense and updating it as well. And then afterward you've got inference, which is kind of operating the model on a going forward basis. And so to your understanding, open AI is currently losing money on just the operations, not even including the kind of the startup expenses. Okay. All the trade. It's not including the training. It's not including the staff, the real estate, the data. None of that is involved in this. Just only inference. This company is annihilating money. Is that your general understanding of the AI industry writ large? Are there companies that as far as we know are making money on a kind of going forward basis, just considering inference, spend versus revenue on just the operations of the model, setting aside training? I don't think it's single one of them is profitable in inference. Okay. I think the whole thing is complete. I think honestly, after seeing what I've seen with open AI's numbers, I don't know who's telling the truth anymore. I'm not saying open AI is misleading anyone. Open AI has made no public statements of any kind about their inference. I'm not saying anything. However, leaks have suggested the inference costs were much lower. So the point is is that I think every single one of these companies is unprofitable just on inference. I can't speak to anyone else. But I will say that a few weeks ago, I published anthropics, sorry, anthropics, Amazon Web Services spend and just their AWS bills were like $2.6 something billion on through September, on like $2.5 something billion worth of revenue. So just the AWS bills. Now with them, I can't tell you if it is inference or not. I don't know what their training spend is. And they spend a bunch of money on Google Cloud as well. So who knows? OK, but the point I'm making is it's very obvious that the cost of inference is not going down. This has been the classic refrain from everyone saying the cost of inference. So creating these outputs, it's going down. And the reason they've said this is the model companies keep making models that they're charging less for. Now it's really easy to say, oh, this means the cost of inference is going down. But what this is is just subsidies. This is corporations subsidizing users. And I think there's going to be a kind of subprime AI crisis eventually. When these companies face with a gruesome margins, eventually have to start charging what it actually costs to run them. And I don't know how that is going to work. But if it happens, and I think it might, you're going to see across the board with start-ups, they're going to see their costs explode. And everybody, I think, yeah, I think basically everybody uses anthropic and basically it's a try to try to intrute tech business strategy to try to grow as fast as you can while subsidizing the customers to a degree with investor money. And then eventually you reach some pivot point where you say, OK, we've really got to be profitable now. You start charging people more. And you also start looking for other revenue streams. So this is my favorite. So the reason I crank my knuckles there is I've had this conversation a lot. So let's talk about the ones that everyone's thinking about. Amazon Web Services. That's the one everyone loves to say. Amazon Web Services. Ed Amazon Web Services. They burn so much money. Burn so much money to make the most profitable thing in cloud computing. Nope. They burned about $68 billion in today's money in the space of nine years. Yeah. That is, let's see, if open AI is maybe 30% of Microsoft's CapEx, we're talking over $100 billion dollars just to build the infrastructure for open AI. Same deal with anthropic. So they have already massively outpaced that. Uber's the other one. People love saying Uber Uber Ed. Uber's the one. Uber's going to know. Yeah. Uber's worst years, I think, were 2020 when they couldn't operate their business. And then they had a future year where they lost, I think, like, seven billion in a year, due to R&D and marketing problems. Like, it's Uber's a terribly wrong company, but they had a path to at least some kind of gap profitability. They also only burned about $33 billion. That's nothing compared to either open AI or anthropic. The scale of this plan to be sure, yes, of open AI and anthropic other AI companies. The scale of the costs are also. Yeah. And investors, patients for that has been very, very striking. And of course, that could run out at any time. But I want to come back to this question of, are there untapped revenue streams that could be turned on in a pinch? One of the things that people have pointed to is advertisements. Chatchy PT, Claude, they don't have advertisements on the models, or as part of the interface. But this is one of the arguments that people will say of, okay, if you really need to turn on the revenue spigots, there's a way to do that. But what do you think? Because of course, Google, Facebook, they make gobs of money from ads. Open AI, we know Chatchy PT does have hundreds of millions of users in some form, whether they're paying or otherwise. That is a substantial asset. Okay, so two things. One, complexity, AI search engine, you're familiar. They turned on advertising in 2024. Do you want to know how much money they made? $20,000. You wouldn't even get a goddamn Super Bowl ticket for that. Wow, what happened there? They made $20,000. Well, I'll tell you what didn't happen. They didn't manage to make much money, and indeed, they ended up shutting down the program this year and their ads head left. Now, here's the question. Why is the most obvious ad-based AI company walking away from AI ads? In my opinion, it's because they don't work so good. Because with advertisers, they require replicability, which LLM is a terrible for, and they require reliability of placement. You may remember when Elon Musk had trouble when he bought Twitter, advertised walking because they couldn't stop their stuff being shown near sexual or graphic. Yeah, brand safety people call this. Yeah, exactly. So that is a massive problem. Sure, OpenAI could turn this on, but where would they go? Would they have banner ads? With any kind of new ad platform, there are always experimental budgets. Companies are not, companies in ads don't tend to just be like, wow, big new thing. $100 million. They'll go, here's a million. Here's a couple hundred thousand. Think about it. When you've got things like TikTok and Facebook and Instagram and Google that all have very good return on investment, why would you take an experimental platform and work with it there? I think you're right that there's a lot of hurdles that would have to be overcome. People would have to be willing to test out a completely new type of advertising construct. I guess your argument would be these are insuperable hurdles that people could not figure this out. There isn't a better way. Maybe they could figure it out, but there is a vast gap between figuring it out and turning this into a massive revenue driver. But they've got to work it all out. It's got to work. Then it's got to actually have a return on investment. Then they've got to wait for advertisers to come in. Then they need to make sure it keeps working. Google has thousands of sales staff for their ads team. They have tons of physical and digital infrastructure. Google owns the publisher, the place you place the ads, and the platform Google search. There is a massive reason that Google does so well. It's their huge monopoly. Same thing with Facebook ads. They have the monopoly on ads on Instagram and Facebook. There is a reason these companies are big. It's not because their products are good. It's because they've created a massive audience and they've worked out how to kind of plumb them. Perhaps if OpenAI had 20 years to work this out, they would be in a better position. But when you're burning billions, I don't think you got that long. Hey, thanks for listening to The World Unpacked. If you want episodes delivered directly to your inbox, use the link in the description. You can also subscribe on our YouTube channel or on popular podcasting platforms. Now back to the show. Maybe there is room for another player serving ads in a different space, just like there's room for both Google and Meta competing against each other, but also basically serving ads. Sure, theoretically. I think it's possible, but I will say, you'll also notice that Google doesn't appear to have worked out ads on AI mode. You don't think, and that's not even being sarcastic or physician-right, I'm just like, if Google can't work it out. No, it's true. It's also a threat to their business too. The more I use, the AI overview at the top of Google. The less I'm clicking on Google's ads, so Google could experience a cannibalization here where their own revenue is getting eaten into. I want to come back to Google and Meta and the big public companies in a moment, but just sticking with the open AI's and anthropics, another argument that people will make as to the path for profitability. You mentioned driving costs down through cheaper inference and training. The other option would be driving revenue up through improved value creation. There, the argument is that the models will get better in the future. The companies are even claiming that we're a year or two away from some kind of recursive self-improvement where the AI's could train and boost themselves, and that the pace of capability growth would not only sustain but radically increase to the point where maybe we might be willing to pay way more for it. You hear stories of how we could have an AI employee in the future. Now, that's obviously a very bold and ambitious vision, but just generally, what do you make of the idea that AI could actually just become much more valuable to its users in the future and we be willing to pay a lot more for it? If my grandmother had wheels, she'd be a bicycle. I mean, sure, you can say that about anything. This table could become a hawk and it could fly out of this room. Nothing about what you're saying is reasonable based on very basics, but these people have been saying this for years. They have hit a wall where they're running out of training data. They are running out of the training data that they need to make the models better. The model's getting better is also something judged by them by benchmarks that are rigged in their favor because large language models are not reliable enough to do tasks. And the user experience to some extent, I mean, I'll say the advent of reasoning models. But we're talking about better. We're talking about better here though. You said better. You said improved. Yeah, you could not do that based on anecdotal data. You actually have to measure efficacy based on tests. You cannot just say, well, I found it good because sure, fine, but the advent of reasoning models allow like test time compute has become the whole new thing that they're claiming will make things better. Why was GPT 5 such such a damp squib then? Why is all of this stuff being so mediocre for so long? Because that's the thing. You're saying, yeah, this could be an AI employee or you can't rely on it. It doesn't seem to be able to do anything. If I needed someone useless, I could find a human to do that. If I like, it's just I'm not even being sarcastic. I'm just saying, let's get tangible here. Let's actually look at the thing and what it can do because if you give them the oxygen to just say in the future, it will be recursively self-learning. I know. I'm not saying that. I know it's a day of sex machina. It's, you know, invisible hand kind of thing. And I will say clearly you and I have different points of view on how valuable current AI is. I feel that it's valuable for me. And so I maybe give it more license when I imagine how it could improve in the future. But I think there's another way that we could look at this. I sometimes think about it as the fake it till you make it theory. And that is that you could believe like Gary Marcus or Jan Lekun that the LLM boom is kind of a bust that people are driving toward a dead end. And yet so much money is pouring into this industry that could provide a bridge towards some shift in the technological paradigm that would allow them to get out of this cul-de-sac. Sure, the big labs must have other types of experiments that they're running on non-LLM paradigms, on new forms of deep learning, on, you know, neuropsymbolic reasoning, other just games that they're playing algorithmic improvements, and maybe just so much money is going into this field that much like the space race, we could kind of manifest technological breakthroughs to some degree. What do you think of that? I don't really know what to say. You're just saying, so just to break down what you said, if they put more money into it, something might happen. I'm not being sarcastic. I'm just saying that's effectively what you said. Yeah, that they surely are working on stuff. Why haven't we heard about it? Open AI said last year that when they have breakthroughs, they pretty much talk about them immediately. Q-star, which eventually became reasoning models, was leaked maybe a couple months beforehand. And the thing is, actually, no, I take the back. Q-star was leaked like eight months beforehand, and they ended up being actually way less important than it was, and they're not really done the same thing. The thing is, you can say surely, but we've actually not seen any sign they're working on anything other than large language models. They don't talk about anything else other than large language models. They train using GPUs on large language models. It's all large language models. It's absolutely speculation on my part. I guess the space race example to me does illustrate that it is possible for humanity to accelerate our progress down a certain tech tree when there is the will and the resources that are put behind it. Some sort of extraordinary acceleration. We weren't destined to invent all of that moon technology. We just we had the will and we put a huge proportion of G.P. and federal spending behind it in a 10 year period. Miracles did come from that. So, yeah, I do think that is the possibility for AI. But that's very different because the moot, first of all, the space race was a long time ago. Dovement funded, not like there was private enterprise involved in it, but there's all sorts of history with like rocket fuel within that that I won't get into. But fundamentally, fundamentally, it's not the same thing and it's way less money. Hundreds of billions of dollars, the attention of every person in tech, every single engineer, the best engineers in the world, the best mathematicians in the world, the best, I don't know. I truly am not a scientist so you can tell I'm running out of words for them. But all of the King's sources and all the King's man, all the money, all the G.P. is every single GPU, every bit of attention from the media and investors and private investors, public investors, retail investors, AI, LLMs, all that. And we are here. I will say one thing you're right about is that we already had the basic science to confirm for us that a moon landing was possible. So we embarked on it with that understanding. That is not the case for AGI or superintelligence. We have not kind of scientifically validated that these goals are reachable in any kind of timeline. If we could pivot from open AI and anthropic, because those are kind of the pure AI plays in some ways, and they're the most attention grabbing, the most famous if you're in this world. But actually, I think when people talk about an AI bubble, they're also talking to a large degree about the big public companies, the so-called Magnificent Seven, Google, Meta, Amazon, Apple to an extent. It's certainly Nvidia, maybe even Tesla. These are the big multi-trillion-dollar companies that have different business models. They're older, more established, have other revenue sources. What's your sense of these companies? Are they equally overvalued or on an unsustainable path? Yes. I think that they don't burn, well, no, they're burning way more because they're capex. But Microsoft, the information is reported repeatedly that they are having trouble selling Microsoft 365 with AI. Microsoft 365 is a cash cow. Over 440 million people pay for it. Based on my reporting, in August of this year, on the 8 million active paid licenses existed. That's do-do. That's terrible. Steve Barmore would throw an error on share at you. You'll also notice something curious. Microsoft stopped reporting their AI revenue in the first quarter of this year. Why? Do you think it's because it's good? Do you think it's because Microsoft is doing great? Why do you think Amazon Meta? Google, none of them talk about their AI revenue. They talk about their capex. I think they're going to do-- Amazon's going to do $116 billion in capex this year. Why aren't they talking about the revenue? And the answer is because they're not doing so good. If they were doing well, they would share it. If they were doing well, it'd be all they'd talk about. When has there ever been a time that a public company is making tons of money off of something where they don't want to share it? They would have absolutely ever been sent up to share it. I guess the question that comes to mind for me is the cloud services aspects of these businesses. Microsoft, Azure, Google Cloud, AWS. I think those are very, very high revenue sources. In the tens or hundreds of billions of dollars. They all put it off for AI. Do we have a sense of what proportion of that is for AI-related activity? Well, I mean, I reported the open AI. I think the 3.10 was $8.6 something billion on inference just on Microsoft Azure. So there's that. The information also reported, but I was not able to confirm that Microsoft sells access to Azure for open AI to run their services at cost, which means open AI is not really making money. They're making revenue for Microsoft. But outside of that, Microsoft doesn't want to talk for even a second for some reason about how much AI revenue they're making. And I think if they're making money off of this, that they're probably not cracking. I'm more than a billion. I'm guessing here, this is all guesswork. I doubt they're making more than a billion or two or quarter each. I think it's probably less. And back a few months ago, an analyst estimated that Amazon would make $5 or $6 billion off of AI this year. And these numbers seem high. They seem really good. Except not in the scale of those businesses. If it's not profitable, and yeah, and when you've got Microsoft making over $70 billion a year on Azure, when you've got, I think, Amazon Web Services, I forget exactly. They're over $120 something billion annualized. I mean, these numbers, the AI stuff is a drop in a bucket, yet they're destroying their balance sheets and they're adding masses of depreciation to their earnings. They're literally burning their income so that they can do this. It's all just, it's a mess. And it's a mess that will eventually come and lie in the, lie in the lapse of retail investors who will be the ones that suffer. Yes. Microsoft, Google, others, they are burning through their balance sheet. But they're largely not financing the data center build out as I understand it with debt. They're beginning to bring some debt to bear on this. But I think a lot of it is being financed with cash flow. These are profitable businesses. And wouldn't they remain profitable even if the AI bubble bursts? I mean, obviously, the stock price would get totally washed out. And a lot of people would lose money as investors. But Microsoft wouldn't go, poof, Google wouldn't go poof. That's actually a really important point. Nothing about what I'm saying is saying any of the hyperscalers are going to go. Okay. Ever I've never said that, I've maintained this is not the case. But Google on November 3rd had a $25 billion bond sale. Metzer, I think, did 17 or 20 billion forgive me for not knowing off the top of my head. Yes. Oracle has massively leveraged their future in debt. And what's crazier as meta just did, I think, a $30 billion special purpose vehicle off balance sheet debt deal with a bunch of people to invest in building the Hyperion data center. It was a cabal of different banks and private equity firms and Blue Owl, who is part of the Crusoe deal for Oracle and Open AI. But nevertheless, this deal is crazy. Because meta has guaranteed they will pay off the debts if they walk away from leasing this data center. So actually, there is going to be, and there is now massive debt associate, but this Microsoft has tons of debt as well. It's just only now is it starting to eat into free cash flow, because wouldn't you know, they're not making a bunch of money for my eye. Yeah, I think we've now seen the first wave of big debt deals. So that does create some risk. Of course, these are huge companies, right? So $25, $30 billion. That is a, you know, it is to some extent the cost of doing business. But I think this gets at part of the bubble question, because when people are worried about a bubble, they're worried about who's left holding the bag if the whole thing comes crumbling down. You mentioned retail investors, but could you just walk us through how we can think about the losers in a bubble scenario? Who would get screwed if there's all of a sudden realization that the AI industry is over extended, overbuilt, over invested, and there's a massive pullback who loses in that scenario. Everyone who invested in Nvidia, that's a great place to start, because Nvidia makes physical things. Physical things, they make actual GPUs. However their strategy may work, Nvidia is in a weird position because last quarterly earnings, and I think they have another one coming up a week from recording this. So forgive me if these numbers change. Last quarter I've seen 55% year-over-year growth. This should lead to everybody popping champagne. The street got pissy. The street said we don't like this, because Nvidia used to grow 100% year-over-year. 100% something, plus 146%, I think one quarter. The market has an unhealthy relationship with Nvidia, which has been an incredible growth vehicle. It was like 8% of the stock market's value, but I think it's higher now. It's astonishing. So you've got, and you've got so much retail money, regular people buying into this, because everywhere they look, CNBC Forbes Bloomberg, everyone's saying, "A.I, this is the biggest thing ever, number go up." And for now that's correct. But because this is all sentiment driven, because remember, there really isn't outside of Nvidia significant revenue in AI. There's not really anything holding this bloody thing up. Everyone's levering into Nvidia. Nvidia is doing really well until it doesn't. And when it crashes, it's gonna be because people get out on the hedge fund level. And then leaving the regular people who can't span Excel at that speed, to panic cell, on top of panic cell, on top of panic cell, at which point they're left with the bag. - Yeah, I think you're right about that. I do worry about retail investors. Even with retail investors, I guess you could still say they made a decision to make that bet. But what worries me even more is the possibility of a financial contagion that could begin in the AI industry and ripple elsewhere, just like we saw in the housing market in 2008. The percentage of the stock market that's in the Mag 7 is so extreme. I think it's been hovering between 35% 40%. Where if you saw, and I think it's conceivable to me, a 50% drop in those stocks, you could then have that translate immediately to a 25% drop almost in the overall stock market and then have all sorts of counterparty risks. Maybe the big banks lose a lot. Maybe the rest of the S&P 500 is somehow exposed. And this could quickly turn into a financial crisis and maybe even a recession. - So based on everything that's come out, we're already in a recession when you remove the money being cloud-entadaisiness. But I want to be clear about something. I don't think this is going to be as bad as the great financial crisis, because the great financial crisis was so heavily levered within people's mortgages they couldn't pay, but also with banks that had been on them and banks on bet on bet with. Watch the big short, it actually does a good job of explaining. But nevertheless, that was also banks and insurance companies started running out of money. I don't think that's going to happen here. I really hope it's not the case. I've not seen signs that this would be the case. So my general thing here is though, is that so much more retail money is in the market than ever. So I think that there will be kind of when E trade first got popular, you had a lot of money. lost from the early retail investors trying to bet on the market. But I think that the contagion to your point, I'm not a stock analyst just to be clear. I think the contagion could be really harsh just because of the amount of weighing weight even the Magnificent 7 has, like you said. I don't know whether it will translate to 25% losses elsewhere, but I do know a lot of people will lose a lot of money, especially, especially the people who would just regular retail investors, users of Robinhood, or what have you, who would just like, oh, I've read CNBC and they say, AI is the biggest thing ever. I'm going to put $10,000 in Nvidia at the top. I feel like we've switched. I feel like we've switched roles in the conversation because earlier you were critiquing AI and I was defending it, but I think I'm actually more worried about the possible consequences of a bubble if one exists than you are just because of the level of market concentration and the fact that we don't really know, or at least I don't know who all the counter parties are, but if it gets into the hedge funds, the big banks, you do wonder what an overall risk retrenchment looks like. And as you said, the AI spend is actually holding up GDP right now. If you subtract that, we could be in a recession in the real economy and that's quite frightening. Typically in the past, when people have worried about whether there is some kind of financial mania or irrational exuberance to use green spans memorable phrase, there is then discussion of should the Fed come in and raise interest rates in order to kind of take some of the wind out of the sales of the market. Now everyone everywhere right now is asking if AI is in a bubble, but I have not heard anyone propose that the Fed raise interest rates and instead there's still this discussion about the Fed being on a path toward lowering interest rates and the only question is how quickly it will do so. But what's your sense of this? Because this is really the main public policy tool that we've used in the past in order to try to pull back on bubbles before they burst violently. So the problem with this bubble, I'm not an expert in this. I've just, I won't talk to interest rates, it's just not my area of expertise, but I will say this. This isn't something that I think interest rate hikes will help with. Sure, it might make more easily available credit, but the scale of the spend required to build an AI data center for example, about $50 billion per gigawatt, two and a half years. So the time horizon of investment on top of these things from private equity and private credit is crazy bad. The amount of money you will have to make back to make this worthwhile is actually kind of insane. Like I calculated recently, the by 2030 big techniques to make $2 trillion of extra revenue on top of what they were already going to make just from AI to make this worthwhile based on their capex and depreciation. This isn't across the board problem. Everybody who has invested in AI data centers needs these things to be at full capacity, full utilization, all the time, paying top rates forever. Otherwise, the whole thing falls apart. Making more credit available only means you're spreading the risk. The risky idea, exactly. In many ways, data centers are a kind of thought contagion themselves. They get into the heads of private equities ago was is the future. This is like, this is buying the future fracking infrastructure. We're just going to print money, except it's not the case, but because we live in a high information low processing society, we have tons of people with tons of money who just don't think too hard. So lower interest rates might free up capital. It might mean there's more credit available for some parties. But at some point, even the most lascivious landers are going to say, hey, when are we going to make money on this? Yeah. How is this going to make money? And I think that that is going to, no matter what the re-interest rates are, that's going to be what people run into. Yeah. Low interest rates are the great facilitator of people making dumb decisions, right? So to the extent that there's already some kind of data center-mind virus, low interest rates make it so much easier to go forward with that. Now, of course, in a way, we really don't have low interest rates today. We have kind of moderately elevated interest rates by the standards of recent history. I just think there's kind of two macroeconomic conversations happening here that really have not collided or integrated yet. Very economists are starting to worry that we might be in a bubble. And yet, there's still this expectation that the Fed will continue to lower interest rates, at least gradually. So if those two things are not reconciled, we could make any bubble that exists far worse or to put it differently because there's this expectation and desire to lower interest rates. We've taken off the table the main public policy tool that we have to ameliorate a bubble before it bursts violently. That's a very difficult situation to be in. I want to pivot a little bit to China if I could because I think the question that comes to mind for me is if there is a bubble, is it just a US bubble or is there a bubble everywhere that Frontier AI is being developed and everywhere that data centers are being built? And of course, China would be the other major global center there. Have you looked into the Chinese AI companies? Are they just as over invested as the American ones? They're actually less invested. There was a, I think it's Joseph Sy, a Chinese billionaire, I believe, who said we're in a bubble. There was a story a few months ago about how there are a bunch of Chinese data centers that are empty. Sorry, they're not being used. And on top of that, we keep getting stories out of China that they're finding ways to train models cheaply, not just deep-seek those. When there's all sorts of different models, it seems China's approach is to make this thing cheaper. If anything, if they have a bubble, which I imagine they do, though I don't have much knowledge of Chinese finances, I think that there's a good chance, there's a good chance they're trying to pivot either away from this or around this in a different way to America. But I will say it is strange. We haven't had a Chinese open AI. And by that, I mean deep-seek, I guess you could call that, but deep-seek isn't signing $1.2 billion worth of compute deals. Making $300 billion deals with Oracle, they're not building, they claim 250 gigawatts of data center capacity by 2030. I think this would really worry me if I were an American policy maker or national security leader who is invested in the AI narrative and beating China because China is, I think, a lot less exposed to bubble risk than we are. There's less stock market concentration in the AI companies in China. I think they're less overbuilt as far as data centers. And there's more of an emphasis on diffusing AI within the real economy and using kind of leaner and meter open source tools to get practical results. So there's a very real situation in which the AI narrative explodes and AI bubble bursts in the United States. But China, whether it's the storm and continues to be more economically sustainable in terms of its AI industry, then it's actually positioned to pull ahead in any future kind of direction this market takes. Yeah, I think that, again, we have the trust what is coming out of China, which of course you should look at with questions. I think there is also, I think the way to look at it is just listen to the story here and listen to the story there. China doesn't appear to be trying to prop up a Chinese company. They're probably up China, but you don't hear them. They might talk about Huawei or what have you or any number of, um, any number of different manufacturers, but you don't have the same rah rah. We must build more China. Sure, you might have Chinese propaganda for sure, but you don't have a, you don't have the same bubble around Nvidia. America's big problem is that we've built this reliance on one vendor that appears to make something with bad gross margins. The information had a report about the B200 GPUs used in a lot of data centers in the GB 200 racks. The GB 200 racks. So 72 GPUs, I believe, negative 100% growth gross margins. Now teach it. You that and business school do they? I think part of what's going on here at is that the US happens to be out in front on the frontier AI race. And so therefore we have built a national narrative around the importance and the glory of that. China is not ahead in the frontier race. They're taking this diffusion strategy and therefore they've built a narrative and strategy around that. That could be the more stable narrative if you're right that this is a bubble about to pop. Well, Ed, you've been very generous as I have thrown objection and objection and argument after argument at you. I want to just end this conversation with a twofold question. One is for people who don't buy the bubble case, what is the single data point that they should be looking at that could change their mind in the future? And then the flip side of that is what data point could change your mind where if some miracle were to occur, this data would actually reveal that we're not in a bubble. We'll start with for the over half trillion dollars that has been put into AI, we're looking at maybe 60, 70 billion dollars of revenue and I don't mean profit. I mean more than likely, I mean 100% the costs of that 70 billion dollars was likely higher than $70 billion. That's my data point. What would change my mind would be, I truly actually don't know because I know so much about the hardware and software behind this. It would have to be, we worked out a way to run large language models for.0001 cent per query. Got it. Some tiny teeny amount. And the truth is, I don't think that'll happen because here's a crazy fact. Three years in, we don't have a good handle on how much these things actually cost to run. That's not something that happens when we're confident in something. Okay, so the data point that could conceivably change your mind as implausible as you might find it is orders of magnitude increase in the efficiency of these models, the training and the inference such that it actually becomes profitable to run them. What about for those people in the audience who aren't convinced of the bubble narrative? What's the thing that they should watch going forward that actually could worsen over time or can approve your case definitively? I mean, the revenue is, as the revenue increases so to the cost that really is. Okay, I've established that very well. That's being established repeatedly and reporting. And I think that story is only going to mature or marinate or I don't know. Get worse. Okay, well, I hope you're wrong Ed, if only for the sake of the global economy and all of our financial futures. But I'll leave that for the audience to decide and we'll find out. Thanks a lot. This has been a really riveting conversation. Thanks for having me. This has been awesome. You've been listening to The World Unpacked, a production of the Carnegie Endowment for International Peace. To get episodes delivered directly to your inbox, use the link in the description or subscribe on YouTube or popular podcast platforms like Spotify or iTunes. Views expressed are those of the hosts and guests and not necessarily those of Carnegie. Learn more at CarnegieEndowment.org.