AI Skeptic: This Business Makes No Sense

0:00Adobe Acrobat, your team's home base, collaborate within a shared PDF space. You've got your docs, your plans, your specs, and then invite the crew to build what's next. They talk off the teamwork, you can update it, render. They think that this design could be a contender. But when somebody wonders, what's the next step? AI helps you finish the rest. Bolts are tight, now your plan's refined. Run a smoother business when you're all aligned. Do that with Acrobat. Learn more at adobe.com slash do that with Acrobat. Study and play. Come together on a Windows 11 PC. And for a limited time, college students get the best of both worlds.

0:40Get the Unreal College Deal. Everything you need to study and play with select Windows 11 PCs. Eligible students get a year of Microsoft 365 Premium and a year of Xbox Game Pass Ultimate with a custom color Xbox wireless controller. Learn more at windows.com slash student offer. While supplies last, ends June 30th. Terms at aka.ms slash college PC. When you need to build up your team to handle the growing chaos at work, use Indeed Sponsored Jobs. It gives your job post the boost it needs to be seen and helps reach people with the right skills, certifications, and more. Spend less time searching and more time actually interviewing candidates who check all your boxes.

1:17Listeners of this show will get a$75 sponsored job credit at indeed.com slash podcast. That's indeed.com slash podcast. Terms and conditions apply. Need a hiring hero? This is a job for Indeed Sponsored Jobs. Today's number, two. That's the number of times Sir David Attenborough has been knighted for his services to British television and conservation. It appears that after Brexit, the UK has run out of people tonight. Money market's mad. If money is evil, then that building is hell. The show goes on!

1:57Welcome to Prof. G Markets. I'm Ed Elson. It is May 14th. Let's check in on yesterday's market vitals. The major indices were mixed as investors digested new inflation data from the producer price index. Tech stocks drove the S &P and the Nasdaq to new records while the Dow fell. We'll talk about that inflation in a moment. Inflation concerns also pushed the yield on 10-year treasuries higher. and on Calci, the odds of a rate hike this year have steadily climbed to 31%. Meanwhile, Apple stock closed at a record high as CEO Tim Cook joined President Trump, as well as several other big tech leaders for a summit in China.

2:34We will be discussing that shortly as well. Okay, what else is happening? Big tech is on track to spend$725 billion on AI in 2026, and it shows no signs of slowing. Yesterday, reports surfaced that Anthropic is in talks to raise$30 billion at a valuation north of$900 billion. And SoftBank posted a$46 billion gain for the year, fueled by the soaring value of its stake in OpenAI. The markets are currently rewarding the AI boom. The Nasdaq is up 27 % since its March 30th low. Much of the gain comes from the AI trade, with chip stocks recording their best monthly performance in decades, but not everyone is buying it.

3:22Big tech's free cash flow has been shrinking as AI CapEx balloons, and some investors are asking when, if ever, the spending will generate a return. So we wanted to check in with one of AI's fiercest skeptics and critics and ask what could the market be missing. So joining us to talk about the state of AI and perhaps some of the stuff that people are not paying enough attention to. We are speaking with Ed Zitron, author of the Where's Your Ed At newsletter, great name, and the Better Offline podcast. Ed, thank you for joining me on the show. Our producer, Claire, mentioned this to you offline, but I will say it now.

4:04You are probably the most requested guest we've ever had on the show. My audience has been asking over and over again. you have to talk to Ed Zittrain, we have to talk to Ed Zittrain. We're finally here. So thank you for joining me. So I want to start with maybe just a summary of your views. You have been writing for a long time that you believe that a lot of what we're seeing in AI is fake, misleading, not what we think it is. If you could just start with a summary of what you believe is really going on in the AI world right now. So across the board with public stocks, nobody's showing a profit from AI.

4:43And for the most part, nobody is showing even the revenue they're making from AI. And when I say revenue, I do not mean profit. Microsoft revealed that a$37 billion ARR run rate for AI, the majority of that comes from OpenAI feeding it money from, well, OpenAI's venture capital funders and probably leftover Azure tokens from the$13 billion in funding. Anthropic similarly is funded entirely with venture capital and their connected counterparties are seeing massive boosts in their remaining performance obligations, so their revenue backlog, mostly from Anthropic and OpenAI. In fact, between Microsoft, Google, and Amazon, $748 billion of their upcoming revenue, over half of it, comes from two companies, either buying compute or, in Google's case, Anthropic buying TPUs from Google to rent back from Google.

5:35The whole thing is very circular. And really, outside of OpenAI and Anthropic, there's not really any sign that there's a real revenue stream here. I have serious questions about the way they count revenue, but as private companies, it's hard to pierce. There's also the other real problem. Big tech is, between Meta, Google, Amazon, and Microsoft, over$800 billion of CapEx. I don't think data centers are being built at anywhere near the rates that people think. I have on good authority that one of the major hyperscalers only has around one and a half gigawatts of actual IT capacity, despite hundreds of billions of dollars of CapEx.

6:10And data centers are just not getting built. 18 to 24 months minimum. People see the AI boom as something that's happened, as in something that's, there's tons of data centers being built. I've seen people saying tens of gigawatts built every year. It's nonsense. I think there may only be a gigawatt or two built every year. The compute constraints people are facing are not a result of incredible demand, but a lack of actual data center capacity coming online. We are yet to get to the real monstrosity that this has created. And when it pops, it's going to be very bad for everyone involved. So I'm going to play a little bit of devil's advocate.

6:46I want to couch this by saying a lot of what you're saying, I agree with. I agree that the revenue is circular in a lot of ways. I agree with you, your points about the data centers, the fact that they're not coming online and there is a lot of questions on, are these data centers even going to get built? Are they even going to work? Are we even going to allow them to get built through regulation and through policy? I agree with a lot of that. At the same time, though, I do look at the growth of some of the revenues of some of these AI labs. I look at OpenAI. And to your point, we haven't gotten the financial audit of these companies.

7:26These are privately reported, but that's the way private companies work. But OpenAI was $2 billion in ARR in 2023. It's up to$25 billion in ARR today. Anthropics reached$30 billion ARR. You know, it took Salesforce about 20 years to get that. They've done it in under three years. I mean, those are dollars that are coming in, and companies are trying to adopt AI as much as possible, and they're paying money to do it. And I guess the way that I feel about it is we're probably very early to this game. They're doing whatever they can to get this through. They want to adopt AI, but I still find those numbers striking.

8:09And to me, it tells me that something about this is real and meaningful. Maybe there's some BS around as well. But overall, I don't want to write it off. What would you make of that? So let's start with that statement about being early. We're not. Early, if we do it by time, it's been four years. Yeah, coming up on four years since ChatGPT came out. And if you think about what constitutes early, early, it really is investment and innovation. We have had all the king's sources, all the king's men trying to make generative AI into something that generates profit. Hasn't happened yet. no one has happened it to i also have serious questions about how anthropic is doing revenue krishna rao their cfo in the department of war lawsuit actually said in an affidavit on march 9th 2026 that they had made five billion dollars in lifetime revenue that does not match up with any of the reported previous annual recurring revenue or arr run rate or whatever they use numbers it just doesn't when you add those all together it's six or seven billion dollars maybe more.

9:11I think Anthropic is counting revenue in a way that is different, fundamentally different to how most companies count it. My theory is that when, so with the largest clients they have, clients don't pay in arrears. Smaller ones do, but the largest don't. They buy massive amounts of tokens in advance. I think Anthropic, for example, is someone who's going to do$50 million worth of tokens and it lasts over six months. They're going to take that money up front. They're going to say, wow, in that month, because all ARR is, is month times 12 or 13, depending on how they do it. So I think both OpenAI and Anthropic are inflating their numbers.

9:47I actually think one or both of them is misleading investors, but that's something we'll find out about in the future. I also fundamentally just don't see evidence outside of these two companies that anything's happening. The largest company within this space was Cursor, and now they've been kind of sort of not really bought by Elon Musk. They also, if there was so much demand, by the way, why did Elon Musk give up an entire data center to Anthropic? There are enough signals here that suggest there is a fundamental weakness and no real major business model here. We still don't know Anthropic's true burn rate, but we do know that if all of the commitments from Amazon and Google come together in this$50 billion round, it's$30, it's$50, it changes every day, it seems, they'll have raised$108 billion in the last year.

10:34What's going on? Where is that money going? And to whom is it going to other than Google, Microsoft, and Amazon? So one of my simplest points as well is, other than Anthropoc and OpenAI, why are there no other AI winners? Why is everyone piddling around 100 to 300 million ARR, which is 10, 20, 30 million dollars a month? It's not actually that much for what is meant to be an industry-changing thingamajig. And there are signs that the economics don't work, the biggest being that Microsoft, GitHub Copilot, by the way, one of the largest clients of Anthropic, is moving to token-based billing on June 1st, 2026.

11:13They have been subsidizing tokens for their users to the tune of, I saw one person who spent$5 ,000 worth of API calls on a$39 a month program. I think this problem is across the industry. I think basically every AI startup is unprofitable at its core, and there is nothing that's going to shift these economics in their favor. There is no sign that inference is becoming cheaper. There is no sign that anyone has any plan to do so, and neither Anthropic nor OpenAI seem particularly concerned with bringing those costs down. So some really interesting points in there. I would still disagree with your point that it isn't early.

11:50I mean, when we're talking about technologies of this size, I still think four years is early. And to the point about profitability, like, you know, you look back to the internet, it took a long time for the internet to get running. Amazon took nine years before it was consistently profitable. And eventually, I mean, you can take decades before your business model works properly. And we have seen that with companies in the past which one specifically just because amazon web services was profitable in nine years the total capex between 2002 and 2017 was 50 52 billion dollars and that was for all of amazon just not not just aws between 20 2002 and 2017 um if we're talking about the dot-com bubble the dot-com bubbles economics were actually fundamentally different fiber was much cheaper and also had more uses, more obvious uses.

12:45Let me put it really simply. I know what you're getting at, but to put it really simply, when the - My point isn't wrong. It took nine years to get profitable for Amazon. And there were many other dot-com companies that took many, many years to get profitable. But Amazon was also not just a cloud compute company. It was a, basically went from a bookstore to a store to a cloud compute company added to the site. Totally different business. Right. Yes. The point I'm making is the way this is being invested in doesn't have the recovery story from the dot-com bubble. Because with the dot-com bubble, the dark fiber that was laid and the interchanges and the various bits of Lucent telecom stuff that was left abandoned, it didn't cost a ton of money to operationalize it.

13:31And its running costs were not incredibly expensive. Yes. If data centers end up not being built, it's going to cost just as much money to finish them in 10 years as it will today. And then on top of it, electricity costs are likely to be more in the future, not less. And there will be less customers. I understand where people get this from. It's just a difficult, difficult to square with me. Yeah, I think so. Those points I agree with. And I think it's something that needs to be talked about more. And I'm glad that you're bringing this up. My point, just to be clear about what my point was, the idea of these businesses, these business models not working yet, to me, doesn't seem to be a huge problem.

14:13But when you look at the amount that they are spending, and when you look at just the sheer size of the spending, to a lot of your points, I do wonder if when we extrapolate this, you know, 10, 15, 20 years down the line, whether it will look like the dot-com boom and whether these internet companies will be comparable. And I am increasingly thinking they won't. And something that I have been thinking, and this is something that Sam Altman has said, is I wonder if it looks like utility services. I mean, Sam Altman has said OpenAI is going to be a utility company. And we've seen this with utility companies before where, you know, having a regular market doesn't really work because the economics don't make sense.

14:58And so you basically just have to have one company that's operating for a specific area. And so I guess what I'm getting at is the business model. I don't write it off entirely, but it does seem that it will have to be different fundamentally from other businesses. And I wonder if you agree with that, or if you think that they just won't work at all, and they will therefore implode. So if we think about the dot-com comparison, and we break down the kind of businesses we had, with the dot-com bubble you had, it was really two bubbles. It was the telecommunications bubble and the software bubble.

15:35The software bubble and the e-commerce bubble, the pets.coms, the Cosmos and such, their economics didn't make sense because they didn't make them make sense. Pets.com, I think it was spending like 250 bucks a customer, something insane like that. But it was being invested and it involved the movement of physical goods. And those are margins you can bring down as Amazon did build their own logistics network. And also Chewy.com is basically pets.com. Cosmo now exists as Instacart. Like there are Webvan, pardon me, I think is Instacart now. Not the literal same company, the same business model. In the AI bubble, you really have three different kinds of companies.

16:11You have AI labs, you have AI startups. so wrappers yeah and then you have infrastructure infrastructure within ai data sense of gpus is one of the most commoditized businesses of all time it's a horrible business but it's also there's really no difference between a core weave and an ebius and a iron or a lambda they're all backed by nvidia they're all feeding money from jensen huang to jensen huang point is is the those companies do not have really, like, there is no changing this business model. Even if, I don't believe this, there was a chip in the future that was theoretically profitable. Blackwell isn't, Vera Rubin won't be, Hopper wasn't, none of the NVIDIA GPUs are clearly profitable for operators, otherwise we would have someone making a profit.

16:58Within the telecoms and dot-com bubble. Just to clarify what you mean there, profitable to use those GPU chips and then generate a return from an AI model. Yes, exactly. More money than they have spent. Yes, but not NVIDIA. Very profitable for NVIDIA. NVIDIA, incredibly. That's the thing. NVIDIA, memory stocks, they're all profitable because they're pre-selling a bunch of capacity, storage and service and all that. But a vast majority, like the dot-com bubble, there was also much smaller than this as far as actual money being invested. Venture capital was a, I think one round from Anthropic this year is bigger than all the venture capital that went in during the dot-com boom.

17:36It's insane. And that's, I mean, even with inflation, it's still pretty close. AI wrapper startups are all dying. And also the way that they are raising and dumping money is not something that's creating any useful residue. Harvey, for example, I think they've raised$900 million and they make 200, 300 million ARR. Ooh, those are stinky margins. But the point is, it's not like Harvey is buying a bunch of stuff and doing proprietary data, or buying servers, like happened a lot during the dot-com bubble because AWS didn't exist, so people were buying their own servers. There's no useful residue. All of that money is being dumped directly into OpenAI and Anthropic.

18:18Anthropic and OpenAI also do not have assets. They don't have CapEx. They have their models, they have their model weights, which, by the way, Microsoft has full and complete access to OpenAI's model weights, I confirmed today. And so they have talent, I guess. They have research, but they don't really have... They just have their models. That's the only thing. Those companies die. There's not really any useful residue to pick up. Perhaps someone could buy the assets of Anthropic, even though Amazon and Google will take them, but there's not really a thing to recover. And then you get to the core weaves, nebbias, irons of the world.

18:52And again, you've got a bunch of old GPUs, probably in data centers that haven't been built yet, which means that there's nothing really to pick up other than you pick up a cheap office chair or a cheap server back in the dot-com bubble. You could do something with that. What are you going to do with a GPU, especially when it's insanely expensive to run? I have a source at Oracle that told me it's about$6.30 an hour of cogs to run a B300 GPU from NVIDIA. These margins do not make sense. These costs don't make sense. I don't see how you pick up the pieces here. I don't see it at all. And indeed, I don't think there's much useful infrastructure left over.

19:31And this is the big argument that people are making, because people are saying, we're in a bubble now, we're in a bubble now. But don't worry, after the dot-com bubble, people picked up stuff for cheap. What stuff? We already have piles of unused GPUs. Supermicro had a multi-billion dollar order cancelled from oracle they have at least a billion dollars a beef 200 chips that no one wants to buy we've already got the fallow inventory nothing's happening i'm sure someone will do something but nothing's going to be cheaper in 10 years things work like you could pick up the servers and you could actually put them in like there's ways that you could make that useful from the dot-com bubble it's not really this story here i think the the part where i i started to disagree is i mean I'd be interested to hear what your views are on AI as a use case.

20:23Like, we've seen what I believe to be sort of BS industries before. One of them, in my view, was the crypto industry that I was very critical of for a long time. That, to me, was people trying to sell us a solution to a problem that doesn't exist, and then hammering it over and over and over again. And my view was, this isn't really a problem. This isn't really helping people. There isn't much of a use case here. With AI, I don't think of that the same way. I mean, I see a lot of what you're saying, which is companies are almost trying to force AI down their employees' throats because they just, they know or they believe that this is the most important thing that they have to do, figure out how to use AI.

21:01And I think a lot of that is BS in a lot of ways. But at the same time, I do find AI useful in certain areas of my life. I know that other people do as well. I know that OpenAI has a billion weekly actives. I know that nearly nine in 10 developers are using AI to code. What percentage of them are being forced to do that? I don't know, but I doubt that the answer is all of them. My belief is that there is some value here. And the question is, how much of it is real? How much of it is fake? And how do we figure out the proportions between the two? But sometimes I get the sense that you believe that there is no value, that it won't work at all.

21:42I think that there is, I have very specifically said there is value in coding. I've said that for years. That's classic misinformation there. Not from you. It's a spread. It's spread from, it's spread all around by certain words I wouldn't say. I'm hearing what you're saying and I'm trying to. No, no, no, no, no. But you're not, you're not saying this. Just being clear. You've been very, you've been very fair to me. Point is there is use, but all of that use and net utility is subsidized. Yes. So GitHub Copilot, I talk a lot about Anthropic. We can talk about GitHub Copilot. They were selling$15 for a dollar.

22:20Like that's fundamentally, there was a person who posted, there's seen someone who spent 39 bucks a month. They spent$1 ,500 worth of tokens. Someone else, five grand. Someone else, three grand. Someone else,$200. This is direct money to Anthropic and OpenAI. This is not something where it's, oh, they can claim they're profitable on inference. This is just straight up token burn. I would love to see how popular this was if people actually paid what it cost. Because I think that right now, this all-you-can-eat nonsense is inflating the value of it. Because if you were paying, Anthropic estimates,$13 a day is what someone will spend using Claude Code.

22:59That's a lot of money. It's hundreds of dollars a month. I don't think most people get that much value out of it. If you had to every little fun little turn you had with OpenAI, a couple cents each time, you're spending several dollars a day, I don't think people would think it was quite as useful. Also, the hallucination problem, the incorrect information, it applies everywhere. It's a very bad thing. But on a fundamental utility level, if this was being charged at cost, I actually think that every journalist should be barred from using it in any way other than token-based billing, because that's the only way you see the realistic cost here.

23:36Because yeah, if you get this thing massively subsidized and you don't know the real cost, you're creating a completely fantastical environment to explain these things. And so, and people will argue, well, the cost of intelligence has come down. No, not at all. That is a very weird aberration where they've said, well, to do a task that you did two years ago it's now 10x cheaper if your dog's awake but your cat's asleep and it's like a weird weird series of asterisks like barry bonds is baseballs the thing is we are yet to actually meet the point when people really have to judge how valuable this is we're seeing it already service now just blew through their ai yearly token budget in what five months we would burn through it in four we're seeing now the real costs start to come in that's where my big argument is And this stuff is not getting cheaper.

24:27It isn't. They might be charging less for some models, but that is not necessarily helpful if the models are burning more tokens. Killer Code had a great piece about this. So in the end, it comes down to, yeah, if this costs nothing, that would be a completely different conversation. But this costs everything. And the argument that people keep making is that training costs will go away. No, they won't. When? How? In fact, very simple question. How does Anthropic become profitable? And my God, if someone says, they're profitable on inference. That comment came from three separate times when Anthropics' Dario Amadei, Mr.

25:02Wario himself, has said, well, here's a stylized fact. Say you're 50 % profitable on inference. It's always this weird duck and weave thing that is decidedly non-gap and also entirely made up. He even said it was made up. So people have come up with this vision that inference is profitable. If it was, 85 % of Anthropic's revenue comes from API calls. Why is Anthropic losing billions of dollars a year then? And if the answer is training, okay, when does training go away? Why is training getting more expensive every year? The answer is, training is a cost of goods. Yeah. There is no profitability for any of this, and you have to keep training the models because of model drift.

25:46I think these are all reasonable points. I would add that this as a concept isn't new, and this is what venture capital is all about. And this is taking it to another level. But the idea of subsidizing a product, making it a lot cheaper, and in some cases free, until it's reached a level of scale and a level of penetration that people are willing to pay for it, and then suddenly you can turn the business on and it works. I think a good example of that would be Uber, for example, where, you know, we were all taking Ubers because it was subsidized and the VCs were essentially paying for our rides.

26:23We keep using it and using it and using it. The network effects grow and then eventually everyone's using Uber. And then eventually it turns into a real business that is profitable. Now, to your point, this would be, this would be like if Uber paid for the cars, paid for the gas, the gas was giraffe blood. They pay for the customer's clothes. Like I'm, I'm being facetious, but like it is the economics of Uber were also, they didn't do a different degree. Yes. Well, also they didn't do the same kind of subsidy. They made a trip cheaper, way cheaper. Way. I remember those days way cheaper, but they didn't charge like they still charge trip to trip, uh, point to point.

27:03Even this is, we are charging. You're paying$20 to drive a hundred miles a month, except in reality, each mile is costing uber 25 bucks or more and there is no compelling argument as to how that changes yeah i mean it seems as though it's a bet on eventually when we flip that switch the prize is going to be so gigantic that it ekes out what has been a gigantic cost and to your point the costs are gigantic and they're not really going down i guess the question becomes will that happen i mean there's a question of the certainty with which the number is going to be negative versus positive when we turn that switch on.