Today on the podcast, I finally bring Ed Zitron around on AI.
Lifeless, empty. Look like a dog's [__] on Thanksgiving. Stupid. You [__] Why don't you like data centers?
Okay, maybe not quite, but I do have a lively debate with Tech's loudest AI skeptic.
So, how are you going to get around the um return on investment thing? He's like, "Yeah, somebody will work it out." Sammy Clammy Sammy, this is your job.
From his podcast, his newsletter, and his 18,000-word takedowns, he has made one thing very clear: AI is a bubble that's about to burst. I'm Eric Newcomer, author of the Newcomer Substack. Let's get into it.
Ed Zitron, welcome back to the Newcomer Podcast, round two.
We're live in person, just you and me, mono. I asked Claude for, I don't know, what did I say? Give some monikers for you. All right, we've got the hater laureate, Cassandra of Compute, the doom sommelier, tech's designated mourner, the bard of the burn rate, chief bubble officer, the last honest man. That's a charitable one, I guess.
[laughter] I just like, I asked the dog, would give me, I'd be just as excited. It would be as good though. The dog would care. I would have more respect for it. It's just like those are things that you get people saying in your mentions every so often. It's like a regular person, like good on you, like a normie, but it's like Claude slop. Okay, the hater laureate. Okay. Just like unex- such, it's so unexceptional. That's the thing. It's not just mediocre in all these different ways. It's just like you read the text, it just feels lifeless. And if you go on Substack, I think it's mostly Claude written.
Yeah. So many.
And we saw, yeah. There were, I forget the company's name, but there's something. Yeah. Yeah. Yeah. And it's just, it's very depressing and soulless. It's lifeless. It's empty. Even human-written LinkedIn slop evokes some kind of reaction. Even if it's just reminded of the feeling of being at a conference, there's still something human about it. Even if it's quite boring. But you read the Claude sloppers, it's not just this. It's this. And that was the twist. That was the turning point. There's these little linguistic tropes. It's just sad. It's not even good. It's not like, we're still in this years in. We're still like, look, yay. It can kind of write like a person and it sucks and you can tell every time.
Yeah. It's just so boring.
So that gives us a taste of where you're coming from. To give context to people coming to this episode, I host an AI conference. I write a Substack newsletter that is not AI slop, AI copywritten. You know, you, I pay. Thank you. Great. Big fan of your newsletter as well. And though it obviously has a point of view, we will get into here.
Yeah. We did one episode before, 9 months ago, sort of disagreed. For context, for the viewer, on me, you know, I feel like I'm somewhere. You know, I host an AI conference. I wrote a case against OpenAI's valuation at I think it was $157 billion, which is March past. I sometimes report data that you're happy to have. So I feel like I sit in some middle ground where sometimes we're allied, sometimes we're clashing heads. And here's the thing with you. I know we don't agree on everything, but you published Tom Dotan, who's a fantastic journalist, who is also quite critical. You yourself are one of the few people who's been even positive. You are still willing to critique Anthropic, the cartoillion. We don't want to say two. That would be really ludicrous.
Yeah. That would just, we couldn't possibly value a company that burns billions of dollars that much. And then Elon Musk comes along, he's like, [laughter] "Watch this. It's so based."
But you at least try, you at least try and find some middle ground of, okay, you think that this will work. I don't, whatever. But you at least try and found that in something and even then you bring in some critical and I think that's healthy. I think it's the healthier way of approaching this for someone who is going to be pro AI or at least willing to accept a pro-AI opinion.
I never want to have a shtick. I feel like my disposition is sort of, I always like to poke and prod at things. So I think this, you know, we'll get into it and I'm sure we'll push back, but at first I just sort of want to give you a chance, like lay out your case for a couple minutes. Like where are you today? Because there's sort of this like, it's a fraud. It doesn't work. Nobody likes it. You know, there's sort of the self-dealing aspect. There are lots of pieces to it. Like where are you right now on the like, it sucks, nobody wants it versus these numbers are made up? Or like what is your core case against the foundation model companies today? So we're really talking about like OpenAI, Anthropic, like the case against them.
So fundamentally there is no ROI in AI. You can dress it up however you want. Look at how Anthropic and OpenAI discuss their products in the sense that they don't, they're always talking in theoreticals, in well the potentials. Yeah, everyone's always talking about the fact that we are still having a conversation about AI's value and whether AI is real or not is kind of the foundation of my argument in that if this was real, we would definitively know. We wouldn't have people saying emphatic like, AI is here and it's real. You don't need to say that. You don't need to say that if something is real. So I have never disputed that LLMs can do some code, but I think people have become a little overwrought about what kind of coding they can do and how reliable or stable or functional that coding is. But the biggest thing by far is that the con is the financial side. The con is the fact that most people's experience of AI is unrealistic when it comes to the cost. So you using ChatGPT, I believe the subscriptions will go away. I think we are just at the dawn of token-based billing. Meaning that most people's use of Claude Code, for example, is based on a relatively unlimited, you can burn between $8 and $13.50 for every dollar of your subscription in tokens. So most people's experience of AI is separate from the cost, divorced from the cost. And this was a deliberate con on the part of Sam Altman and Dario Amodei because they both knew that if they actually made people pay on a per million token basis to use these things, they'd go, "Wait, how much? What? Whoa, this thing messed up and I still had to pay." And you're kind of seeing that already with the enterprise customers, the Ubers and Walmarts of the world who are starting to have to pull back.
This is the token maxing soul-searching we're seeing going on.
Well, the thing is Uber wasn't token maxing. They, I don't think they had a leaderboard. I think they might have at some point, but they burned through their entire annual token budget in 4 months, in the space of the first quarter. And then their COO, Andrew Macdonald, said, "It's getting hard to justify because you can't connect it to outcomes." This is, we are years in. We can't, why are we all acting like this is going well? It isn't. And so the experience of AI, the value judgments of AI are all based on phony metrics, an outright con about the costs, and quite frankly, a very weird SCOP. The whole thing is a SCOP. The cult of personality around AI, the very much if you are in the out group and you don't like it, you're othered, you're different, you're a hater, you're a skeptic. And it's frustrating because Silicon Valley is meant to be a place of meritocracy. It's meant to be a place where you evaluate based on your value. Theoretically speaking, remember the, I don't know if you ever saw the GitHub thing, the meritocracy doormat they had.
No, they had like a White House.
It looked like a CIA thing, which is also good. We love that. But the thing is, Silicon Valley is meant to be a meritocracy. Yeah. Everyone coddles these things. Everyone talks about these things like a gifted child, like they're afraid that they'll get turned into an ear of corn like that Twilight Zone episode for saying the wrong thing. And that's...
Oh, I watched your Bloomberg interview recently and you know, I used to work at Bloomberg. I'm very sympathetic to them, but the response to some of your criticism was very, uh, but they're smart people who believe this, like surely. [laughter] Yeah. I felt a little bit like, is that the best AI defense they're going to muster? So I just want to make sure, I mean, I think, you know, argument number one is that AI is talk, you know, the AI companies talk about what they can do in the future, not what they can do today. Your assessment of what they can do today is not that exciting. To the extent companies are using it, it's often subsidized or and/or there's sort of this delusional sense of like, clearly we're supposed to get some value out of it, we're going to spend, that's going to come to a stop. What's your view? You know, Anthropic has now confidentially filed to go public. I think you've, you said in our last conversation there wouldn't be exits. Is now the view that they're going to try to exit, but they're going to pull one over on us? Like, what's your view on, I guess, when does this all come to an end? Like, does it survive an IPO?
So I said this on Bloomberg. I don't think regulatory bodies should let Anthropic or OpenAI go public. I think that they are both irresponsibly run companies run by craven liars. Both Dario Amodei and Sam Altman sell their products based on what they might do theoretically rather than what they can do. Every time anyone asks Sam Altman how this will work out on CNBC, David Faber was like, "So, how are you going to get around the um return on investment thing?" He's like, "Yeah, somebody will work it out." Sammy, clammy, Sammy, this is your job. You are the AI industry. But to your greater point, I think one of them is going to try and go public. I think they're going to bull rush this thing. I think it's inevitable OpenAI files to go public. And based on some stuff I might have read recently, they sure shouldn't. They're losing too much money. Anthropic, on the other hand, is also losing just as much money. They had this, one of the most disgusting things I've seen done by these companies outside of the environmental damage, the stealing, the cam that Grok generated. Putting all that aside, Anthropic leaking that they are profitable in the two months that Elon Musk discounted their compute. Do we not have an SEC anymore? Do we not have any kind of regulatory body? We certainly don't have a conscience in the valley because everyone was like...
Your view is that they made a profit because Elon gave them a deal on their compute and that's in the SpaceX S-1.
Well, they specifically, the May and June in Q2, Anthropic will become profitable, will because it's June now. It was not when that story came out and mysteriously those are the two months, May and June, when Elon Musk discounted Colossus 1 and 2 for Anthropic. It's just like...
And what sucks is that's publicly available information. This is in the S-1 for SpaceX and people are still like, they're profitable now. It's like, yeah, you're profitable as long as you don't count the costs. And so...
I mean, my view on this is we're hope, we will get the answers when the S-1 public and you know, we'll talk about the Uber analogy because I covered Uber very closely.
You were one of the, and you were actually one of the few people who are really critical of them.
I was very critical, but in some ways I learned the opposite lesson that you did, which is they ended up turning a profit. People over-extrapolated from some of the private financials and drew these complicated models that sure, like...
Uber was spending more aggressively and they were, they were like, Uber Pool never worked, right? So there were things about Uber that didn't work that were worthy of criticism, but...
I just, there are just limits to how much you can sort of do financial calculuses off of these like limited private numbers that we see.
Do you want to wait to discuss it because I...
No, I don't. Let's talk about the Uber case because it forms both of our views. So...
Exactly. So with Uber, for example, a lot of their money was marketing and R&D. I like to equate Uber to Groupon more than anything. Heavy, heavy marketing and that was where they did the subsidies for the rides and all that. Also, R&D efforts within autonomous driving that I think went nowhere. I forgive me for not having looked recently. Also, the burn rate on Uber was just a different planet. I think they burned...
Way lower. It was scary then, but now it seems...
$32 billion in history and Anthropic has raised over $75 billion in the last 6 months. So I think that that's altogether different and also the level of burn, the scale of burn and the linearity of it. It's a word I think basically as Anthropic's revenue increases, so do its costs, except the costs are always up here. Same with OpenAI. Uber had a messy cost structure, but they didn't change the unit of work, the unit of revenue. So they didn't have a thing where it wasn't, they tried the ride pass where you could just pay one fee and drive as much as you want, but you still paid for a ride. A customer was still paying to go from point A to point B. With the subsidized AI labs' subscriptions, they are training people and they've trained millions of people to use these things in a way that does not make sense and they pay the costs. The reason that you're seeing companies burn so much money is imagine if you were driving your car and you just never really had to think about fuel. Fuel was just included with it perfectly. And then one day, and so you actually, no, let's reframe this. Use a cab service and it will drive you as much as you want. You get 300 rides a month. You can drive from LA to San Francisco. You can drive from just the Upper West Side to the Upper East Side. No matter what, it's always the same costs. Then suddenly one day they say, actually going to have to pay on a mile per mile basis. You would have no idea how to plan your life. You had planned your life around this subsidized unlimited thing. And perhaps you could say the, oh, the ride point to point from the original Uber was like that. No, this is like if every Uber ride used to cost $3 and the new ones cost $300, $500, $1,000. People do not know how much their AI costs. They have no idea how to measure a unit of work because they've never had to think about it. Like if you were used to just loading up Claude, going, fix these bugs, and it kind of does it, and then it messes some stuff up, and you poke it, bonk it on the head a few times, maybe it gets it right, maybe you mess around with it, but you don't know what it costs and you weren't paying the costs, suddenly doing that would give you CTE. Look at the people. So GitHub Copilot, which is Microsoft's coding tool, they just moved from a premium request model to a token-based billing. Go on the subreddit. People are screaming like they're being stung to death by bees because they are finding they're burning through their entire monthly allotment, which is just the price of tokens, in like a day. There's someone who did it in three prompts.
There's so much I want to respond to this. So I'd like to take a moment. I do the specific one major contradiction that seems to appear to me in your argument is that people see the cost being high and they freak out because they want the product desperately. Like how do you reconcile with the idea that it's like people who are bullish on AI see the fact that people want to spend money on it, want access to it as evidence that it is doing well, that they see some value in it. Like do you not see that?
I do see value in it. I'm not disputing that. I'm saying that the value they get out of it is based on the idea of subsidies because they have evaluated it based on it not costing what it really does. So if you use the product and you could do all this stuff with it, however accurate and inaccurate it is, you can say, "Well, that's a great deal at 200 bucks a month. Is it a good deal at 300 bucks a week?"
Probably not. You wouldn't be able to do as much. And I also think psychologically you look at it differently. If something costs 200 bucks for a month and you run up against rate limits, fine. You can accept that. So I want to talk through the Uber analogy as I see it. I mean, yes, I think Uber was a similar case where people felt like, oh, it might work at one price but it won't work at another. And I think what happened was people got really upset in this amazing period where Uber was overpaying drivers relative to what they needed to pay them to get them to do the work. Discounting rides to keep the riders from going to Lyft. And basically we had this like free giveaway and there was like a crowd of people online who are like very angry about this even though they were getting stuff at a big discount, right? They were getting it. And so then that period ends basically. Uber and Lyft sort of come to a sort of equilibrium where it's, I don't know, it's probably 85-15 or even more aggressive on Uber's side. And yeah, I mean basically they have to squeeze drivers more, especially on airport trips, but there is a price where people take Uber, it works. It's not like the insane Google-level business that people thought. And things like Uber Pool, which were supposed to allow it to sort of like get to scale, don't work. So there are pieces, I just don't think there's this like something's valuable at one, you know, it's sort of a sliding scale. And [snorts] I feel like we didn't learn the lesson from Uber that, you know, that this sort of like big hustle was had on us.
Right? But here's the thing. When you take from Brooklyn to Grand Central, 50, 60 bucks, it's still within the realm of what you would pay for a cab. I am saying that what's happening with AI now is people went from, just for the GitHub Copilot example, this is not hyperbole, there were people who were paying 39 bucks a month and being able to burn 100 to $3,000 worth of tokens. That is not the same. That's not the same at all. That would be if like Uber rides were $5 and then became $1,500 or $300. It's just a different economic scale and also you can plan around that still even though it's more expensive. You can say, okay, this ride was 10 bucks. Damn, I wish it was still 10 bucks, but it's 50 bucks now. With AI, it's you don't know how to measure a unit of work. That's the biggest thing. Like it's tough across models, across harnesses. You use OpenCode, using a DeepSeek model, using GPT-5.5x High Plus Alpha, you're using whatever you're using, each prompt, oh, what are you having it do? It might act differently because hallucinations are a mathematical guarantee per OpenAI. So you don't know how to measure the unit of work. You also pay regardless of whether there is a mistake or not. If it gets it right, if it doesn't get it right, you are still paying on a per million token basis. So it's not like you can say, well, okay, I was spending 200 bucks a month. Now I'm spending 500 bucks a month. You're unaware. You have no idea. The more complex the work you do, and because AI code is very verbose, the more AI you use, the more convoluted things are going to become and the more reliant you are on the machine to kind of work out what it's doing. And those costs are dramatic. They are unrealistic for most people. They're unrealistic for most organizations.
The models that we have today will be cheaper in the future, right? We see that over and over again. What do you mean? Like the models of 12 months ago fall in price. The new models obviously are new, have new prices, but you're able, the old models become either offerings from Anthropic and OpenAI themselves and/or open-source competitors, basically copy versions of that model and offer them at a discount. So I can see the argument that, and you've talked about this sort of like the capex expenditure, like the situation with Uber and OpenAI and Anthropic are very different in the sense that like OpenAI's costs are because they're spending a lot training a new model.
Well, but before we get on from the token thing, there's an important distinction I need to make which is models are burning more tokens. So even if the cost of intelligence comes down, the biggest myth of it all, even if a model might be cheaper in general, they are burning more tokens. Opus 4.8 burn more tokens than ever. It's making, and every time Anthropic has released a new model, it seems to get worse. Don't know if you noticed that. They seem to be more convoluted. They do things differently. Again, burning tokens. So you can...
You're saying every Anthropic model is getting worse.
If you go and look at Twitter right now and see how people are reacting to 4.8, see how they react...
But they're comparing it to Codex.
No, they're comparing it to Opus 4.6, man. Like, they're literally comparing it. And but this is the thing, good example actually. Even if it's better, it is still different, which means you have an adjustment to it, which means you are going to burn tokens. But anyway, they are burning more tokens. Chain of thought reasoning models burn more tokens. So even if the model is cheaper on a per million token basis, you're still burning way more tokens. So you're burning more tokens and spending more money. Same thing with open-source models. This has been a consistent problem. It's a problem with Grok. It's a problem with all of them. And the thing is you can, if you're happy as a company with the model today, you can basically find a version of that model, you can make a version of that model yourself. Like...
What do you mean to make a version?
Like people can build their own models. There are open-source models that people use that they sort of fine-tune themselves. There are companies that really that build it. They're in-house models. Like it's not like you're forced by Anthropic and OpenAI to shift off the model architecture that you like. The thing is you're speaking in generalities here. Who has successfully done that? Microsoft just farted out six new models from Mustafa himself, Mr. Fussell man, who is apparently quite a horrible boss. I hear he goes nuts on [laughter] Slack. Putting all my gossip all over. Obviously, there are a lot of Chinese companies that have...
Well, DeepSeek is running the Chinese wear them down process. They discount it permanently just to drag down American models. That's tactical. I love that. It's just funny to watch them do that because it's the kind of thing American AI labs have no idea how to do, which is political rat-fucking. I'm not going to swear today. But back to the core thing, which is it's kind of like the mileage thing. Okay, gas is cheaper, but you're driving an extra 50 miles, so it's not, the cost of the journey is more expensive. So every model is burning more because the more complex they get, the more things they, the whole mixture of experts thing that was meant to bring cost down. It is not more expensive to do the same task. It gets more expensive because people get more ambitious about what task they're going to do.
No, it's actually that because everything I've read is that...
Go on the Claude Code sub. I love going on these subreddits because you can get a real taste test for it. They will find these things, new models going in crazy loops. The things that they could do before magically don't work. This is a common experience with large language models. Speak to these users even if you're...
Look, yeah, certainly it is true that when new models come out, they have like all these different edge cases and their personalities change, their syntax...
Are they delivering? I don't either because I don't think they're incentivized to. I actually don't think Anthropic has any incentive to make their models more efficient. Same with OpenAI. The markets are rewarding them for being big, lossy, messy companies. And on top of that, they have the various business idiots and various companies who will, Zillow. I've been talking to people at Zillow for a few weeks. Crazy AI psychosis over there burning over a million dollars a month in tokens. But the thing is these companies are not incentivized to be more efficient and they're not incentivized to bring cost down because people, Anthropic believes that they have the mandate of heaven. I believe that they're a deeply cynical organization and also they're burning more tokens. Like they obviously are doing it. You can, people will say, oh, we're just picking out a few edge examples. Are any of the people who are excited about this actually talking about the real costs? Probably not. And it's...
I agree with you that there's, you know, public companies, all sorts of companies have an incentive to show that they're getting the most out of AI and therefore they have a reason to spend aggressively independent of what they're delivering. I think you underestimate that companies do find value in AI. Like I feel like you...
Well, literally my own company, my own company, you know, our ticketing infrastructure was vibe coded. Like we could have paid, you know, for HubSpot or whatever, or some SaaS product and you build your own version so that you can quickly change it. Like...
Quick question, did you pay on a per token basis?
Uh, no, I doubt it. You're, you're like, it's not, that's the thing. You are, you are...
You are living in a dream, but who is the dreamer here? Like this is the thing. You are experiencing it through the larger con, which is, yeah, it burped out some open-source software. You were like, it burped out a copy of an open-source CRM. You have it, you have it privately. So there's not really, I mean, I would make sure it's checked by a real coder, but you also did not pay for it. You paid for the subsidized rate. I'm guessing the $20 a month plan, maybe the $100. I'm sure it's $100.
But that's the thing though. This is my point. You didn't experience the reality of AI. You were, you're the victim of a con. You have been convinced that something is a price that it is not. And Anthropic only moved companies to token-based billing along with OpenAI, who do it much quieter, about 3 months ago. It was announced by The Information in April, but it's been happening longer. And already they're screaming. They're like, "Ah, I don't know about this." Because it's one thing to say, "Wow, I'm spending 200 bucks a head." and they get rate limits and they can go nuts, nuts, bananas on this. It's another when you have people spending $1,500, $2,000 a month. Uber just had to limit their engineers to $1,500 a month, which makes me wonder, how much were they burning before? Because that's the thing at 200 bucks a month, very different product to one that is variable [snorts] cost of 100 bucks a day, 200 bucks a day. Anthropic's own documentation says that they think that people spend an average of 13 bucks a day in tokens on Claude Code. That's not sustainable for the average person. And I think ethically speaking, any journalist writing about AI should not use the subscriptions. I think they should only use token. I'm, no, no. Why are you laughing? I'm not even being, I'm not being aggressive here. I'm saying why aren't they doing that? Because to have a realistic perspective on the actual efficacy of these models, you have to experience the horrors of this thing just went on a little loop and burned five bucks.
Yeah. Most consumers are not going to want, consumers don't want an experience where the product could accidentally, you know, way overspend. And they'll, and enterprises are sophisticated. I...
I mean, enterprises are not sophisticated.
Not always, but at least...
I mean, in general. I mean, but the point I'm making is I think token-based billing is coming for everyone. I think it is an inevitability. The economics do not make sense for subscriptions. Everyone does this funny fusion dance of, oh, inference is profitable. No one has proof of that. Maybe, maybe if they didn't pay any training, there would be economics that may be lined up. But in truth, inference and training are one and the same at this point. You have to keep post-training. You cannot stop. I mean, we're going to see the... Do you think these IPOs are not going to happen? They're going to fail. Make a specific prediction.
Oh, man. The thing is things are so wacky because Anthropic was like, "We've confidentially, as they announced it, filed a night for S-1, but we're not going to say when we're doing it." OpenAI, the journal reported a few weeks ago since Burbagen was like, "Oh, they're going to file soon, but they haven't yet." I think that there is going to be a reaction based on the SpaceX IPO. If it rips, OpenAI lists. If it doesn't, OpenAI slow rolls. But also, I could see Sam Altman getting real mad and saying, "I got to get, I got to get the Sarah," like screaming at Sarah Frier.
I mean, to be clear, I think the SpaceX IPO to me is uniquely worrying. I guess my, to me it's like the problem with SpaceX is it has failed to be the AI company it wants to be and the space company has a limited, you know, there's just like you can value the space business and it doesn't reach, you know, $2 trillion. But so where we differ is...
Well, I don't know what the correct price of OpenAI and Anthropic is to say, yeah. I mean, but that's, it's not my job, you know, it is not to pull the Bloomberg, but it is sort of the business of investors to speculate about the price and make bets and have sort of this competitive dynamic and that is ultimately what sets prices.
So I think whoever goes public out of Anthropic and OpenAI stops the other one. I think that...
You think it's so bad that they will only...
Oh, I'm confident that OpenAI is really bad, very confident. And I think that Anthropic...
Like the numbers will appear in a so abysmal...
I think that the losses for OpenAI are going to look like a dog's [__] on Thanksgiving. I think that, I don't know about Anthropic, but the fact that Dario Amodei, I have to give him credit. He is one of the best media manipulators in history. He's better than Jobs. He just, he knows how to do the financial stuff with the annualized revenue. He's so good at it.
But you complain about this a lot. Annualized revenue is useful to convey if your revenue is growing and so you're trying for something like API calls though. Those are not monthly recurring expenses. They are, there was some numbnuts who spent $500 million on Claude in April, I think it was reported by Matt. Your point of view, your point of view, right? So this, like if I report my ARR this month and then I have a problem next month, like I'm going to be screwed if that number comes out because I'm going to go from a great number to a bad number. So you've been complaining about the ARR over and over and over again, but the revenue keeps going up. Like the problem with using ARR is that like the chickens can come home to roost.
The problem with ARR is not that they can come home, it's that you can manipulate it in any which way you like. Anthropic has never defined what they mean by run rate, not once. And I've read every single story. The Information, I think three over there reported that they do it by taking the day subscriptions, well like the monthly subscriptions they have times 12, then their API revenue for the last four weeks times 13. API revenue is a really weird one to do on a recurring basis because that is going to flow and as you see these cost cuts happening, yeah...
That's coming down, so it's just like revenue is how you should report revenue, and you should also—all these things are non-GAAP. It's the non-GAAPiest non-GAAP that's ever non-GAAPed. And it's frustrating because they share the run rate so that it's—it's a really simple manipulative trick. You're meant to go, 'Oh, $47 billion of run rate. That's how much they made.' They made $47 billion. That's what it's meant to—
I mean, there are sophisticated investors doing it, right? I mean, what? I'm sorry, mate. No, I'm just saying, like, to a regular person, the idea that it's like forward-looking revenue might seem unusual, but investors are very aware of what, like, 2025 revenue is versus, like, taking—
Are they? Yes. I mean, are they getting audited financials or are they getting a PowerPoint deck that Dara's doing the Elizabeth Holmes CEO voice show? It's got—look at this.
Do you think the books are cooked?
I don't think the books are cooked. I think that at the end of every presentation is like the—like a font-size-three series of disclosures that are like, 'on a non-GAAP basis,' if they even show—
I mean, they're going to go public, so they will GAAP.
Financials—I cannot wait because—my god, would it be nice to stop having to—no, I don't mean this one, but you know what I mean—it's constantly defending this. But here's the thing—
You have been saying that we are in a bubble since at least 2024, right? I mean, what have you gotten wrong? Like, you've been negative in a period where everybody else has gotten more positive.
Oh, I definitely flopped how quickly it would end. I will fully admit, like, 2024, I was like, 'It'll end in three months.' I was just wrong.
So what? Like—like over and over again, you're like, 'This is the peak. This is the peak.' Right.
And then early in 2025, I very specifically was like, 'I'm not going to call the peak anymore,' because, you know—
You just did call the peak again, right? Didn't you do, like, 'It's a bubble'?
The top is the IPOs, 100%. I mean, Paul Kedrosky agrees with me. He's excellent. Like, you just had him on the show. Like, the Google equity raise—Google raising $84 billion on the public markets like a goddamn mid- or low-tier cap company. Come on, man.
But the thing—that is, I agree with you. That is a potential bubble sign. I think—but not in the cynical way you're putting it. So, like, Google's like, 'Wow, people really think we're worth a lot of money. Like, let's get some cash.' So, I mean, it's their share. They're allowed to get money, you know, to raise money.
I mean, they're allowed to do it. That's perfectly legal. I'm just saying that they're doing it because the credit markets are drying up. That's 100% what's happening.
Well, their equity is super inflated, I guess we'd both say. So, they can get a good deal raising money at a high equity price.
Sure. But also, the credit markets are drying up. Like, that's what's happening here. And that's one of the big things that people really don't want to talk about, which is—let's say Nvidia does the thing they're meant to, which is they're meant to sell a trillion dollars of GPUs by the end of 2027. Yeah. Sell, by which I mean put in a warehouse. That would—I think that's like 15 or something gigawatts of capacity. Data centers aren't being built at that scale. We are watching one of the biggest Kickstarters, a pre-order campaign of all time. For there to be a trillion dollars of sales, we're going to need like $800 billion worth of debt. We're going to—putting aside cynicism or anything, just dollars to donuts, we're going to need $800 billion worth of debt. And for Nvidia to keep growing at their current rate, in a couple years, they're going to have to be selling a quarter of a trillion dollars worth of GPUs in a quarter. $250 billion in a quarter. And again, the cash flow required to do this is so severe that Microsoft, Google, Amazon, and Meta all having to raise their—all having to do dodgy little, um, what's it called, like, off-balance-sheet deals. Putting aside however we feel about AI, that is more money than I think is out there. People say, 'Oh, the money's on the sideline.' There was a story in the FT saying banks are worried about choking on data center debt. For this to continue requires an astronomical amount of money. And then there's the messy little question of why. What does building more capacity do? If OpenAI had 10 gigawatts more capacity today, what would they do? What did Anthropic do when they got access to Colossus 1 from Elon Musk? Oh, they upped the rate limits on Claude. Yay.
I mean, the model—the consensus is certainly that the models have improved. Most people think the scaling laws are holding true, that the more—the larger training runs you run, the better output you get. I mean, they severely pulled back on pre-training. Isn't it more post-training? I mean, they're—there are scaling still the training runs, obviously, but they're scaling on inference with chain-of-thought reasoning, and also pre-training is—it's still happening. Right, it's all of the above. They are coming up with new techniques to improve the models in addition to trying to run larger models.
Then why—here's the thing, though—then why are we only really talking about code? Like, that's—that's a very—like, we really only talk about code, and there's—it's because there's a ton of training data for code.
There's a lot of people that can post-train the models. I mean, I do think, like, a key—there's clearly a receptivity among, like, consumers to the argument you're making. And my view on it is that AI requires, like, proactive work. You know, you have to go out, you have to use the models, you have to find things you want from them. Still, I mean, when we talked last time, you were really emphasizing that OpenAI was, like, the only one, and now the top three apps, I believe, are, like, ChatGPT, Claude, and Gemini in some order. So it's not just one, like you've argued in the past. So that is another thing that changed. But—but the core issue is that there's a large audience of people that are like, 'I'm not—I don't—it's stupid. I used it a couple months ago, like, hallucinated,' and they're not, like, investing, like, the energy and time to, like, get value out of the models to the extent they experience it. They get it through Google, and they—they underestimate the extent that they're getting LLM answers in the first place. So they're sort of using Google results, forgetting how the product used to work, how it is now, and sort of—it's pretty easy if you don't pick up the app, use it, and spend time on it to say nothing's happening. And so people like this story of they don't drive any value. But then I—I feel like I'm using a fairly—like, it's not like I'm building, you know, like, I have, like, my own server at home. You know, I'm using Claude in a pretty straightforward way, and I find that I derive a lot of value from—
From the subsidized thing. But also, were you just—this is not an insult—you are describing—no, no, no, I really want to be clear, this is not—sure—doesn't sound very intelligent. You are—you're frontloading this by saying, 'Well, they haven't made the effort to make these things useful.' That doesn't sound like AI to me. That sounds like a kind of scam. It sounds—it sounds like something that someone at these model companies would say because they know that they can't just give you something and use it. When I used the first iPhone, I was one of the first people to buy it in State College, Pennsylvania, Singular Wireless, that thing immediately—like, 'Holy crap.' Like, 'This is—this is the future.' Using AI to this day, and I've played around with it. I've used it. I've really tried. I've thrown a few models in it. Messes things up all the time. Still does it.
This is the same fight we had last time, where, I mean, you sort of presented AI as if it had been sold to the world and we've been, like, conned into it. But—but the iPhone was significantly marketed. Like, it took the world a long time. Like, most people did not use the first million of them in the first year.
Most of the—yeah, because it was limited to a single carrier.
I mean, I'm sure ChatGPT has more users than the first iPhone.
Yes. It was limited to a single carrier, Singular Wireless, and didn't come to Verizon, I think, for at least—I think it was six months to a year.
But Apple was a great marketing company. Like, the iPhone was sold to America in a way that, like, the AI providers are fairly clumsy marketers. Like, you think AI—the AI companies are good marketers?
What are you talking about? We're talking about AI non-stop for three years.
Because as a culture, we're interested in it. It is—because it is the stuff of science fiction. Like, it is not the companies forcing us to do this. It is something that humanity is in. Every country in the world is interrogating this.
Since 2023, we have stories in mainstream media saying we should be scared of these things. We've had stories in the New York Times saying turn it off because it conned a TaskRabbit when it didn't actually do it. We have had it shoved in Google. We've had it shoved in every app we use. Every single goddamn app. Summaries, generations where we don't need them, shoved. It's taken over Google Search. It's been crammed in everything. Every company has been saying, 'Well, if AI is coming along, it's going to take your job.' Every news outlet has been talking about AI, AI, AI, AI. And it's been done on the terms of the company. Every product is forcing you to Apple Intelligence. That was not a consensual—
Apple Intelligence, I'm negative about. Certainly.
They forced GPT on people. They forced AI on people. Google, Microsoft puts Copilot everywhere. How—how can you say with a straight face that this was a consensual exploration of this tech? People were not told, 'Take your Nokia 3210, throw it in the trash or you'll lose your job. Take the iPhone.' That never happened. There was no pressure campaign like there has been.
There's been this thing of—but my point is that it is a sort of—certainly there is pressure, like, bosses want employees to use AI because they see value. But it is, to me, when you—you say it's sort of like a marketing effort on the part of the companies. Like, I think this is an organic idea that people are excited about something.
I think you are ignoring things. I think you are ignoring the fact that both Altman and Amodei—the Super Bowl ads that they ran, like, Anthropic's Super Bowl ad was, like, not effective. It was, like, one of the worst ads that has ever been run. Like, bad marketers—except they managed to pretend they weren't involved in the war in Iran and then managed to get Katy Perry and a bunch of other people to say, 'Yes, sweet Claude,' because nobody paid attention to the fact they've been working with the government since 2024. But the point I'm making is Altman—
That was a pretty big veer.
I mean, that's what happened. But let's—we can get back to that in a second. The point I'm making is Altman has been doing the big scary technology thing. Mira Murati was on The Daily Show in November 2022. Like, this is not something where—this is manufactured consent at scale. The iPhone didn't happen. I promise you, because I remember being at PC Magazine and turning to my editor and going, 'We should write about apps on smartphones.' He went, 'Nah, mate, it's not going to happen.' I'm serious. I don't push back on this because I'm just being a hater. I—I push back on it because it is—it is manufactured consent for a technology, and an anti-labor technology at that. And more importantly, the media fell for it.
For it to be anti-labor, it needs to work.
No, it doesn't. Why? It just needs to be an excuse.
But if—if the jobs weren't needed in the first place, then—what do you mean?
