AI Bubble: ‘Business idiots’ are finally seeing the downside of uncapped AI | Ed Zitron

Ajouté : 2026-06-04

[00:00:00]No one really knows how to use less tokens. Nobody really knows how many tokens they're using. Uber COO even said, "It's really easy to be impressed by AI when you're not paying the cost, but someone eventually has to pay that bill." And I think we are going to start seeing the true value of AI cuz everyone has been using the subsidized stuff.

[00:00:20]Everyone has been high on the hog using AI effectively for free. like and the argument is of course people say well they're profitable on inference zero proof none there's no proof on that what I think is likely is that Anthropic is going to see their revenue burst and then contract violently hello and welcome to the tech report.

[00:00:49]I'm Natasha Banal and joining me today is Ed Citroron, the author of Where's Your Edat and the host of the Better Offline podcast. Welcome Ed.

[00:00:58]>> Hey, thanks for having me.

[00:00:59]>> Thank you for being here. So, we're going to start out with what seems to be one of the biggest stories at the moment uh going on, which is all about AI token maxing and how things aren't really working out. Can you tell us what's the latest?

[00:01:11]>> So, important thing to know about AI, you don't really know how much any task will cost. You truly don't. depending on the model, the prompt, the harness using like clawed code. So just whatever you're plugging the models into and how you're using them means the task could cost anything. It could be cheaper with some models, but some part could involve a more expensive model. You have to tinker with that which all the way ends up saying you don't really know how to set your AI budgets. Then there's the other problem which is uh we also don't know how to measure the ROI which is why Uber COO Andrew McDonald said this week that he is having trouble justifying the AI budget as there's not really a direct way of comparing it to some sort of outcome. This is the worst possible scenario for these companies because it isn't just a case where they're saying, "Oh, we're not getting a ton of ROI out of it. We're not seeing as much as we'd like." it is. We don't know. We don't know how much it will cost and we don't know what spending gets us. This means that it's very likely we will see cutbacks. I have been just before this conversation talking with a major payments processor, some an engineer over there. There is someone in that organization who spent over $100,000 in tokens in a week. in a single week. And so organizations are doing this and I think the natural next step is they're going to start saying I don't know about this and that's because leading up to this point everything's being subscriptionbased.

[00:02:44]They be so when you use a you or I would use claude or chat GPT don't know why I'd be doing that. uh you you would be just charged a monthly subscription, you'd have rate limits. But enterprises only a few months ago, may I add, have been moved over to tokenbased billing.

[00:03:00]So they're paying on a per million token basis and months in only like less than a quarter in cuz I think they only started token based in February for organizations only a few months in they're already squealing. They're already crying uncle.

[00:03:14]They don't like suddenly these people who were saying the power of AI are here is here and it's amazing are saying oh wait a second did we need an ROI from this anyone check >> that's terrible and it's it's one of those things where you think if this is happening at Uber which is a huge organization what kind of ramifications will there be for the rest of the sector I mean this is this is a basically a failure in the model of trying to appeal to the corporates right this is this is it just shows it's not working >> yeah and I think The the most alarming part is like I said there's no real way of measuring how much you should spend.

[00:03:48]So your budgets are Uber blew through their token their annual token budget in 4 months. I think it was I was speaking with someone at Zillow. They are I think as of the end of May they're going to be at 98% of their cursor budget for the year. Everyone's spending way more than they expected and that's because nobody knows how to measure AI. You can say whatever you want about its efficacies.

[00:04:11]I don't think it's very good at many things. I think that it's questionably useful in most applications and even in coding. It makes the easy things easier and the hard things harder. But this is the first largecale examination of the actual value of AI.

[00:04:27]And they are or these people who were talking about it like they were speaking in tongues like they were saying that this is a new god. They're now saying ooh oh I don't know about this. Oh why are we doing oh should we be doing this?

[00:04:40]And I think the natural next step is for Anthropic's revenue to start declining.

[00:04:46]They because think of it like this.

[00:04:49]Anthropic needs to keep growing at remarkable rates. They are I think they have 300 billion plus in cloud commitments. They Dario Amade himself said if they misallocate the funds for inference and for compute they will go bankrupt. That's from a Darkh Patel interview. But everyone's excitement over anthropics growth is coming from a bad place. It's coming from what is inherently experimental revenue. Do you really think companies are going to be spending the payments processor was talking about over $1.5 million a week on tokens without really knowing why.

[00:05:26]And sure, at first they're going to do it, but this means that Anthropic's growth is coming from a very unstable place, from a from a place of executive incompetence far more than it's coming from any kind of ROI. I don't even know why I said far more. None of them I've talked to so many organizations about this now. Not a single one of them has a single metric to measure actual ROI. They have pull requests which is just AI AI produced things to say hey we should make these updates to the codebase or to whatever we're working on. They there are some companies that say we should raise those we should have more requests to change the software. Doesn't mean it's good doesn't mean they're good changes. All of the metrics are geared towards one thing which is token maxing which is spending as much as you can on tokens to do AI. No one has an explanation as to why we're doing AI or indeed what doing AI results in. And that is kind of the core of the AI bubble.

[00:06:25]>> Do you think there's a part of it because I saw reporting talked about it on on the show um earlier today reporting about Microsoft kind of falling into that same bucket as Zillow and Uber saying, you know, we encouraged people to use AI kind of too many people used what what we'd budgeted for. We've blown through the budget really quickly.

[00:06:43]We're going to swap AI tools. Do you think that's going to be a thing that people will just be like, "Okay, now we're going to shop around. Instead of going with Claude, we're going to go with Chat GPT. We're going to go with C-Pilot. Whoever's cheapest." Are we going to see a race to the bottom where you've gone from being in a situation where you're trapped in adding extras and adding tokens and now the reverse will start happening and there'll be a competitive edge going on with people trying to offer people more for less.

[00:07:06]>> The thing is, all of these companies are horribly unprofitable. Even Anthropic Anthropic leaked that they had a profitable quarter and it was specifically because Elon Musk gave them a discount on compute just for two months. Amazing stuff. But nevertheless, these companies all lose billions. They all are terrible businesses. But fundamentally, it's not really possible to do a race to the bottom here. We can get to Chinese models in a second. Like OpenAI right now is handing out $1,000 of API credits to business customers for codecs. So that's interesting. That's OpenAI's plan. Just give people money and hope that that works out. But the thing is, this isn't a situation where a cheaper provider would necessarily fix the problems. Sure, people say Deep Seek, V4, or what have you. Sure, it's cheaper. But realistically, I don't think Deep, maybe it is, maybe this is damage to them, I don't know.

[00:07:58]But I think realistically, what people eventually do is not use AI at all or use it with a very aggressive cap. And the problem with that is is that let's say you cap engineers at hundred bucks a day. This engineer that you're capping has been spending thousands a week, tens of thousands a week. Their workflow is now completely different. They it just doesn't function in the same way. you.

[00:08:26]They have through token maxing, through incentivizing token maxing, they have intentionally created a situation where nobody really knows token austerity. No one really knows how to use less tokens.

[00:08:39]Nobody really knows how many tokens they're using. Uber's COO even said, "It's really easy to be impressed by AI when you're not paying the cost, but someone eventually has to pay that bill." And I think we are going to start seeing the true value of AI because everyone has been using the subsidized stuff. Everyone has been high on the hog using AI effectively for free. Like and the argument is of course people say well they're profitable on inference.

[00:09:08]Zero proof. None. There's no proof on that. If they were profit if Anthropic was profitable on inference they'd be profitable every month because they have 85% of their revenue in API calls. What I think is likely is that Anthropic is going to see their revenue burst and then contract violently. But I don't think that money goes elsewhere. I don't think people are going to go, "Well, I spent $200,000 a month or a million dollars a month on anthropic. I'm going to go spend 150 grand on uh on OpenAI."

[00:09:40]I don't really see who like I don't see who's doing that. I think the question is going to increasingly become why am I doing this at all? How do I measure this? And it's so funny. You go on Twitter and there's like post after post written by AI of course of people saying these 2,000word posts that mostly come down to, yeah, what if we measured what kind of return we got on our investment in something?

[00:10:07]And people responded going, holy crap, wow, it's the most important thing I've read all day. The the truth is we're our businesses are run by imbeciles. They're run by imbeciles that do not do real work. Uh business idiots as I call them who are disconnected from production. So yeah, if you use a chatbot that tells you everything you want to hear and can burp out code sometimes, yeah, you think anything can happen. Just let anyone use this as much as you can. Except that's not working and I think it's going to go poorly for everyone involved. Yeah, it's a really interesting thing that you gave the example of the engineer and how um their work um has materially changed.

[00:10:46]Well, I guess you know that's a kind even optimistic in of itself Ed like you know from what from what I've been seeing in the world of work often it's the extremely low-level work that is being given to AI or trying to be outsourced to AI and that's never going to create that ROI that makes it worthwhile. Right? So you got a situation where if the if the people who are being even moderately savvy about it are realizing that it's not worthwhile.

[00:11:09]You can imagine the cohort of people who are like, "Wait a second. I'm paying for what? For how much?" Right? It could it could be massive. When you think about it on those scales, >> well, I'll explain from the perspective of this payment processor I talked to.

[00:11:22]Apparently, they didn't have a way of tracking how much people were spending. They had a vague way. I know. I know. It's so cool.

[00:11:31]Horrible. So funny.

[00:11:32]>> It's so cool. I love But then someone on the financial side said, "Hey, hey, quick question. We're spending millions and millions and millions of dollars on this. Why?" Or may you know what? If not, why who is spending all this money?

[00:11:48]And then maybe and they don't even have a month's information. They don't it's a huge company. There are huge companies doing this. And I think what this is is this is kind of the apex of the business idiot. This is the ultimate form of an economy run by people that don't do anything. Because think of it like this.

[00:12:09]If you knew how a business worked and you weren't an imbecile, for example, and it made tens of millions, billions of dollars a year, for example, and you were like, "The power of AI is here."

[00:12:20]Would you not think to immediately be like, "But let's make sure we know what it is. Just make sure we know what it costs, how good it is." Nobody. Not Not one of them. None of them. None. Not once. I can find no examples. And I've looked. I've been talking to tons. My whole week has basically been writing and emailing people saying, "Hey, hey, do you have any way of measuring ROI?"

[00:12:45]And they don't. They don't have it. They have number of pull requests by AI which is basically saying how much stuff are you trying to push into the system? How much potential AI code are you trying to cram in there? And um days that you've used AI a number of sessions I believe it's just it's truly weird. It it's it's not even I it's not even a I can't even think of an analog because I was about to say oh measuring a restaurant based on how many people go to it versus the checks. But even then that would be a better measurement of popularity than how much we're using it. Not how much we're using it in outputs. And the thing is there is no way to measure this because it's actually quite difficult to measure ROI of anything. It's very difficult to measure knowledge work ROI which is why you don't massively increase the cost for no reason. That's why you do that when people do a computer job because it's quite difficult to measure. It's not impossible. OKRs as which are just the measure and or KPIs, whatever you choose, the measurement of how good you are at something or what you're doing.

[00:13:55]You can measure it, but you also need managers and executives who know what's going on and actually do work. And I don't know, I most CEOs don't seem to want to do any work. Most managers don't. Middle managers should be eradicated from our society. I don't mean violently. I just mean removed as a role. But the this is just we have an economy captured by middle managers.

[00:14:18]Middle managers who want to do things based on what makes them feel good. And that's all that AI does. Whatever you say for bloody years, people have been saying, "Oh, coding models, they're changing software engineer. They they they're writing all their their code with AI." And I kept saying the whole time, "Okay, okay, but what what what's the outcome? Are you shipping things faster? Is the stuff good? Does it work?

[00:14:46]Is your software better? Look at all of software right now. Look at every app you use and tell me, is software better now than it was in 2022 when Chat GPT came out. And the answer is unilaterally, no. Every single app I use is worse. Every single website is clunkier and more broken. Blue Sky, a website I love, which is also, they have said multiple times, run with a lot of Claude code stuff. Buggy. I can't crop images on mobile right now. The average piece of software you use, stinks. It's worse. So, yeah, that's the only output I can actually find from AI, other than the fact that we helped make Daario goddamn Amade a billionaire for shame.

[00:15:31]It's really interesting because you're you're you're absolutely right in the sense that, you know, plugging I feel like AI has been plugged into loads of different things that we used to use and it's not necessarily making it any better. But it but it is interesting that that dynamic of you know the way people think because I genuinely think companies first of all managers didn't ever think that the that the tap was going to be turned off. Um I think there's been a huge amount of support for big AI investments. We saw all the AI related layoffs or like the this the excuse used >> AI washing.

[00:16:02]>> Exactly. We saw all that. We saw the wave of that and now it's it's like that the the the I suppose the penny's dropping where where they're going, wait a second, you know, the stuff that we put in place to like measure things. The stuff that you mentioned is absolutely spot on. But you've got people whose performance is linked to how much they use AI. The more you use AI, the better you're going to perform, the better you'll be paid, the more your bonuses are linked to it. So the entire system that has been set up is is pretty broken. But also you think about that and you think that's going to completely collapse, right? And um yeah, it's it's this huge problem. It's it's a cultural reimagining of like what you need to think about right at work.

[00:16:40]>> What it is is that all these companies, as I've been saying for years, made it so that you could just use them as subscription, including with enterprises. So everyone's experience including every single journalist has been from these subsidized products with Claude. You can I think burn anywhere from $3 to $13.5 uh for every dollar of subscription revenue in tokens. They will claim this is profitable. It's not. Don't believe them. They're they're just saying stuff and also the companies haven't confirmed that. But so everyone's experience was based on this illusion. It's a mirage.

[00:17:15]It's this idea that this stuff is so cheap and so easy to use and just you can use it as much as you want.

[00:17:23]The moment they actually had to pay, suddenly everyone's talking like flipping John Major is back we got we got all be it's all back to bases all got austerity measures. We've got to pull everything in. What is going on?

[00:17:38]And the answer is they did not know what they were talking about before and they certainly don't now. The business idiots, the CEOs, the managers, the VPs, the people saying, "You need to use AI for everything, didn't know why they were doing it. They had no idea. They used it. They went, "Wow, I'm a I'm an imbecile. I'm an imbecile that only sends emails, goes to lunch, and leaves lunch. I love taking other people's work, and saying I did it, which makes AI the best thing ever." I assume because I can make this do something unspecified that everyone when given this will magically do something. Now you may think that they think make more money, make better margins, ship better software. None of these are the actual ROI because no one can measure them and no one can connect them because the actual way to measure it would be to use the product and talk to customers and that's just not what it is. That's what what a CEO does. Sachin Nadella doesn't make $90 million a year so that he has to talk to ah customers. Oh, he has to he has to use co-pilot to do something. And that's the thing. It's executive hysteria meeting with the fact that our economy is run by middle managers, MBAs and imbeciles, people that don't know anything about work. They don't do the job. They don't do the job of a software engineer. They don't do the job of anything. They go on podcasts. They complain that people aren't masculine enough. They spend 12 hours on Twitter a day slowly drifting towards the right.

[00:19:13]They post that they had a quadrillion tokens and then they have no idea what any of this means because they don't do work. And if you don't do work, how the hell would you know if people are being productive? Well, then you'd have to talk to them. Ah, no. AI was sold as this panacea for the management class, for the work to fully suppress the working class to fully ri wrap them up and take what they had, their creativity, their code, their writing and then sell it to other people effectively without paying anyone other than the model provider. Except the costs never lined up. The subscription economy was never ever sustainable. And so as a result, the moment that they ticked them over to actually paying the cost, everyone's freaking out because we've never really had good ways of measuring productivity. Organizations are too big. Google should be 13 companies. Apple should be what five or six like all these companies need to be broken up. But even then, you'd think, you'd want to think at least that all of these executives, the hyperscaler, Sundar Pashai of Google, Sachin Nadella of Microsoft, Andy Jesse of Amazon, Elon Musk of course of very of miscellaneous um and of and Tim Cook to a much lesser extent. But the point I'm making is you think executives and people generally think executives are smart. Executives, you wouldn't get to the top of a company if you weren't smart. And that is just fundamentally untrue. It's completely untrue. These people have no idea why they're doing anything other than what makes them feel good. They have no idea why they're making everyone use AI coding other than the fact that it personally impressed them and the people that they pay to agree with them on everything. And so they put AI everywhere. I am leading into the capex because people say these are the biggest software companies in the world. They couldn't possibly they wouldn't dare make such a huge mistake except now they're spending what $800 billion$900 billion in capex just this year trillion dollars next year and I think it's time to start asking do they have a plan is there a plan do they have a road map beyond we will keep expending as much as we possibly can and being the pay pigs for Jensen Wong of Nvidia or Micron or sending everyone's stocks like I guess that's what we do. No one has a plan.

[00:21:37]Sundapeshai of Google has no plan and neither does Sachinadella. That's the important message here. Nobody has a plan. AI is an exploitation of executive ignorance. And we're seeing it on the hardware side, the semi side, and the software side. And it all ends in tears.

[00:21:55]Yeah, it feels like with these um sort of CFO interventions and just looking at the numbers, we're kind of having a first peak behind the curtain, right?

[00:22:03]Where it kind of all felt rather obvious for those who were looking at it and being like this this doesn't stack up, but I guess it just becomes so that the rug's completely pulled, right? It becomes so obvious that this cannot continue at the pace that that that it's at. I mean, I know you talked about Anthropic and you talked about the problems that it might have with its revenue given these huge glaring issues that that it's facing with some of its clients, but I know you you wrote recently about um you know, open AI and what might happen there. I mean, that that seems to be another big problem that's just happening in front of our very eyes and unveiling itself at the moment.

[00:22:38]>> Yeah. And that's the thing. So, OpenAI per the information Q1 2026 they had an a non-GAAP. So non- generally accepted accounting principles uh operating margin of negative 122%. Which is uh not good. That means that they lose about $122 for every dollar they make which is insane. Openai's margins have got worse.

[00:23:04]Anthropic margins have got worse. They try and do this funny game of oh it's an operating margin so we leave this cost off. We add this cost remove this around. The problem is is that the same problems have been there for years.

[00:23:17]These products only lose money. The training costs are never going away.

[00:23:21]Inference is not profitable. If it was, these companies would only run inference. So creating the outputs. They would only sell model access. Why in the world would they have subscriptions at all? If that if they were profitable in inference, especially with Anthropic, who makes most of its money out of API calls? And the answer is cuz none of it makes sense. None of it is economically viable. It's just because we live in the era of the business idiot, the epoch of imbeciles. We are controlled by people that don't actually look. They don't think well this doesn't make any money loses a lot of money and it's not really replicable or measurable in any way, shape or form.

[00:24:01]But what if we kept doing this until it did? And that is the I believe that is the entire logic. I think OpenAI and Anthropic's plan is to keep raising money and keep burning it until they work out a business model. I think Microsoft, Google, Amazon, and Meta's plan is to keep spending money on capex until something works out. I mean, Mark Zuckerberg said yesterday in the shareholder meeting that if they think that they need all this capacity, they think, which is really bad because they're spending $125 billion on capacity this year. But in the event they build too much, they can just rent it to anyone else. They can rent it.

[00:24:39]They could become a Neocloud.

[00:24:42]What? What? That's your plan? Your plan is to become an infrastructure provider.

[00:24:47]You're an ad network attached to a dying social network. What do you mean infrastructure? That's because no plans.

[00:24:53]Open AAI no plans. OpenAI does not have one. They sign of where the tech media is right now. There was a story a few weeks ago that Open AI was expecting to lose I think like 75% of its paying customers for its $20 a month plan. But don't worry, they're going to replace them with over a hundred million in this year. By the way, this is what they're saying. A 100 million of their $5 or $8 a month ones. It's just that easy. you know, just yeah, you know, how you going to fix this? Well, 100 million people.

[00:25:26]>> But it's not just that. It's it's also advertising, right? They were saying that their plan was to try and and get that to to to happen um so that everyone who's using the free version of chat GBT has to now navigate through I don't know of that, >> right? But the funny thing is with ads is they're going up against Google, Facebook, and Tik Tok. three of the most three of the most performant ad platforms though Facebook has had some serious issues with advertisers losing money on the platform of Facebook doing bucker all nevertheless you don't just like add ads it's not easy to do it's also more difficult in a generative model because you don't know big thing with ads is placement is what are you seeing this next to it's one of the things that Musk had to deal with Twitter because posts from Nazis were appearing next to ads in this case, it's where are the ads appearing? How are the ads monetizing? All signs point to chat GPT being a pretty terrible channel.

[00:26:24]Walmart's shopping vertical that they did with them, the shopping um integration even, they ended up ending it because it wasn't making them any money, which is again, we keep coming back to this thing about like are we making money? It's the return on investment and that's because nobody really knows. So, they're going to try and add ads. they're going to try that.

[00:26:46]They're kind of already doing it. And then um profit, they don't really have an answer for them to scale to the level that OpenAI needs to. And I mean this is based on their own projections and based on um what they owe Oracle and Amazon and Microsoft. Yeah, they need $852 billion of cash flow in the next four years. Otherwise, Oracle dies. Oracle runs out of money and uh Coreweee likely does too. when actually a whole bunch of companies get materially harmed.

[00:27:17]And and what's funny is when you read these articles, no one asks how that will happen. No one is like, "Yeah, you know, um they'll raise more money than anyone's ever raised in a smaller period of time than anyone's ever raised it within $852 billion. No problem." Oh, also other problem, Anthropic needs 300 billion as well. Don't know where that's coming from also.

[00:27:43]That's 50% of the remaining performance obligations of Amazon, Google, and Microsoft. No problem. Though I'm sure they'll work that out somehow. Again, it's everyone is just like, "How are we going to work this out?"

[00:27:56]You know, don't worry about don't worry about it.

[00:28:00]Don't worry about it, mate. Uh, you know, I've known enough dodgy people in my life to to know when people are acting in a way where nobody's making any money. This is what this is. This is just ex exploitation of executive hysteria and it's hilarious but also terrible for our markets and economy.

[00:28:18]>> It kind of feels like the first person to like you know lose their nerve might be the one that that falls the first. I mean it's really remarkable as you say I completely agree there's there's been this sort of growth at all costs huge amount of optimistic support huge amount of lending. there's been so much creative borrowing happening and look I'm intrigued because I mean for for me I've always I've always felt that the the perhaps the the one that stands to gain the most has been Nvidia all of the you know circular economy that it's created it's become kind of like the de facto central bank for AI um it's still waving money around like the latest this week was a huge investment in Taiwan saying this is going to be the epicenter of AI it's a huge amount of money I mean they've been very transparent about everything that's going on. It's like they're hiding the fact that they're giving money so that people can then buy their their chips. I mean, I I don't know. I don't know. Oh no, they they are they are mostly hiding it. But people keep ignoring something which is in their latest earnings they have over the next 6 years pledged to rent eight I think it's $30 billion over the next five or six years uh just renting GPUs renting the GPUs that they are selling to people back to Nvidia that's not what you do when a business is working but the really funny thing is just Dario great j IO fantastic writer crazy website. Um he had a great point which is in their last quarter, Nvidia despite making what $81 billion in revenue that quarter only had $600 billion 600 million sorry more dollars in cash flow. Like their cash only increased 600 million cuz they've been spending tens of billions of dollars on these circular financing things. And it's complete. It's should not be legal. None of this should be legal. This is crazy. But the thing is, it's like, yeah, Nvidia is the only one really making money other than the other bits that go in a data center. All for data centers that are taking 18 to 24 months to build. And the demand doesn't exist for them. The demand doesn't exist for any of this. All of this demand is honestly like executive AI psychosis.

[00:30:34]>> Yeah. And all this kind of relies on deadlines that is very hard to see how they could happen. Um, and it relies on something not coming out of the woodworks. I mean, I know this week again there was a bit of reporting saying that Huawei might, you know, scupper some of Nvidia's plans that it's it's grown massively recently because it's managed to capture the China market, whereas Nvidia is really struggling to get a foothold there. So, it feels like if you're spending with one hand, which Nvidia definitely is, well, maybe with both hands, some might argue, but if you're spending at that level, you need to be confident that you're going to keep up that that growth and that expectation from investors because there's huge pressure on Nvidia.

[00:31:10]People think it's the weather vein for the entire industry. So, if it doesn't carry on growing, if Huawei is able to swoop in and take its crown, that could that could be the thing, right? And I I'm not sure. What do you think? Could that be the thing that that tumbles everything down if someone were to just come in and swoop Nvidia's crown?

[00:31:27]>> Well, the problem is with Nvidia, there's a few of them. The first one is is that data centers are not getting built at the level they're under construction, but they're not being built. They're not being finished. And by my own analysis, I went and looked at every single Microsoft data center a few weeks ago that was said to be under construction. None of them are finished.

[00:31:47]Not one. Not since 2023. Uh so this means that Nvidia is likely pre it's the largest Kickstarter in history. They are pre-selling I think tens of billions of GPUs and warehousing them. That is a problem because if that's happening now, eventually people might say, "Hey, I'm not installing these things. they're sitting on my balance sheet. I need to get rid of them or I don't need more of these. I bought too many which is what everyone should be saying. That is a huge problem for them. But the thing is uh there was a short seller called Kulpa Research who recently reported that they believe that 20% of Nvidia's revenue in fiscal year 2026 which ended at the beginning of this year confusingly um came from China and a company called Megape and intermediaries that they were taking the GPUs and sending them to China. So Nvidia is already cut off from China. I think that Huawei their whole situation is happening. I think the bigger question there is if China doesn't need Nvidia, does anyone else need Nvidia? Like there's there's just a general purpose question. But then it gets to a simpler point. Why do we need all these GPUs?

[00:32:59]Why what are they doing? Why? Like what what are they for? Inference training.

[00:33:05]Training what? Inference for whom? What does more capacity add to the AI industry? What does the AI industry lack right now that capacity will give them?

[00:33:14]Oh, there's compute constraints. I don't know about that. I think the actual compute constraints are coming from the fact that 70 to 80% of all AI capacity is taken up by two companies, Anthropic and Open AI. I also think that there is no proof that having more capacity makes these models better, just that having capacity constraints makes them worse. I don't know what the point of the capacity is, but I know for a fact that there is nowhere near the demand from customers, AI startups, businesses, whomever, for this for these data centers being built. Nowhere near like there's probably 10% of what they need, maybe less. So Nvidia's problem and why they're doing all of this circular financing is they know this. Jensen Hong 100%. He know I'm sure that Jensen Hong knows GPUs are getting to China. There was a Bloomberg story that showed last year that he had he was sat at a a club table with a woman called Alex Alice Wong, no relation, CEO of Mega Speed, which is sending the GPUs to China. Like there's obviously connections here that's going to come down. So what Nvidia is doing right now is getting as much money as possible, shoving as much money into their into their coffers as they can to keep the con going. It can't go forever because you've got on the buyer side, the people running these GPUs, you've got a business model that is mostly experimental. You've got people spending millions of dollars without knowing why. These aren't happy customers. These aren't people who are like, "Gee golly whiz, I love spending $500 million a year on this." They're like, "What is what's going on?" They feel they're being stung to death by bees the way they react to it. You've got your two largest customers by proxy, OpenAI and Anthropic. only lose money, worst met margins you've ever seen, and no real way out. Oh, and also hyperscalers who are mostly beholden to those companies. So, on the demand side, you have a big question. And on the supply side, Nvidia has made tens of billions of dollars of commitments to TSMC that um I don't know how they'll pay if there's cancellations of GPUs. That's actually a big unanswered question. Can people cancel GPU orders? Could somebody cancel $2 billion worth? Because that happened to Super Micro, which is a reseller. They take they build the AI servers, they put the GPUs in them, they ship it to a customer. Oracle cancelled a billion or two worth of B200 GPUs.

[00:35:36]That story kind of went under the radar.

[00:35:37]In fact, most stories about weird stuff with Nvidia just get aggressively ignored. Like potentially 20% of their revenue came from China last year, which would be vastly against sanctions.

[00:35:50]Just just no one talking. Yeah, don't worry about it. It's fine. It's kind of like how 2 years ago we found out that 10% of Meta's revenue came from scams.

[00:36:00]Like it's just there are these stories that people just will not internalize because once they do, they have to break the whole AI bubble. They have to the kayfabe has to end. You can't pro when you start accepting these weird realities. None of it makes sense.

[00:36:17]>> No, that's the that's the main problem, I suppose. And it's it's something that again you kind of think about the ecosystem in general and you're like okay if you actually think about it and you look at things closely that's when things just start looking more and more bizarre. And I think like your point about whether there could be huge cancellations. Great point. I think that's that's something that you know we haven't seen masses of uh but we h what we have seen is delays right. So even if everything goes really well we've seen what happens when Nvidia's chips are delayed even by a small amount. So you can imagine what might happen if you you see a situation where they might have over spent. You know, you never know. I don't know, you know, what they might have in their coffers for this quarter, but they might have overspent. They might not be able to, you know, deliver on on the on the promises of of this of the chips that you're arguing people don't even need to begin with. So all these all these things it's it's like the the the nerve of the the shareholders is also playing into this because as long as you produce what you're saying you're going to produce and you supposedly have buyers uh which may be the their own clients or whatever you know um but you know in these circular economy as long as people are still functioning as long as it still works on paper there's not a problem right you're still making money still all good >> well it's um Kakashi a great Substack writer he made he's made this point repeatedly that Nvidia is likely doing buy and hold or bill and hold, which is you say, "I want $10 billion worth of GPUs." Brilliant. Um, I don't have anywhere to put them. No problem. We just won't put them anywhere. There's an accounting term called transfer of ownership. They just go duh. These are technically yours. We're They're not going anywhere. We'll ship them when you need them. This way, Nvidia can book revenue and everything's fine. This is right on the edge of legal. Like, I think I don't think it's illegal. It's the kind of thing you have to disclose to shareholders. The other problem is it just occurred to me is that the next generation GPUs the Vera Rubins are going to be like twice as expensive and at some point you just got to ask how long can this go. It >> is interesting that we we operate in a system where you can just operate at a loss for the entirety of a company's life and it's absolutely fine. It is it's a bizarre it doesn't make any sense but it's it's totally the way it works.

[00:38:27]>> Well, that's the thing though. It's the mythology of Uber. Uber lost a lot of money. Except that's not what happened.

[00:38:33]Uber did lose about $30 billion. Uh OpenAI raised $122 billion this year.

[00:38:38]Anthropic has already raised 50, I think. Uh they're trying to raise even more. That's good. We love that. Um anyway, Uber did that and then became a kind of messy profitable because they're it was a lot of R&D and marketing expenses rather than what Open AI and Anthrop are doing which is just burning infrastructure costs. These it's fuel effectively. People say Amazon Web Services, Amazon Web Services cost a lot of money, cost $57 billion over 13 years. It was profitable in nine. The the mythologies that people have is that a company can stay unprofitable, but it will work out because again, the executive imbecile myths have traveled through everything to rationalize bad behavior and it's bad for everyone involved other than pretty much Sam Orman and Dario Amade and Jensen Hong.

[00:39:27]is really interesting when you look at the way things could fall apart as well.

[00:39:31]I mean, I know that there has been a lot of conversation about demand versus supply, right? And we know that there are lots of data centers out there that are not being used that we know that there are a load more that haven't even been built. Um, so, you know, the capacity isn't, you know, it's like we're we're going on to chat GPT premium version and it's going, "Oh, sorry, you can't use it right now because there's too many people using like that's not something that actually happens, right?"

[00:39:55]So we know that that that you know the demand is is not you know surpassing their supply. you've said it here, you know, it's it's very clear, but there's also a talk about, you know, what actually happens and you kind of mentioned it with the chips, but there was there's never been a very clear in my in my view um view of how long these chips, these Nvidia chips and other people's chips will actually last, their lifespan because it feels very much like that's another element where there's huge costs involved. You know, from what I understood, I saw some sort of graphic maybe a couple years ago about maybe things lasting about four years depending on usage. But if you if you picture in your mind, you know, huge data centers not being used, equipped with loads of stuff that might have an expiry date, that's a big problem as well. And I don't necessarily think it's it's part of the conversation where you go actually there's inventory. It's not just that you're spending loads on infrastructure that you might never need to use. It's that you might have already spent on infrastructure that's not being used that might have an expiry date, which is that's worrying, right? That's that's another part of the puzzle. It's worrying for those who own the GPUs because e even even if four to five years or six years and it's I've heard failure rates of as much as 10% on these things. It's hard to find them because no one wants to leak these ever. They they keep them close to their chest.

[00:41:13]Eitron 76, if you're listening, you'd like to talk. Um, but nevertheless, the other thing is that OpenAI Anthropic, they're spending all of this money and it's not an investment. They I guess they're training models, but the majority of their money goes towards running their operations. It's not like they are expending the capex. All of their infrastructure is being built by Microsoft, Google, and Amazon. Open AAI and Anthropic have very little actual IP beyond their models. And Microsoft has full access, can fully touch up uh OpenAI's IP as long as they want. I think they might have kept it at 2030, but OpenAI will be dead by then. So, who cares? The thing is, no real value is being created here. Six, even if it's 6 years for these GPUs and they die. Okay, 6 years from now, I guess we just spent $2 trillion.

[00:42:06]Why?

[00:42:08]Well, we lost a lot of money and we had to hear from terribly boring, annoying guys like Dario Amade and Sam Wman every day. Every day we had to hear Dario Amade go and 50% of jobs are going away due to the chat bots. Pretty good. Um, and I think that it's frustrating because nothing has really been built from this. The best we can say at the end of this is that more power got built. But in truth, I don't think much more power got built as a result of this. I think this is just a massive capital misallocation. And I think that we like let's say in let's say these things last for 7 years. It's still crap. It's still bad because GPUs can't really be used for other things. People want to tell themselves this comfortable bedtime story. Let's say OpenAI dies and all these GPUs are for well don't worry in the dotcom era all of the dark fiber got used and we're able to except dark fiber carries the internet GPUs carry AI analytic like deep analytics and uh 3D modeling those are not big industries and they're also going to be as expensive to run in a few years as they they are right now any incomplete data center is going to cost just as much to build and when you turn them on who are you going to rent them to? Aquaman like what what's the plan here? And there is none. No one has a backup plan. That's the craziest thing about this. No one appears to on the top end have thought what if we're wrong. And that's the only that's only something you can do from privilege. That's only something you will do if you've never got your teeth kicked in. You've never suffered. And that's what happens when you disconnect all management from production. When you let McKenzie freaks like Sundap Pashai and MBAs, all of them are MBAs other than Mark Zuckerberg. All of them. It's crazy. These people must be stopped.

[00:44:03]MBAs must be banned from business. I think we need to get rid of all the managers. We need to get rid of middle managers. We need to stop people that don't know how things work from touching the money. These people don't understand problems. They don't experience problems. Thus, they can't come up with solutions. And this is what happens when they're allowed to run wild. When they're allowed to pump the entire industry and society and the media full of propaganda, you get this trillions of dollars of waste, annoying people made rich, and what will probably amount to one of the largest destructions of value in shareholders, both private and public, in history, worse than the dot bubble. I am genuinely afraid of how this ends.

[00:44:45]>> My god, that's a really ble bleak picture um that you that you paint. And I think like I I want to delve into the people who are who are benefiting from this because I mean you're looking at not just the the companies that you've that you've spoken about but I mean this this week there were a couple of other companies chip companies that joined the trillion dollar club which you know it's it's just it feels insane because it's driven by this AI boom which you know as you have argued is a huge kind of fabrication kind of of a collective hysteria I imagine of people who have just decided this is going to happen at all cost. But you've got, you know, SKH Highix, which is from South Korea and and the American company Micron, basically, you know, riding that wave of AI data centers. You know, this is obviously revenue is not the same as like, you know, what's going to actually lie in anyone's pockets, but it certainly shows that, you know, there's there are there are people and there are companies that are benefiting from this movement, whether it's something that's going to end in tears or or not. And then I'm sure you saw the TSMC um news this week again that you know they're going to be giving loads of bonuses because you know they've they've done so well in the world of AI and and all that sort of stuff. It's it's fantastic. So there are people that are profiting off of this. Um I I I guess you know I'm just wondering how many more could join the club at this point? Is it now is it getting harder for for for you know companies and and for for for people to actually say okay I'll I'll do this while this lasts. So just to be clear, the value of a stock is nothing to do with its revenue. Just they are not connected. Perhaps the stock price is inspired by revenue perhaps. But it is all interpretive dance. Nevertheless, with Micron for example, Micron, SKH, Samsung, all these companies are see are pricing up their RAM. they are charging more because of the supply constraints and also selling more of this high bandwidth more expensive RAM which means that there is just flatout less space on foundaries uh to actually build these things to actually build more RAM for the rest of us which is why we're facing constraints. The thing is is that this is happening because of data centers cuz those NVL72 racks from Nvidia, I think they have like one and a half two terabytes of RAM on top of the high bandwidth RAM on the GPUs itself.

[00:47:07]Probably off a bit. Someone in the comments is going to be annoying about that. But point is they are the RAM. All of these sales of RAM, all of these sales of hardware and storage of SanDisk, it's all because of AI data centers. And that's where the scary thing is for me. That's where the that's where the scary thing is because um I don't know these data centers are getting built for demand that doesn't exist. We need at this point two or three more open AI and anthropic sized customers for AI compute. Otherwise we don't have an AI comput industry. Mo I think 90 easily 90% of all AI compute revenue is those two companies or Microsoft buying it for open AI. So of course these company like the the shovels are being bought for more money than ever. It's the most beautiful shovels we've ever seen. And it's frustrating because people see this and they go, "Oh, it's all proof. It's all proof that AI is here." No, not at all.

[00:48:05]This is proof that we have the largest real estate and industrial construction bubble of all time. It's bigger than it's bigger than any we've ever seen.

[00:48:13]This has more in common with the ghost cities in China and uh the the Dubai 2008 luxury car lots, the lots full of Bentleys and the like and Hummers unsold, unsullied but by people's hands that couldn't afford them. Right now, debt is easy to get for AI data centers.

[00:48:32]If you have a plausible enough plan, you can get debt. Which means where does that debt go? Right into building AI servers. Who do they buy the RAM from?

[00:48:42]Samsung, SKHEX. They buy it from Micron.

[00:48:46]Stock goes up, revenue goes up. The money may be real. Again, don't know if people can cancel. What happened in the com bubble was people cancelled a bunch of orders at once and then everything kind of fell apart. That could happen here. But even then at some point someone because look the fundamental theme of today's conversation is executives going wait did anyone check if we're doing this for any reason just did we did anyone check if this was good because it made me feel very important but did anyone check if this was good that is terrifying the AI industry now imagine that on the data center level say Goldman Sachs or JP Morgan SNBC or MUFG Okay, just whoever it is that's lending. What if someone there goes, "Wait a second. Is there demand for AI data centers?"

[00:49:36]And you may think, "That sounds like I'm being a little flippant."

[00:49:42]Right now, the biggest AI story is supposedly smart people going, "Did we check if this was good?" I think that the exact same thing will happen with data centers. I think someone will start saying maybe we want to wait to lend more money until we prove that any of this is worthwhile.

[00:50:00]All that's all it takes. And it's stupid. It should I don't know. I'm I'm crazy, right? I would have asked that question before spending the money.

[00:50:08]>> Yeah, you would think. Of course, >> but they're not. They didn't. And I think that the fact is the AI bubble may burst because someone will go, "Did we do this for any reason?

[00:50:20]That would be the worst the lamest way for the bubble to burst.

[00:50:26]That would be I've been saying for months like maybe it'll be new technology that comes in kind of similar to you remember when that everyone was obsessed with the metaverse and it was like oh the meta I remember oh let's oh it's amazing oh it's fantastic you know oh let's all buy a house next to Snoop Dog's virtual house and the metaverse it's going to be fantastic and then yeah and then that completely fell apart suddenly everyone's focused on something different I I've been saying for months I I really do think something else will come out of the woodworks and distract us and it will suddenly be like Oh yeah.

[00:50:56]Do you remember that we, you know, and they will stay as part of the overall technological landscape but not necessarily be as big a priority to the point where people are concerned that it's going to [ __ ] economies and plunge us into, you know, some sort of darkness. But but this this feels almost worse cuz it's like it's a sad a sad outcome where it's just like, oh wait, should we have done that just because we could? That's >> I think it it's not the end of burn after reading. We didn't learn anything.

[00:51:25]It's just like why' we do that? I don't know. I And I and I think I think people need to realize this is kind of the fundament of society. Great financial crisis. Hey, did anyone check if these mortgages would pay?

[00:51:40]And then the greatest the AIGs of the world. Hey, are we capitalized enough to pay out all of these insurance contracts we're selling? It turns out they weren't. The moment that they checked, everything fell apart. like it it it's the same old story. Every single bubble just ends in a very sad and gruesome way. It sucks for regular people the most. This one may be one which really hurts rich people too because private credit Oh yeah, it's also going to hurt because private credit is funded by insurance and pension funds which is bad. But the private creditors are going to suffer here. the asset managers, the Blackstones of the world because they're heavily levered in data centers too and they've doubled and tripled and quadrupled down on them. It's just I don't think like something is going to die. That's how these things end. Like something falls apart. An AI startup that can't afford to exist, for example.

[00:52:33]But I think it does end with just people asking really simple questions that I asked in 2023, like, "Hey, hey, does this make more money than it costs?" All right. If it does, is it good? Is it good enough to do it? Oh, you just you just made up something. That's what everyone did. They The craziest thing about the AI bubble is that everyone is talking about something that doesn't exist. They're like, "Oh, it's changing software. It could do." No, it's not.

[00:52:59]It's helping. It is helping some coders do some things and it's giving brain damage to many of them, making it impossible for them to write code anymore without having a chatbot that is not writing good code, but they don't know what good code looks like anymore because they haven't written it for a while. The inherent defensiveness of some LLM users is just proof that yeah, you're not confident in this at all. You know, I'm right. And all of those people were so defensive. A lot of the people in the media like The Atlantic did the big glossy story about Claude Code a few months ago. All of them are writing from the perspective of not paying the real cost for AI. Now that they're going to eventually have to eventually consumers, eventually regular people are going to have to pay the regular cost of AI and that's also going to make the feathers fly. Yeah, I think this is extremely good advice from Ed for all of you watching. Um, ask questions. That's basically ask the question and see what happens, right? Um, thank you so much, Ed, for joining me today. Um, expect to see all of you commenting in our comment section about what you think is going to happen next. And, um, thank you so much.

[00:54:03]>> Thanks a lot.