Transcription
Anytime someone attempts to make the bull case for AI, they talk in the future tense. We actually have no profitable AI companies. We are three or four years into this. How are we still discussing if it's productive? This isn't an industry. It's two subsidiaries of the largest tech companies in the world doing battle to see who can lose the most money.
The artificial intelligence revolution is helping drive the stock market to record highs. But a growing debate on Wall Street is whether companies like OpenAI and Enthropic will ever generate enough revenue to justify the massive amount of money being spent to build AI. I'm joined by Easy Primary Research CEO Ed Zitron. He's also the host of Better Offline podcast and writes a newsletter called Where's Your Ed. He came on to discuss why he sees AI as the next major market bubble. So you've argued that the economics behind AI don't make sense, but some investors who own Nvidia, Broadcom, and some of these other bigger AI infrastructure trades have made a fortune. So in your view, when you say AI isn't working, what exactly in your view is not working?
So I think people conflate an equities bubble and equity returns with a functional industry. All of the money flowing into Broadcom especially uh Nvidia all of these AI data centers comes from speculative investment from heavily debt based or alternatively hyperscaler capex. So Microsoft, Google, Amazon, Meta buying a bunch of GPUs in the hopes that it will turn into something when you actually look at the raw revenues none of them actually disclose their AI revenues. They use run rate. Microsoft $37 billion of annual run rate sounds like a lot of money. Actually that's 3.08 08 billion a month and I think they spend $31.9 billion in capex in the last quarter. So it's not really a real revenue story anywhere. And indeed the two largest companies, OpenAI and Anthropic, both horribly unprofitable, both lose tens of billions of dollars. And on top of all of that, they are the largest consumers of AI compute. They are the largest customers. 70 to 80% of all compute revenue is them. They take up the majority of hyperscaler infrastructure. They are the largest clients of corewave because core wave's largest client is Microsoft for open AAI. The other one is Nvidia. Then the next one is Google for open AAI and of course open AAI. It's the same story everywhere. It's either speculative buildouts for data centers that will be used they hope by somebody except the demand doesn't really exist underneath. Um or hyperscaler buildouts for AI services that no one really wants to buy. No one has a good revenue story. And when you look at the returns on some of these AI companies, do you think it could be fair to say that the AI trade may be working right now, but the AI business isn't?
I actually think that that's a very fair thing to say. The AI trade, so just what is speculation just on stocks. Yeah, that's been working for people. That is completely separate, entirely separate from the AI industry at this point. I mean, Micron, the third largest company on the stock market. Now, the reason they're that is because AI data centers are being built and you need a ton of RAM. Okay, to simplify actually, high bandwidth RAM for GPUs, the things going in the data centers takes up more of the fab. So, there's less space for regular RAM, which means they can jack up the price on that. Point is, that's only happening because of the speculation on AI data centers. It's not happening because there's genuine diverse revenues. In fact, there are no diverse revenues. Most I think the information reported a few months ago that OpenAI and Anthropic make up 89% of the largest of the revenues of the largest AI companies. This isn't an industry. It's two subsidiaries of the largest tech companies in the world doing battle to see who can lose the most money.
And so you're arguing that the economics aren't working right now. Do you think that they could work in the future or are you saying it's just it's not going to work at all?
I don't think they're going to work in the future in the way they wanted to. I don't think that there is much of a future even with specialist silicon with etched with Broadcom's specialist jalapo chip with open AAI. I don't think any future exists for this at scale sold via the cloud. I think there could be very expensive kind of a boring Oracle style hardware licensing business. You buy an on-prem system for 100 grand 150 grand probably more now with the RAM crisis. I think the problem is is that when you even open source models, we've not proven that those are profitable either. We actually have no profitable AI companies. The problem there is when you let's say you're having a service to offer GLM 5.2, which is the the current model to Jour, the opensource model that's competitive with Opus 4.8. We if you set that up, you need to buy GPU capacity to serve customers. It only makes sense economically if you can successfully guess how much demand you have. If you buy too little capacity, you're leaving money on the table and you're going to provide a bad service for your customers. If you buy too much, you have a bunch of compute you have to pay for. Well, that no one's using. So, it's this constant thing if you're trying to knife catch, you're trying to work out how much you need. And I just don't think that there is an economic way of providing inference, unless you're willing to take a massive loss, in which case fine. But this is not like Amazon Web Services. This is not like Uber. Uber $33 billion over the course of its life before becoming this messy kind of profitable. Amazon Web Services between 2003 year it was launched and 2015 the year it became profitable the total capex normalized for inflation was $29.7 billion that is 3 billion $300 billion less than anthropic raised in February and little less than half of what they raised in May or June I think it was. Point I'm making is this doesn't align with any historical precedent. The only thing it aligns with is just the bust part of the dotcom bubble. And really, I mean the telecommunications part.
And you mentioned OpenAI's new chip that they're making. Do you think being able to do that inhouse could help their valuation at all?
I think well, there's two problems there. One, it's in-house, but Microsoft has complete access. Sachin Nadella said a few months ago, they have complete access to these chips. Indeed, Microsoft is buying a chunk of the chips. They're also not totally doing it in-house. They're building it with Broadcom. Broadcom is building the TPUs for Google as well, but which they're selling to Anthropic. There's a horrible circular financing thing there. I don't think that it gives them much of an advantage because well, they're yet to actually prove whether it will save them a significant amount of money. They are in early testing. This chip was meant to come out last year. It's been on the books for since 2023. People misreported it as, oh, it took us 9 months to get here. Now, there were reports in Reuters in 2023 about it. And I just don't see I don't see that scaling to help them anytime soon. And indeed, I don't know if there's any proof that it'll actually help significantly.
So, every earning season, we hear CEOs talk about AI investments, right? But where today can you actually point to measurable ROI?
Nowhere other than for the semiconductors company. Micron, SKH Highix, Samsung, they are living high on the hog. They're doing great. Nvidia, same deal. Broadcom kind of same deal though. Broadcom has to do weird debt deals. They have a $35 billion deal where Anthropic borrows $35 billion to buy TPUs from them. It's a mess. You're not really seeing any returns anywhere. Most most companies do not disclose their AI revenues. As I mentioned, Microsoft 37 billion run rate, which is month divided by 12. Maybe they don't define it. Amazon $15 billion run rate. Same deal. Those are not their AI revenues. That's a unspecific month times 12 for the across the board. People don't report AI revenues other than Salesforce which has some pathetic amount of annualized revenue. I think 900 million or a billion annualized for a company that makes 40 billion or more a year. And IBM very specifically last quarter said, "We are no longer disclosing these." Now, here's here's the thing. Public companies love good news. They love telling you things. When a public company stops telling you something, it's usually because it's bad. usually because they don't want you to think about it too much. And that's the story across the board. The only thing we hear out of companies about AI and money is how much it's costing them.
A recent Wall Street Journal survey found that about half of economists don't expect AI to significantly change overall employment over the next 5 years. Could the bull case sort of be that it could help workers, not necessarily replace them?
I think that's what they want us to believe and indeed that I guess LLMs can help people, but pretty much every study of the productivity of LLM has found out they make you less productive. They make you feel more productive cuz you're pointing at agents and making them do stuff even though they mathematically guaranteed to hallucinate, will make mistakes, but they're kind of an adult busy box. They make you a busy board even the thing they give to babies with the clickers. It's a thing that makes you feel productive and in kind of like a classic social media product engages you. It's meant to keep you using it. The actual you know what I'll make a simple point. We are three or four years into this. How are we still discussing if it's productive? I don't mean this as a criticism of you. I just mean in general there would be something you could point to with a really solid use case. The internet. People love to make this comparison. People doubted the early internet and the question productivity gains. I've read a lot of those articles. A lot of them were asking very reasonable questions, but also the thing that burst the dot bubble was speculation on very small internet startups and telecommunications companies that were doing the wackiest stuff with roundtpping. I mean, there was a deal with the Lucent Technologies where they loaned $2 billion to WinStar, a now dead company, uh, to make $100 million of revenue. Bizarre stuff. But point I'm making is there should not be a debate about this anymore. The fact that we're this much money over a trillion dollars in that we've done all of these things and we're still debating it is kind of the sign that it's not a real it's still Pinocchio. It's never becoming a real boy.
And a lot of experts like you mentioned like to point to the dot bubble. What do you think the biggest misconception is right now between if AI were to be a bubble and the dot bubble comparing the two?
So the.com bubble was really two bubbles. It was a telecommunications bubble. So the Lucents of the world and uh the Nortell and such and of course the the dotcom startups. The funny thing is the dotcom startups were pretty small. They were actually I think like the globe.com they got like a billion dollar market cap maybe. Point is the actual underlying layer the money was not really like money was lost there and the equities loss was significant but the actual spend wasn't really lost there. It was lost in the telecommunication stack. Now the misconception is this idea of a post.com bubble resolution similar as in there is no AI bubble recovery a an AI GPU that you buy today or you bought a year ago will cost just as much to run in 2030 the problems I've mentioned with inference will be there in 2030 perhaps they'll have better AS6 but who knows if that funding who knows if the companies that were building the specialist compute will finish their mission if the venture falls out and I think the big thing is is the idea of useful infrastructure. That cable in the ground, the fiber was easy to light up. The telecom's gear was expensive up front, but the capex to realize it wasn't crazy expensive, at least not compared to AI data centers. There's a compelling argument to be made that it's going to be more expensive to finish an incomplete data center in 2030 than it is today because of construction costs. There are limits, sorry, limits, supply chain, um, limits and shortages on the talent, on the electrical grade steel, on the transformers, on the turbines, on everything, everything. There's a shortage and construction is only going to get more expensive. It's naturally going to. So, I think that people kind of misremember the com bubble. They think that, oh, it was there was all this money spent and it was mostly on the startups. No, it's mostly on the gear. Okay, but we could turn that stuff into useful infrastructure. It was really easy to do that. AIG GPUs pretty much have no other use case there. People will say, "Oh, what about 3D modeling? What about data analytics?" And the answer is those are teeny teeny tiny industries. I mentioned earlier a number a few hundred million. Nvidia has said that they have a trillion insight into a trillion dollars of specifically Vera Rubin and Blackwell GPU sales through the end of 2027. work that out to about 40 GW of data center capacity, about 30 GW of chips. That means that they need by 2030, assuming these things get built, $400 and something billion dollars of annual compute spend to make that worthwhile. That won't change in a few years. That won't become much worse in a few years indeed if the bub when the bubble bursts. And I think that people need to realize is if we're having this conversation a year and a half ago, we could have done something. We could have slowed down. The problem is the market got addicted to Nvidia. The market got addicted to the quarterbyquarter stock salvation that is Nvidia. Nvidia will save us all. They're always growing. The only way that Nvidia can keep growing is if there's infinite debt, which there's not. Banks are, the FT reported about a month ago, afraid of choking on AI data center debt. And because these things are so much money up front, well, they need that debt. And if the debt runs out, it's over. And I don't think people get how levered various economies are. The three largest com three of the largest companies on the Japanese stock market are a RAM company, Soft Bank and uh MUFG, Mitsubishi, the bank. They are heavily leveled leveled in data centers. That RAM company is dependent on AI data centers. SoftBank is dependent on OpenAI going public to continue growing to well actually to continue existing in some ways and maybe Masoshi's son's end. So we're in this weird situation where I think people want it to be like the dotcom bubble. They desperately want it to because there was a nice story at the end. I don't see one and I don't take any joy in this. It's genuinely going to be a global catastrophe and it really makes my heart sink because we could have stopped this. We could have stopped this a year ago. We could have solved this in 2024. We could have pushed back and said, "Okay, let's let's not buy all this stuff up front." But the problem was is that the tech industry has run out of hyperrowth ideas. There is no smartphone. There's no Google search. There's no new app store. There's no new cloud computing. They needed AI to work. So, everyone funneled every dollar they had into this. And I worry that well, they've got no no more dollars after that. And then they have no ideas left.
Do you think that could be why Anthropic and Open AAI are pushing to go public right now?
Yes, I think that while OpenAI, New York Times, Mike Isaac reported that they're apparently considering pushing to 2027, which may as well be in 2040 at this point. Both of them are rushing because they want to dump their stocks onto the markets. They want to make it someone else's problem, which is pretty much their whole existence. Amazon, Google, and Microsoft built their capacity. They funded them. Everyone else has kind of dealt with their problems for them and now they want to dump them onto the public markets. The problem they have is SpaceX. SpaceX's bond sale. The bond buyers now underwater. SpaceX's stock has gone up. It's gone down. It's not really ripping in the same way they'd want it to. The fact that OpenAI delayed heavily suggests that Anthropic will face the same problem. It's and I think what it is is they know the markets are going to see these numbers and say, "I don't what do you mean? How much money do you need? when will you become profitable? And they don't have an answer. None of them do. They have been making this completely false assertion that they're profitable in inference. When you look at I reported OpenAI's uh audited financials a few weeks ago, they bundle their inference cost for free users in sales and marketing. They're intentionally doing accounting shenanigans. kind of gets back to the thing I've been saying, which is if this was a healthy, rigorous industry that was going to work out, they wouldn't be doing accounting shenanigans, they wouldn't be doing funny things. We wouldn't be asking these questions. We would be saying, "Okay, wow, that makes sense." Even during the growth of the internet. Even then, the questions that were being asked were about what would happen when the internet grow, where would the money come from, where would the productivity come from? And all of those were kind of solved by using the internet. I've used plenty of AI services. Trust me, none of them answer this grander question of how does this become a real business.
And why do you think CEOs and experts aren't talking about this like you are?
I think that I'm going to word this very precisely. I think a lot of executives don't do any work. I think a great many CEOs go from lunch to meeting to lunch and they read and sometimes answer emails. If that's what you do, this is magical. Yeah. Wow. It summarizes my emails for me. I don't have to do that. I think also a lot of executives are heavily levered in this. I think the Sachin Nadella, Sundar Pashai, Mark Zuckerbergs and Andy Jassis of the world, they have no other ideas. They don't. They have been trying for years to find a new thing, a new platform to sell consumer and enterprise services and they haven't. So, they're very desperate to push it. I also think that a lot of CEOs are fully disconnected from production. I don't think they contribute to the bottom line other than taking the most compensation and saying the most things. And I also think that AI is an ingratiation machine. It will make it will tell you every idea you've ever had is brilliant. And when you do something wrong, it'll give you a convenient excuse. People with the money, the people controlling the companies, the people making the calls. We are finally seeing what happens when you fully disconnect management from production. And the answer is you get grifted. You get grifted on an economic scale we've never seen. The internet, the original internet, the early internet, the telecom boom and the.com boom and bust were there were some grifters. Don't get me wrong, there were some real especially the web ones. There were some ridiculous ones. Yes, there was griffs with the financials, but there was no fundamental argument against the internet being useful and important. The poor Krugman thing about it being less influential than the facts. Yeah, he made a bad call. Wow. You've got one thing on this. I think AI is still trying to prove its rigor. You're still trying. You could immediately use the internet and say, "Wow, I can buy stuff online. I can game online. I can download things." Instead of receiving a letter, I can like there's an obvious thing there. With this, it's constantly people trying to put different harnesses and different scripts around large language models to get away from the fact that they are mathematically certain to hallucinate and that their outcomes are full of errors and that they are incredibly computationally expensive and they require constant training. Otherwise, they drift. And so we're in this horrible situation where the people with the money don't really know what's going on and they are the ones spending the money and the workers are being tortured and terrorized and forced to use these products and some of them are jumping up and saying I love this because they know they'll lose their job if they don't. It's the opposite of profit. It's the opposite of a functional economy. It is the most evil kind of speculation which is when you've removed reality from the markets. when you've removed reality even from when the companies are how the companies are run
And so we really look at this from the bare case different angles on the other side what would you say the strongest bull cases that you've seen?
I hate to be this way but I don't the only plausible bullcase I see is on device and I don't mean on your phone I mean specialist machines I already hear I've heard tale of Nvidia selling chips within the like home black home office-based on-prem Blackwell. I think that that's realistic. But I am talking specialist. We have this for a few coders. We use this. We they know what they're doing. They understand the outcomes. They understand the risks. They understand the limits. I've yet to find a compelling bull case because a lot of them, and this is a very important thing, I think every investor needs to know. Anytime someone attempts to make the bull case for AI, they talk in the f the um future tense. They can never talk in the present terms. They've immediately drifted when it will, if it will, how it will, it will will. They never talk about the today. And I think that that is the underlying problem. They can't talk about today cuz what we have today is pretty mediocre.
And you've argued that the media often reports AI projections without questioning them. If you were sitting down with Sam Alman, what would be the one question you would want to ask him?
H when you going to become profitable, Sam? Like that's that's the easiest first question. and he will dither and go I don't know and the next question would be why don't you know because you should at this point
To wrap things up here what's one metric that you'll be watching over this next year?
Capital expenditures from Microsoft Google and Amazon and Meta to an extent when one of them and it is a when pulls back on capex that is the trigger and I've been saying it for years that will send everything to the end it's got when one of them and it needs to be a commitment to it needs to We are reducing capex with regards to AI. That is the thing. It's not because Microsoft technically dropped quarter of quarter. I'm saying they need to say that this is the thing they're doing. And they'll probably couch it as AI efficiency or uh efficient use of AI. We found they'll say we've found ways to make it cheaper. They haven't. It's more expensive than ever thanks to the RAM. But they're going to make that case. And I saw someone over at Goldman Rich something who said that actually he believes one of the Delta analysts believes that the first hyperscaler to do this will be rewarded by the markets. Once that happens, it's over. Everything will change because the only reason they invested the capex was because they got rewarded. And when they get rewarded for stopping, they will because ultimately big tech is run by people that don't have any vision.