There's no other way to read this other than Meta has to—they built too much computer capacity and now they're selling it. I think what we're going to see now is kind of the H Bomber Guy thing of sell the compute to who? Aquaman. Meta is flat out of reasons to have this capacity. They said on that annual shareholders meeting they thought they had a reason. They thought they had a reason to have all this compute capacity, but they might sell it if they don't. And I guess they don't. I'm not sure how else to read this. This is very bad for the AI bubble. This is exactly what I feared, which is that these companies bought way more capacity than they could ever, ever need.
Welcome to Prof Markets. I'm Ed Elson. It is July 2nd. Let's check in on yesterday's market vitals. The Dow was roughly flat to start the third quarter. Meanwhile, chip stocks dragged the Nasdaq and the S&P 500 into the red. Brent crude dropped as the US signaled talks with Iran were productive. And the yield on 10-year Treasuries rose after Fed Chair Powell said inflation was, quote, too high. Okay, what else has happened? The cloud computing industry has just gained an unexpected new arrival. Meta is reportedly planning a cloud business to sell its excess AI capacity. The company has poured billions into its AI infrastructure, and until now Meta has maintained that all of those investments and the resulting compute capacity would be used for internal purposes. But clearly now something has changed. Meta stock closed up nearly 9% on the news, but other NeoCloud compute providers such as CoreWeave dropped on this news. So, for more on why Meta is getting into the compute game and what it means for the rest of the AI industry, we are joined again by Ed Zitron, author of the Where's Your Ed At newsletter and host of the Better Offline podcast. Ed, good to see you. I'll just start with the good news on my end, because for those of the listeners who follow me on social media, I actually bought Meta last week and it's up 10% since then. So technically, I'm happy because my view was that the price was pretty cheap relative to the rest of tech and relative to the S&P. That's the good news. The bad news potentially is we thought that Meta was going to build their own AI products. Now they're saying no, we're going to sell the compute to someone else for someone else to build their AI products. And that's why I wanted to talk to you, because it doesn't seem like a great signal for the AI industry, and you've been all over this. So your reactions to this, Meta?
Well, I think it's a sign that Meta is walking away from its AI play. It's Baron Barovski over at the Wall Street Journal made this point earlier. Meta was meant to have an API for its Muse Spark AI model weeks ago. Just hasn't happened. And now we got these rumors of Meta allegedly selling off its compute capacity. There's no other way to read this other than Meta has to—they've built too much compute capacity and now they're selling it. I think what we're going to see now is kind of the H Bomber Guy thing of sell the compute to who? Aquaman at some point, because who are the people that are buying this compute? Meta is, I think, the second or third largest buyer of AI compute. They have a $17 billion contract with Nebius, a $22.4 billion contract with CoreWeave. I'm not sure how this works. And the bulls are already trying to frame this as, oh, this is Meta intelligently monetizing their AI stack, when what it actually is is Meta is flat out of reasons to have this capacity. They said on their annual shareholders meeting they thought they had a reason. They use the word think. They thought they had a reason to have all this compute capacity, but they might sell it if they don't. And I guess they don't. And I'm not sure how else to read this. This is very bad for the AI bubble. This is exactly what I feared, which is that these companies put way more capacity than they could ever, ever need.
Yeah. Well, just to go back to the quote from Mark Zuckerberg. This is from May. He said, quote, when asked about whether he would do what he's doing now, which is sell the compute, he said, quote, we haven't done that yet because we think we have a use for the compute, but obviously if we get to a point where we feel that we have overbuilt, then that is an option that we have, and that is partially what gives us confidence in investing in building this out. So this seems to be the answer to that point, which is we did overbuild thinking that we were going to use all of the compute to build our own stuff, but now we're kind of admitting defeat and saying no, we're just going to sell the compute to someone else for someone else to build the stuff. To be fair to the bulls and, you know, explaining why the stock is up right now, I think Wall Street was concerned that they were building this stuff and they had no plan to monetize whatsoever. Why are you building all these data centers? Why are you building up all this compute? They've come out with an answer now. It's just not a great answer. It's probably the laziest answer you could come up with. And so this brings us to the question of who is going to monetize this stuff? Because if Meta had all those engineers and they had all of this, all of the resources, and they are deciding no, we're not going to monetize AI ourselves with a front-end product, who's going to do it? Will it be OpenAI? Will it be Anthropic? I mean, what do you think?
Well, I said on Twitter a few weeks ago, you know, that we're at the end when Anthropic buys compute from Meta. I think that if we see a deal signed with OpenAI and Meta or Anthropic and Meta, we're at the end. They're out of ideas because right now there are no large buyers of compute other than OpenAI and Anthropic. Meta was the other one. This whole time I've been saying, wow, Meta has no AI story. Meta doesn't really have a use for this compute. Turns out that Meta has no use for this compute because they don't have an AI story. So this will be actually an interesting demonstration of how much actual compute demand there is, because Meta—I don't imagine us as much of the—and they have a large amount. They have almost as much capacity as I think Microsoft. They were one of the largest buyers of H100s and H200s at the beginning. They're building that vast Hyperion data center now, which PIMCO owns most of the bonds of, which is interesting. It's great to hear where our retirement funds are going. But it's weird though. I thought that they needed all this compute. I thought it was very important they had this compute. And it just makes me wonder whether they ever had a strategy, because people want to say, oh, their strategy failed. I don't think they had one. I think that they just bought all the compute because that's all Mark Zuckerberg can do. He sees someone else do something and he goes, I'm going to do that. And now he's becoming what? A really big boring cloud infrastructure provider. I mean, it's also the question of how much money they'll make of this, because I've seen some fanciful projections, some truly ludicrous ones, with people saying they'll get $20 billion a gigawatt. That has no precedent, because Oracle, which is building Stargate Abene, the 1.2 gigawatt, I think it's 880 megawatt of critical IT for OpenAI, they're only expecting to make about $10 billion a year from that once it's fully built. So, okay, let's say that Meta gets lucky and makes $15 billion a year from their capacity. All right, I mean, for how long? Who can afford that? Because really, the answer is OpenAI and Anthropic. No one else is buying that much. I think Jane Street is getting 100, 200 megawatt maybe. But who are the other large buyers? Because there certainly isn't an aggregate of gigawatts of capacity demand. And this is the larger point I've been making about the buildout, which is that the demand does not exist at scale. And the largest consumers are two bulbous fail sons that only lose billions of dollars. And I don't know—I'm seeing people try and rationalize this. I must warn them that this is a very, very bearish sign.
Yeah, I think that's exactly right. And I mean, going back to the parallels to the dot-com bubble, I mean, the trouble that we keep running into is we keep hearing that there's all of this demand for compute and there's all of this demand for chips, which there is, which is why all the prices of this stuff is going up. But when you go to the demand for the AI products on the front end from the consumer perspective or from, you know, a B2B perspective, yes, they're technically spending money on this thing, but not nearly enough compared to the losses that we have seen you've reported on from Anthropic and from OpenAI. And so it does certainly seem that this is an admission of that point. I mean, I really—you'd really think that if someone was going to figure out how to make AI an extremely profitable consumer product or business-to-business product, Meta would be that company. But they have decided with this move, it seems, that they can't figure it out. And we'll see what happens. We'll see if they maybe go back on this. This is kind of initial reporting. Just going to that phrase though, excess compute. Mhm.
That's a crazy phrase to hear, and that was exactly what was written in the article, that there is excess compute and that's why they're making this move. I mean, what do we know about compute supply? I thought that there was no excess. I thought that we were completely constrained. So based on my analysis across all of these companies, I would say 80% of all AI compute is owned by or used by OpenAI or Anthropic, with the rest of it being Meta. Meta is—they have those multi-billion dollar deals. They've been—I sorry, I actually forgot—they are also doing 1.2 gigawatt with Crusoe. Just slipped my mind there. So what's happening is it's a mirage of demand. What it is is two large companies are taking up an absolute crap ton. They're buying anything they can get, leaving scraps for the rest of us. Well, well, not me, but you know what I mean. So these companies taking up all this space. So, of course, the excess compute. Meta isn't using it. I don't even think Microsoft is using all of their compute. I don't think anything that's being—like Google recently also said, well, it was reported that Google couldn't give Meta all of the access they needed for Gemini and PE because they have capacity constraints, and people were saying, well, that's a bullish sign because it shows the demand. No, it's Anthropic. It's all Anthropic. Anthropic is just rapacious in their demands for compute. So the excess compute is whatever is left over, the dregs that are left over from Anthropic and OpenAI. So I think that the moment there are large bumps of compute that enter the market, such as Meta leasing, it really comes down to how much they—if it's 10% or 30%, maybe if they do more than 50%, this is just going to flood the market because aggregate customers—I've really looked, I've done a lot of analyses on this. I've looked and looked, I can't find companies providing inference or people paying for AI compute in anything more than tens or maybe a hundred million dollars a year in spend. The demand is not there. And that was before we got to this weird thing where, yeah, everyone's cutting back. There was the UBS study that said 60% of organizations are token minimizing. They're limiting their token spend. We're not in a supply constraint situation. It's that just two companies are taking it all up. And if Meta dumps this onto the market, oh boy, we'll finally get to see whether we're truly in a supply chain crisis. Really just having two oaths taking it all up.
Yeah. Just thinking about how this might end if it is true that the ROI or that the business model doesn't work. This has been your contention. That's what you've been saying for a long time. I mean, I just think about how it would play out and who would get hurt. Clearly, if you're spending hundreds of billions of dollars on capex to build out the data centers and it turns out that the ROI isn't as large as we thought back down the chain, then that's going to hurt for you, but you might figure out a way to monetize those data centers eventually. That's sort of a question. The people, the companies for whom this will really hurt, clearly in my view, it's the companies that are offering the AI product whose job is to do the difficult task of figuring out how to make more money than they spend on the AI and who have not proven their business models yet. And that would be OpenAI and Anthropic. And if those are the only two companies that are essentially buying up all of that compute when you've done your analyses and tried to figure out who's buying this stuff and it's just those two, it seems that those are the ones who get hurt at the end of the day in a very significant way. Would that be your view?
I mean, it's everyone gets hurt. I think it's all of the construction firms. I think the construction firms underlying the data center. I think there are going to be just a bunch of data centers that get built and never get used, which will be hellish for the private credit funds which are backed by people's pensions and insurance annuities. I think that OpenAI and Anthropic will be able to find the compute they need, but then they will run out of money, and then that will leave all of this compute sitting. I think that this still hurts Meta because look, if Meta drops all this compute onto the market and they can't sell it, that's also bad. That means that Meta has got a bunch of compute it can't use and will, depending on how bad things are, actually have to take an impairment if they don't think they're going to be able to use all that compute. They're not going to be able to monetize it. Meta is actually in a position to mothball much more of this because they can just claim, oh yeah, we're working on ad models, just ads, just doing some ads on the back end. No one's ever going to find out. It will get bad for the Microsofts and Googles of the world. And I think that the apocalyptic conditions that there will be, if I'm right, will be quite severe. I think we're going to see a lot of incomplete data centers. I think we're going to see data centers that just cannot pay. They'll be assumed by lenders. And if that happens at scale, it's going to affect anyone who is invested in this. It's going to affect the Japanese stock market because of SNBC, Sumitomo Bank, two of the largest banks in Japan who are heavily levered in the data centers. I think the spread of this will be hellish, but it's going to kill the NeoClouds. I'm so tired of people on Twitter being like, oh, Nebius and Iron, they're so good. No, they're all invested in by Nvidia. They're all backstopped by Nvidia, and their customers are Microsoft for OpenAI, OpenAI, Anthropic, and Meta. So it's just—I don't think people realize how bad it is. If I'm right, I'm not even saying that this is not a bit. This is just if the demand is not there, there is allegedly over 100 gigawatt of compute capacity under construction. If we can only sell 6 gigawatt, if there's only demand for six, and that demand is mostly coming from two unprofitable companies that pretty much only make money because of a hype cycle and also subsidizing a lot of their products, what does that look like in the future? It looks like a disaster. And unlike the dot-com bubble, there is no useful infrastructure here. AI GPUs do not have other monetizable use cases anywhere near even a fraction of what's been promised of AI.
Final question for you before we let you go. Some news has occurred since I last had you. Specifically, OpenAI reportedly delaying their IPO. I have a tinfoil hat theory that the reason they did that is because of you, specifically the financials that you leaked on OpenAI and kind of the storm that that caused. Reactions to OpenAI apparently not going to go public in 2026 anymore.
Ratat, that's what happens when people see your stinky numbers. I'm serious. My tinfoil hat theory is that OpenAI's numbers in their S-1 differed from mine. I don't mean fraudulently, just to be clear. They just presented them in a different GAAP variant or they just didn't show all of them. I also think that just there was a very large loud discussion of the numbers, and I don't know, I imagine because the whole reason that they delayed, per the Times, that they say was their investors, their advisers wanted to get a trillion dollar valuation and they couldn't. That's very bad because their last valuation pre-money was 750 something billion, I think, something around there. Like a 30% premium. That shouldn't be a huge ask unless of course bankers looked at this and said, mate, that's not worth 700 or even 500 billion. Which kind of lines up with something from a few weeks ago. SoftBank tried to get a margin loan in which they used all of their OpenAI shares, which on paper are worth over a hundred billion, as collateral, a $6 billion loan, and the banks would not give it to them. That probably suggests that banks might have some issues with them. And people will say, well, Anthropic will be fine. Anthropic is exactly the same kind of business. They run in the same way. Oh, they use the enterprise more. Everyone's on the enterprise now, even OpenAI. So my reaction here is I don't think the future is going to be better for these companies to list at. And I think that this is—I don't know if OpenAI ever goes public now. Maybe they will. Maybe they slop this thing out there. Maybe they find a weird way in, but I don't know. If they want a trillion dollars, it's very clear that valuation isn't realistic, at least at this point in time.
Well, it's a shame because I was looking forward to seeing those financials. That was going to be my big—that was going to be my Super Bowl. And eventually, yeah, that S-1 when it happens, you and I will have to do some live stream or something and react in real time, but we'll have to wait if it ever happens. We'll see. Ed Zitron is the author of the Where's Your Ed At newsletter and host of the Better Offline podcast. Ed, appreciate your time. Thank you.
