We have this incredibly bloated financial sector. It's a phony crisis that they don't have to be cuts. We know how to get out of the Great Depression. You spend money. If we take a look at the typical worker, they're not doing fine.
Hi, I'm Dean Baker. I'm your host for Mostly Economics. Today we have with us Ed Zitron. It's been about four or five months since we had him back. Ed is the CEO of EZPR and also the writer-editor of the great newsletter, Where's Your Ed At, which I strongly recommend if you want to keep up to date on the latest goings-on in the AI industry. And Ed, as many of you probably know, he's probably an even bigger critic of AI than I am. So, we got a lot to talk about. Thanks for coming on, Ed.
Thanks for having me back.
So, I'll start with where you broke a story. You broke the story, but you reported on the story this morning with Amazon pseudo-selling some of its chips. So, do you want to tell us what Amazon's latest brilliant move was?
Well, I'll tell you what it appears. So, what Amazon is doing is they have chips that are either deployed or are deploying in data centers. Not really clear which one it is, and they are putting them in an SPV and then leasing them back, and these investors are going to give them $8 billion. I assume these are investors without object permanence, which would be great for chips that aren't really installed anywhere. I will be honest, the response to this has been one of my favorite AI booster things because their response has been, 'Oh, leasebacks are fine. Leasebacks happen all the time. What are you talking about? Leasebacks are fine.' Yeah, like I think Enron did them. I'm pretty sure that Excite@Home did during the dot-com bubble, but they leased back their capacity from AT&T. I had a quick look at this last night. Either way, I only have a few questions. Okay, Amazon, why are you leasing back $8 billion of chips? And people are saying, okay, it's a balance sheet thing. Great. But I assume that Amazon is going to go right back this quarter and buy more GPUs, right? Why? What is going on? Is this just short-term cash flow? Amazon's a very big company to be doing weird stuff like this, and I had a brief look. I can't find even a leaseback transaction this large. I'm sure someone's done one at some point, but have you ever seen anything like this? Because I'm...
I haven't. I mean, I'm not in the habit of running through corporate books closely, but it strikes me as bizarre. You just ask the obvious question, you know, being the economist here, that you just go, why? You know, as you said, like, you know, this is costly just from the simple fact that you had to have people spend time negotiating this, work out the details. Why would you do that? And Amazon, at least in principle, is not a company that's short of cash. So, you could imagine a startup company would, you know, say, 'Oh, we need some capital. We need money right away. So why don't we sell...' you know, in effect it's a loan, you know, so you're selling something and then you're leasing it back, so you're in effect getting a loan. It's a different, you know, call it what you like, but I mean it's effectively a form of a loan. So again, you could imagine a startup that really needs capital, that might be a good way to get capital. You could at least understand them doing that, whether it's the best way or not, you know, you could argue, who knows. But that's a way to get capital, but Amazon in principle shouldn't need capital.
Except they're cash flow negative as of their last quarter.
Well, exactly. That's, you know, they're cash flow negative, and what they're concerned about, at least I'd imagine, I don't have an inside line there, what I imagine they're concerned about is having their bonds downgraded.
And also, but the thing as well is for a company like this, $8 billion, which, and they're very clearly going to buy more of these things. Is this a short-term cash flow issue? And I mean, this seems kind of obscene to say this about a company the size of Amazon, but when the cash flow is negative and it's 8 billion, I don't know. I'm not... sorry to talk over you. It's just the whole thing completely is just completely bizarre. And bringing up the article from the FT, it says thousands of Blackwell chips it is deploying in data centers. And then you look down in the article, it says and they have been deployed in more than a dozen US data centers. Which is it? It can't be both.
Yeah. And you know, and again, this is something you've been very much on top of. I think you even mentioned when I had you on the show last time, and it was news to me at the time, that a very, very large share, and maybe you have a sense of what the proportion is, but a very large share of the chips that Nvidia has sold and obviously booked the profit on aren't actually even in data centers. So on the one hand, we have the story that, oh, we've built all these data centers and they have all this capacity and they have to keep buying more and more chips because they're building more data centers. And what you pointed out is a very large share are not in data centers. They're sitting in warehouses.
Well, since then I did a big story with the Guardian where I got Microsoft GPU numbers. They got about 2.2 million chips, and I couldn't report at the time, but it was about two... it's a little bit under 2 GW worth of capacity, of AI capacity that Microsoft has. Since then, Bloomberg's Brody Ford reported that it is about 2 gigawatts of AI capacity despite Microsoft saying they have 12 gigawatts of data center capacity. All of this is to say that Microsoft has a teeny... they spent $265 billion on capex so far since 2022, but only installed about 50, 60 billion worth of chips. Then I went and looked up the construction in process part of all of the balance sheets across SpaceX, Oracle, Amazon, Google, Meta, and so on. There's about 374 billion of CIP right now. My estimate is over $200 billion of Nvidia chips are nowhere. I think they're sitting in warehouses or data centers without power. And then Bryce Elder from the FT reported the other day, I think it was a Jefferies or Morgan Stanley analyst, I can't remember, said that 50% of the chips sold in 2026 and 2027 have nowhere to plug in. This is ridiculous. This is... it's beyond dark fiber at this point. It is something altogether uglier.
Yeah. And just to, you know, to be clear to people who might not be following that closely, the reason why this is incredible: one is that they're acting as though all these chips are being gobbled up and put to use, not sitting in a warehouse. They're being put to use as fast as Nvidia could produce them. And obviously that's not true. But the other point about chips sitting in a warehouse is they become obsolete. So on the one hand, if you actually have the chips in operation, it turns out, this was news to me, that they burn out very quickly. So it's only like two years or so. So I think even as short as 18 months in some cases, that's their active life. So once they're in use... but needless to say, if they're not in use, they do become obsolete fairly quickly. So if you envision these chips are going to be sitting in warehouses for... again, some of them might be there six months, a year already. If they're going to be there another year or two years, no one's going to want them. So the obsolescence question is also weird because...
Sure, they may burn out in three years or four years or what have you, but how does obsolescence work when most of the chips aren't even plugged in anywhere? Because Blackwell was, I think, they started shipping it beginning of 2025, maybe late 2024, but I don't think more than half of those are installed. Though while they might be obsolete in some way, I think it's also the other way of just... obsolete or not, there's just this overwhelming supply waiting to come online. And the question is how quickly is it happening? And we actually don't know. I understand that usually companies don't do asset-by-asset breakdowns in the balance sheet and all, but I feel like this is just too large for them to just be like, 'It's in the PP&E, go. It's in a big pile of stuff. Don't worry about it.' I am worried about it. I worry about it all the time.
Yeah. Yeah. So, again, that's not something they've been anxious to highlight. I mean, you uncovered it and others have looked into this now, but they haven't been anxious to highlight that. 'Oh, look at this. We have these warehouses filled with chips that no one's, you know, they don't know when they'll be able to use.' Um, let me get to Oracle. I know you have a fondness for Oracle and we've seen some action with Oracle in the last, I guess, two weeks. I forget, when was it when they issued force majeure? That was the week.
About a week and a half ago. Yeah.
Yeah. So, you want to say a little bit about what's going on? I may have my own speculation, but I'll let you go. Go ahead and I'll share my two cents for what it's worth.
So, let me set the scene. September 2025, Oracle announces a $300 billion five-year contract with OpenAI. Oracle doesn't have the capacity to serve it. OpenAI doesn't have the cash to pay for it. Nevertheless, this begins the construction of about 7.1 gigawatts worth of capacity, data center capacity, I should say, in Stargate Abilene, which already started in June 2024, but New Mexico, Doña Ana, New Mexico; Saline, Michigan; Port, Washington; Shackelford, Texas, and there's another one I can't remember, there's just... I think it might be another Texas one. Nevertheless, all of these have faced the big problems that every data center faces, which is the power issues. Just getting the power, because you can have the power plant around, but actually getting the power in a stable fashion to the data center, well, that's difficult. There's also the fact that it requires a bunch of electrical-grade steel, which is in short supply, shortage of talent, there's all sorts of reasons. Anyhow, so they've broken ground on all these things now, but out in New Mexico, they have had a hell of a time getting power to the place, and they wanted to use a natural gas pipeline, I believe. But then the state said, 'Hell no, you're not doing that at all. We don't like... we don't want this to happen. You need to resubmit the permitting.' They did it. They resubmitted. New Mexico said, 'No, no, no. That's not good enough. Do it again.' So, this data center... there's so much to this, so I'm going to simplify a bit. This data center was funded by a group of bankers and led by a company called Blue Owl, the private credit fund. Now, it's also a Stack Infrastructure, which is connected to Blue Owl. I can't remember if they own it or just have a majority share, but while working on this thing, they have faced numerous delays, primarily power-related, but some construction-related and weather-related. Nevertheless, the other day, Oracle... the way it was worded was so strange. It was they actioned a force majeure clause within the contract that allows them to delay initial payments due to late power delivery. However, this announcement came out. Oracle said force majeures are super normal. They're not, but that's what they said. These are regular things that people do all the time. But then the FT published a piece saying that actually Oracle has to pay for this thing whether or not it has power. So, it appears that what Oracle has been doing is saying, 'Hey, the power doesn't get there. I don't have to pay for this, right? It's not going to work. I'm not going to have to pay for it. Is that okay?' And it appears everyone said no. Actually, Larry, you're going to have to pay that one regardless. Now, the other fun thing about this is that this is not...
Larry Ellison. Yes. Pardon me.
No, no, no. I'm going off now.
But the other problem is that all of these things are done through these nebulous special purpose vehicles. So, technically, this is not an Oracle data center. It is a project LLC that then Oracle leases from. Now, I can't remember if this one is, but a few of these are non-recourse, which means that if it falls apart, you can sell the pieces, but Oracle doesn't have to pay you back. I know Michigan actually is recourse, that the bankers there said, 'No, no, no. If this goes under, you're going to have to pay Oracle.' But this one is just very strange because I can find no data center on record that's had a force majeure-related thing. Like, I literally can't. I actually genuinely... I've talked to a few people within power planning. They've said they've never heard of it with relation to this. Usually force majeure refers to, for example, the... SAS have said they've called a force majeure on liquefied natural gas because of the war in Iran, which makes sense because it's a... it obviously there is a war going on. How do you get the power? I think it's Pakistan and India might be the ones, but they're not sending it. But the big thing as well with this is that Oracle's bonds have also been selling off already. So it's not just about the project. It's about the overall credit health of this company. And I can stop for a bit because I know I've thrown a hundred things at you.
Yeah. No, that's obviously... the credit health is a very serious issue. So yeah, I mean there's been a big selloff in Amazon. Oracle's bonds, maybe they will be with Amazon too, but Oracle's bonds, their interest rate was up around 8.7, I think, which, you know, puts them in junk territory. They aren't rated as junk. They're still rated a grade above junk, unless they've taken some steps very recently.
No, if they had been rated junk today, I'd have party streamers. I'd have a hat on.
