It means that Nvidia is strongarming their customers and saying, "You have to take these now or you won't get an allocation in the future." And the pay pigs over at Amazon and Google and Microsoft saying, "Of course, here's my credit card right away," even though they have nothing to do with them. This is bad because it's created a global consumer electronics price inflation. The cost of RAM is up everywhere. The cost of basically every imaginable consumer electronic is up so that these people can put GPUs in warehouses. It's actually morally reprehensible.
Welcome to Prof G Markets. I'm Ed Elson. It is October 21st. Let's check in on yesterday's market vitals. The S&P 500 and the Dow climbed. The Nasdaq was flat as chip makers fell again. Brent crude dipped just below $109 per barrel. The yield on 10-year Treasuries decreased, and Netflix shares fell in after-hours trading following earnings that showed the slowest revenue growth in three years. Okay, what else is happening?
Amazon is spending $220 billion this year to build AI. However, some of what it bought may be sitting unused. According to new reporting from Business Insider, despite the fact that Amazon is spending billions of dollars a month on Nvidia server racks and GPUs, it can't actually immediately install them in their data centers. Because these chips are so scarce, Amazon and other companies are forced to buy more than they currently have the capacity to use, meaning that some of these racks are currently sitting inside of a warehouse. An Amazon spokesperson said that this is routine and, quote, "best business practice," but others fear that this is proof that the hyperscalers have overestimated AI demand. Our next guest is one of those people. Last month, he wrote an investigative piece estimating that roughly half of AI chips in America are, quote, "sitting in warehouses." He believes that hyperscalers significantly overbought GPUs relative to data center construction. And while this news doesn't necessarily prove fully that he's right, it certainly suggests that he might be. So, here to discuss what is happening in AI, what is the meaning of these chips in warehouses, we're speaking with Ed Zitron, CEO of EZPR and author of the Where's Your Ed At newsletter.
Ed, thank you for joining us on the show. I want to start with a report that you wrote last month that you titled "Where Are All the AI Chips?" And you asked this question. You say, where is all the money that these hyperscalers are spending? Where is it actually going? And you said it's going into, quote, "a warehouse" or, quote, "into a data center without power." That is the implication of the news that we just got from Business Insider. Before we get to that, what was the impetus for this original investigation? Why is this warehouse problem relevant to the AI story?
So a few months ago, I did a story with Pressure Down over at The Guardian where I got a source who had given me a fairly precise list of all the AI chips that Microsoft has in operation, and I didn't publish it. It was about 1.993 gigawatts worth. The reason they didn't publish it is we couldn't ascertain whether it was inclusive or exclusive of OpenAI. Bloomberg then reported about a month later that Microsoft only had 2 gigawatts worth of actual AI data center capacity despite saying it brought on a gigawatt every single quarter, a single gigawatt for three quarters straight. So they, I think they said they had 12 gigawatts total, but only 2 gigawatts of that is AI. This made me think, okay, I've got to start looking around. And when you look across the various hyperscaler earnings, Amazon only reports it on an annual basis, but there's a thing called construction in progress, which is basically where chips go if you have this segment of your balance sheet. It's where chips go before you use them. And it's where data centers in progress. It's exactly what it sounds like. And I found that there was hundreds and hundreds of billions. I think it's $374 billion of construction in progress across CoreWeave, SpaceX, Amazon. I would Google as well. Google has a remarkable over 100 billion themselves. And the thing is, people will say, oh, it could involve data centers as well, but this is the thing. The big reason I looked into this is everyone's so cloak and dagger about this despite this being the biggest, most hugest thing ever that blah blah blah blah blah. They're so proud of buying the chips, but they're not proud of where they're going. They don't really want to communicate even how much capacity they have, which means, in my opinion, and I think it's kind of being proven true, that they don't have that much actually going on. They don't actually have that much installed. They will say power. They'll say power is the problem. I think it's also construction delays. I think it's a lack of electrical grade steel, things connected to power perhaps. But I don't know. I think there's a lot more going on here than just, oh, we don't have them, don't have the space to plug them in. And the Amazon story kind of elaborates on that too.
Yeah. Yeah. It sounds like what's happening is it's not that they're not buying the chips, they are buying the chips. The question is where do the chips go once they've bought them. And the implication of this news, and to be clear, we actually don't know how many chips are in the warehouse. We don't know for how long they're going to stay in the warehouse, but it does align quite closely with something that you have been putting forward, which is that it seems like the data centers are running into problems and it means that they have all of these chips that they can't actually plug in and then use. Clearly, there seems to be some sort of mismatch between the rate at which data centers are being constructed and the rate at which hyperscalers are buying chips. My question, why is that bad? Why is it a problem in your view if there is this mismatch, which could, you know, some would argue this is what happens in construction, you buy things, timelines don't match up. Why is this a problem in your view?
So the Business Insider story is written by Eugene Kim, who's been an Amazon reporter for donkey's years. He's a really great reporter, and there were some curious details in there that suggest that Nvidia is channel stuffing. Channel stuffing is when you force sales through the channel without needing to. And the reason I believe that is part of the story says that there was a period when Amazon didn't have to pay for the racks before they were put in the data center. Very standard. You would pay closer to when they're going in there. You'd pay, say, when they're about to be installed. Except something called the early rack receipt program started. It wasn't clear when. And suddenly Amazon has to start receiving racks early. So early that there was a quote in there that they had no foreseeable path to installing some indeterminate amount. I believe because there's this also thing you said earlier that they say, oh, because of the scarcity of Nvidia GPUs. There's no scarcity. If there was scarcity, like, why would there be scarcity? People can't install them, unless, of course, the scarcity is created by Nvidia saying you have to buy these and you have to receive them immediately even if you don't need them, which is obviously what's happening. This isn't a theory. This isn't a grievance. This is what was said in the story. It's obvious that Nvidia is saying you have to take these early, unless Amazon just happens to want them early, which I don't think is the case. And the fact of the matter is this is bad for a number of reasons. One, Nvidia is artificially inflating its revenues. Now, people will say, like, the CEO in Margin Call, "I'm selling to willing buyers." Oh, this is fine. And perhaps it is. It's fully legal, I guess. Though I don't think channel stuffing is well looked upon by the SEC. But nevertheless, it means that Nvidia is strongarming their customers and saying, "You have to take these now or you won't get an allocation in the future." And the pay pigs over at Amazon and Google and Microsoft saying, "Of course, here's my credit card right away," even though they have nothing to do with them. This is bad because it's created a global consumer electronics price inflation. The cost of RAM is up everywhere. The cost of basically every imaginable consumer electronic is up so that these people can put GPUs in warehouses. It's actually morally reprehensible. And they will say, "Oh, it's about planning." It's clearly not about planning. Because if it was about planning, you'd have planned for this. It wouldn't be, there wouldn't be emails going about being like, "Yeah, I got no idea where to put these, mate." If you were planning, you would say, "Okay, we will have a, we're planning to get in an allocation three, six months late, early. We're going to put them here and here." No, this sounds like, because again, the original statement was Nvidia didn't, they used to have to not pay for the racks until they put them in the data center. Now, early rack receipt, we got to do it. I think Nvidia is channel stuffing to pump its earnings. That's a bad sign for Nvidia, but also has created material harm throughout the world, has made the price of it, and people say, oh, you don't need consumer electronics. Yeah, as long as you don't want to live in society. You want to buy a phone, you want to buy a laptop. Even the cheap ones, they're more expensive now. Every single one, even the cheapest Apple MacBook Air is more expensive as a result of this. And it's all because Jensen Huang is greedy and his hyperscaler pay pigs will do whatever they can to make sure that they can keep this up. Because if they stop buying GPUs, the markets will immediately reward them. And then they'll say, "Why'd you do that?" As long as they keep buying them, they can keep kicking the can and they can keep saying, "Oh, AI will pay off in the future, will pay off at some point in the future."
I think I am 100% in agreement with you that there is a pricing or a production mismatch. There is something, there is some misalignment in the market here which is clearly causing memory prices to soar and it is why iPhones and consumer electronics are soaring, 100%. I'm not sure I agree with you on what exactly is causing it because in my view, I mean, on the one hand there's this idea that maybe Nvidia and Jensen Huang are trying to force the hyperscalers' hand and say, okay, you need to buy our chips early and you basically need to overorder, you know, ahead of schedule. But at the same time, I'm looking at this other thing, which is that it does appear that data center construction is being massively delayed. It does appear that investors and the builders of these data center companies just didn't really know how difficult and expensive it would be to build. Possibly because a lot of, we're seeing a lot of this political backlash, which might be a real problem in terms of the AI buildout. And to me, that is a plausible explanation as to why there is a mismatch right now. Maybe Amazon goes and they buy the chips. They think that the data centers are going to come online within the next few months or so, and then they learn actually no, because a local community said we don't want you to do this, or for whatever reason. To be clear, I'm not actually sure what the reasons are, but it seems to me more like a supply problem of the data centers. That to me seems more plausible than an artificially intentional decision by Jensen Huang. What would you say to that?
I would agree with you if it wasn't for the explicit use of a term of early rack receipt program. If it wasn't for the fact that within that story they say they are taking $3 billion worth of deliveries every single month for the foreseeable future. If they know that there's tons of delays and they've known a while, the delays have been obvious for six to twelve months, then they could say, hey, Nvidia, we won't need these until here.
Could it not be that they thought that there was going to be huge demand and Amazon decided themselves, we want to try to get ahead of the curve here. We're going to sign up for some program where we get early chips, and now they learn that was a stupid idea because we have all of these chips in a warehouse.
I think early rack receipt is an Amazon program versus an Nvidia one. And again, I would agree if it wasn't for the fact that the story also said that they previously did not have to pay for racks until they went in. That's where the disconnect is. If it was, because if it was consistently that Jensen Huang was saying, you pay for these things, you take them when they're ready, that would be, it would be bad. It would suck, but it would make, it would fit in. It would be a standard business thing. However, it's obvious that something has changed. It's in the story. And remember, Nvidia has ballooning accounts receivable as well. So who knows what is going on there. Their accounts receivable, I think for the first half of the year was five companies with 70% of it. How much of that is Amazon? We don't know. And I agree there are absolutely delays. But the artificial increase in the cost of RAM is directly a result of Nvidia taking up so much of the fabs. It's a direct result of the high bandwidth memory going into Nvidia GPUs that are sitting in warehouses. I think hundreds of millions, at least $200 billion worth of them are sitting in warehouses.
And here's the thing. Again, I would agree with you if it wasn't for the fact there's so much.
If it was, okay, we've hit a wall, and Nvidia goes, "All right, yeah, you don't like, we'll stop shipping them to you." No, I think Jensen has these morons over a barrel because outside of Anthropic and OpenAI, there's really no large customers of AI compute. You can listen to whatever insane guy on Twitter all you want, but the analysts across the financial markets all are in agreement that 70% plus is Anthropic and OpenAI. So, it's not like they have insane demand outside of that they need to capture. I think that they are just kind of doing what they can to keep up. I think that they're doing this so that they can make sure they don't piss off Jensen. There was a story in The Wall Street Journal a few months ago saying that if you as a neocloud were to buy AMD GPUs, that you would get put in something called Jensen jail, which is when Jensen Huang will no longer sell to you. General story. I think it's Robbie Wheel and Katherine Blunt, I believe he wrote, is about Google's TPUs. So Jensen Huang is very, he's done this before. He's done it. I think that this is a case where he is channel stuffing, where he's saying to his partners, "No, you will take delivery or you won't get Vera Rubin." And I know for a fact, I have a source who told me that there are Vera Rubins making their way to Amazon. So, it's obvious that they're buying so many things. They have Blackwells that they haven't installed, but they're buying and taking receipt of Vera Rubins. This is very strange. And it strikes me as that Nvidia is just channel... I can see no other way. And people really want, I'm not saying you here, but people really want to take a charitable view of Jensen goddamn Huang, who misstated that gaming revenue was growing about a decade ago when it was actually crypto sales, who was consistent.
I want to know what the real truth is. And I mean, I think that I can see a world in which there was a gold rush. I mean, this story to me is quite indicative of what we saw with, like, the railroads as an example, and we just had Leek Ahmed on to explain the similarities, where to me, I'm less jumping to Jensen manufactured this, but we know he had a lot of power. There was a chip rush that was akin to a gold rush. Everyone wanted to buy chips. Suddenly everyone's trying to buy them. And to me this is an example of, oh yeah, we overspent. This is sort of the physical example of we put too many railroad tracks on the ground and now we're not using them. Now they're sitting in a warehouse because we didn't fully figure out how to match our timelines. And let me just be explicit about what I'm saying. I think you were right for the most part. I think, yeah, he got a ton of orders that backed up.
Like, that's one, like, there was a gold rush. It happened. I now think with this early rack receipt program he's taking advantage of that because this whole thing about scarcity just does not make sense to me. What scarcity? When you've got Microsoft, Google, Amazon, Meta, all of these companies with hundreds of billions of dollars of GPUs they can't install, how is there scarcity? Scarcity? Like, there's clearly enough of them. There's clearly so many of them you can't install them. What's making the scarcity happen? Or indeed maybe the question is what would make these scarce? And it seems what would make them scarce was not taking delivery. We may never know unless someone gets actual reporting around that. And Eugene Kim did a magnificent job with this. Eugene has been around forever. He should get his roses over this one. He really kicked ass on this report. But I think even if you don't agree with me on the channel stuffing, this is a sign that hyperscalers are overburdened with these assets and we are so close to this thing bursting and at some point investors are going to turn on them and start demanding they pull back on capex. I mean, justifiably, because if you're an investor in a hyperscaler right now and you're saying, okay, I invest and I want the value of my equity to go up because you're buying these GPUs, and you hear these things aren't even being installed, the natural question is why? Why are you still doing this? Oh, you claim there's demand. Why not stop buying them until you've taken your fill? Is there a reason until the data centers are able to be operational?
Well, and that to me is less of a they pulled a fast one and more of a they just haven't figured this problem out. And so, I mean, as we start to wrap up here, thinking about the next chapter in this story, I mean, it seems to me that if we have physical examples of hyperscalers overbuying chips, the next chapter in the story is they're going to buy fewer chips because for whatever reason it is, whether it's because of channel stuffing, I'm not sure about it, but I mean, what is certainly clear is the data centers aren't coming online as quickly as they had hoped. That is a good enough reason, I think, to say, "Okay, let's reduce our order next month. Let's buy fewer chips than we did last month." And we're starting to see some signs, some bad signs, in the markets. We've seen CDS spreads on some of the big data center names widening. Oracle's five-year credit swaps hit a record high last week. We have Ramp data showing that token spending is not growing. In fact, it's been flat for two straight months. One of these AI data center companies just went public. Elise fell more than 30%. There are bad signs brewing. I guess as we start to wrap here, what are you focused on? What data or what signs do you think are most important in terms of understanding what's next in this story?
I think capex. I think what they do with capex. If capex is flat, that's a problem. If capex is higher, that's a bigger problem because the question is why? But I think it ultimately comes down to does Anthropic go public? Does Anthropic go public at 15? Does it go public at two? Because if it goes public at 15, that screws OpenAI's 1.4 trillion private round. And also, Anthropic's numbers came out just before we went through it. These companies are clearly not economically viable unless you really squint. And so, I think it's going to come down to is the capex going to keep up? Because you can remove Anthropic and OpenAI fully from the picture. Hyperscalers don't disclose their AI revenues. So it really comes down to will hyperscalers keep spending? Because if they stop, I don't think that there are any parties big enough to fill the gap with Nvidia other than SpaceX and Oracle. And we're at a point when even still, the 10-year Treasury's at 5.1 still. So, I mean, things are so expensive and they're only seeming to get more expensive. And the more of the chips they buy, and these chips that are waiting in warehouses, the more expensive the RAM gets, the more expensive the servers get. At some point something has to break. So, I'm looking at the numbers that are kind of the enumeration of spend on tokens, token spend as you brought up, but also on capex. And if anyone pulls capex, that's a bad scene for everyone.
So, on that note, The Wall Street Journal just reported more details from Anthropic's S-1. This happened just before we started recording and I want to get your reaction to some of these numbers that we learned. So H1 2026 net loss of $81 billion. That's crazy. $72 billion of it from fair value accounting charges on convertible notes. So we'll give him, we'll give him a big asterisk there. Compute spend of $9 billion in the first half. And two customers account for about 15% of H1 revenue. Okay. So we have some more color on Anthropic's financials. Your initial reaction, Ed.
Well, so that's an $8 billion net loss-ish. Journal, I love him to death. Love Ana over there. The numbers in that piece are put in the world's worst chart. It's malpractice. I don't know why you'd do that. $8 billion net loss. Also, within the second quarter, Elon Musk was discounting SpaceX's compute. But nevertheless, this company lost $8 billion in the first half of the year. This company had a net loss of $80 billion, but like the cash loss about $88 billion. This is not a company that's improving its economics. It's got about $413 billion of non-cancellable commitments. I'm going to be honest. I don't think any amount of growth can make up for the fact that Anthropic's business just isn't particularly good. And also, we have now, despite this, this is about the first half of the year, we now have tangible proof revenue growth's slowing. The run rate, as bollocks as that number is, is plateauing. And the reason that they are pushing, Anthropic is pushing so hard on this drug discovery stuff is they know that they are running out of coding revenue. There's only so far they can go on it. And also Ramp has data that says 80% of their revenue comes from the 1% of their customers, most of which are unprofitable, unsustainable AI startups funded by venture capital.
Yeah. And this company, as big as their revenues may be, and they are big because everyone's focused on them, everyone's sending them money, this company is a load-bearing part of the AI trade and also of the AI economy. 70% of compute spend is Anthropic and OpenAI. I don't know what this company's going to do. I think that they should get rated CCC by Moody's and S&P. They are junk, junk, junk. And this is a test of financial regulators and also honestly the market's overall intelligence. It's a dog of a company.
Predictions for when it goes public. Do you think it goes out at 2 trillion?
I think 15. I think I just, I get the sense 5. No. Everyone I've talked to in finance is just like, two trillion? No thank you.
Everyone's jittery. Yeah.
Yeah. It's like the nerves are bad right now. The vibes are sour.
Bad vibes in AI. That's the headline. Ed Zitron is host of the Better Offline podcast and author of the Where's Your Ed At newsletter. Ed, appreciate your time. Thank you.
We'll be right back. And if you're enjoying the show, be sure to subscribe to the Prof G Markets YouTube channel at the link below.
Support for the show comes from LegalZoom. If knowing your way around a search engine was enough criteria for expertise, the world would be chock full of geniuses. But it's not. A quick answer is not the same as the right answer. That's especially true when it comes to legal stuff. Whether you're starting or running a business or protecting your family with a will or trust, check out LegalZoom. Say you're registering a business yourself. Looks easy, sure, but getting one thing wrong can lead to penalties down the line. LegalZoom has 25 years and 5 million businesses' worth of experience, so they can help you start off on the right foot. Once your business is up and running, they can also help you stay compliant, tracking deadlines and filing reports so you're not hunting down answers yourself. And when it comes to creating a will or trust, LegalZoom helps you protect what matters and they're accepted in all 50 states. LegalZoom is America's number one online legal services company. When you need a legal hand, visit LegalZoom. Go to legalzoom.com to get started. LegalZoom provides access to independent attorneys and its law firm, LZ Legal Services. LegalZoom is not a law firm and does not provide legal advice.
We're back with Prof G Markets. SpaceX just delivered a blowout quarter in its second earnings report as a public company. Revenue rose 161% year-over-year, largely driven by strong performance in its connectivity and AI businesses. Profit came in above expectations and capital expenditures came in slightly below what investors...
anticipated. The stock rose more than 4% in after-hours trading. So to help us unpack these results and what they mean for SpaceX, we are joined by Nicholas Owens, equity analyst at Morningstar. Nicholas, thank you for joining us. We just got these numbers, so we're reacting pretty much in real time here, but from what I can see, pretty much good across every business line. I mean, there are certainly some underlying concerns about customer concentration, potentially about capex, which maybe we can get into, but first, just your initial reactions to this quarter.
Thanks for having me again, Ed. Yeah, good, and you know, to be honest, probably about a billion and a half more profit than I expected in the quarter. But that said, I'm not really pacing this company at that quarter-to-quarter basis. You know, we are in the hockey stick in a way, and so it's hard to be precise, you could say. You know, these things like big chunky capex and so forth, those they can space those out, you know, throughout the year. So it certainly was a good quarter, and they have a very strong trajectory. Perhaps the highlight was Starlink. They had a lot more subscriber growth than I had anticipated at this point.
So you mentioned this idea that it's sort of chunky, that they can space out their capex. Why is that relevant? Why does that make you feel not as excited as perhaps others might?
I mean, it's exciting. The point I think I'm trying to make is you can't take a linear plot from two quarters and say this is the path, right? So if it in either direction. So they might have pushed out one project, and you know, I think they're building out this Louisiana facility which is going to, you know, soak up quite a bit of capex, and that's really putting them on pace to spend quite a bit of money there this year. Those types of things are simply lumpy. I mean, that's the bottom line.
Yeah. Just looking at the AI revenue, $7.6, $6 billion, up more than 800% year-over-year. My initial reaction is like, wow, that's big. But then I kind of remember that they have this customer concentration problem where we know that two customers make up around 40% of the revenue. So I assume this is just that Anthropic's paying a lot of money and Google's paying a lot of money. And there is a question, an open question as to how long is this compute boom going to actually last? Is Anthropic going to continue to pay billions of dollars for this compute? From a business perspective, that they're actually not profitable and they're basically taking venture capital dollars to fund this thing. So I guess when you look at that just from an underlying fundamental perspective, what is the trajectory of SpaceX, and is that kind of revenue and that kind of trajectory, is it sustainable?
