This is the worst possible scenario because it means that Nvidia does not have crazy demand for GPUs. This is a very, very bad thing. It runs contrary to what everyone is saying. The measurement of AI's success and the insatiable demand for AI compute has mostly been measured in Nvidia's GPU sales. If customers are being forced to take early delivery, that is not demand.
Hi, I'm Isaac, and today on the Tech Report I'm joined by the writer of Where's Your Ed At and the host of the Better Offline podcast, Ed Zitron. Great to have you on as always.
Almost one month to the day from when you estimated about 50% of all AI chip sales were sitting somewhere in warehouses, Business Insider has reported that Amazon has billions of dollars of GPUs stranded in warehouses. According to an internal Amazon email, the company received $5 billion of Nvidia GPUs in about August and has been receiving $3 billion every month after that, and according to them, for the foreseeable future. So just quickly, how much of Amazon's planned $200, $220 billion capex this year is just stranded in warehouses currently?
That's the thing. That's a bloody good question because the piece is so strangely written, and I understand because a lot of it is just quoting emails directly, but Amazon's been spending, I think their last capex, 40, 50 billion in a quarter.
What's 40 or 50 between friends? I guess it's strange that they're only taking delivery of 3 to 5 billion a month. It makes me wonder, like, the breakdown of capex is so strange because where is all this money going? How, like, is it more than I thought? Because if it's so slowly, if it's coming in so slowly, I don't know. But the big thing about this story that stood out to me within that is that it's clear that Nvidia has started forcing these things on Amazon in advance. Because there was a comment in the story, this is the early rack receipt program. There was a comment earlier in the story that said that previously Amazon didn't have to pay for servers, which is a way of saying GPUs, by the way, because the GPUs are in the servers before they arrived, which is a normal way of doing business. Usually you put a deposit down and then when the thingy is ready, you get the thingy. Perhaps you might pay upfront for some things, but like for a GPU especially, but you'll notice that Nvidia's sales have grown parabolic every quarter for like the last year.
I think that Nvidia is doing something called channel stuffing, which I thought was a form of illegal accounting, but this is Trump's America, so who knows? But it appears because it used to be that they didn't have to pay until they received them. Now they suddenly both have to pay and receive them earlier. So I think what's happened is Nvidia has said if you want these things in the future, you have to take them immediately. And this is channel stuffing. There's no way around it. People will say, 'Well, Amazon's choosing to do this.' Sure. But within the story, there was this term scarcity that they used. They said, 'Due to the scarcity of these GPUs, that's why Amazon's having to do this.' Doesn't really... riddle me this, Batman. If a chip is arriving months before it's installed, you're taking it early. It's an early rack receipt program. Everyone across the industry is taking way more of these things than they currently need. What is scarce about that? Unless, of course, the vendor in question, Nvidia, is saying, 'If you don't do this now, we will not sell you this in the future.' And I think that's what's happening. It's classic Jensen Huang nonsense.
I mean, there's a lot more to get into around a lot of that stuff, but I just want to touch on something a little bit, which is that the 'I told you so' moments for you are coming a bit more frequently at the moment. And I'm obviously curious what you think that means in general, like what that increasing pace means. But I'm also curious how it feels to have the 'I told you so' moments coming quicker.
Oh, the GPUs one, that was a glass of crystal for me. Ooh. Oh, I sat there going, 'Very good,' because what's funny about it is when I did that piece, 'Where Are All the GPUs, AI Chips,' even back, what, about a month to the day now, I got so much flak. I had a bunch of people saying, 'You're just wrong, you're misreading construction in progress, they wouldn't do this, these companies are, these are the smartest people in the room, these companies would never get these things in advance, Nvidia would never...' For all these people were extremely aggressively angry. And the reason they were is this is the worst possible scenario because it means that Nvidia does not have crazy demand for GPUs. Nvidia has a crazy demand program where they force it on you whether or not you really want them at that time. And people will say, 'Oh, they're planning in advance.' But it's like, yeah, they are planning in advance, and Nvidia is saying, 'Oh, I don't care about your planning. I care about my planning, which is growth every single quarter without fail.'
I think the speed up of these events is partly that journalism is catching up, but also just the passage of time is turning over every stone. We're finally seeing the emails leak out that kind of prove these things. Also, Eugene Kim, the guy at Business Insider who did this, he's been doing Amazon beat for what, over a decade. He's a great reporter and he's probably been looking and he finally found it. And I think we're going to see others pop up with these as well. What really grinds my gears though is that outside of Eugene, I've not seen a lot of popular journalism bothered to get on this. And the reason is, is this is a very, very bad thing. It runs contrary to what everyone is saying, which is the measurement of AI's success and the insatiable demand for AI compute has mostly been measured in Nvidia's GPU sales. Now, if Nvidia was selling all those GPUs because hyperscalers were plugging them in and then printing money, that's one thing. What's obviously happening is that Nvidia is pushing these things saying, 'You have to take them now, you have to take them now. If you don't take them now, I won't give them to you in the future.' That means there's not actually demand like Nvidia is saying. There actually is not the demand that Nvidia is saying. We don't know how big the difference is between the sales and the actual demand. We never will. But it's very, like, just very simply, if customers are being forced to take early delivery, that is not demand.
You can say, 'Oh, well, I guess they want it.' I guess there is a way of framing it as demand, but it's not demand measured in the revenues we're seeing today. The actual revenues we're seeing today are the compressed demand over what is likely months or years. And the real question for me is what happens when this becomes too big to ignore, because at some point these hyperscalers have to say, 'Jensen, I got an entire warehouse of these bloody things. I got an entire warehouse full of your goddamn Grace Blackwell chips gathering dust, and there's a bloke from Super Micro who keeps sneaking in with a hair dryer.' No, that's a joke. That already happened. But that's the thing, and people will get, I'm sure, in the comments and people responding to this will say, 'Well, whoa, the hyperscalers do want these GPUs.' They think they want them. They clearly have ordered too many, or at least they've ordered as many as they need, but way too early. I don't know. And it just shows that there is a fundamental disconnection between what people perceive as AI demand and what is actual AI demand, and this is a problem across the industry. It's a problem that I've been analyzing for three years, which is we cannot actually measure the true demand for AI because even when it comes to, and I know we're going to get into this as well, that ramp stat of 80% of OpenAI and Anthropic's revenues, 1% of their customers, except most of those customers are venture-backed AI startups who are selling their tokens, basically letting you burn $1,200, $1,300 worth of tokens for $200 a month. Cognition was in that article, as in Cognition lets you burn $1,200, $1,300 worth of Claude or ChatGPT tokens. Again, is that demand for AI? Is it demand for AI if you're being forced to take the GPUs early, as in you don't actually, just fundamentally, you don't need them right now, but you are being forced to take something you don't need? Is it strong demand for AI if you have to sell $5 for a dollar or $10 for a dollar? Or is it inflated demand? Is it artificial demand? And it's obviously the latter. It's obvious that it's fake demand. Now, there is some demand, but it's not commensurate with the sales we've seen and indeed the media headlines we've been reading for three or four godless years.
If we go back to the rate at which Amazon is buying these racks, these GPUs, a delay of a few weeks would mean that the stack of warehouse GPUs would presumably then grow by billions of dollars. A four-week delay, that's another 3 to 5 billion you have waiting.
Yeah, exactly. But I've seen in places, and I'm sure you have as well, that wait time for energy components is going up exponentially, let alone the construction itself. There's loads of these things which are just delay. There's the moratorium. There's lots of things causing delays which are a little bit longer than just a month or two.
Yeah. And how much of a delay are these projects facing? And then what does that mean for the amount that is then presumably a growing amount of GPUs that are sitting around growing in the background? If we take Amazon as the standard bearer, they're taking billions of dollars of these things in every month. Let's call it three billion. Oracle's, the one out in Project Jupiter, the New Mexico one, that one has been delayed six to 12 to 18 months. So we're talking tens of billions of dollars of GPUs just sitting rotting. Now, what's crazy about that is by the time those things get installed, will they? Will, won't they? We don't know yet. We'll be on to Vera Rubin. By the time Vera Rubin's installed, we'll be on to Feynman if we get that long. If we go that far, which I don't know if we will.
I think it's just, I think that there is something strange going on that we don't know, like there is a data point we don't know, which is how is the actual delivery mechanism of a GPU working, because Nvidia sells them to Amazon or Google or Microsoft, those GPUs then wind their way to Taiwan, they get one of the original design manufacturers, the Foxconns, the Quantas of the world, they then put them in the server and ship them. Now it's clear that those servers are shipping, and then Nvidia is saying, 'Well, I guess those things are shipping,' and then they're just sitting in a pile somewhere. I don't know. There's something missing in this chain because even 3 billion a month worth of GPUs doesn't line up with the amount of capex being spent. Think of it like Amazon spends 40, 50 billion a quarter, or 30 billion, 40 billion. So many, it's got to a point where numbers no longer mean anything. That's also the other problem I have. But let's say they've been spending tens of billions of dollars a quarter, but only a few billion worth of GPUs are coming in. Where is the money going? Why don't we know? And I think that is a problem too. And it's crazy how big this thing has got despite us knowing comparatively nothing about the overall health of the AI bubble. We don't really understand where the money's going from or going to. Could it be that Amazon isn't, well, I mean, no, they're doing early rack receipt, but it could be that Nvidia is forcing the ODMs, the ones building the servers in Taiwan, to buy the GPUs at the moment rather than Nvidia, I don't know, putting them on consignment. It isn't really clear what's happening. It's crazy we don't know. I know I'm kind of all over the place at this point, but it's crazy how much money, a trillion or more dollars, has gone into this and we don't know really anything about it. I read a report the other day that says that Microsoft, who claims, Satya Nadella claimed they don't have enough power to plug these things in, actually has the power in Georgia where one of the Fairwater data centers is, but they're not pulling it, suggesting low utilization. I think at the end of this we need like an actual referendum on corporate disclosures because I don't think you should be allowed to spend this much money without giving investors significantly more clarity in what the hell it is you're actually doing.
