Like half of Nvidia's revenue could literally be sitting in warehouses. I thought it was tens of billions of GPUs, but based on this Microsoft story, it could be hundreds of billions. If I'm right here, this is going to be one of the greatest financial shenanigans — this is going to be an absolute scandal in a way that will humiliate some of the largest companies in the world. Nvidia's claiming they're going to double sales next year. If that's the case, okay, how many of those are installed in data centers? Because if it's half or a quarter or a tenth, or it's going to take three or four or five years to install them, how do we actually value this boom? Is this a boom?
Hi, I'm Isaac, and on the Tech Report with me today is writer of Where's Your Ed At and the host of the Better Offline podcast, Ed Zitron. Welcome back.
So, just before we started, we completely did away with the original question list that I had because you sent through some interesting stuff from Bloomberg, who's reported that Microsoft only has about 2 gigawatts of actual AI compute online out of an expected 12 gigawatts that it's supposed to have by now.
I mean, just tell us about that.
So, Bloomberg said 2 GW of IT load of just AI chips with Microsoft. Microsoft has been saying they've been adding one gigawatt of capacity every single quarter for like three quarters, I think it is. I think they said they doubled their capacity in one of them. And it's funny, a few months ago, Aisha Down and I at The Guardian reported a very similar story and said that Microsoft only had a few million chips. It was less than they expected. And when we went to Microsoft, they said, 'Oh, this is not indicative. You don't know what you saw.' So the source I talked to, and that we talked to, actually had this number. I have 1.993 GW and a precise breakdown of the chips, and the amount that they're worth is about $50 billion. Microsoft has spent over $265 billion in capex, and if we assume that capex — like 50% of capex is GPUs — it's actually impossible to tell. These companies don't disclose. That means that Microsoft has like $82.5 billion of GPUs not installed at all.
Now, this is significant for Microsoft. I'm actually going to do a newsletter on this next week. Significant for Microsoft, but also for the rest of the world, and also begs a few really worrying questions. So, Microsoft was the first company, I believe, to build an AI supercomputer. By which I mean, in 2020, they built one out of A100 chips from Nvidia. There were other AI supercomputers, but it was the first big one using the Mellanox-powered — Mellanox was a networking startup that Nvidia bought — to network together GPUs. The first one of those in 2020 for OpenAI. Now, they have the most experience. They have the most logistical experience. They have the most money to do this. Amazon technically has done more cloud for longer, but Microsoft is way more in AI. And Microsoft has what looks like less than half of their purchased GPUs installed.
This means that I believe personally that this is an across-the-industry phenomenon. And if that's the case, we likely have hundreds of billions of dollars of Nvidia GPUs in warehouses waiting to be installed. Which begs the question: why are people still buying them? And how is Jensen Huang doing it? Now, big boy Jensen Huang, CEO of Nvidia, might just be fending them. He might just be saying, 'You pig, if you don't buy these today, I won't sell you the next ones.' But this is fundamentally insane.
This means that GPU capacity is coming on extremely slowly, and I believe personally that there is an industry-wide attempt to lie about available capacity. I think Oracle claimed they brought on several hundred gigawatts — sorry, megawatts — worth of capacity last quarter. They claimed there were six out of eight buildings in Stargate Abilene, based on discussions with a source less than a month ago. That's not true. There are less than — I don't even think the fourth was brought online. And so everyone is talking about capacity. We brought this much capacity on, then that's this much capacity. I don't think they're being honest. I think what's actually happening is there is capacity being brought online, as in CPUs and infrastructure related to GPUs, but not the GPUs themselves, because of power, because of data center construction times. But here's the thing. Is Nvidia running the largest Kickstarter of all time? Is Nvidia just selling GPUs years in advance? How many GPUs have come online? Because if it turns out that everyone is like this, everyone is just hoarding GPUs, that is actually — that should be an SEC investigation, because put it like this: I don't think it's illegal. But it is something that I think shareholders need to know, because there's a big difference between Nvidia is selling a bunch of GPUs because of demand for data centers that are getting built, and Nvidia is selling a bunch of GPUs because everyone's worried that if they don't buy them now, they won't be available in the future. And I must be clear: Microsoft is so experienced at doing this. If Microsoft's got less than half of their GPUs installed, do you think Meta's much better? Do you think Amazon? Google? I don't think so.
Google's apparently building 4.5 gigawatts worth of capacity for Anthropic's TPUs. So, Google's AI chips. To which I say, no, they're not. Like, I actually question how many data centers have been finished here. I brought this up a few months ago, but these numbers really bring it home. Only 2 GW of capacity. That's crazy. And I must be clear, when I'm talking about the capacity, having seen the numbers I have, a lot of that is H100 GPUs and H200 GPUs. It's older GPUs. And this includes data center capacity, I think, from either Nscale or at Nebius. It doesn't include CoreWeave. Long and short of it is: is Nvidia just pre-selling GPUs years in advance? Because if that's the case, there isn't AI demand at the scale. The reason that we have capacity constraints is stuff is not coming online at all. Like if Microsoft — I mean, this story, this 2 gigawatt one, came out a couple weeks ago. I must be clear, the numbers I saw were from months ago, and they were 1.993 GW. Now, this was my own mathematics based on their TDP, so the power that the GPUs pull. Pretty close, though. But I think it begs the question of what exactly is going on with the AI data center buildout, because if it's years delayed, if it's truly years delayed, that means that any of the gigawatt figures we've seen are nonsense, because they claim they brought 12 GW online, only 2 GW is AI data center capacity, like actual chips. Though I'm guessing it must be CPUs. But okay, are we meant to be impressed by a bunch of CPUs? Like, what is it that's actually going on here? We need clarity, and we're not going to get it because we don't have a functional regulator. We don't have functional financial regulators, and we certainly don't have journalists who are going to push much further than this. And it's frustrating because this here is a scandal. And surely it's going to only get worse, because the longer these GPUs are just left sitting around collecting dust, the less relevant they become. And then it's going to reach a point where everyone's like, 'Oh, we've got this amount of gigawatts. We've bought all of these GPUs. Oh, actually, yeah, 50% of that was actually — we never installed them. We never used them.'
That's the thing as well. It's okay, so, you've bought tens of billions of dollars of Blackwell. By the time you install all the Blackwell, we'll be on Feynman, which is the thing after Vera Rubin. By the time you install Vera Rubin, it will be whatever LMFO seal. I don't know what they're going to call the next ones, but they're going to call them something.
And also, how much — like, I really must, I'm repeating myself — how much capacity is actually online? Because if it's — we've — because I've seen, oh, 17 gigawatts of capacity has come online. I've seen 12. I've seen 10. But if that 10 gigawatts is only one gig — just to be clear, this 2 gigawatt number is all time for Microsoft. This isn't just brought online. This is all they've got. If the total amount of AI capacity that's come online to date is only 3 or 4 GW, that's crazy. And this is why I think all of these capacity stories are so ridiculous. Because if it turns out most of it isn't AI chips, then those numbers are functionally useless.
So just take us through: what is the significance of the fact that instead of it being 12 GW, it's just 2 GW?
Well, it means that Microsoft has been boasting about building gigawatts of data center capacity with everyone thinking, oh, they mean they're building AI data center capacity, and they're not. Means the AI data center capacity just isn't coming online. It means that they have been sinking hundreds of billions of dollars into something mostly to warehouse it. Now, the argument would be, okay, we bought the GPUs and the construction's in process and it will come online at some point, but when? And if Microsoft, the apex predator of AI infrastructure, cannot do this, you think Meta isn't doing the same thing? Meta, what, 130-something billion, 160 billion they're going to spend this year, little under 100, I think, last year. Amazon was, I forget, maybe 200 last year. Are those chips coming online? How many chips? How much of that capacity is actually AI infrastructure? Because if it's not much, then we have seen the trillion dollars of capex spent on bringing online what, $100 billion, $200 billion worth of chips, as Nvidia has sold $535 billion worth of them. And I think that when you tell people this story — I've tried to tell a few reporters this — they go, 'Ah, you know, it's not that big.' Nah, it's not a big deal. It's a huge deal on actually an insurance level. If you've got hundreds of billions of dollars or tens of billions of dollars of chips sitting in a warehouse somewhere in Taiwan and in America somewhere, you need to ensure that. And I don't know, I hope the warehouse is good. I hope nothing happens to the cooling in there. I hope that nothing's wrong with the chips when you turn them on. Also, why buy so many in advance?
Like, that is the real question here. Why have you pre-ordered so many? And sure, you could say, well, I need to make sure I have the capacity in time. But if you're taking years to build the thing, well, you could have waited. You could have waited a while. Unless, of course, Jensen Huang is just finding them and he's saying, 'If you don't buy this, pig, I won't sell you GPUs in the future.' And I think that's been very effective. But if it turns out like half of Nvidia's revenue could literally be sitting in warehouses. I thought it was tens of billions of GPUs, but based on this Microsoft story, it could be hundreds of billions. It could be the vast majority of its 2025 revenue. Like, that's a crazy number. And Nvidia and Microsoft and all of these companies have been working really hard to obfuscate the actual numbers. Just think of it like this. Microsoft every single quarter, I think it's three straight, said one gigawatt capacity added. One gigawatt capacity added. It's very clear that within that one gigawatt, there's what, 50 megawatts of capacity or 100 megawatts of capacity. It's horrifyingly misleading. And it also misleads shareholders into believing that this capital expenditure fiasco is leading to actual revenue growth or will lead to anything in the next few years. And the question is as well, is why do you buy Blackwell when Vera Rubin's coming along? Why do you buy Vera Rubin when Feynman's coming along? And if the answer is Nvidia won't sell it to us, well, that sounds like a prisoner's dilemma situation to me. That sounds like something that's entirely created by everyone buying the GPUs. If everyone went, 'All right, I've got plenty. I'm going to build some in a year. I'm going to need them then.' And if everyone just kind of backed off buying these in advance, everything kind of ends for Nvidia. And I think that that kind of happens anyway because the amount of debt they're going to need at a time when interest rates are spiking is going to be kind of monstrous.
That's one of the — I mean, this all kind of raises one of the big questions that I've had in the whole data center buildout, which is that even if you're super optimistic and you get everything built in two years, which is —
Yeah, very optimistic. Not going to happen already. It's already past two years.
Then you're buying GPUs from right now to put in data centers that are going to be built in two generations' time.
Two GPU generations' time, which to me doesn't make any sense. You don't want to be buying today's GPUs at today's prices when you could be buying them in two years at two years' prices, and you'll still be plugging them in at the same sort of time anyway. So, and then it also begs the question of why spend even more on future capex if you've got all of this backlog waiting to be plugged in that you're, yeah, not even plugging in apparently.
