Monologue: The AI Data Center Overbuild

0:02 Hello and welcome to this week's Better Offline Monologue, I'm your host Ed Zitron. As ever, subscribe to the newsletter, premium, you know all that good stuff, I'll have the links in there. And this week I want to talk about actually a free newsletter I put out called Where Are All The AI Chips? And it turns out that the answer is either in warehouses or unpowered data centers. In an investigation with The Guardian, Aisha down over there worked on it with me, she's awesome, I reported that Microsoft only had 2.2 million GPUs in service, far less than people believed. And while I couldn't report it at the time because I still had to run down some leads, the total power of the chips was estimated at about

0:47 1.993 gigawatts of capacity, with the overall cost of those GPUs installed somewhere in the region of $ 50 to $ 60 billion. A few weeks ago, Bloomberg reported that despite reporting that Microsoft had added a gigawatt in each of the last three quarters, and by reporting I mean the things that Microsoft literally said on their earnings call, that only two gigawatts out of their total 12 gigawatts of data center capacity was specifically for AI. In other words, despite having spent over $ 265 billion on capital expenditures since the beginning of 2022, Microsoft has only put about $ 50 billion worth of GPUs into service, and I estimate that as much as $ 106 billion worth of GPUs and associated hardware, are

1:31 now sitting either unpowered in data centers, or incomplete data centers, or in warehouses, and that very little AI compute is coming online throughout the entire industry. Microsoft CEO Satya Nadella had sort of admitted this in November of last year, when he told an interviewer that he had a bunch of chips sitting in inventory that he couldn't plug in, and that's a quote, though he didn't mention the sheer scale of the warehousing or indeed how little he'd actually turned on. Well, I'm a curious little critter, so I went looking for more evidence of the problem outside of Microsoft and found it buried in the balance sheets of multiple hyperscalers and neoclouds under the construction-in-progress line on the

2:10 balance sheet, which is specifically where companies bury their uninstalled GPUs and incomplete data centers. And this number has grown across a ton of them, well, by a remarkable amount. Google, Meta, Oracle, Amazon, SpaceX, and Tesla, neoclouds like Core, Weave, and Iron, and co-location companies like Core Scientific and Applied Digital all have about $ 374 billion of construction in progress, a figure that's likely lower than the true number because Amazon's contribution, which is about $ 71 billion, is only current as of the end of 2025. And there have been two, now very soon three more quarters of that. And the company only reports annual, like I said. A large chunk of that larger CIP number is Google's, which sits around $ 122 billion, which is

2:56 truly shocking. And in every case, by the way, the number has grown every single time it's been reported for the last 12 quarters, with some fluctuation in the case of Amazon because of their logistics operations. But really, right now, it's only growing. Now, not everybody reports construction in progress, so that number doesn't include Microsoft, Firmus, Sharon, Equinix, Nebius, Poolside, private operators like Vantage, any of the sovereign AI build-outs in the Middle East or Europe, the private projects built for OpenAI and Anthropic, or any of Oracle or Meta's off-balance sheet construction projects, which were likely, I think, numbering in tens of billions of dollars, and because these are

3:34 special purpose vehicles kept off of their balance sheet, that wouldn't be in construction in progress. Now, if I had to guess, the total construction in progress number is somewhere between $ 400 and $ 500 billion. And based on everything I've analyzed, I estimate that approximately 50% of all AI hardware and chips that have been sold is being warehoused. And will take more than two years to fully ingest. Which means that basically everything you've seen NVIDIA sell has been a pre-order campaign that arrives sometime in 2028, 2029, or maybe beyond that. In other words, I think NVIDIA has made somewhere between $ 200 and $ 300 billion in revenue on stuff that's been sold probably 24 to 36 months in advance,

4:16 and that any scarcity around AI compute is a result of most of the capacity being sold immediately to Anthropic and OpenAI, leaving very little for the rest of the world, with more capacity taking agonizingly long to build and not being prioritized unless you're one of the AI labs. So what does all of this mean? First of all, basically any announcement of how much capacity a hyperscaler has built is suspicious. The fact that Microsoft is intentionally obfuscating how much capacity, AI capacity I mean, it has built, is a sign that the entire AI industry is doing so, and that the available AI capacity is much, much, much smaller than we think. All of those estimates of 12 or 15 gigawatts coming online? Bollocks. Wank. Tosh. I don't know,

5:00 I can't remember other terms for it. Second, if Microsoft, one of the most well-capitalized and experienced data center developers in the world, is having trouble bringing the majority of its AI capacity online, everyone is. Really, I think Amazon is the only company, maybe Google, that has comparable experience, let alone more. And I think Microsoft was also the first AI supercomputer powered by GPUs back in 2020 for OpenAI.

5:27 There's probably someone who did it before just to mess with me. Nevertheless, if they're failing, I think everyone is. And lastly, any beliefs that anyone has about the insatiable demand for AI compute are completely distorted because very little actual capacity is coming online, likely gigawatts less than people believe, and almost all of it is flowing directly to anthropic and open AI, creating an illusion of scarcity when it's actually just two unsustainable venture-backed AI labs sucking up whatever exists.

5:55 And this also means we're in a terrible overbuilt situation. Right now, everyone building more AI data center capacity is doing so because of the massive revenue backlogs across Microsoft, Google, Amazon, Oracle, and CoreWeave, ignoring the fact that anywhere from half to 75% of those backlogs are directly from Anthropic and OpenAI, neither of whom can actually pay for them, and any hyperscalers signing contracts to give that capacity to them can probably cancel the agreements. To the outside world, everything looks like a thriving industry with tons of demand both for AI GPUs, which they think are turning into money immediately, and for AI compute. With Microsoft's clearly crooked statements around adding a

6:35 gigawatt of capacity every single quarter, creating the illusion that its revenue growth is from adding all of that capacity and from spending all of those capital expenditures and from diverse demand on top of it. Rather than mostly from OpenAI's compute spend, which made up 70%, of Microsoft's AI revenues in fiscal year 2026 and 7% of its overall revenues in that same fiscal year. To be very specific about what I mean, people are assuming that lots of computers coming online every quarter and that all of that compute is being immediately rented out for high rates. As a result, they think that building more AI data centers is an obvious choice and akin to printing money because they believe there's unbelievable amounts of pent-up demand waiting to

7:19 be met and that meeting that demand will be as simple as bringing a data center online, which Microsoft is, from the outside, proving can be done a gigawatt a quarter. Except it's not AI data center capacity that's coming online. What's actually happening is very little capacity is coming online for AI services, and what little does is going straight into the hands of two companies with near-infinite resources provided by venture capital, and in some cases the hyperscalers themselves, which allows them to sign contracts to take up massive amounts of capacity, and indeed fill those revenue backlogs. Microsoft is saying that it's bringing data center capacity, not AI-specific data center capacity, online a gigawatt at a time, but

8:02 it appears the actual AI capacity is, at best, coming on at a rate of maybe tens or hundreds of megawatts a quarter, and that really is the best case. Perhaps there are others that are doing it faster, but I actually can't find evidence of them. I'll be honest that I have had this story in my head for weeks, and was hesitant to jump the gun because of its ramifications.

8:24 As it appears that everyone's got this story wrong. This is magnitudes worse than the dot-com bubble. To make the obvious comparison, this is akin to the fiber optic cable not even being laid in the ground, or only being halfway connected, except the actual demand for AI infrastructure is so thoroughly distorted by OpenAI and Anthropic that there's anywhere from 5 to 50 times the amount of capacity in planning than there is demand for it. And even then, that demand is inflated. To make matters worse, it's not going anywhere after this. There is no situation where this is going to be used at the scale it's built.

9:01 It will also be just as expensive to finish an AI data center in 2 or 3 or 5 years, and just as expensive to run those GPUs. And who knows how many generations will be in there, but I don't know, in my opinion, My opinion, my esteemed opinion, I don't think we make it past Vera Rubin or Feynman, which is the generation after Vera Rubin. Even then, I have questions about whether Feynman actually ships. Who knows, though?

9:24 Things are crazy. I'll also be clear that, as I discussed with Paul Kudrowski this week, that there is no post-bubble economy for GPUs, at least at the scale they're being built, because the cost of serving inference is similar to an airline with high fixed costs. You as an operator must buy capacity, hundreds or thousands of GPUs over a certain period of time, whether or not you actually have the demand. While there might be a situation where you could eke out a profit of some sort with the perfect amount of customers, anything above the demand you foresaw means an unstable and unreliable service and turning customers away at the door, and anything below can every ounce of margin you might have. Dario Amadei himself said as

10:04 much in an interview back in February with Dwarkish Patel, saying that, in a theoretical scenario, he could buy $ 1 trillion of compute that starts at the end of 2027, but if his revenue wasn't $ 1 trillion, if it was even $ 800 billion, there was, and I quote, No hedge on earth that could stop him from going bankrupt. These are the underlying economics of every single company running an AI service, and they're equal parts brittle and volatile. Thanks for listening. I'll see you next week.