WEBVTT - Monologue: The AI Data Center Overbuild

0:00:02.310 --> 0:00:06.830
<v Speaker 1>Hello and welcome to this week's Better Offline Monologue, I'm

0:00:06.850 --> 0:00:18.140
<v Speaker 1>your host Ed Zitron. As ever, subscribe to the newsletter, premium,

0:00:18.180 --> 0:00:20.160
<v Speaker 1>you know all that good stuff, I'll have the links

0:00:20.260 --> 0:00:22.520
<v Speaker 1>in there. And this week I want to talk about

0:00:22.640 --> 0:00:24.840
<v Speaker 1>actually a free newsletter I put out called Where Are

0:00:24.880 --> 0:00:27.020
<v Speaker 1>All The AI Chips? And it turns out that the

0:00:27.100 --> 0:00:31.630
<v Speaker 1>answer is either in warehouses or unpowered data centers. In

0:00:31.670 --> 0:00:34.550
<v Speaker 1>an investigation with The Guardian, Aisha down over there worked

0:00:34.590 --> 0:00:37.150
<v Speaker 1>on it with me, she's awesome, I reported that Microsoft

0:00:37.270 --> 0:00:40.940
<v Speaker 1>only had 2.2 million GPUs in service, far less than

0:00:40.950 --> 0:00:43.120
<v Speaker 1>people believed. And while I couldn't report it at the

0:00:43.159 --> 0:00:44.979
<v Speaker 1>time because I still had to run down some leads,

0:00:45.320 --> 0:00:47.860
<v Speaker 1>the total power of the chips was estimated at about

0:00:47.860 --> 0:00:51.540
<v Speaker 1>1.993 gigawatts of capacity, with the overall cost of those

0:00:51.580 --> 0:00:54.090
<v Speaker 1>GPUs installed somewhere in the region of $ 50 to $ 60 billion.

0:00:56.880 --> 0:01:00.080
<v Speaker 1>A few weeks ago, Bloomberg reported that despite reporting that

0:01:00.120 --> 0:01:02.560
<v Speaker 1>Microsoft had added a gigawatt in each of the last

0:01:02.600 --> 0:01:04.920
<v Speaker 1>three quarters, and by reporting I mean the things that

0:01:04.940 --> 0:01:09.080
<v Speaker 1>Microsoft literally said on their earnings call, that only two

0:01:09.160 --> 0:01:11.860
<v Speaker 1>gigawatts out of their total 12 gigawatts of data center

0:01:11.900 --> 0:01:16.020
<v Speaker 1>capacity was specifically for AI. In other words, despite having

0:01:16.080 --> 0:01:20.539
<v Speaker 1>spent over $ 265 billion on capital expenditures since the beginning

0:01:20.560 --> 0:01:24.360
<v Speaker 1>of 2022, Microsoft has only put about $ 50 billion worth

0:01:24.380 --> 0:01:27.450
<v Speaker 1>of GPUs into service, and I estimate that as much

0:01:27.510 --> 0:01:31.650
<v Speaker 1>as $ 106 billion worth of GPUs and associated hardware, are

0:01:31.709 --> 0:01:36.009
<v Speaker 1>now sitting either unpowered in data centers, or incomplete data centers,

0:01:36.090 --> 0:01:38.810
<v Speaker 1>or in warehouses, and that very little AI compute is

0:01:38.850 --> 0:01:43.890
<v Speaker 1>coming online throughout the entire industry. Microsoft CEO Satya Nadella

0:01:43.940 --> 0:01:46.460
<v Speaker 1>had sort of admitted this in November of last year,

0:01:46.800 --> 0:01:48.580
<v Speaker 1>when he told an interviewer that he had a bunch

0:01:48.620 --> 0:01:51.000
<v Speaker 1>of chips sitting in inventory that he couldn't plug in,

0:01:51.040 --> 0:01:53.600
<v Speaker 1>and that's a quote, though he didn't mention the sheer

0:01:53.660 --> 0:01:56.340
<v Speaker 1>scale of the warehousing or indeed how little he'd actually

0:01:56.380 --> 0:02:00.070
<v Speaker 1>turned on. Well, I'm a curious little critter, so I

0:02:00.130 --> 0:02:02.270
<v Speaker 1>went looking for more evidence of the problem outside of

0:02:02.310 --> 0:02:05.640
<v Speaker 1>Microsoft and found it buried in the balance sheets of

0:02:05.700 --> 0:02:10.000
<v Speaker 1>multiple hyperscalers and neoclouds under the construction-in-progress line on the

0:02:10.040 --> 0:02:13.840
<v Speaker 1>balance sheet, which is specifically where companies bury their uninstalled

0:02:13.860 --> 0:02:17.110
<v Speaker 1>GPUs and incomplete data centers. And this number has grown

0:02:17.169 --> 0:02:23.490
<v Speaker 1>across a ton of them, well, by a remarkable amount. Google, Meta, Oracle, Amazon, SpaceX,

0:02:23.530 --> 0:02:27.169
<v Speaker 1>and Tesla, neoclouds like Core, Weave, and Iron, and co-location

0:02:27.230 --> 0:02:30.460
<v Speaker 1>companies like Core Scientific and Applied Digital all have about

0:02:30.460 --> 0:02:35.000
<v Speaker 1>$ 374 billion of construction in progress, a figure that's likely

0:02:35.060 --> 0:02:38.120
<v Speaker 1>lower than the true number because Amazon's contribution, which is

0:02:38.139 --> 0:02:41.260
<v Speaker 1>about $ 71 billion, is only current as of the end

0:02:41.300 --> 0:02:45.120
<v Speaker 1>of 2025. And there have been two, now very soon

0:02:45.240 --> 0:02:48.720
<v Speaker 1>three more quarters of that. And the company only reports annual,

0:02:48.760 --> 0:02:52.080
<v Speaker 1>like I said. A large chunk of that larger CIP

0:02:52.139 --> 0:02:56.010
<v Speaker 1>number is Google's, which sits around $ 122 billion, which is

0:02:56.150 --> 0:02:59.150
<v Speaker 1>truly shocking. And in every case, by the way, the

0:02:59.190 --> 0:03:01.550
<v Speaker 1>number has grown every single time it's been reported for

0:03:01.570 --> 0:03:04.840
<v Speaker 1>the last 12 quarters, with some fluctuation in the case

0:03:04.919 --> 0:03:08.419
<v Speaker 1>of Amazon because of their logistics operations. But really, right now,

0:03:08.860 --> 0:03:13.130
<v Speaker 1>it's only growing. Now, not everybody reports construction in progress,

0:03:13.270 --> 0:03:18.210
<v Speaker 1>so that number doesn't include Microsoft, Firmus, Sharon, Equinix, Nebius, Poolside,

0:03:18.270 --> 0:03:21.570
<v Speaker 1>private operators like Vantage, any of the sovereign AI build-outs

0:03:21.610 --> 0:03:24.109
<v Speaker 1>in the Middle East or Europe, the private projects built

0:03:24.150 --> 0:03:27.530
<v Speaker 1>for OpenAI and Anthropic, or any of Oracle or Meta's

0:03:27.669 --> 0:03:32.190
<v Speaker 1>off-balance sheet construction projects, which were likely, I think, numbering

0:03:32.210 --> 0:03:34.070
<v Speaker 1>in tens of billions of dollars, and because these are

0:03:34.130 --> 0:03:37.130
<v Speaker 1>special purpose vehicles kept off of their balance sheet, that

0:03:37.170 --> 0:03:40.710
<v Speaker 1>wouldn't be in construction in progress. Now, if I had

0:03:40.750 --> 0:03:43.290
<v Speaker 1>to guess, the total construction in progress number is somewhere

0:03:43.330 --> 0:03:47.490
<v Speaker 1>between $ 400 and $ 500 billion. And based on everything I've analyzed,

0:03:47.590 --> 0:03:51.070
<v Speaker 1>I estimate that approximately 50% of all AI hardware and

0:03:51.110 --> 0:03:55.370
<v Speaker 1>chips that have been sold is being warehoused. And will

0:03:55.410 --> 0:03:58.330
<v Speaker 1>take more than two years to fully ingest. Which means

0:03:58.370 --> 0:04:01.090
<v Speaker 1>that basically everything you've seen NVIDIA sell has been a

0:04:01.130 --> 0:04:03.330
<v Speaker 1>pre-order campaign that arrives sometime in 2028, 2029, or maybe

0:04:03.350 --> 0:04:08.880
<v Speaker 1>beyond that. In other words, I think NVIDIA has made

0:04:08.940 --> 0:04:12.640
<v Speaker 1>somewhere between $ 200 and $ 300 billion in revenue on stuff

0:04:12.660 --> 0:04:16.020
<v Speaker 1>that's been sold probably 24 to 36 months in advance,

0:04:16.060 --> 0:04:19.670
<v Speaker 1>and that any scarcity around AI compute is a result

0:04:19.710 --> 0:04:22.650
<v Speaker 1>of most of the capacity being sold immediately to Anthropic

0:04:22.690 --> 0:04:25.630
<v Speaker 1>and OpenAI, leaving very little for the rest of the world,

0:04:25.730 --> 0:04:29.310
<v Speaker 1>with more capacity taking agonizingly long to build and not

0:04:29.350 --> 0:04:33.190
<v Speaker 1>being prioritized unless you're one of the AI labs. So

0:04:33.210 --> 0:04:36.250
<v Speaker 1>what does all of this mean? First of all, basically

0:04:36.450 --> 0:04:39.310
<v Speaker 1>any announcement of how much capacity a hyperscaler has built

0:04:39.570 --> 0:04:43.510
<v Speaker 1>is suspicious. The fact that Microsoft is intentionally obfuscating how

0:04:43.550 --> 0:04:46.919
<v Speaker 1>much capacity, AI capacity I mean, it has built, is

0:04:46.960 --> 0:04:49.960
<v Speaker 1>a sign that the entire AI industry is doing so,

0:04:50.300 --> 0:04:52.880
<v Speaker 1>and that the available AI capacity is much, much, much

0:04:52.960 --> 0:04:55.820
<v Speaker 1>smaller than we think. All of those estimates of 12

0:04:55.820 --> 0:05:00.870
<v Speaker 1>or 15 gigawatts coming online? Bollocks. Wank. Tosh. I don't know,

0:05:00.880 --> 0:05:05.060
<v Speaker 1>I can't remember other terms for it. Second, if Microsoft,

0:05:05.300 --> 0:05:08.539
<v Speaker 1>one of the most well-capitalized and experienced data center developers

0:05:08.560 --> 0:05:11.140
<v Speaker 1>in the world, is having trouble bringing the majority of

0:05:11.200 --> 0:05:15.900
<v Speaker 1>its AI capacity online, everyone is. Really, I think Amazon

0:05:15.940 --> 0:05:19.820
<v Speaker 1>is the only company, maybe Google, that has comparable experience,

0:05:19.860 --> 0:05:23.260
<v Speaker 1>let alone more. And I think Microsoft was also the

0:05:23.350 --> 0:05:27.210
<v Speaker 1>first AI supercomputer powered by GPUs back in 2020 for OpenAI.

0:05:27.230 --> 0:05:29.130
<v Speaker 1>There's probably someone who did it before just to mess

0:05:29.150 --> 0:05:31.890
<v Speaker 1>with me. Nevertheless, if they're failing, I think everyone is.

0:05:32.710 --> 0:05:36.710
<v Speaker 1>And lastly, any beliefs that anyone has about the insatiable

0:05:36.770 --> 0:05:40.700
<v Speaker 1>demand for AI compute are completely distorted because very little

0:05:40.900 --> 0:05:44.460
<v Speaker 1>actual capacity is coming online, likely gigawatts less than people believe,

0:05:44.779 --> 0:05:47.300
<v Speaker 1>and almost all of it is flowing directly to anthropic

0:05:47.320 --> 0:05:50.380
<v Speaker 1>and open AI, creating an illusion of scarcity when it's

0:05:50.490 --> 0:05:54.830
<v Speaker 1>actually just two unsustainable venture-backed AI labs sucking up whatever exists.

0:05:55.870 --> 0:05:58.310
<v Speaker 1>And this also means we're in a terrible overbuilt situation.

0:05:59.490 --> 0:06:02.950
<v Speaker 1>Right now, everyone building more AI data center capacity is

0:06:03.010 --> 0:06:08.080
<v Speaker 1>doing so because of the massive revenue backlogs across Microsoft, Google, Amazon, Oracle,

0:06:08.120 --> 0:06:12.080
<v Speaker 1>and CoreWeave, ignoring the fact that anywhere from half to 75%

0:06:12.080 --> 0:06:15.799
<v Speaker 1>of those backlogs are directly from Anthropic and OpenAI, neither

0:06:15.839 --> 0:06:18.500
<v Speaker 1>of whom can actually pay for them, and any hyperscalers

0:06:18.560 --> 0:06:21.630
<v Speaker 1>signing contracts to give that capacity to them can probably

0:06:21.670 --> 0:06:25.650
<v Speaker 1>cancel the agreements. To the outside world, everything looks like

0:06:25.690 --> 0:06:28.650
<v Speaker 1>a thriving industry with tons of demand both for AI GPUs,

0:06:28.670 --> 0:06:31.270
<v Speaker 1>which they think are turning into money immediately, and for

0:06:31.350 --> 0:06:35.560
<v Speaker 1>AI compute. With Microsoft's clearly crooked statements around adding a

0:06:35.630 --> 0:06:39.800
<v Speaker 1>gigawatt of capacity every single quarter, creating the illusion that

0:06:39.839 --> 0:06:42.420
<v Speaker 1>its revenue growth is from adding all of that capacity

0:06:42.440 --> 0:06:45.520
<v Speaker 1>and from spending all of those capital expenditures and from

0:06:45.580 --> 0:06:49.039
<v Speaker 1>diverse demand on top of it. Rather than mostly from

0:06:49.100 --> 0:06:53.000
<v Speaker 1>OpenAI's compute spend, which made up 70%, of Microsoft's AI

0:06:53.060 --> 0:06:57.640
<v Speaker 1>revenues in fiscal year 2026 and 7% of its overall

0:06:57.700 --> 0:07:02.620
<v Speaker 1>revenues in that same fiscal year. To be very specific

0:07:02.660 --> 0:07:05.300
<v Speaker 1>about what I mean, people are assuming that lots of

0:07:05.360 --> 0:07:08.160
<v Speaker 1>computers coming online every quarter and that all of that

0:07:08.220 --> 0:07:11.460
<v Speaker 1>compute is being immediately rented out for high rates. As

0:07:11.520 --> 0:07:14.140
<v Speaker 1>a result, they think that building more AI data centers

0:07:14.180 --> 0:07:16.760
<v Speaker 1>is an obvious choice and akin to printing money because

0:07:16.780 --> 0:07:19.740
<v Speaker 1>they believe there's unbelievable amounts of pent-up demand waiting to

0:07:19.780 --> 0:07:22.010
<v Speaker 1>be met and that meeting that demand will be as

0:07:22.090 --> 0:07:24.900
<v Speaker 1>simple as bringing a data center online, which Microsoft is,

0:07:24.980 --> 0:07:28.140
<v Speaker 1>from the outside, proving can be done a gigawatt a quarter.

0:07:28.760 --> 0:07:31.480
<v Speaker 1>Except it's not AI data center capacity that's coming online.

0:07:32.340 --> 0:07:36.410
<v Speaker 1>What's actually happening is very little capacity is coming online

0:07:36.450 --> 0:07:39.830
<v Speaker 1>for AI services, and what little does is going straight

0:07:39.870 --> 0:07:42.830
<v Speaker 1>into the hands of two companies with near-infinite resources provided

0:07:42.870 --> 0:07:46.040
<v Speaker 1>by venture capital, and in some cases the hyperscalers themselves,

0:07:46.820 --> 0:07:49.560
<v Speaker 1>which allows them to sign contracts to take up massive

0:07:49.600 --> 0:07:54.860
<v Speaker 1>amounts of capacity, and indeed fill those revenue backlogs. Microsoft

0:07:54.920 --> 0:07:59.570
<v Speaker 1>is saying that it's bringing data center capacity, not AI-specific

0:07:59.690 --> 0:08:02.870
<v Speaker 1>data center capacity, online a gigawatt at a time, but

0:08:02.910 --> 0:08:06.230
<v Speaker 1>it appears the actual AI capacity is, at best, coming

0:08:06.310 --> 0:08:08.250
<v Speaker 1>on at a rate of maybe tens or hundreds of

0:08:08.290 --> 0:08:11.110
<v Speaker 1>megawatts a quarter, and that really is the best case.

0:08:12.380 --> 0:08:14.620
<v Speaker 1>Perhaps there are others that are doing it faster, but

0:08:14.680 --> 0:08:18.040
<v Speaker 1>I actually can't find evidence of them. I'll be honest

0:08:18.060 --> 0:08:20.410
<v Speaker 1>that I have had this story in my head for weeks,

0:08:20.810 --> 0:08:23.530
<v Speaker 1>and was hesitant to jump the gun because of its ramifications.

0:08:24.530 --> 0:08:28.470
<v Speaker 1>As it appears that everyone's got this story wrong. This

0:08:28.530 --> 0:08:31.690
<v Speaker 1>is magnitudes worse than the dot-com bubble. To make the

0:08:31.750 --> 0:08:34.700
<v Speaker 1>obvious comparison, this is akin to the fiber optic cable

0:08:34.760 --> 0:08:37.100
<v Speaker 1>not even being laid in the ground, or only being

0:08:37.160 --> 0:08:40.619
<v Speaker 1>halfway connected, except the actual demand for AI infrastructure is

0:08:40.660 --> 0:08:44.829
<v Speaker 1>so thoroughly distorted by OpenAI and Anthropic that there's anywhere

0:08:44.870 --> 0:08:47.130
<v Speaker 1>from 5 to 50 times the amount of capacity in

0:08:47.190 --> 0:08:50.189
<v Speaker 1>planning than there is demand for it. And even then,

0:08:50.510 --> 0:08:55.070
<v Speaker 1>that demand is inflated. To make matters worse, it's not

0:08:55.210 --> 0:08:58.830
<v Speaker 1>going anywhere after this. There is no situation where this

0:08:58.870 --> 0:09:00.740
<v Speaker 1>is going to be used at the scale it's built.

0:09:01.059 --> 0:09:03.120
<v Speaker 1>It will also be just as expensive to finish an

0:09:03.160 --> 0:09:05.439
<v Speaker 1>AI data center in 2 or 3 or 5 years,

0:09:05.920 --> 0:09:09.179
<v Speaker 1>and just as expensive to run those GPUs. And who

0:09:09.220 --> 0:09:11.780
<v Speaker 1>knows how many generations will be in there, but I

0:09:11.780 --> 0:09:14.800
<v Speaker 1>don't know, in my opinion, My opinion, my esteemed opinion,

0:09:15.360 --> 0:09:17.939
<v Speaker 1>I don't think we make it past Vera Rubin or Feynman,

0:09:18.020 --> 0:09:21.400
<v Speaker 1>which is the generation after Vera Rubin. Even then, I

0:09:21.480 --> 0:09:24.230
<v Speaker 1>have questions about whether Feynman actually ships. Who knows, though?

0:09:24.540 --> 0:09:28.400
<v Speaker 1>Things are crazy. I'll also be clear that, as I

0:09:28.420 --> 0:09:31.490
<v Speaker 1>discussed with Paul Kudrowski this week, that there is no

0:09:31.650 --> 0:09:34.610
<v Speaker 1>post-bubble economy for GPUs, at least at the scale they're

0:09:34.630 --> 0:09:37.650
<v Speaker 1>being built, because the cost of serving inference is similar

0:09:37.670 --> 0:09:40.760
<v Speaker 1>to an airline with high fixed costs. You as an

0:09:40.820 --> 0:09:44.400
<v Speaker 1>operator must buy capacity, hundreds or thousands of GPUs over

0:09:44.440 --> 0:09:46.420
<v Speaker 1>a certain period of time, whether or not you actually

0:09:46.480 --> 0:09:48.990
<v Speaker 1>have the demand. While there might be a situation where

0:09:49.000 --> 0:09:51.570
<v Speaker 1>you could eke out a profit of some sort with

0:09:51.590 --> 0:09:54.610
<v Speaker 1>the perfect amount of customers, anything above the demand you

0:09:54.650 --> 0:09:57.929
<v Speaker 1>foresaw means an unstable and unreliable service and turning customers

0:09:57.970 --> 0:10:00.739
<v Speaker 1>away at the door, and anything below can every ounce

0:10:00.760 --> 0:10:04.920
<v Speaker 1>of margin you might have. Dario Amadei himself said as

0:10:04.960 --> 0:10:07.420
<v Speaker 1>much in an interview back in February with Dwarkish Patel,

0:10:07.520 --> 0:10:10.720
<v Speaker 1>saying that, in a theoretical scenario, he could buy $ 1

0:10:10.480 --> 0:10:13.900
<v Speaker 1>trillion of compute that starts at the end of 2027,

0:10:13.900 --> 0:10:16.060
<v Speaker 1>but if his revenue wasn't $ 1 trillion, if it was

0:10:16.080 --> 0:10:20.679
<v Speaker 1>even $ 800 billion, there was, and I quote, No hedge

0:10:20.740 --> 0:10:24.500
<v Speaker 1>on earth that could stop him from going bankrupt. These

0:10:24.540 --> 0:10:27.619
<v Speaker 1>are the underlying economics of every single company running an

0:10:27.679 --> 0:10:32.329
<v Speaker 1>AI service, and they're equal parts brittle and volatile. Thanks

0:10:32.350 --> 0:10:33.970
<v Speaker 1>for listening. I'll see you next week.