WEBVTT - Monologue: What The Hell Are The Hyperscalers Doing?

0:00:02.800 --> 0:00:07.760
<v Speaker 1>Zone, media, superstition, fear, and jealousy. Welcome to this week's

0:00:07.760 --> 0:00:22.599
<v Speaker 1>Better Offline monologue. I'm your host ed ze Tron. As ever,

0:00:22.640 --> 0:00:25.120
<v Speaker 1>please subscribe to the newsletter. We're running a one year

0:00:25.120 --> 0:00:27.240
<v Speaker 1>anniversary deal right now. The link will be in the

0:00:27.280 --> 0:00:29.760
<v Speaker 1>episode notes, and we're giving you ten bucks off the

0:00:29.800 --> 0:00:32.760
<v Speaker 1>annual rate for life and even if you die and

0:00:32.760 --> 0:00:34.519
<v Speaker 1>come back to life, I'll give it to you anyway.

0:00:34.760 --> 0:00:37.040
<v Speaker 1>The deal runs through midnight on Sunday. But if you

0:00:37.040 --> 0:00:39.120
<v Speaker 1>happen to hear this after this point, email me at

0:00:39.120 --> 0:00:41.880
<v Speaker 1>easy at Better offline dot com or ez at better

0:00:41.920 --> 0:00:44.479
<v Speaker 1>Offline dot com for the British and Canadians in the audience,

0:00:44.640 --> 0:00:46.800
<v Speaker 1>and I'll sort you out. I'm generous like that, and

0:00:46.880 --> 0:00:48.920
<v Speaker 1>I love my premium mates. Really some of the best

0:00:48.960 --> 0:00:52.080
<v Speaker 1>analysts work out there. It's very deep. I got the

0:00:52.159 --> 0:00:54.160
<v Speaker 1>terminal going now and pulling all sorts of data. The

0:00:54.240 --> 0:00:57.320
<v Speaker 1>many numbers I read every week are driving me truly insane.

0:00:57.960 --> 0:01:01.000
<v Speaker 1>And with that in mind, to day's monologue starts with

0:01:01.000 --> 0:01:05.520
<v Speaker 1>a simple question, what the fuck are hyperscaler is doing? Now?

0:01:05.560 --> 0:01:08.160
<v Speaker 1>I realize I've asked this question maybe one hundred thousand

0:01:08.200 --> 0:01:10.200
<v Speaker 1>times in the last two years, but it's begun to

0:01:10.240 --> 0:01:15.280
<v Speaker 1>get so incredibly stupid that I'm actually a little exasperated. McKinsey.

0:01:15.319 --> 0:01:19.280
<v Speaker 1>Spawn and Google CEO Sundar Pashai announced in Wednesday's earnings

0:01:19.280 --> 0:01:22.200
<v Speaker 1>that Google would raise its year end capital expenditures to

0:01:22.240 --> 0:01:24.640
<v Speaker 1>somewhere between one hundred and ninety five billion dollars and

0:01:24.680 --> 0:01:27.880
<v Speaker 1>two hundred and five billion dollars after burning eighty point

0:01:28.000 --> 0:01:30.000
<v Speaker 1>six billion dollars in the first half of the year,

0:01:30.319 --> 0:01:32.720
<v Speaker 1>and this suggests that Google will spend another one hundred

0:01:32.760 --> 0:01:35.000
<v Speaker 1>and fourteen billion dollars by the end of twenty twenty six.

0:01:35.040 --> 0:01:37.200
<v Speaker 1>They're going to be doing what fifty five sixty billion

0:01:37.240 --> 0:01:41.199
<v Speaker 1>dollars a quarter. It's fucking insane, it's fucking stupid. I'm

0:01:41.240 --> 0:01:46.040
<v Speaker 1>sorry for swearing, but it's absolutely deranged. What are you doing, Sunda.

0:01:46.160 --> 0:01:48.720
<v Speaker 1>You've already spent two hundred and eighty eight billion dollars

0:01:48.720 --> 0:01:51.040
<v Speaker 1>in capex since twenty twenty two, and all you've got

0:01:51.040 --> 0:01:53.760
<v Speaker 1>to show for it is several also RAN llms and

0:01:53.840 --> 0:01:58.080
<v Speaker 1>one indeterminately large customer in Anthropic, who you've also invested

0:01:58.120 --> 0:02:02.600
<v Speaker 1>thirteen billion dollars in. Now. One answer might be that

0:02:02.840 --> 0:02:06.240
<v Speaker 1>Anthropic helped choose Google cloud's growth by eighty two percent

0:02:06.320 --> 0:02:08.720
<v Speaker 1>year of a year in this qure But considering that

0:02:08.760 --> 0:02:10.840
<v Speaker 1>Google doesn't actually break out how much of that is

0:02:10.840 --> 0:02:14.040
<v Speaker 1>from AI, or from Anthropic, or really from anything, it's

0:02:14.080 --> 0:02:16.280
<v Speaker 1>difficult to say what exactly this money is doing, other

0:02:16.320 --> 0:02:18.480
<v Speaker 1>than turning its free cash flow negative for the first

0:02:18.560 --> 0:02:21.840
<v Speaker 1>time since Google went public. While I can't say for

0:02:21.880 --> 0:02:24.440
<v Speaker 1>sure how much Anthropic is spending, I do believe it's

0:02:24.480 --> 0:02:27.480
<v Speaker 1>taking up the vast majority of Google's AI infrastructure and

0:02:27.560 --> 0:02:31.680
<v Speaker 1>representing most of its cloud growth. Google Cloud was, up

0:02:31.760 --> 0:02:34.119
<v Speaker 1>until fairly recently the last year or so, are much

0:02:34.120 --> 0:02:37.680
<v Speaker 1>smaller business than Microsoft. As euroamason web services have been

0:02:37.720 --> 0:02:40.639
<v Speaker 1>growing somewhere between twenty two percent and thirty percent year

0:02:40.680 --> 0:02:43.240
<v Speaker 1>of a year, hitting thirty three point nine percent year

0:02:43.280 --> 0:02:44.839
<v Speaker 1>of a year growth in the third quarter of twenty

0:02:44.919 --> 0:02:49.840
<v Speaker 1>twenty five. That very same quarter, Google and Anthropic announced

0:02:49.840 --> 0:02:52.120
<v Speaker 1>the deal to expand its use of TPUs, which is

0:02:52.160 --> 0:02:55.280
<v Speaker 1>Google's special GPUs, in a contract worth than I quote

0:02:55.360 --> 0:02:59.680
<v Speaker 1>tens of billions of dollars. Mysteriously can't imagine how Google's

0:02:59.720 --> 0:03:01.880
<v Speaker 1>next time three quarters of cloud revenue growth showed forty

0:03:01.919 --> 0:03:04.560
<v Speaker 1>seven point seven percent sixty three point four percent and

0:03:04.600 --> 0:03:06.840
<v Speaker 1>eighty two point one percent year of a year growth.

0:03:07.080 --> 0:03:10.240
<v Speaker 1>Whatever could that be? Where did that come from? I

0:03:10.320 --> 0:03:13.440
<v Speaker 1>can't work it out, can you, dear listener? What could

0:03:13.480 --> 0:03:18.120
<v Speaker 1>it possively? It's anthropic, It's it's obviously anthropic. Except in

0:03:18.200 --> 0:03:22.680
<v Speaker 1>those same quarters, Google's capital expenditures went up ninety five percent,

0:03:22.919 --> 0:03:25.079
<v Speaker 1>one hundred and seven percent, and one hundred percent year

0:03:25.120 --> 0:03:28.240
<v Speaker 1>of a year, absorbing every dollar of cloud revenue and

0:03:28.280 --> 0:03:30.720
<v Speaker 1>then some. In fact, outside of the fourth quarter of

0:03:30.720 --> 0:03:34.560
<v Speaker 1>twenty twenty four, Google's capital expenditures have outpaced cloud growth

0:03:34.600 --> 0:03:36.520
<v Speaker 1>by at least twelve percent year of a year for

0:03:36.560 --> 0:03:40.680
<v Speaker 1>two goddamn years. And if your pushback is, well, Google

0:03:40.760 --> 0:03:44.280
<v Speaker 1>spending that CAPEX on something else, No, they're not. They're not.

0:03:44.680 --> 0:03:46.320
<v Speaker 1>Maybe they've got a couple of billy coming in from

0:03:46.400 --> 0:03:51.200
<v Speaker 1>Google Workspace and cucking their various subscribers with annoying things.

0:03:52.000 --> 0:03:55.680
<v Speaker 1>It's coming from anthropic. Google is not making that money

0:03:55.680 --> 0:03:58.560
<v Speaker 1>from Gemini. Your ass must be crazy if you think

0:03:58.640 --> 0:04:01.560
<v Speaker 1>that's that's the point. But all of this begs the question,

0:04:02.160 --> 0:04:06.480
<v Speaker 1>when exactly does all of this payoff? How does it payoff?

0:04:06.800 --> 0:04:08.960
<v Speaker 1>What does the payoff even look like? And why am

0:04:08.960 --> 0:04:13.360
<v Speaker 1>I one of the few people asking these questions, hyperscalas

0:04:13.360 --> 0:04:15.960
<v Speaker 1>have now spent I asked them at about one point

0:04:15.960 --> 0:04:20.240
<v Speaker 1>one trillion dollars. I haven't seen Amazon and Microsoft's quarters yet,

0:04:20.279 --> 0:04:21.719
<v Speaker 1>and I'm going to guess they spend a good amount

0:04:21.720 --> 0:04:24.560
<v Speaker 1>on capex. They now spent one point one trillion bloody

0:04:24.600 --> 0:04:27.120
<v Speaker 1>dollars on capital expenditures. And when you ask them to

0:04:27.160 --> 0:04:29.880
<v Speaker 1>tell you how much revenue they made from said expenditures,

0:04:29.920 --> 0:04:32.039
<v Speaker 1>they look at you like that meme of Brendan Fraser

0:04:32.040 --> 0:04:35.120
<v Speaker 1>from the whale. Anyway, here's a fun fact for you.

0:04:35.640 --> 0:04:38.720
<v Speaker 1>Thanks to my own reporting and open ais auditive financials,

0:04:39.000 --> 0:04:42.040
<v Speaker 1>we know that It's spent seventeen point two billion dollars

0:04:42.080 --> 0:04:44.640
<v Speaker 1>on Microsoft a zure in colendar year twenty twenty five,

0:04:44.960 --> 0:04:47.960
<v Speaker 1>a year when Microsoft's intelligent cloud segment made one hundred

0:04:47.960 --> 0:04:50.640
<v Speaker 1>and twenty point four billion dollars in revenue. The segment

0:04:50.680 --> 0:04:54.360
<v Speaker 1>grew as a result twenty six percent year of a year. Yeah,

0:04:54.400 --> 0:04:56.719
<v Speaker 1>I'm a curious little Chris or around the numbers, and

0:04:56.839 --> 0:05:00.640
<v Speaker 1>open Ay's revenue represented fourteen point three percent of all

0:05:00.839 --> 0:05:04.240
<v Speaker 1>Microsoft's cloud revenues for that period. And to be clear,

0:05:04.440 --> 0:05:07.080
<v Speaker 1>it's calendar year twenty twenty five, they have a fiscal

0:05:07.120 --> 0:05:09.520
<v Speaker 1>core thing. But I went and worked it out. But

0:05:09.600 --> 0:05:13.359
<v Speaker 1>this means the open Aiy's money represented eighteen percent of

0:05:13.400 --> 0:05:16.520
<v Speaker 1>the year of year growth of Microsoft's cloud. Without it,

0:05:17.040 --> 0:05:19.479
<v Speaker 1>their cloud revenue growth was a mere eight percent year

0:05:19.520 --> 0:05:22.680
<v Speaker 1>of a year. That's pathetic. That's a bonable. That's barely

0:05:22.760 --> 0:05:26.320
<v Speaker 1>keeping up with inflation. That's dog shit. Satya, Satchy, you

0:05:26.320 --> 0:05:29.080
<v Speaker 1>should be ashamed of yourself. Sacha, come on my show

0:05:29.200 --> 0:05:32.400
<v Speaker 1>so I can ask you very reasonable questions. I'll be nice.

0:05:32.480 --> 0:05:35.640
<v Speaker 1>I swear you won't. You're scared, just let the rest

0:05:35.680 --> 0:05:39.520
<v Speaker 1>of them, just like Clammy Sam on one Clamuel, your

0:05:39.520 --> 0:05:42.159
<v Speaker 1>public relations people said you're too scared to speak to me. Well,

0:05:42.200 --> 0:05:43.599
<v Speaker 1>they said you wouldn't speak to me, and I know

0:05:43.640 --> 0:05:46.599
<v Speaker 1>you're scared. Same with weird Daria. Where are you boys?

0:05:46.640 --> 0:05:49.240
<v Speaker 1>You're too scared of better offline? Anyway, Enough of the

0:05:49.279 --> 0:05:53.320
<v Speaker 1>Trump voice, though, there is actually a reason that hyperscalers

0:05:53.320 --> 0:05:57.000
<v Speaker 1>are still spending all that capex, open AI and Anthropic

0:05:57.040 --> 0:06:00.000
<v Speaker 1>allow hyperscalers to boost growth by feeding their own money

0:06:00.120 --> 0:06:03.400
<v Speaker 1>back into their businesses, a short term solution to the

0:06:03.480 --> 0:06:06.520
<v Speaker 1>long term problem of the end of the era of hypergrowth.

0:06:06.760 --> 0:06:09.400
<v Speaker 1>One of the reasons they're so shy about AI revenues

0:06:09.520 --> 0:06:12.279
<v Speaker 1>is that they know it'll become obvious where those revenues

0:06:12.320 --> 0:06:16.680
<v Speaker 1>are coming from two unprofitable and unsustainable AI labs and

0:06:16.720 --> 0:06:20.920
<v Speaker 1>a little trickle, little spurt from AI services they dump

0:06:21.000 --> 0:06:24.279
<v Speaker 1>on their enterprise customers through confusing subscriptions up cells and

0:06:24.400 --> 0:06:28.839
<v Speaker 1>just adding AI to random things and being like, look, folks,

0:06:29.120 --> 0:06:31.479
<v Speaker 1>don't we love this? Don't you all love this? You

0:06:31.560 --> 0:06:36.520
<v Speaker 1>little shits? And hyperscalers also know that the second they

0:06:36.560 --> 0:06:41.200
<v Speaker 1>cut their capex spend, their AI trade unwinds. If Microsoft, Google,

0:06:41.320 --> 0:06:44.320
<v Speaker 1>Amazon and Meta aren't buying hundreds of billions of dollars

0:06:44.320 --> 0:06:47.880
<v Speaker 1>of GPUs a year, that means Taiwanese server manufacturers like

0:06:47.960 --> 0:06:50.640
<v Speaker 1>Quanta and on High won't sell as many servers and

0:06:51.000 --> 0:06:53.919
<v Speaker 1>thus won't need as many GPUs broad common in video

0:06:54.000 --> 0:06:57.039
<v Speaker 1>won't need as much high bandwid memory for those TPUs

0:06:57.080 --> 0:06:59.960
<v Speaker 1>and GPUs, and they are sent speculators across the world

0:07:00.160 --> 0:07:03.600
<v Speaker 1>will suddenly realize that they conflated hyperscalers being stupid with

0:07:03.640 --> 0:07:06.360
<v Speaker 1>their money with the existence of massive amounts of demand

0:07:06.400 --> 0:07:09.960
<v Speaker 1>for AI compute. But another way, hyperscalers are doing k

0:07:10.160 --> 0:07:12.560
<v Speaker 1>fape to keep the crowd riled up, but everybody slowly

0:07:12.560 --> 0:07:14.680
<v Speaker 1>working out that the wrestlers aren't really hitting each other

0:07:14.720 --> 0:07:18.360
<v Speaker 1>and everything is scripted. I also should add I can

0:07:18.440 --> 0:07:20.600
<v Speaker 1>just anticipate I'm going to get an email from someone

0:07:20.640 --> 0:07:23.400
<v Speaker 1>being like, well, actually, this is good. The cattle expenditures

0:07:23.400 --> 0:07:26.440
<v Speaker 1>are good because it showed growth, even though it's short term.

0:07:26.600 --> 0:07:30.400
<v Speaker 1>That's a silly, silly perspective to take, considering they're feeding

0:07:30.440 --> 0:07:32.720
<v Speaker 1>all the revenues back into the business in the capex

0:07:32.720 --> 0:07:35.480
<v Speaker 1>for stuff that won't be useful in the future. Mathematics.

0:07:35.480 --> 0:07:41.040
<v Speaker 1>My friend try it sometime. Anyway. It is a funny

0:07:41.040 --> 0:07:42.760
<v Speaker 1>old time for me right now, because I feel like

0:07:42.760 --> 0:07:45.960
<v Speaker 1>I'm on three or four podcasts or TV shows or

0:07:45.960 --> 0:07:48.800
<v Speaker 1>interviews a week, as the system appears to be trying

0:07:48.840 --> 0:07:51.200
<v Speaker 1>to cover its ass by having me on to explain

0:07:51.240 --> 0:07:53.800
<v Speaker 1>stuff I said two years ago, but a little more

0:07:53.840 --> 0:07:57.920
<v Speaker 1>insistently and clearly than before. While the mainstream media seems

0:07:57.960 --> 0:08:00.360
<v Speaker 1>hesitant to fully commit to the graveyard small of the

0:08:00.360 --> 0:08:03.160
<v Speaker 1>AI bubble. It feels as if there's a conscious attempt

0:08:03.200 --> 0:08:05.840
<v Speaker 1>to represent my work as more than just the other side,

0:08:05.960 --> 0:08:08.440
<v Speaker 1>or a hater or a skeptic or a cynic, which

0:08:08.480 --> 0:08:12.080
<v Speaker 1>is both edifying and enraging because we are now so

0:08:12.240 --> 0:08:17.120
<v Speaker 1>far past past the point of no Returnemy. Yeah, we

0:08:17.240 --> 0:08:20.560
<v Speaker 1>really can't avoid what's coming. As I wrote in my

0:08:20.600 --> 0:08:23.120
<v Speaker 1>newsletter this week, I think there's a subprime data center

0:08:23.200 --> 0:08:26.160
<v Speaker 1>crisis brewing right now, with over five hundred billion dollars

0:08:26.160 --> 0:08:28.920
<v Speaker 1>in data center SPVs that can only ever make investors

0:08:28.920 --> 0:08:31.360
<v Speaker 1>whole if there's hundreds of billions of dollars of AI

0:08:31.720 --> 0:08:34.280
<v Speaker 1>compute demand in the next two years. And to be clear,

0:08:34.320 --> 0:08:36.520
<v Speaker 1>when I say hundreds of billions of dollars demand, I

0:08:36.600 --> 0:08:39.480
<v Speaker 1>mean a year. I mean literally a year. I ran

0:08:39.520 --> 0:08:42.000
<v Speaker 1>the numbers, and for the what I think it's one

0:08:42.040 --> 0:08:44.839
<v Speaker 1>hundred ninety giga wats of supposedly planned data centers, you'd

0:08:44.880 --> 0:08:47.240
<v Speaker 1>need one point six trillion dollars a year in annual

0:08:47.280 --> 0:08:51.200
<v Speaker 1>compute revenue to make it worth worthwhile. And by the way,

0:08:51.200 --> 0:08:54.360
<v Speaker 1>the global software market is under eight hundred billion dollars

0:08:54.400 --> 0:08:56.480
<v Speaker 1>a year, So I don't know where that fucking money's

0:08:56.520 --> 0:08:59.040
<v Speaker 1>coming from, and no one else seems to really want

0:08:59.040 --> 0:09:01.480
<v Speaker 1>to think about him. And I deeply worry about what

0:09:01.559 --> 0:09:04.080
<v Speaker 1>happens when the demand doesn't arrive. As these data center

0:09:04.120 --> 0:09:06.679
<v Speaker 1>debt deals are often funded by private credit funds, which

0:09:06.760 --> 0:09:10.240
<v Speaker 1>are in turn funded by both public and private pensions

0:09:10.240 --> 0:09:14.360
<v Speaker 1>and insurance funds. Anyway, do read the free newsletter to

0:09:14.440 --> 0:09:16.800
<v Speaker 1>learn more. It is free this one. I have a

0:09:16.800 --> 0:09:20.440
<v Speaker 1>premium one coming out today that will be about well,

0:09:20.520 --> 0:09:22.800
<v Speaker 1>the fact that Oracle's dying, which I mean, hey, you

0:09:22.840 --> 0:09:26.080
<v Speaker 1>in the back, stop laughing. Stop it, poor Larry Ellison

0:09:26.280 --> 0:09:30.079
<v Speaker 1>Paul who gives a fuck in all seriousness. Though, it's

0:09:30.120 --> 0:09:32.040
<v Speaker 1>going to be a banger. The Free ones a banger.

0:09:32.080 --> 0:09:34.760
<v Speaker 1>There's bangers everywhere. And I'll be back next week with

0:09:34.800 --> 0:09:37.120
<v Speaker 1>my good friend Nick Saresh to talk about how AI

0:09:37.280 --> 0:09:40.520
<v Speaker 1>mania is eviscerating global decision making, creating a cult lug

0:09:40.640 --> 0:09:44.400
<v Speaker 1>environment where once participation in AI is almost a religious act,

0:09:44.480 --> 0:09:47.640
<v Speaker 1>and if you talk badly about AI, it's considered a

0:09:47.679 --> 0:09:50.440
<v Speaker 1>cardinal sin or a cardinal sin, one of those two.

0:09:50.640 --> 0:09:53.880
<v Speaker 1>Sure one of you will correct me on this one. Nevertheless,

0:09:54.240 --> 0:09:56.600
<v Speaker 1>it's a really he it's based on this incredible piece

0:09:56.640 --> 0:09:58.600
<v Speaker 1>he wrote. I can't wait for you to read it.

0:09:58.640 --> 0:10:00.960
<v Speaker 1>I can't wait for you Timmy talk to him. Oh no,

0:10:01.160 --> 0:10:04.000
<v Speaker 1>can't wait for you to consume my content. What a

0:10:04.000 --> 0:10:07.160
<v Speaker 1>peculiar thing to say. Anyway, I'll got you Next week

0:10:07.559 --> 0:10:08.240
<v Speaker 1>Zitron out