WEBVTT - Hater Season: Corey Quinn 0:00:02.400 --> 0:00:03.560 All Media. 0:00:04.559 --> 0:00:07.400 Hello on, Welcome to Better Offline. I am, of course 0:00:07.440 --> 0:00:22.880 your host ed zitron. Now we're in hate a season. 0:00:22.880 --> 0:00:25.040 Hate a season is when I bring people on allowed 0:00:25.079 --> 0:00:27.880 to get rude, because you know, this podcast is usually 0:00:27.920 --> 0:00:30.960 so reserved and nice in the way we refer to people, 0:00:31.040 --> 0:00:34.120 especially tech executives, which we've never referred to as funck 0:00:34.159 --> 0:00:37.839 wits or morons or dumbasses or shittheads or shit heels 0:00:37.920 --> 0:00:41.160 or ass wipes or fuck nuts or nob ends. We've 0:00:41.200 --> 0:00:44.240 never done any of that ever. Nevertheless, joining me today 0:00:44.280 --> 0:00:47.080 is chief cloud economist that Doug Bell Corey. Quinn Corey, 0:00:47.159 --> 0:00:47.800 how are you doing? 0:00:48.280 --> 0:00:50.720 I have delighted to be here because normally in all 0:00:50.760 --> 0:00:53.040 the other places I find myself, I have to be 0:00:53.120 --> 0:00:53.840 so reserved. 0:00:54.480 --> 0:00:59.440 You have to pull your punches. Cory, today's Hatter season. 0:00:59.680 --> 0:01:02.720 Tell me about your feelings on Amazon Web services. 0:01:03.360 --> 0:01:07.320 Oh, dear lord, it's it's like it's weird. I don't 0:01:07.360 --> 0:01:10.280 actually hate Amazon. People think I do because of the 0:01:10.319 --> 0:01:12.160 things I say and the things that I do. But 0:01:12.200 --> 0:01:15.000 if I hate it, then this directions, but ten years 0:01:15.040 --> 0:01:17.360 in I write a newsletter about it. If I actually 0:01:17.400 --> 0:01:19.640 hate the company and want nothing but ill things to 0:01:19.680 --> 0:01:22.080 happen to them. First off, that's a pathology and I 0:01:22.120 --> 0:01:25.160 need a restraining order. And secondly, it always bugs me 0:01:25.200 --> 0:01:27.200 when people think that I have an ax to grind 0:01:27.280 --> 0:01:30.679 against the company, because ten years in, are you seriously 0:01:30.760 --> 0:01:33.319 suggesting this is the best I would be able to 0:01:33.400 --> 0:01:35.919 do if I actually wanted to hit them where it hurts? 0:01:36.040 --> 0:01:36.399 Come on? 0:01:37.000 --> 0:01:39.000 But yeah, they've been annoying the piss out of me lately. 0:01:39.480 --> 0:01:42.520 So AWS is weird because maybe you can explain this 0:01:42.600 --> 0:01:46.480 to me. Over my entire tech career, I've heard this 0:01:46.680 --> 0:01:50.000 constant story that your AWS bill just it just goes up. 0:01:50.440 --> 0:01:53.640 It just arbitrarily increases. It's not really clear whether they're 0:01:53.640 --> 0:01:56.920 doing price increases, though I know those exist too. 0:01:56.680 --> 0:01:59.200 Rarely open but occasionally increasingly these days. 0:01:59.320 --> 0:02:02.400 Yes, but why what is it that makes the bill 0:02:02.480 --> 0:02:04.720 go up? Why? And why does this happen so often? 0:02:05.160 --> 0:02:07.240 I used to think it was a conspiracy, and then 0:02:07.280 --> 0:02:09.480 I realized Amazon is in no way, shape or form 0:02:09.560 --> 0:02:12.600 organizationally competent enough to pull something like that off. It 0:02:12.680 --> 0:02:17.519 is the nature of charging per usage and honestly, assholes 0:02:17.800 --> 0:02:20.359 where Okay, we're going to launch a service and we're 0:02:20.360 --> 0:02:22.920 going to charge you for every gigabyte you put in it. 0:02:22.919 --> 0:02:25.880 That's all we're gonna do. Great, straightforward humans can rationalize 0:02:25.919 --> 0:02:28.520 around that, great eazy psy. They did that when they 0:02:28.560 --> 0:02:31.200 first launched S three and beta back in two thousand 0:02:31.200 --> 0:02:33.760 and three. Is storage right, it's a big object store 0:02:33.840 --> 0:02:36.880 picture it as an infinite sized disc and you're pretty close. 0:02:37.200 --> 0:02:40.239 But they would that Originally they only charged for the 0:02:40.320 --> 0:02:43.320 data you stored there. Great, so people stored a bunch 0:02:43.360 --> 0:02:46.720 of objects that were zero byte in size, so they 0:02:46.840 --> 0:02:49.800 just used the names of those objects, and the ability 0:02:49.840 --> 0:02:51.640 to query with that object existed or not in what 0:02:51.680 --> 0:02:55.680 its name was. As a free database, this at the time, also, 0:02:55.800 --> 0:02:58.600 according to legend, didn't just it wasn't just about taking 0:02:58.600 --> 0:03:02.120 free services. It was also just crippling the performance of 0:03:02.360 --> 0:03:04.200 S three at the time. So what they did is 0:03:04.240 --> 0:03:07.959 they started also charging every request you make with jarj 0:03:08.040 --> 0:03:10.200 a small fee. I think it's something like a penny 0:03:10.240 --> 0:03:12.840 per thousand requests of memory served. Don't book me on that. 0:03:12.880 --> 0:03:15.760 I look it up when it matters, and that solves 0:03:15.760 --> 0:03:18.280 that problem. But then you have things like that that 0:03:18.320 --> 0:03:20.680 are patched on and tied to things, and tied to 0:03:20.720 --> 0:03:24.600 things and tied to more things, and in any complex organization. 0:03:25.080 --> 0:03:28.280 What happens is is now suddenly you have this morass 0:03:28.480 --> 0:03:31.400 of one hundred million dollars a year and spent great, 0:03:31.639 --> 0:03:33.720 what's in there that's costing all the money? 0:03:33.840 --> 0:03:34.080 Well? 0:03:34.520 --> 0:03:37.520 Everything? But if you look at anything individual and you 0:03:37.600 --> 0:03:40.400 don't think it's being used, Okay, am I turn I'm 0:03:40.400 --> 0:03:42.839 going to turn it off. If I'm right, it's going 0:03:42.880 --> 0:03:44.880 to save me a bit of money. If I'm wrong, 0:03:44.920 --> 0:03:47.280 it's going to take down production. And now I have 0:03:47.360 --> 0:03:50.520 a serious problem trying to serve my customers. So the 0:03:50.640 --> 0:03:56.120 safe bias is don't turn things off. Also, you're an engineer. Hypothetically, 0:03:56.160 --> 0:03:57.880 I realize that can be an insult to your part 0:03:57.920 --> 0:04:00.920 of the world. Roll with it for a moment. Exactly, 0:04:01.040 --> 0:04:03.680 you're an engineer. You want something to spin up, and 0:04:03.720 --> 0:04:05.880 no one is letting you spin it up. You will 0:04:05.920 --> 0:04:09.040 take up residents, annoying the hell out of them every 0:04:09.080 --> 0:04:11.680 ten minutes until you get the access that you need. 0:04:12.720 --> 0:04:15.560 You do your job. Things are done. Two things now 0:04:15.560 --> 0:04:17.960 work against you. One, you had to ask and bag 0:04:18.000 --> 0:04:19.680 and plead them to spin that up. You want to 0:04:19.680 --> 0:04:21.080 hang on to it in case you need it again 0:04:21.080 --> 0:04:23.040 because you don't want to go through that. But this 0:04:23.200 --> 0:04:27.520 is an organization exactly exactly. This is not Amazon's fault. 0:04:27.560 --> 0:04:29.880 This is human nature's fault. They haven't patched that yet, 0:04:29.920 --> 0:04:32.240 though they're I think they're trying their best iteration of 0:04:32.279 --> 0:04:35.360 a Laxa plus it. Why not, I'm kidding that thing 0:04:35.400 --> 0:04:38.080 just sells ads. We're good boy way. 0:04:38.120 --> 0:04:41.120 So I'm an engineer and I've advocated for my thing. 0:04:41.920 --> 0:04:44.599 Yes, and now you're done using it. No one is going, 0:04:44.640 --> 0:04:47.000 there's no there's no call to action to get you 0:04:47.080 --> 0:04:48.640 off your ass to go, and no one off or 0:04:48.640 --> 0:04:50.520 no one will come and badger you about it. It 0:04:50.640 --> 0:04:54.120 just it tends to exists. It's why people cynically say 0:04:54.120 --> 0:04:55.719 you're not charged for the things you use in the 0:04:55.720 --> 0:04:57.479 clouds so much as you're charged for the things you 0:04:57.520 --> 0:04:58.360 forget to turn off. 0:04:58.880 --> 0:05:01.240 And Amazon loves this. I'm guessing they love the fact 0:05:01.240 --> 0:05:04.200 that there's just this cloud anchor system. 0:05:05.160 --> 0:05:08.000 I would say that, on some level, contrary to popular belief, 0:05:08.160 --> 0:05:10.440 they don't love it as much as you'd think, because 0:05:10.520 --> 0:05:12.880 all right, I'm trying to run on my blog and 0:05:12.920 --> 0:05:15.040 I've spun up eight attempts at this blog, so it 0:05:15.040 --> 0:05:18.279 now costs eight times what it should. Great and the 0:05:18.400 --> 0:05:20.880 narrative becomes if they're not careful with all that wasted 0:05:20.960 --> 0:05:24.760 usage that's doing nothing is huh, the cloud is pants 0:05:24.760 --> 0:05:27.840 shittingly expensive. We should go back to data centers. There's 0:05:27.880 --> 0:05:30.760 really no one on the other side of this issue. 0:05:30.839 --> 0:05:33.640 So why doesn't Amazon help people turn tune these things 0:05:33.680 --> 0:05:35.760 down and turn it off? They made some a board 0:05:35.800 --> 0:05:38.080 of attempts at part of the organization. But this is 0:05:38.120 --> 0:05:41.360 going to surprise you. Amazon is not the most organizationally 0:05:41.520 --> 0:05:43.440 competent company you're ever going to meet. 0:05:43.880 --> 0:05:46.000 But what does that mean in practice? Is it just 0:05:46.080 --> 0:05:48.360 that they don't know they're ass from the air hole. 0:05:48.440 --> 0:05:51.640 They don't really know how things work. All of the above. 0:05:51.800 --> 0:05:54.839 Silos, it's all silos. They're over two hundred services. And 0:05:54.880 --> 0:05:56.960 we long ago across the point where I can speak 0:05:57.040 --> 0:06:01.520 incredibly convincingly about AWS service is that do not exist 0:06:01.800 --> 0:06:05.160 two Amazon employees and not get called out on it 0:06:05.200 --> 0:06:08.200 because who in the world knows everything that Amazon does? 0:06:08.320 --> 0:06:10.720 AND's just th guy, I sound as plausible as AI 0:06:10.839 --> 0:06:11.360 these days. 0:06:11.640 --> 0:06:16.600 Yeah, So it's just a labyrinth, the Leviathan of different 0:06:17.080 --> 0:06:18.960 SKUs and product categories that. 0:06:18.960 --> 0:06:22.880 Make million SKUs. If you want to be precise, Yes, Jesus. 0:06:22.600 --> 0:06:25.360 And those are all just different kinds of virtual machines 0:06:25.400 --> 0:06:25.880 and the like. 0:06:26.680 --> 0:06:29.320 Okay, we'll start there. Sure, virtual machines. You want to 0:06:29.320 --> 0:06:32.839 spin up a virtual machine in Virginia, US East one? Okay, 0:06:32.880 --> 0:06:34.840 which one are you going to pick? Because there are 0:06:34.920 --> 0:06:38.600 approximately seven hundred different types you can spin up. Are 0:06:38.600 --> 0:06:40.440 you going to pick the right one? Of course you're 0:06:40.480 --> 0:06:40.840 fucking not. 0:06:41.720 --> 0:06:44.599 Oh so how do you pick the right one? This 0:06:44.760 --> 0:06:48.200 is the thing because it seems that the people making 0:06:48.240 --> 0:06:51.039 these decisions of which virtual machines to get or whatever 0:06:51.279 --> 0:06:54.960 spin up on aws or engineers, right, Usually sulfware engineers 0:06:55.440 --> 0:06:58.600 software engineers aren't necessarily infrastructure experts, are they. 0:06:58.760 --> 0:07:02.520 You've noticed very often companies will standardize on one particular 0:07:02.560 --> 0:07:05.000 size of instance across the board. Why well, because in 0:07:05.040 --> 0:07:07.360 development it's the one that the original founding engineer picked, 0:07:07.360 --> 0:07:09.760 and we'll come back and revisit this later ten years ago. 0:07:10.080 --> 0:07:11.080 No one is revisited. 0:07:11.520 --> 0:07:12.520 M take that. 0:07:13.000 --> 0:07:15.320 They do have a tool that I like, it's even free, 0:07:15.360 --> 0:07:18.160 which I'm sure someone loses sleepover called compute optimizer that 0:07:18.200 --> 0:07:20.200 looks at you're running workloads and says, hey, this one 0:07:20.200 --> 0:07:22.840 should be bigger, this one should be smaller. I was 0:07:22.880 --> 0:07:26.240 extraordinarily skeptical when it came out. It is better at 0:07:26.240 --> 0:07:28.120 figuring that out than I am with a tooling that 0:07:28.160 --> 0:07:30.480 I built internally. So once again I'm thrilled to turn 0:07:30.480 --> 0:07:33.760 it off. It's one of the Amazon at its best things. 0:07:33.920 --> 0:07:36.600 I can only assume that the people responsible for that 0:07:36.680 --> 0:07:38.240 are being eyed for layoffs or. 0:07:38.160 --> 0:07:43.880 Pips, because Andy Jesse and several NBA snipers are outside 0:07:43.920 --> 0:07:45.360 at the location right now. 0:07:45.960 --> 0:07:51.400 It's frust stay, but I like AWS truly. Something changed. 0:07:51.440 --> 0:07:52.960 He got a job that no one in their right 0:07:53.000 --> 0:07:56.120 mind could possibly want but also could not refuse. And 0:07:56.240 --> 0:07:58.560 I feel like I sould a soul. Yes, absolutely, but 0:07:58.680 --> 0:08:00.480 he just he was the CEO of a double us. 0:08:00.800 --> 0:08:03.640 And he was he when he gave keynote talks, when 0:08:03.680 --> 0:08:07.160 he gave presentations, when I was fortunate enough to basically 0:08:07.200 --> 0:08:09.840 sneak my way into them, you could see the irrepressible 0:08:09.880 --> 0:08:12.800 humanity leaking out around the edges. He was a person. 0:08:13.440 --> 0:08:16.560 Now he is a figurehead. He doesn't want to customer, 0:08:16.640 --> 0:08:17.720 just to be clear to Congress. 0:08:18.440 --> 0:08:21.560 This shift was when he became CEO of Amazon or 0:08:21.640 --> 0:08:22.800 CEO of a WS. 0:08:22.920 --> 0:08:26.080 CEO of Amazon. AWS was his own Fifdom, and he 0:08:26.120 --> 0:08:30.280 could do whatever he wanted to my understanding within that company, 0:08:30.360 --> 0:08:33.080 it was the only people that mattered, Brother Garrett who 0:08:33.080 --> 0:08:35.280 had an opinion on this were Andy Jassey and God. 0:08:35.320 --> 0:08:37.240 And God was sort of absence, so great, We're just 0:08:37.240 --> 0:08:39.480 gonna leave it to Andy. Andy went deep and he 0:08:39.559 --> 0:08:42.320 understood every aspect of this stuff. Factly, a year ago 0:08:42.400 --> 0:08:44.320 he went to Reinvent in Las Vegas and gave a 0:08:44.360 --> 0:08:45.080 talk on stage. 0:08:45.200 --> 0:08:46.160 I think because he missed it. 0:08:46.200 --> 0:08:49.120 He missed being able to get up there and talk 0:08:49.200 --> 0:08:51.920 about computers with people because now he has to ship 0:08:51.960 --> 0:08:53.440 underpants all over the world instead. 0:08:54.080 --> 0:08:57.080 Right, But is he technical because he's an MBA? Is 0:08:57.080 --> 0:08:57.320 he not? 0:08:58.320 --> 0:09:00.400 This is one of the dangerous things that I have 0:09:00.480 --> 0:09:04.960 found people underestimate Amazon leadership on is it is never 0:09:05.000 --> 0:09:07.520 the smart bet to assume that they don't understand the 0:09:07.559 --> 0:09:10.640 technical nuances of the things that they are working with. 0:09:11.120 --> 0:09:14.959 They go so sarcastically deep in so many different ways 0:09:15.320 --> 0:09:17.440 that it is never a smart move to say that 0:09:17.440 --> 0:09:20.800 they don't understand it is unlike virtually any other tech 0:09:20.800 --> 0:09:23.320 company I've ever spoken with, other than small scale ones 0:09:23.320 --> 0:09:27.760 with technical founders, he very clearly understands the technology and 0:09:27.840 --> 0:09:29.040 used it to build things himself. 0:09:29.559 --> 0:09:34.400 Right, But here's the reason I question that. Please, he's 0:09:34.400 --> 0:09:38.600 talking about general IVAI. I've just pulled up his notice. 0:09:38.720 --> 0:09:40.440 Oh no, one understands how that works under the hood. 0:09:40.520 --> 0:09:42.920 Let's be very clear on that. And yeah, he lost 0:09:42.960 --> 0:09:44.880 the fucking plot when it came to gen AI. 0:09:45.040 --> 0:09:47.400 That's the thing, because he's like, we're also using GENERATIVEAI 0:09:47.480 --> 0:09:51.520 broadly across our internal operations in our fulfillment network. We're 0:09:51.600 --> 0:09:54.680 using AI to improve in ventur replacement. So not generate IVEAI. 0:09:54.880 --> 0:09:58.200 Demand forecasting definitely not generative AI. And the efficiency of 0:09:58.200 --> 0:10:00.760 our robots not generative AI. I don't know, maybe he 0:10:00.800 --> 0:10:02.560 does know what he's talking about. He's just a fucking 0:10:02.600 --> 0:10:04.040 carnival barker now he's. 0:10:03.920 --> 0:10:06.040 Doing to say that it all seems to be. And 0:10:06.080 --> 0:10:09.800 the trick to disprove everything you just said about that, 0:10:10.000 --> 0:10:12.000 rather everything that he said that you relate to us 0:10:12.120 --> 0:10:15.760 is very simple. Go out with an Amazon employee for 0:10:15.960 --> 0:10:19.600 several drinks and then ask them about their experience of 0:10:19.760 --> 0:10:23.760 using AI within the company. Oh, I have two different companies. 0:10:24.280 --> 0:10:28.120 There is no there is no congruity between some of 0:10:28.200 --> 0:10:29.239 those two perspectives. 0:10:29.280 --> 0:10:32.360 So Corey, maybe you can help me with this because you, certainly, 0:10:32.400 --> 0:10:35.360 without saying exactly what we're talking about, you knew about 0:10:35.360 --> 0:10:37.800 a contract that I knew about way before I knew, 0:10:37.840 --> 0:10:40.640 and I knew about it months early. Yes, with a 0:10:40.679 --> 0:10:44.320 large company we discussed, so you're very well connected. It 0:10:44.480 --> 0:10:48.480 seems up until jeneralf Ai, like AWS was at least 0:10:48.520 --> 0:10:52.600 a somewhat sensible operation, that it was something that we had, 0:10:53.240 --> 0:10:55.760 if not focus, at least a focus on scale and 0:10:56.040 --> 0:10:59.200 useful cloud computing. Yes, what the fuck is it about 0:10:59.240 --> 0:11:01.080 GPUs that's having everyone insane? 0:11:01.280 --> 0:11:03.800 Because I want to point something out that I think 0:11:03.880 --> 0:11:06.640 is being lost in the streetwide because I've talked to 0:11:06.679 --> 0:11:10.400 folks at AWS about this, and folks who have left AWS, 0:11:10.760 --> 0:11:13.160 who have I used to dabble working in data centers, 0:11:13.160 --> 0:11:14.920 which means I know just enough to know what I 0:11:14.960 --> 0:11:17.960 don't know, and I've talked to these folks. When Amazon 0:11:18.040 --> 0:11:20.959 commits to being able to bring certain amounts of capacity online, 0:11:21.000 --> 0:11:24.400 it is real capacity. It is each data center they 0:11:24.440 --> 0:11:26.240 build out, each region they build out, is a multi 0:11:26.240 --> 0:11:29.320 billion dollar investment, and it takes years to go from 0:11:29.480 --> 0:11:33.319 sign contract to its serving customer traffic. They are real 0:11:33.440 --> 0:11:36.120 when it comes to this stuff. There are a number 0:11:36.120 --> 0:11:38.920 of other companies building well, yes, and they agree that 0:11:39.080 --> 0:11:41.600 running it. They have operationalized how to run these things 0:11:41.640 --> 0:11:43.920 better than you or I would ever be able to do. 0:11:44.000 --> 0:11:47.960 That is their superpower. Other companies and they're spinning up 0:11:47.960 --> 0:11:49.959 a data center, do not do this. They are, well, 0:11:49.960 --> 0:11:52.560 we've we've rented a warehouse out, we round up, throwing 0:11:52.640 --> 0:11:55.319 some generators in the back and were check. Call Comcast 0:11:55.400 --> 0:11:56.720 and see what they can do to get it hooked 0:11:56.800 --> 0:12:00.319 up to the internet netney to go. Yeah, you say 0:12:00.320 --> 0:12:02.880 other companies, You don't mean Google and Microsoft here, No, 0:12:03.080 --> 0:12:06.400 I'm taught, well, I am talking Oracle, but yes, Microsoft 0:12:06.679 --> 0:12:11.240 kind of Microsoft. Microsoft is a shitty at building data centers. 0:12:11.280 --> 0:12:14.880 Azure has been a disappointment across the board. Capacity shortballs 0:12:14.880 --> 0:12:18.640 plagued them for a long time during COVID. They generally 0:12:18.640 --> 0:12:20.960 tend to view two racks in a data center somewhere 0:12:21.000 --> 0:12:24.439 as a region, which is weird and messed up. Google 0:12:24.600 --> 0:12:27.920 is real, but Google has made other interesting technical trade 0:12:27.960 --> 0:12:31.480 offs that I can see why they did it. But 0:12:31.520 --> 0:12:33.959 I would say when it comes to infrastructure, in my experience, 0:12:34.000 --> 0:12:35.200 AWS is number one. 0:12:35.520 --> 0:12:37.720 Google is a close second. I have heard some crazy 0:12:37.760 --> 0:12:40.600 shit about Google that they're doing these things called stalks 0:12:41.360 --> 0:12:47.920 where they just like like create these pop up data centers. 0:12:46.559 --> 0:12:48.079 For a long time. Every company does that for a 0:12:48.120 --> 0:12:49.240 point centering stuff. 0:12:49.320 --> 0:12:51.880 Yes, anyway, but back to the major question, which is 0:12:52.320 --> 0:12:56.199 AWS seems like a sensible business that does real things. 0:12:56.240 --> 0:12:57.160 I'm not disputing that. 0:12:57.640 --> 0:13:01.280 Yes, it prints money. Let's be clear on despite all, remember, 0:13:01.640 --> 0:13:05.719 I help companies negotiating money cloud contracts and understand what 0:13:05.760 --> 0:13:08.120 they're spending and where the money is going, which means 0:13:08.120 --> 0:13:11.240 that companies can say a lot. People say a lot 0:13:11.240 --> 0:13:13.280 of things. But one of my prime rules about this 0:13:13.320 --> 0:13:17.760 stuff is that customers lie mostly to themselves. They're not 0:13:17.880 --> 0:13:20.360 intentionally trying to mislead me when they tell me what's 0:13:20.400 --> 0:13:22.959 going on. But I find the bills of the ultimate 0:13:23.000 --> 0:13:24.920 source of truth, because if you're not paying for it 0:13:24.920 --> 0:13:28.040 doesn't really exist. And the AI spend in most of 0:13:28.080 --> 0:13:31.239 these companies hovers, in my experience with a few outliers 0:13:31.480 --> 0:13:36.280 between five and seven percent of their total infrastructure spend. Right, 0:13:36.440 --> 0:13:39.320 it is not three times. Companies that are doing a 0:13:39.360 --> 0:13:42.559 three hundred million dollars a year contract with AWS are 0:13:42.600 --> 0:13:45.120 not turning around saying better make it four hundred because 0:13:45.120 --> 0:13:47.679 of all the GENAI that is not happening. 0:13:58.520 --> 0:14:01.240 That's kind of my point though. It's like Amazon Web 0:14:01.280 --> 0:14:03.839 Services doesn't seem like a business that chases fats, or 0:14:03.840 --> 0:14:06.880 if it does chase them in a more sustainable way. 0:14:07.000 --> 0:14:07.800 They now they's credit. 0:14:07.840 --> 0:14:10.599 They docked they Dutch blockchain almost entirely, and I was 0:14:10.640 --> 0:14:12.320 worried they were going to fall down that rabbit hole. 0:14:12.760 --> 0:14:15.600 But I mean, they doing two hundred billion dollars in 0:14:15.679 --> 0:14:18.200 capex this year. What is it they see in AI? 0:14:18.240 --> 0:14:20.680 Because to your point, just now, what like seven to 0:14:20.720 --> 0:14:23.560 eight percent of a three hundred million contract, that's dog 0:14:23.640 --> 0:14:26.520 shit like that's that's not going to pipe out that 0:14:26.560 --> 0:14:27.760 Capex Sandy. 0:14:27.960 --> 0:14:31.040 I believe this is my belief here. I can't base 0:14:31.080 --> 0:14:33.200 it on anything other than what I've seen Amazon do, 0:14:33.840 --> 0:14:36.960 uh I, and what I've seen from customers. The Amazon 0:14:37.120 --> 0:14:40.760 is selling the capacity that they are spinning up. They 0:14:40.800 --> 0:14:43.000 are spinning up as fast as they can spin it. 0:14:43.400 --> 0:14:46.800 The question that I have is twofold one. Okay, you 0:14:47.080 --> 0:14:49.320 invest in all these AI data centers in the AI 0:14:49.400 --> 0:14:53.440 bubble pops to flates whatever, how repurposable are those facilities, 0:14:53.480 --> 0:14:56.600 and that capital expenditure if it's data centers and networking, Yes, 0:14:56.800 --> 0:14:59.840 very repurposable. Yeah, GPUs, well, we're all going to get 0:14:59.840 --> 0:15:01.840 real in online gaming in a few years. I don't 0:15:01.840 --> 0:15:02.400 except they. 0:15:02.320 --> 0:15:05.440 Don't have video out, they don't have DisplayPort. 0:15:05.600 --> 0:15:08.000 Yes, yeah, we still call them graphics, which I love. 0:15:08.240 --> 0:15:11.640 I think I like the people are I've also heard 0:15:11.640 --> 0:15:14.400 people say general purpose units, which is the opposite of 0:15:14.400 --> 0:15:16.640 what a GPU is like. It is very much not 0:15:16.760 --> 0:15:18.320 general purpose. It was. 0:15:18.320 --> 0:15:20.280 It became public through a press release that they have 0:15:20.360 --> 0:15:24.080 signed a thirty eight billion dollar contract with open AI 0:15:24.160 --> 0:15:28.560 with insane. There is zero doubt in my mind that, 0:15:28.760 --> 0:15:32.160 because I did some digging on this, is there line 0:15:32.200 --> 0:15:36.080 of sight for AWS to bring all on line thirty 0:15:36.080 --> 0:15:41.120 eight billion dollars of capacity to provide to open AI. Yes, yes, 0:15:41.160 --> 0:15:44.280 there is where my doubts come in. Is open AI 0:15:44.400 --> 0:15:45.280 good for the money? 0:15:45.680 --> 0:15:45.800 No? 0:15:45.880 --> 0:15:48.440 I have some questions here, mostly from reading your work. 0:15:48.800 --> 0:15:53.560 And that's the thing. It seems like something about GPUs 0:15:53.600 --> 0:15:56.560 and large language models have broken everyone. That's kind of 0:15:56.600 --> 0:15:58.720 the point I'm getting too, because like everything I've read 0:15:58.720 --> 0:16:01.120 about AWS, and I've read because I'm a dickhead, I 0:16:01.200 --> 0:16:04.280 went and I read all of the previous coverage of 0:16:04.320 --> 0:16:08.320 the lead up to AWS becoming profitable. Kevin Rust, by 0:16:08.360 --> 0:16:10.400 the way, literally a month before they did so, wrote 0:16:10.400 --> 0:16:12.800 an article saying it would be a while. What fucking 0:16:12.800 --> 0:16:16.960 three points in that ass wipe? Anyway, everything I read 0:16:17.040 --> 0:16:20.360 was everyone's saying, oh, Amazon's being really spendy. They're spending 0:16:20.560 --> 0:16:24.760 sixty seven billion in kapex a quarter. Oh no, and 0:16:24.760 --> 0:16:27.320 then the total capex of sixty nine billion. But everything 0:16:27.320 --> 0:16:29.560 from Amazon was like, we're building something sustainable, we see 0:16:29.560 --> 0:16:33.240 the revenue potential behind it. It was all very boring. Now 0:16:33.280 --> 0:16:35.800 with AWS, everything they talk about with large language models, 0:16:35.800 --> 0:16:38.720 it sounds like they've all been huffing ayawasco or something. 0:16:39.080 --> 0:16:41.360 They sound insane they. 0:16:41.200 --> 0:16:43.760 Do, which tells me one of two things is true 0:16:43.880 --> 0:16:47.600 or honestly, this is the real world. Unlike on Twitter, 0:16:48.080 --> 0:16:50.680 two complex competing things can in fact be true at 0:16:50.680 --> 0:16:53.320 the same interest time. One of them is they're still 0:16:53.320 --> 0:16:55.960 doing a lot of that boring stuff, but the publicity 0:16:56.080 --> 0:16:59.120 all accrues to the things that are far flung. This 0:16:59.200 --> 0:17:01.760 is also two hundred billion dollars across the entirety of 0:17:01.880 --> 0:17:06.560 Amazon that includes distribution centers, that includes Low Earth Orbits 0:17:06.920 --> 0:17:10.119 and basically Amazon Basic Starlink. I think they're calling it 0:17:10.200 --> 0:17:16.320 Leo now. It includes yes, their AI nonsense, their factory boondoggles, 0:17:16.359 --> 0:17:20.160 their robotic stuff, office buildouts, et cetera, et cetera, et cetera. 0:17:20.440 --> 0:17:22.800 They don't give clear breakdowns on how much of that 0:17:22.920 --> 0:17:25.320 is AI services, and frankly, I wouldn't trust them if 0:17:25.359 --> 0:17:27.879 they did not, since I saw a job ad a 0:17:27.880 --> 0:17:32.119 few years back on Amazon dot Jobs saying experience with 0:17:32.160 --> 0:17:38.080 other AI services like S three storage service, that's not AI. 0:17:38.480 --> 0:17:41.400 You can use it to feed AI. But my hairanthropic 0:17:41.760 --> 0:17:44.440 anthropics one of the largest S three customers I've heard. 0:17:44.960 --> 0:17:47.080 I've heard that. I've heard from seven different companies so 0:17:47.119 --> 0:17:49.560 far that they are the largest customer really three. Yeah, 0:17:49.560 --> 0:17:51.240 that is funny how that tends to work out. 0:17:51.320 --> 0:17:54.800 God, I miss I miss working in cloud software so much, 0:17:54.880 --> 0:17:58.280 where everyone is the everyone's the exclusive partner, everyone's the 0:17:58.359 --> 0:17:59.960 number one partner, and everyone's the larger. 0:18:00.600 --> 0:18:03.399 It's so but the qualifiers get They're like, we're the 0:18:03.480 --> 0:18:06.639 largest S three customers in Australia, great in the Northeast, 0:18:06.880 --> 0:18:09.440 in the Northeast region between the hours of six am 0:18:09.480 --> 0:18:11.880 and twelve am. 0:18:11.920 --> 0:18:14.959 Okay, So here's the thing with the capex. I'm sure 0:18:15.000 --> 0:18:18.320 a lot of it is for is for AI, just 0:18:18.359 --> 0:18:20.720 because if you look at their previously as of Capex, 0:18:20.720 --> 0:18:24.840 it's like four A fifty sixty Yeah, four A sixty 0:18:24.880 --> 0:18:28.200 one sixty three fifty two eighty three one hundred and 0:18:28.240 --> 0:18:29.520 thirty one two hundred billion. 0:18:30.560 --> 0:18:32.879 Gerald Over at Platform and Nomics has been tracking capec 0:18:32.920 --> 0:18:35.439 span to the hyperscalers for a long time now, and 0:18:35.800 --> 0:18:37.399 he was one of the first people to call out 0:18:37.520 --> 0:18:41.560 the cloud pretenders IBM and Oracle specifically, because they're making 0:18:41.600 --> 0:18:43.760 all the cloud noises, but they're not spending anything on 0:18:43.880 --> 0:18:45.800 Capex and you kind of have to build the data 0:18:45.800 --> 0:18:48.520 centers for this to work. Not to worry, Oracle's taking 0:18:48.560 --> 0:18:51.040 care of that. What I want to know, my question is, 0:18:51.040 --> 0:18:52.840 when you start, okay, we're going to spend even assume 0:18:52.880 --> 0:18:55.639 it's all AI whatever, fine, two hundred billion dollars on 0:18:55.720 --> 0:18:57.960 AI spent, how much of that is just in the 0:18:57.960 --> 0:18:59.399 form of a check written to Nvidio. 0:19:00.600 --> 0:19:06.320 And that's the question because from my understanding, my understanding 0:19:06.880 --> 0:19:10.280 the I should bring up the chart. I'm just gonna 0:19:10.320 --> 0:19:12.280 do some live on air. I just bringing up an 0:19:12.320 --> 0:19:15.919 image and of course can't find it. Nevertheless, there is 0:19:15.960 --> 0:19:19.320 an image I have the Amazon doesn't. Well, Amazon builds 0:19:19.320 --> 0:19:22.119 its own the trainium GPUs. And do they still use 0:19:22.160 --> 0:19:23.119 inferentia or. 0:19:23.119 --> 0:19:26.000 They That sounds like that that my grandpa got diagnosed with. 0:19:26.720 --> 0:19:29.040 Yeah, but do they liked inferentia? I thought was there? 0:19:29.080 --> 0:19:30.760 Oh, they talk about it a fair bit. They talk 0:19:30.800 --> 0:19:32.920 about trainium. And the thing is is what you don't 0:19:32.920 --> 0:19:35.840