WEBVTT - Part Three: NVIDIA Isn't Enron - So What Is It? 0:00:02.800 --> 0:00:03.159 Media. 0:00:04.840 --> 0:00:20.079 Hi, I'm ed Zitron, and welcome back to Better Offline 0:00:20.840 --> 0:00:23.000 And this is our third and final part of our 0:00:23.040 --> 0:00:26.480 Better Offline and Vidia special, where we're talking about well, 0:00:27.120 --> 0:00:30.240 the shakiness behind its growth and how the company, despite 0:00:30.280 --> 0:00:33.879 being on incredibly infirm ground, is definitely not Enron or 0:00:33.960 --> 0:00:37.120 Nortail or Wolcom or Lucien or any other dot com 0:00:37.120 --> 0:00:39.600 bubble era affirm that imploded under its own way and 0:00:39.760 --> 0:00:43.520 well quite dodgy accounting. The thing is, even if Enron 0:00:43.600 --> 0:00:46.920 is nothing like them, there's still quite a few causes 0:00:46.920 --> 0:00:49.919 for concern, and that's largely driven from the fact that 0:00:49.960 --> 0:00:53.080 in Vidio makes the majority of its money selling GPUs 0:00:53.120 --> 0:00:56.360 to a handful of customers, and so well, some of 0:00:56.400 --> 0:00:58.959 those also look to be on some of their own 0:00:58.960 --> 0:01:02.840 incredibly shaky ground. And yeah, I'm talking about Oracle now. 0:01:02.880 --> 0:01:05.880 In Vidia's health saying nothing of its growth isn't just 0:01:05.920 --> 0:01:08.559 tied to these customers. It's also tied to whether these 0:01:08.600 --> 0:01:11.360 customers can actually turn a profit from their capex spending. 0:01:12.040 --> 0:01:14.920 And even that's not even certain so due to the 0:01:14.920 --> 0:01:17.360 fact that so much money has been piled into building 0:01:17.400 --> 0:01:20.960 AI infrastructure and big tech has promised to spend hundreds 0:01:21.000 --> 0:01:23.600 of billions of dollars more in the next year. Big 0:01:23.640 --> 0:01:25.240 tech has found itself in a bit of a hole. 0:01:26.319 --> 0:01:28.240 How big of a hole? Well, by the end of 0:01:28.280 --> 0:01:31.280 the year, Microsoft, Amazon, Google, and Meta will have spent 0:01:31.319 --> 0:01:34.560 over four hundred billion dollars in capital expenditures, much of 0:01:34.600 --> 0:01:37.160 it focused on building AI infrastructure, on top of two 0:01:37.280 --> 0:01:40.120 hundred and twenty eight point four billion dollars in capex 0:01:40.160 --> 0:01:42.679 in twenty twenty four and around one hundred and forty 0:01:42.720 --> 0:01:46.120 eight billion in capital expenditures in twenty twenty three, for 0:01:46.160 --> 0:01:48.800 a total of seven hundred and seventy six billion in 0:01:48.840 --> 0:01:53.120 the space of three years, and they expect to spend 0:01:53.320 --> 0:01:56.480 more than four hundred billion dollars more in twenty twenty six. 0:01:56.520 --> 0:01:59.280 Every time I read these numbers, I feel a little crazy. 0:02:00.640 --> 0:02:03.400 As a result, based on my own analysis, big tech 0:02:03.440 --> 0:02:06.080 needs to make two trillion dollars in brand new, brand 0:02:06.160 --> 0:02:10.440 spanking new revenue, specifically from AI by twenty thirty. All 0:02:10.480 --> 0:02:13.160 of this was effectively for nothing. Now, I go into 0:02:13.320 --> 0:02:15.440 detail about this in the premium newsletter I did on 0:02:15.480 --> 0:02:17.880 October thirty first, but I'm going to give you a 0:02:17.919 --> 0:02:22.080 short explanation here. First, though, we have to talk about depreciation, 0:02:22.280 --> 0:02:24.480 and because I'm lazy, I'm going to quote myself in 0:02:24.480 --> 0:02:28.000 that newsletter I just mentioned a couple of seconds ago 0:02:28.280 --> 0:02:31.680 at hem. So when Microsoft buyers say one hundred million 0:02:31.720 --> 0:02:33.960 dollars worth of GPUs, it immediately comes out of its 0:02:34.000 --> 0:02:36.880 capital expenditures, which is when a company uses money to 0:02:36.960 --> 0:02:40.520 invest in either buying or upgrading something. It then adds 0:02:40.560 --> 0:02:43.760 to its property, plants and Equipment assets PPE for sure, 0:02:44.080 --> 0:02:46.680 although some companies list this on their annual and quarterly 0:02:46.720 --> 0:02:50.360 financials as property and equipment PPE sits on the balance sheet, 0:02:50.440 --> 0:02:52.840 it's an asset as it's stuff for the company that 0:02:52.919 --> 0:02:56.880 it owns or as least, GPUs depreciate, meaning they lose 0:02:56.960 --> 0:03:00.120 value or over time, and this depreciation is represented on 0:03:00.120 --> 0:03:03.119 a balance sheet and the income statement. Essentially, the goal 0:03:03.200 --> 0:03:05.200 is to represent the value of an asset that a 0:03:05.240 --> 0:03:09.080 company has on the income statement, and we see how 0:03:09.160 --> 0:03:11.840 much the assets have declined during the reporting period, whether 0:03:11.880 --> 0:03:13.760 that be a year or a quarter or something else, 0:03:13.880 --> 0:03:16.799 Whereas the balance sheet shows the cumulative depreciation of every 0:03:16.840 --> 0:03:20.280 asset currently in play. Depreciation does two things. And I 0:03:20.280 --> 0:03:22.240 know this sounds like a lot, but I'll break it 0:03:22.280 --> 0:03:24.880 down for you First, it allows a company to accurately 0:03:25.200 --> 0:03:27.840 to an extent, represent the value of things it owns 0:03:27.880 --> 0:03:30.920 over their useful life. Secondly, it allows a company to 0:03:31.000 --> 0:03:34.040 deduct the value of an asset across said useful life 0:03:34.360 --> 0:03:37.240 right up until its eventual removal, versus having to take 0:03:37.280 --> 0:03:40.320 a big hit up front. The way this depreciation is 0:03:40.320 --> 0:03:43.520 actually calculated can vary. There are several different methods available, 0:03:43.640 --> 0:03:46.160 with some allowing for greater deductions at the start of 0:03:46.200 --> 0:03:48.240 the term, which is useful for those items that will 0:03:48.240 --> 0:03:50.960 experience the biggest drop in value right after buying them 0:03:51.240 --> 0:03:54.040 and their initial use. An example you're probably familiar with 0:03:54.160 --> 0:03:56.400 is a new car which loses a significant chunk of 0:03:56.400 --> 0:03:58.520 its value the moment is driven off a dealership. Long 0:04:00.000 --> 0:04:03.160 creation has become a big ugly problem with GPUs specifically 0:04:03.200 --> 0:04:05.800 because of that useful life to find either as how 0:04:05.800 --> 0:04:07.560 long the thing is able to be run before it 0:04:07.640 --> 0:04:11.160 dies or how long before it becomes obsolete, and nobody 0:04:11.160 --> 0:04:12.920 seems to be able to come up with a consensus 0:04:12.920 --> 0:04:15.680 about how long this should be. In Microsoft's case, the 0:04:15.680 --> 0:04:18.560 appreciation for its service is spread over six years, a 0:04:18.600 --> 0:04:21.200 convenient change it made in August twenty twenty two, A 0:04:21.200 --> 0:04:24.239 few months before the launch of chat GPT and before 0:04:24.240 --> 0:04:26.719 it bought a bunch of fucking GPUs. This means that 0:04:26.760 --> 0:04:28.919 Microsoft can spread the cost of tens of thousands of 0:04:28.920 --> 0:04:31.120 a one hundred GPUs brought in twenty twenty or the 0:04:31.400 --> 0:04:33.640 four hundred and fifty thousand, h one hundred gupus it 0:04:33.680 --> 0:04:36.360 bought in twenty twenty four across six years, regardless of 0:04:36.400 --> 0:04:39.800 whether those are the years they'll be generating revenue or 0:04:40.120 --> 0:04:44.320 naturally functioning. Corwy for what It's worth says the same thing, 0:04:44.640 --> 0:04:46.560 but largely because it's betting that it'll still be able 0:04:46.600 --> 0:04:49.600 to find users for older silicon after its initial contracts 0:04:49.600 --> 0:04:52.599 with the companies like OpenAI expire. The problem is is 0:04:52.600 --> 0:04:56.039 that aigpus are fairly new concepts, and thus all of 0:04:56.080 --> 0:05:00.159 this is pretty much untested ground. Whereas we know how long, 0:05:00.279 --> 0:05:02.440 say a truck or a piece of heavy machinery can 0:05:02.520 --> 0:05:05.080 last and how long it can deliver value to an organization, 0:05:05.400 --> 0:05:07.120 we don't know the same thing about the kind of 0:05:07.200 --> 0:05:10.160 data center GPUs that hyperscalers are spending tens of billions 0:05:10.200 --> 0:05:13.520 of dollars on each year. Any kind of depreciation schedule 0:05:13.560 --> 0:05:17.039 is based on at best assumptions and at worst hope. 0:05:17.160 --> 0:05:20.800 Now this is important. The concept of an AI data 0:05:20.839 --> 0:05:24.000 center is super new. We maybe saw the first ones 0:05:24.040 --> 0:05:29.080 in twenty ninety. In question, it's kind of hard to say, 0:05:29.120 --> 0:05:31.719 but even at the scale we're seeing today a Gigawa 0:05:31.880 --> 0:05:35.200 data center pretty much brand new, maybe a couple years old. 0:05:35.240 --> 0:05:37.320 I don't even think they've even built any but we'll 0:05:37.360 --> 0:05:39.560 get to that in a bit. There are a lot 0:05:39.600 --> 0:05:42.560 of assumptions at play. There's the assumption that the cards 0:05:42.600 --> 0:05:45.040 won't degrade with heavy usage, or the assumption that future 0:05:45.040 --> 0:05:47.640 generations of GPUs won't be so powerful and impressive that 0:05:47.680 --> 0:05:50.800 they'll render the previous ones more obsolete than expected, kind 0:05:50.800 --> 0:05:52.760 of like how the first jet powered planes of the 0:05:52.839 --> 0:05:56.040 nineteen fifties did to those manufactured just a decade prior. 0:05:56.560 --> 0:05:59.400 The assumption that there will be in fact a market 0:05:59.400 --> 0:06:01.160 for Alder cards, than that there'll be a way to 0:06:01.240 --> 0:06:06.320 lease them profitably. What if those assumptions are I don't know, wrong. 0:06:06.760 --> 0:06:09.560 What if that hope is ultimately irrational. So there's a 0:06:09.640 --> 0:06:13.279 quote from the Center for Information Technology Policy framing this problem, well, 0:06:13.279 --> 0:06:15.880 that'll link to in the notes. Here is the puzzle. 0:06:16.240 --> 0:06:18.160 The chips at the heart of the infrastructure build out 0:06:18.240 --> 0:06:20.400 have a useful lifespan of one to three years due 0:06:20.400 --> 0:06:24.320 to rapid technological obsolescence and physical wear, but companies appreciate 0:06:24.400 --> 0:06:27.040 them over five or six years. In other words, they 0:06:27.040 --> 0:06:29.880 spread out the cost of their massive capital expenditures over 0:06:29.880 --> 0:06:32.599 a longer period than the facts warrant what the economist 0:06:32.720 --> 0:06:35.680 is referred to as the four trillion dollar accounting puzzle 0:06:35.720 --> 0:06:38.520 at the heart of the AIICLOUD. This is why Michael 0:06:38.560 --> 0:06:41.279 Burry brought it up recently because spreading out these costs 0:06:41.320 --> 0:06:43.880 allowed spigtech to make their net income i e. Their 0:06:43.880 --> 0:06:46.840 profits look better in simple terms. By spreading out the 0:06:46.880 --> 0:06:49.680 costs over six years rather than three, hyperscalers are able 0:06:49.720 --> 0:06:51.840 to reduce the line item that eats into their earnings, 0:06:51.880 --> 0:06:55.039 which makes their companies look better to the markets. So 0:06:55.120 --> 0:06:58.040 why does this create an artificial time limit? Well, let's 0:06:58.080 --> 0:07:00.520 start with a horrible fact. It takes it's two point 0:07:00.600 --> 0:07:03.080 five years of construction time in about fifty billion dollars 0:07:03.120 --> 0:07:06.719 per gigawa of data center capacity. No matter when the 0:07:06.760 --> 0:07:09.600 GPUs for a giga what data center are bought one 0:07:09.600 --> 0:07:12.440 way or another, these GPUs are depreciating in value either 0:07:12.480 --> 0:07:15.200 through death or reduced efficacy through wear and tear, or 0:07:15.200 --> 0:07:17.720 becoming obsolete, which is very likely as in Video is 0:07:17.720 --> 0:07:21.440 committed to releasing a new GPU every single year. Newer 0:07:21.480 --> 0:07:25.120 generation GPUs, like in videos Blackwell and Verra Reuben require 0:07:25.280 --> 0:07:28.160 entirely new data center architecture, meaning that one as to 0:07:28.160 --> 0:07:30.360 why they build a brand new data center a retrofit 0:07:30.400 --> 0:07:32.760 an old one. Essentially, we have facilities that are being 0:07:32.760 --> 0:07:35.960 built around a GPU design or product that may change 0:07:36.000 --> 0:07:38.240 in a year or two. Now I hear that in 0:07:38.600 --> 0:07:42.080 the Oberon racks that they use for the Blackwells will 0:07:42.120 --> 0:07:44.840 be used with some Verra Rubin. But even then there's 0:07:44.880 --> 0:07:47.440 going to be an even bigger, more huger Vera Ruben 0:07:47.480 --> 0:07:49.880 that comes that might even that I read somewhere that 0:07:49.920 --> 0:07:52.040 there might even be like killer what level ones just 0:07:52.880 --> 0:07:56.800 in like one hundred killer what ones that this company's insane. Nevertheless, 0:07:56.920 --> 0:07:58.640 at some point Wall Street is going to need to 0:07:58.680 --> 0:08:00.920 see some sort of return on this investment, and right 0:08:00.960 --> 0:08:03.640 now that return is negative dollars. I break it down 0:08:03.680 --> 0:08:06.760 on my October thirty first premium piece, but for your sake, 0:08:06.800 --> 0:08:09.240 I'll just say it. I estimate the big techniques to 0:08:09.280 --> 0:08:11.480 make two dollars for every dollar of CAPEX they've spent, 0:08:11.720 --> 0:08:15.120 and this revenue must be new, brand new, As this 0:08:15.160 --> 0:08:18.320 CAPEX is only for AI. This CAPEX is useless for 0:08:18.360 --> 0:08:21.200 everything else. It does not help it. And no, it 0:08:21.240 --> 0:08:24.360 doesn't help that they bolted copilot onto fucking everything that 0:08:24.520 --> 0:08:28.120 is not working, and in fact, the Australian Competition Commission 0:08:28.160 --> 0:08:31.840 is suing them. Maybe I mentioned that later, but whatever, Meta, Amazon, 0:08:31.880 --> 0:08:34.520 Google and Microsoft are already years and hundreds of billions 0:08:34.559 --> 0:08:36.520 of dollars in and are yet to see a dollar 0:08:36.559 --> 0:08:39.440 of profit, creating a one point two to one trillion 0:08:39.559 --> 0:08:42.760 dollar hole just to justify the expenses. So around six 0:08:42.840 --> 0:08:44.920 hundred and five billion dollars of CAPEX all told at 0:08:44.920 --> 0:08:47.640 the time I calculated it. Much of this CAPEX has 0:08:47.640 --> 0:08:49.640 been committed or spent before they've even turned on a 0:08:49.720 --> 0:08:53.360 single goddamn GPU. You might argue that there's a scenario 0:08:53.480 --> 0:08:56.920 here where say, an A one hundred GPU is useful 0:08:56.920 --> 0:08:59.439 past the three or six year shelf life. Even if 0:08:59.440 --> 0:09:01.280 that were the case, the average rental price of an 0:09:01.280 --> 0:09:03.839 A one hundred is ninety nine cents an hour. This 0:09:03.920 --> 0:09:06.400 is a four or five year old GPU, and customers 0:09:06.440 --> 0:09:08.120 are paying for it like they would a five year 0:09:08.120 --> 0:09:10.680 old piece of hardware. The same fate awaits the H 0:09:10.720 --> 0:09:13.439 one hundred, which was released in twenty twenty two but 0:09:13.520 --> 0:09:15.960 was still sold in great volume through twenty twenty four, 0:09:16.160 --> 0:09:18.000 and I hear the H two hundred of the same 0:09:18.080 --> 0:09:21.959 generation is still selling to this day. Every year in 0:09:22.080 --> 0:09:24.679 Vidia releases a new GPU, lowering the value of all 0:09:24.720 --> 0:09:27.240 the other GPUs in the process, making it harder to 0:09:27.280 --> 0:09:29.520 fill in the holes created by all the other GPUs 0:09:29.600 --> 0:09:33.200 is capex and costs. This whole time, nobody appears to 0:09:33.240 --> 0:09:34.880 have found a way to make a profit, meaning that 0:09:34.880 --> 0:09:37.920 the hole created by these GPUs remains unfilled, all while 0:09:37.960 --> 0:09:43.600 big tech firms by more GPUs, creating more holes to fill. 0:09:43.760 --> 0:09:46.160 So now that you know this, there's a fairly obvious 0:09:46.320 --> 0:09:48.880 question to ask, why in the hell are they still 0:09:48.920 --> 0:09:52.560 buying GPUs? Well, so, where the fuck are these GPUs going? 0:10:05.800 --> 0:10:07.800 So a few weeks ago I wrote a piece Premium 0:10:07.840 --> 0:10:09.600 one called the Hater's Guide to Nvidia, and I asked 0:10:09.600 --> 0:10:12.319 the basic question in there, where have all the GPUs 0:10:12.360 --> 0:10:15.000 that Invidia has sold actually gone? In particular, the six 0:10:15.080 --> 0:10:18.160 million Blackwell GPUs that Jensen one keeps banging on about 0:10:18.600 --> 0:10:20.839 Now there's little evidence that these are being used in 0:10:20.880 --> 0:10:23.360 the volume which they're sold, suggesting that they're either languishing 0:10:23.360 --> 0:10:26.079 in the supply chain or being warehoused by hyperscalers, or 0:10:26.120 --> 0:10:28.960 even in Nvidia themselves. Now there's the argument that this 0:10:29.040 --> 0:10:32.800 could be and this is wanky Nvidian bullshit, this could 0:10:32.800 --> 0:10:36.320 actually be two GPUs per GPU sold because there's two 0:10:36.800 --> 0:10:39.920 chips on each GPU. Even if that was the case, 0:10:40.280 --> 0:10:44.520 three billion GPUs Blackwell, specifically the brand new ones, they're 0:10:44.559 --> 0:10:45.320 not in service. 0:10:45.960 --> 0:10:46.080 Now. 0:10:46.120 --> 0:10:47.839 While I'm not going to go and copy paste an 0:10:48.000 --> 0:10:50.840 entire premium piece into this script, I am, however, going 0:10:50.880 --> 0:10:52.960 to go into detail about what I found. And the 0:10:53.040 --> 0:10:56.280 truth is, I can only really see in this inclodes, 0:10:56.320 --> 0:11:00.800 looking over like bunches of data center maps, reading hundreds 0:11:00.800 --> 0:11:04.319 of press releases, documents, earning statements. I've only been able 0:11:04.400 --> 0:11:09.439 to find maybe a couple hundred thousand Blackwell GPUs in existence, 0:11:09.520 --> 0:11:12.840 maybe half a million to seven hundred and fifty thousand 0:11:12.840 --> 0:11:15.200 if you include the stuff that hasn't even been built yet. 0:11:15.640 --> 0:11:18.840 Let's go into it. So Stargate Apilene allegedly four hundred 0:11:18.880 --> 0:11:22.600 thousand Blackweld gps A, going there now Oracle CEO Coco, 0:11:22.679 --> 0:11:26.200 I should say, Clay McGurk, that's probably not how you 0:11:26.240 --> 0:11:29.400 say that. He claimed very recently there were ninety six 0:11:29.440 --> 0:11:33.120 thousand of them in stored, so not great. There's theoretically 0:11:33.160 --> 0:11:35.760 in one hundred and thirty one thousand Blackwell GPU cluster 0:11:35.840 --> 0:11:38.360 owned by Oracle that they announced in March twenty twenty five, 0:11:39.360 --> 0:11:42.960 so that should be online. Never five thousand blackweld GPUs 0:11:42.960 --> 0:11:46.200 at the University of Texas Austin, which sound like they're online, 0:11:46.280 --> 0:11:49.199 more than fifteen hundred in a Lambda data center in Columbus, Ohio. 0:11:49.280 --> 0:11:52.319 Those are online. The Department of Energy is still in 0:11:52.360 --> 0:11:56.280 development one hundred thousand GPU supercluster as well as ten 0:11:56.320 --> 0:11:58.840 thousand in Vidia blackweld GPUs that are expected to be 0:11:58.880 --> 0:12:02.680 available in twenty twenty six, and it's Equinox cluster. Really 0:12:02.720 --> 0:12:05.160 can't establish how many of those are actually in operation. 0:12:05.760 --> 0:12:08.880 Fifty thousand of these blackweld GPUs going into the still 0:12:08.960 --> 0:12:13.520 unbuilt musk Run Colossus to supercluster, Corby's largest GB two 0:12:13.640 --> 0:12:16.520 hundred Blackwell cluster of two four hundred and ninety six 0:12:16.559 --> 0:12:21.280 Blackwell GPUs, tens of thousands of them deployed globally by Microsoft, 0:12:21.280 --> 0:12:25.120 including forty six hundred Blackwell Ultra GPUs and two hundred 0:12:25.160 --> 0:12:28.120 and sixty thousand of them, these Blackwell GPUs going into 0:12:28.120 --> 0:12:32.000 five AI data centers for the South Korean government, and yeah, 0:12:32.080 --> 0:12:33.640 I just want to be clear that that is also 0:12:33.679 --> 0:12:37.840 fairly recently announced, so probably not even not even built, 0:12:38.440 --> 0:12:41.600 let alone powered on. I goan to be honest, I'm 0:12:41.679 --> 0:12:46.680 genuinely unable to find one million Blackwell GPUs like inexistence. Now, 0:12:46.960 --> 0:12:48.880 some of you might say, oh, there's a bunch of 0:12:48.960 --> 0:12:51.800 secret ones. There's a bunch of them. They don't announce 0:12:51.840 --> 0:12:54.600 every single one. Here's the thing. Three million of these 0:12:54.640 --> 0:12:57.920 fucking things have allegedly been shipped. I can't find a 0:12:57.960 --> 0:13:01.360 million of them. And considering everybody always talks about their 0:13:01.400 --> 0:13:05.480 GPU purchases, I'm kind of shocked at calm. Now. I 0:13:05.600 --> 0:13:07.960 do not know where these six million black Weld GPUs 0:13:08.000 --> 0:13:10.200 have gone, but they certainly haven't gone into data centers 0:13:10.200 --> 0:13:12.480 that are powered and turned on. In fact, power has 0:13:12.520 --> 0:13:14.640 become one of the biggest issues of building these things, 0:13:14.760 --> 0:13:16.679 and the fact it's really difficult and maybe impossible to 0:13:16.679 --> 0:13:18.360 get the amount of power these things need to the 0:13:18.400 --> 0:13:22.720 goddamn data centers. In really simple terms, there isn't enough 0:13:22.800 --> 0:13:25.400 power or built data centers for those black world GPUs 0:13:25.400 --> 0:13:27.680 to run, in part because the data centers aren't built, 0:13:27.679 --> 0:13:29.560 and in part because there isn't enough power for the 0:13:29.559 --> 0:13:33.240 ones that are. Microsoft CEO Sacha Nadella recently said in 0:13:33.280 --> 0:13:36.160 a podcast that his company and I quote didn't have 0:13:36.240 --> 0:13:39.679 the warm shells to plug into, meaning buildings with sufficient power, 0:13:40.400 --> 0:13:43.320 and heavily suggested that Microsoft may actually have a bunch 0:13:43.360 --> 0:13:47.800 of chips sitting in inventory that they couldn't plug in now. 0:13:48.120 --> 0:13:50.160 Just to give you an estimate here, even if we 0:13:50.200 --> 0:13:53.839 say three million GPUs, even if we're going with the moonmath, 0:13:53.880 --> 0:13:56.640 the Vinvidia, if we're going into the make believe world, 0:13:56.679 --> 0:14:01.120 the twisted mind of Jensen Wang, still three million GPUs. 0:14:01.160 --> 0:14:04.120 We'll look at still like five or six gigawatts of capacity. 0:14:04.679 --> 0:14:06.800 It's not being built. I don't even think two gigawats 0:14:06.840 --> 0:14:08.880 of data center capacity have been built. And I swear 0:14:08.920 --> 0:14:10.840 to fucking God, if one of you emails me and said, 0:14:10.880 --> 0:14:14.240 and Eric is built, and twenty gigawatts or something, power 0:14:14.240 --> 0:14:17.160 can get built. Power can get built. You can build power, 0:14:17.600 --> 0:14:20.480 getting it to the data center and actually powering the 0:14:20.560 --> 0:14:24.960 data center correctly, as in things turn on, everything works, 0:14:25.200 --> 0:14:29.160 nothing overloads, nothing blacks out, and the power is consistently done. 0:14:29.400 --> 0:14:32.960 Takes what's just months of surveys and scientific stuff and 0:14:33.000 --> 0:14:35.760 then years to just get it done. Stargate Appleine only 0:14:35.760 --> 0:14:38.600 has two hundred megawatts. They're gonna need over one point 0:14:38.600 --> 0:14:41.360 four gigawats just to turn the fucking thing on. I'm 0:14:41.400 --> 0:14:45.320 so tired of He's God damn GPUs bo. With all 0:14:45.360 --> 0:14:48.760 this said, why pray tell? Is Jensen Huang of Nvidia 0:14:48.960 --> 0:14:51.960 saying that he has twenty million Blackwell and Vera Rubin 0:14:52.040 --> 0:14:54.600 GPUs ordered through the end of twenty twenty six. Where 0:14:54.600 --> 0:14:57.480 are they fucking going? Jensen? Now, I think that number 0:14:57.480 --> 0:15:00.680 also includes the six million. And also, to be clear, 0:15:01.520 --> 0:15:03.600 I know a lot of you aren't technical, which is awesome. 0:15:03.640 --> 0:15:06.000 I love. I want you all to know about this. 0:15:06.400 --> 0:15:08.160 You need to know that this is part of the 0:15:08.200 --> 0:15:12.680 course within Nvidia in video loves schmushing accountancy things together 0:15:12.720 --> 0:15:15.360 and coming up with random numbers. Credit the case could 0:15:15.360 --> 0:15:17.240 go our friend of the show for telling me the story. 0:15:17.280 --> 0:15:20.880 But during the early twenty twenties, so that there is 0:15:20.960 --> 0:15:24.800 twenty twenty two during the Big Crypto rush. In Vidia 0:15:24.880 --> 0:15:28.080 classified gaming GPUs that were sold to bitcoin miners as 0:15:28.080 --> 0:15:31.160 gaming revenue. They got digged by the SEC. Wasn't fraud, 0:15:31.760 --> 0:15:34.320 but just so you know, in video will move shit around. 0:15:34.600 --> 0:15:37.120 And I truly do not know where these GPUs are. 0:15:37.280 --> 0:15:41.200 I do not know even why anyone is still buying GPUs. 0:15:41.600 --> 0:15:41.760 Now. 0:15:41.840 --> 0:15:45.160 AI bulls will tell you that there's this insatiable demand 0:15:45.160 --> 0:15:47.760 for AI and these massive amounts of orders are proof 0:15:47.800 --> 0:15:51.720 of something or rather, and you know what, I'll give 0:15:51.760 --> 0:15:53.920 them that it's proof that people are buying a lot 0:15:53.960 --> 0:15:57.400 of GPUs. I just don't know why nobody has made 0:15:57.440 --> 0:15:59.880 a profit from AI, and those making revenue aren't really 0:16:00.080 --> 0:16:02.640 making that much. Let me give you an example. My 0:16:02.760 --> 0:16:05.360 reporting on open ai from November twelfth suggests that the 0:16:05.360 --> 0:16:07.640 company I only made four point three two nine billion 0:16:07.720 --> 0:16:10.720 dollars in revenue for the end of September, extrapolated from 0:16:10.760 --> 0:16:13.520 the twenty percent revenue share that Microsoft receives in the company. 0:16:14.080 --> 0:16:18.000 And now some people who write really shit our substacks 0:16:18.400 --> 0:16:20.840 have argued with the figures, claiming that they're either delayed 0:16:21.320 --> 0:16:23.240 or are not inclusive of the revenue that open Ai 0:16:23.440 --> 0:16:26.520 is paid from Microsoft as part of being's ai integration 0:16:26.640 --> 0:16:29.360 and sales of open AI's models throughout Microsoft as ure. 0:16:29.600 --> 0:16:31.000 So I want to be clear of two things. Some 0:16:31.160 --> 0:16:35.120 a deeply bitter person. This is a crual accounting, meaning 0:16:35.160 --> 0:16:37.720 that these numbers are revenue booked in the quarter I 0:16:37.800 --> 0:16:42.119 reported them. Any comments about quarter long delays or naive approaches, 0:16:42.160 --> 0:16:44.400 and you know who I'm fucking talking about, if you're listening, 0:16:45.040 --> 0:16:49.560 are incorrect and a riboso. Also, Microsoft's revenue share payments 0:16:49.560 --> 0:16:54.080 to open ai kind of pathetic, totally based on documents 0:16:54.120 --> 0:16:57.680 reviewed by this newsletter publication whatever you call me, media 0:16:57.880 --> 0:17:01.400 entity floating blob in the podcast, for sixty nine point 0:17:01.440 --> 0:17:04.080 one million dollars in counting the year Q three, twenty 0:17:04.119 --> 0:17:07.840 twenty five. And by the way, the actual number for 0:17:08.040 --> 0:17:11.120 that three month period, including all royalties, is about four 0:17:11.200 --> 0:17:15.439 point five two seven billion dollars of revenue. I just 0:17:15.480 --> 0:17:18.080 want to be clear about something with open Ai. I'm 0:17:18.080 --> 0:17:21.080 not saying they're misrepresenting their numbers to anyone. I hope 0:17:21.080 --> 0:17:23.960 that open ai is being honest with their revenues. But 0:17:24.040 --> 0:17:26.400 if it comes out, I'm right, if it comes out 0:17:26.440 --> 0:17:28.679 that it turns out that they've been telling investors completely 0:17:28.680 --> 0:17:33.119 different numbers, I'm gonna be absolutely fucking insufferable. I'm going 0:17:33.200 --> 0:17:35.600 to bring in I'm going to be playing Tommy Trumpets 0:17:35.600 --> 0:17:38.160 a walk around cheering that you can get five minutes 0:17:38.200 --> 0:17:41.200 of monologue about that. Also in the same period, Open 0:17:41.240 --> 0:17:43.760 Ai spent eight point six seven billion dollars on inference, 0:17:43.760 --> 0:17:45.680 which is the process in which an l them creates 0:17:45.720 --> 0:17:49.240 its output. This is the biggest company in the generative 0:17:49.280 --> 0:17:52.080 AI space, with eight hundred million weekly active users in 0:17:52.119 --> 0:17:54.919 the Mandate of Heaven in the eyes of the media, Anthropic, 0:17:54.960 --> 0:17:58.160 it's largest competitor, Allegacy, will make eight hundred and thirty 0:17:58.160 --> 0:18:00.760 three million dollars in revenue in December tween twenty five, 0:18:00.960 --> 0:18:03.000 and based on my estimates, will end up having about 0:18:03.000 --> 0:18:05.119 four and a half to five billion dollars if revenue 0:18:05.160 --> 0:18:07.520 by the end of the year. Based on my reporting 0:18:07.560 --> 0:18:10.640 from October Andthropic spent two point six six billion dollars 0:18:10.680 --> 0:18:13.400 on Amazon Web Services through the end of September, meaning 0:18:13.440 --> 0:18:15.959 that it, based on my own analysis of reported revenues, 0:18:16.000 --> 0:18:18.720 spent one hundred and four percent of its revenue up 0:18:18.760 --> 0:18:21.679 to that point just on AWS likely spent as much 0:18:21.720 --> 0:18:25.159 on Google Cloud. Now, the reason I'm bringing up these 0:18:25.240 --> 0:18:28.760 numbers is these are the champions, the champions of the 0:18:28.840 --> 0:18:33.920 AI boom, yet their revenues kind of fucking stink. Wow. 0:18:34.320 --> 0:18:36.800 Even if open ai made thirteen billion dollars this year, 0:18:36.840 --> 0:18:40.040 even if Anthropic made five billion dollars, okay, wow, so 0:18:40.119 --> 0:18:45.240 that's not even twenty billion dollars. That's like nineteen billion 0:18:45.320 --> 0:18:48.679 dollars less than Microsoft spent on GPU's and other capex 0:18:48.880 --> 0:18:53.199 in the last quarter. That's dogshit. I'm sorry, I'm just 0:18:53.280 --> 0:18:56.640 tired of I am tired of humoring this. I'm sure 0:18:56.680 --> 0:18:59.159 all of you are too. I find it loathsome that 0:18:59.200 --> 0:19:02.440 we have to pretend these people are gifted somehow, they 0:19:02.480 --> 0:19:05.399 have shit. Ask businesses that burn billions of dollars, and 0:19:05.440 --> 0:19:08.840 you know what. Another thing I'm tired about is everybody 0:19:08.880 --> 0:19:13.600 telling this story about Anthropic being more efficient and only 0:19:13.640 --> 0:19:16.920 burning two point eight billion dollars this year. Now one 0:19:16.960 --> 0:19:19.440 has to ask a question about why this company that's 0:19:19.480 --> 0:19:23.480 allegedly reducing costs had to raise thirteen billion dollars in 0:19:23.520 --> 0:19:26.919 September twenty twenty five, after raising three point five billion 0:19:26.960 --> 0:19:30.080 dollars in March twenty twenty five, after raising four billion 0:19:30.119 --> 0:19:33.320 dollars in November twenty twenty four. Am I really meant 0:19:33.320 --> 0:19:36.480 to read stories about Anthropic hitting break even in twenty 0:19:36.480 --> 0:19:39.600 twenty eight with a straight face, especially as other stories 0:19:39.640 --> 0:19:42.119 say that we cash flow positive as soon as twenty 0:19:42.160 --> 0:19:45.640 twenty seven. This company's as big a pile of shit 0:19:45.720 --> 0:19:48.840 as open Ai. Open Ai raised eighteen point three billion 0:19:48.880 --> 0:19:51.960 dollars this year. That's less than two billion dollars more 0:19:52.000 --> 0:19:54.960 than Anthropic, who makes a bunch less revenue. Can't believe 0:19:54.960 --> 0:20:00.000 I'm defending open Ai. But these companies are the two 0:20:00.000 --> 0:20:02.240 two largest ones in the generative of AI space, and 0:20:02.280 --> 0:20:06.160 by extension, the two largest consumers of GPU COMPUW Both 0:20:06.160 --> 0:20:08.960 companies burn billions of dollars and require an infinite amount 0:20:09.000 --> 0:20:11.120 of venture capital to keep them alive. At a time 0:20:11.160 --> 0:20:14.119 when the Saudi Public Investment Fund is struggling and the 0:20:14.200 --> 0:20:16.720 US venture capital system is set to run out of 0:20:16.720 --> 0:20:18.880 cash in the next year and a half, The two 0:20:19.000 --> 0:20:22.080 largest sources of actual revenue for selling AI compute are 0:20:22.119 --> 0:20:25.320 subsidized by venture capital and debt. What happens if these 0:20:25.320 --> 0:20:28.240 sources dry up. They're not paying out of cash flow, 0:20:28.600 --> 0:20:32.280 and in all seriousness, who else is buying AI compute? 0:20:32.440 --> 0:20:35.639 What are they doing with them? Hyperscalers other than Microsoft, 0:20:35.680 --> 0:20:37.919 which chose to stop reporting its AI revenue back in 0:20:37.960 --> 0:20:40.200 January when it claimed it made about a billion dollars 0:20:40.200 --> 0:20:43.720 a month in revenue, don't disclose anything about their AI revenue, 0:20:43.840 --> 0:20:45.840