1 00:00:02,800 --> 00:00:03,159 Speaker 1: Media. 2 00:00:04,840 --> 00:00:20,079 Speaker 2: Hi, I'm ed Zitron, and welcome back to Better Offline 3 00:00:20,840 --> 00:00:23,000 Speaker 2: And this is our third and final part of our 4 00:00:23,040 --> 00:00:26,480 Speaker 2: Better Offline and Vidia special, where we're talking about well, 5 00:00:27,120 --> 00:00:30,240 Speaker 2: the shakiness behind its growth and how the company, despite 6 00:00:30,280 --> 00:00:33,879 Speaker 2: being on incredibly infirm ground, is definitely not Enron or 7 00:00:33,960 --> 00:00:37,120 Speaker 2: Nortail or Wolcom or Lucien or any other dot com 8 00:00:37,120 --> 00:00:39,600 Speaker 2: bubble era affirm that imploded under its own way and 9 00:00:39,760 --> 00:00:43,520 Speaker 2: well quite dodgy accounting. The thing is, even if Enron 10 00:00:43,600 --> 00:00:46,920 Speaker 2: is nothing like them, there's still quite a few causes 11 00:00:46,920 --> 00:00:49,919 Speaker 2: for concern, and that's largely driven from the fact that 12 00:00:49,960 --> 00:00:53,080 Speaker 2: in Vidio makes the majority of its money selling GPUs 13 00:00:53,120 --> 00:00:56,360 Speaker 2: to a handful of customers, and so well, some of 14 00:00:56,400 --> 00:00:58,959 Speaker 2: those also look to be on some of their own 15 00:00:58,960 --> 00:01:02,840 Speaker 2: incredibly shaky ground. And yeah, I'm talking about Oracle now. 16 00:01:02,880 --> 00:01:05,880 Speaker 2: In Vidia's health saying nothing of its growth isn't just 17 00:01:05,920 --> 00:01:08,559 Speaker 2: tied to these customers. It's also tied to whether these 18 00:01:08,600 --> 00:01:11,360 Speaker 2: customers can actually turn a profit from their capex spending. 19 00:01:12,040 --> 00:01:14,920 Speaker 2: And even that's not even certain so due to the 20 00:01:14,920 --> 00:01:17,360 Speaker 2: fact that so much money has been piled into building 21 00:01:17,400 --> 00:01:20,960 Speaker 2: AI infrastructure and big tech has promised to spend hundreds 22 00:01:21,000 --> 00:01:23,600 Speaker 2: of billions of dollars more in the next year. Big 23 00:01:23,640 --> 00:01:25,240 Speaker 2: tech has found itself in a bit of a hole. 24 00:01:26,319 --> 00:01:28,240 Speaker 2: How big of a hole? Well, by the end of 25 00:01:28,280 --> 00:01:31,280 Speaker 2: the year, Microsoft, Amazon, Google, and Meta will have spent 26 00:01:31,319 --> 00:01:34,560 Speaker 2: over four hundred billion dollars in capital expenditures, much of 27 00:01:34,600 --> 00:01:37,160 Speaker 2: it focused on building AI infrastructure, on top of two 28 00:01:37,280 --> 00:01:40,120 Speaker 2: hundred and twenty eight point four billion dollars in capex 29 00:01:40,160 --> 00:01:42,679 Speaker 2: in twenty twenty four and around one hundred and forty 30 00:01:42,720 --> 00:01:46,120 Speaker 2: eight billion in capital expenditures in twenty twenty three, for 31 00:01:46,160 --> 00:01:48,800 Speaker 2: a total of seven hundred and seventy six billion in 32 00:01:48,840 --> 00:01:53,120 Speaker 2: the space of three years, and they expect to spend 33 00:01:53,320 --> 00:01:56,480 Speaker 2: more than four hundred billion dollars more in twenty twenty six. 34 00:01:56,520 --> 00:01:59,280 Speaker 2: Every time I read these numbers, I feel a little crazy. 35 00:02:00,640 --> 00:02:03,400 Speaker 2: As a result, based on my own analysis, big tech 36 00:02:03,440 --> 00:02:06,080 Speaker 2: needs to make two trillion dollars in brand new, brand 37 00:02:06,160 --> 00:02:10,440 Speaker 2: spanking new revenue, specifically from AI by twenty thirty. All 38 00:02:10,480 --> 00:02:13,160 Speaker 2: of this was effectively for nothing. Now, I go into 39 00:02:13,320 --> 00:02:15,440 Speaker 2: detail about this in the premium newsletter I did on 40 00:02:15,480 --> 00:02:17,880 Speaker 2: October thirty first, but I'm going to give you a 41 00:02:17,919 --> 00:02:22,080 Speaker 2: short explanation here. First, though, we have to talk about depreciation, 42 00:02:22,280 --> 00:02:24,480 Speaker 2: and because I'm lazy, I'm going to quote myself in 43 00:02:24,480 --> 00:02:28,000 Speaker 2: that newsletter I just mentioned a couple of seconds ago 44 00:02:28,280 --> 00:02:31,680 Speaker 2: at hem. So when Microsoft buyers say one hundred million 45 00:02:31,720 --> 00:02:33,960 Speaker 2: dollars worth of GPUs, it immediately comes out of its 46 00:02:34,000 --> 00:02:36,880 Speaker 2: capital expenditures, which is when a company uses money to 47 00:02:36,960 --> 00:02:40,520 Speaker 2: invest in either buying or upgrading something. It then adds 48 00:02:40,560 --> 00:02:43,760 Speaker 2: to its property, plants and Equipment assets PPE for sure, 49 00:02:44,080 --> 00:02:46,680 Speaker 2: although some companies list this on their annual and quarterly 50 00:02:46,720 --> 00:02:50,360 Speaker 2: financials as property and equipment PPE sits on the balance sheet, 51 00:02:50,440 --> 00:02:52,840 Speaker 2: it's an asset as it's stuff for the company that 52 00:02:52,919 --> 00:02:56,880 Speaker 2: it owns or as least, GPUs depreciate, meaning they lose 53 00:02:56,960 --> 00:03:00,120 Speaker 2: value or over time, and this depreciation is represented on 54 00:03:00,120 --> 00:03:03,119 Speaker 2: a balance sheet and the income statement. Essentially, the goal 55 00:03:03,200 --> 00:03:05,200 Speaker 2: is to represent the value of an asset that a 56 00:03:05,240 --> 00:03:09,080 Speaker 2: company has on the income statement, and we see how 57 00:03:09,160 --> 00:03:11,840 Speaker 2: much the assets have declined during the reporting period, whether 58 00:03:11,880 --> 00:03:13,760 Speaker 2: that be a year or a quarter or something else, 59 00:03:13,880 --> 00:03:16,799 Speaker 2: Whereas the balance sheet shows the cumulative depreciation of every 60 00:03:16,840 --> 00:03:20,280 Speaker 2: asset currently in play. Depreciation does two things. And I 61 00:03:20,280 --> 00:03:22,240 Speaker 2: know this sounds like a lot, but I'll break it 62 00:03:22,280 --> 00:03:24,880 Speaker 2: down for you First, it allows a company to accurately 63 00:03:25,200 --> 00:03:27,840 Speaker 2: to an extent, represent the value of things it owns 64 00:03:27,880 --> 00:03:30,920 Speaker 2: over their useful life. Secondly, it allows a company to 65 00:03:31,000 --> 00:03:34,040 Speaker 2: deduct the value of an asset across said useful life 66 00:03:34,360 --> 00:03:37,240 Speaker 2: right up until its eventual removal, versus having to take 67 00:03:37,280 --> 00:03:40,320 Speaker 2: a big hit up front. The way this depreciation is 68 00:03:40,320 --> 00:03:43,520 Speaker 2: actually calculated can vary. There are several different methods available, 69 00:03:43,640 --> 00:03:46,160 Speaker 2: with some allowing for greater deductions at the start of 70 00:03:46,200 --> 00:03:48,240 Speaker 2: the term, which is useful for those items that will 71 00:03:48,240 --> 00:03:50,960 Speaker 2: experience the biggest drop in value right after buying them 72 00:03:51,240 --> 00:03:54,040 Speaker 2: and their initial use. An example you're probably familiar with 73 00:03:54,160 --> 00:03:56,400 Speaker 2: is a new car which loses a significant chunk of 74 00:03:56,400 --> 00:03:58,520 Speaker 2: its value the moment is driven off a dealership. Long 75 00:04:00,000 --> 00:04:03,160 Speaker 2: creation has become a big ugly problem with GPUs specifically 76 00:04:03,200 --> 00:04:05,800 Speaker 2: because of that useful life to find either as how 77 00:04:05,800 --> 00:04:07,560 Speaker 2: long the thing is able to be run before it 78 00:04:07,640 --> 00:04:11,160 Speaker 2: dies or how long before it becomes obsolete, and nobody 79 00:04:11,160 --> 00:04:12,920 Speaker 2: seems to be able to come up with a consensus 80 00:04:12,920 --> 00:04:15,680 Speaker 2: about how long this should be. In Microsoft's case, the 81 00:04:15,680 --> 00:04:18,560 Speaker 2: appreciation for its service is spread over six years, a 82 00:04:18,600 --> 00:04:21,200 Speaker 2: convenient change it made in August twenty twenty two, A 83 00:04:21,200 --> 00:04:24,239 Speaker 2: few months before the launch of chat GPT and before 84 00:04:24,240 --> 00:04:26,719 Speaker 2: it bought a bunch of fucking GPUs. This means that 85 00:04:26,760 --> 00:04:28,919 Speaker 2: Microsoft can spread the cost of tens of thousands of 86 00:04:28,920 --> 00:04:31,120 Speaker 2: a one hundred GPUs brought in twenty twenty or the 87 00:04:31,400 --> 00:04:33,640 Speaker 2: four hundred and fifty thousand, h one hundred gupus it 88 00:04:33,680 --> 00:04:36,360 Speaker 2: bought in twenty twenty four across six years, regardless of 89 00:04:36,400 --> 00:04:39,800 Speaker 2: whether those are the years they'll be generating revenue or 90 00:04:40,120 --> 00:04:44,320 Speaker 2: naturally functioning. Corwy for what It's worth says the same thing, 91 00:04:44,640 --> 00:04:46,560 Speaker 2: but largely because it's betting that it'll still be able 92 00:04:46,600 --> 00:04:49,600 Speaker 2: to find users for older silicon after its initial contracts 93 00:04:49,600 --> 00:04:52,599 Speaker 2: with the companies like OpenAI expire. The problem is is 94 00:04:52,600 --> 00:04:56,039 Speaker 2: that aigpus are fairly new concepts, and thus all of 95 00:04:56,080 --> 00:05:00,159 Speaker 2: this is pretty much untested ground. Whereas we know how long, 96 00:05:00,279 --> 00:05:02,440 Speaker 2: say a truck or a piece of heavy machinery can 97 00:05:02,520 --> 00:05:05,080 Speaker 2: last and how long it can deliver value to an organization, 98 00:05:05,400 --> 00:05:07,120 Speaker 2: we don't know the same thing about the kind of 99 00:05:07,200 --> 00:05:10,160 Speaker 2: data center GPUs that hyperscalers are spending tens of billions 100 00:05:10,200 --> 00:05:13,520 Speaker 2: of dollars on each year. Any kind of depreciation schedule 101 00:05:13,560 --> 00:05:17,039 Speaker 2: is based on at best assumptions and at worst hope. 102 00:05:17,160 --> 00:05:20,800 Speaker 2: Now this is important. The concept of an AI data 103 00:05:20,839 --> 00:05:24,000 Speaker 2: center is super new. We maybe saw the first ones 104 00:05:24,040 --> 00:05:29,080 Speaker 2: in twenty ninety. In question, it's kind of hard to say, 105 00:05:29,120 --> 00:05:31,719 Speaker 2: but even at the scale we're seeing today a Gigawa 106 00:05:31,880 --> 00:05:35,200 Speaker 2: data center pretty much brand new, maybe a couple years old. 107 00:05:35,240 --> 00:05:37,320 Speaker 2: I don't even think they've even built any but we'll 108 00:05:37,360 --> 00:05:39,560 Speaker 2: get to that in a bit. There are a lot 109 00:05:39,600 --> 00:05:42,560 Speaker 2: of assumptions at play. There's the assumption that the cards 110 00:05:42,600 --> 00:05:45,040 Speaker 2: won't degrade with heavy usage, or the assumption that future 111 00:05:45,040 --> 00:05:47,640 Speaker 2: generations of GPUs won't be so powerful and impressive that 112 00:05:47,680 --> 00:05:50,800 Speaker 2: they'll render the previous ones more obsolete than expected, kind 113 00:05:50,800 --> 00:05:52,760 Speaker 2: of like how the first jet powered planes of the 114 00:05:52,839 --> 00:05:56,040 Speaker 2: nineteen fifties did to those manufactured just a decade prior. 115 00:05:56,560 --> 00:05:59,400 Speaker 2: The assumption that there will be in fact a market 116 00:05:59,400 --> 00:06:01,160 Speaker 2: for Alder cards, than that there'll be a way to 117 00:06:01,240 --> 00:06:06,320 Speaker 2: lease them profitably. What if those assumptions are I don't know, wrong. 118 00:06:06,760 --> 00:06:09,560 Speaker 2: What if that hope is ultimately irrational. So there's a 119 00:06:09,640 --> 00:06:13,279 Speaker 2: quote from the Center for Information Technology Policy framing this problem, well, 120 00:06:13,279 --> 00:06:15,880 Speaker 2: that'll link to in the notes. Here is the puzzle. 121 00:06:16,240 --> 00:06:18,160 Speaker 2: The chips at the heart of the infrastructure build out 122 00:06:18,240 --> 00:06:20,400 Speaker 2: have a useful lifespan of one to three years due 123 00:06:20,400 --> 00:06:24,320 Speaker 2: to rapid technological obsolescence and physical wear, but companies appreciate 124 00:06:24,400 --> 00:06:27,040 Speaker 2: them over five or six years. In other words, they 125 00:06:27,040 --> 00:06:29,880 Speaker 2: spread out the cost of their massive capital expenditures over 126 00:06:29,880 --> 00:06:32,599 Speaker 2: a longer period than the facts warrant what the economist 127 00:06:32,720 --> 00:06:35,680 Speaker 2: is referred to as the four trillion dollar accounting puzzle 128 00:06:35,720 --> 00:06:38,520 Speaker 2: at the heart of the AIICLOUD. This is why Michael 129 00:06:38,560 --> 00:06:41,279 Speaker 2: Burry brought it up recently because spreading out these costs 130 00:06:41,320 --> 00:06:43,880 Speaker 2: allowed spigtech to make their net income i e. Their 131 00:06:43,880 --> 00:06:46,840 Speaker 2: profits look better in simple terms. By spreading out the 132 00:06:46,880 --> 00:06:49,680 Speaker 2: costs over six years rather than three, hyperscalers are able 133 00:06:49,720 --> 00:06:51,840 Speaker 2: to reduce the line item that eats into their earnings, 134 00:06:51,880 --> 00:06:55,039 Speaker 2: which makes their companies look better to the markets. So 135 00:06:55,120 --> 00:06:58,040 Speaker 2: why does this create an artificial time limit? Well, let's 136 00:06:58,080 --> 00:07:00,520 Speaker 2: start with a horrible fact. It takes it's two point 137 00:07:00,600 --> 00:07:03,080 Speaker 2: five years of construction time in about fifty billion dollars 138 00:07:03,120 --> 00:07:06,719 Speaker 2: per gigawa of data center capacity. No matter when the 139 00:07:06,760 --> 00:07:09,600 Speaker 2: GPUs for a giga what data center are bought one 140 00:07:09,600 --> 00:07:12,440 Speaker 2: way or another, these GPUs are depreciating in value either 141 00:07:12,480 --> 00:07:15,200 Speaker 2: through death or reduced efficacy through wear and tear, or 142 00:07:15,200 --> 00:07:17,720 Speaker 2: becoming obsolete, which is very likely as in Video is 143 00:07:17,720 --> 00:07:21,440 Speaker 2: committed to releasing a new GPU every single year. Newer 144 00:07:21,480 --> 00:07:25,120 Speaker 2: generation GPUs, like in videos Blackwell and Verra Reuben require 145 00:07:25,280 --> 00:07:28,160 Speaker 2: entirely new data center architecture, meaning that one as to 146 00:07:28,160 --> 00:07:30,360 Speaker 2: why they build a brand new data center a retrofit 147 00:07:30,400 --> 00:07:32,760 Speaker 2: an old one. Essentially, we have facilities that are being 148 00:07:32,760 --> 00:07:35,960 Speaker 2: built around a GPU design or product that may change 149 00:07:36,000 --> 00:07:38,240 Speaker 2: in a year or two. Now I hear that in 150 00:07:38,600 --> 00:07:42,080 Speaker 2: the Oberon racks that they use for the Blackwells will 151 00:07:42,120 --> 00:07:44,840 Speaker 2: be used with some Verra Rubin. But even then there's 152 00:07:44,880 --> 00:07:47,440 Speaker 2: going to be an even bigger, more huger Vera Ruben 153 00:07:47,480 --> 00:07:49,880 Speaker 2: that comes that might even that I read somewhere that 154 00:07:49,920 --> 00:07:52,040 Speaker 2: there might even be like killer what level ones just 155 00:07:52,880 --> 00:07:56,800 Speaker 2: in like one hundred killer what ones that this company's insane. Nevertheless, 156 00:07:56,920 --> 00:07:58,640 Speaker 2: at some point Wall Street is going to need to 157 00:07:58,680 --> 00:08:00,920 Speaker 2: see some sort of return on this investment, and right 158 00:08:00,960 --> 00:08:03,640 Speaker 2: now that return is negative dollars. I break it down 159 00:08:03,680 --> 00:08:06,760 Speaker 2: on my October thirty first premium piece, but for your sake, 160 00:08:06,800 --> 00:08:09,240 Speaker 2: I'll just say it. I estimate the big techniques to 161 00:08:09,280 --> 00:08:11,480 Speaker 2: make two dollars for every dollar of CAPEX they've spent, 162 00:08:11,720 --> 00:08:15,120 Speaker 2: and this revenue must be new, brand new, As this 163 00:08:15,160 --> 00:08:18,320 Speaker 2: CAPEX is only for AI. This CAPEX is useless for 164 00:08:18,360 --> 00:08:21,200 Speaker 2: everything else. It does not help it. And no, it 165 00:08:21,240 --> 00:08:24,360 Speaker 2: doesn't help that they bolted copilot onto fucking everything that 166 00:08:24,520 --> 00:08:28,120 Speaker 2: is not working, and in fact, the Australian Competition Commission 167 00:08:28,160 --> 00:08:31,840 Speaker 2: is suing them. Maybe I mentioned that later, but whatever, Meta, Amazon, 168 00:08:31,880 --> 00:08:34,520 Speaker 2: Google and Microsoft are already years and hundreds of billions 169 00:08:34,559 --> 00:08:36,520 Speaker 2: of dollars in and are yet to see a dollar 170 00:08:36,559 --> 00:08:39,440 Speaker 2: of profit, creating a one point two to one trillion 171 00:08:39,559 --> 00:08:42,760 Speaker 2: dollar hole just to justify the expenses. So around six 172 00:08:42,840 --> 00:08:44,920 Speaker 2: hundred and five billion dollars of CAPEX all told at 173 00:08:44,920 --> 00:08:47,640 Speaker 2: the time I calculated it. Much of this CAPEX has 174 00:08:47,640 --> 00:08:49,640 Speaker 2: been committed or spent before they've even turned on a 175 00:08:49,720 --> 00:08:53,360 Speaker 2: single goddamn GPU. You might argue that there's a scenario 176 00:08:53,480 --> 00:08:56,920 Speaker 2: here where say, an A one hundred GPU is useful 177 00:08:56,920 --> 00:08:59,439 Speaker 2: past the three or six year shelf life. Even if 178 00:08:59,440 --> 00:09:01,280 Speaker 2: that were the case, the average rental price of an 179 00:09:01,280 --> 00:09:03,839 Speaker 2: A one hundred is ninety nine cents an hour. This 180 00:09:03,920 --> 00:09:06,400 Speaker 2: is a four or five year old GPU, and customers 181 00:09:06,440 --> 00:09:08,120 Speaker 2: are paying for it like they would a five year 182 00:09:08,120 --> 00:09:10,680 Speaker 2: old piece of hardware. The same fate awaits the H 183 00:09:10,720 --> 00:09:13,439 Speaker 2: one hundred, which was released in twenty twenty two but 184 00:09:13,520 --> 00:09:15,960 Speaker 2: was still sold in great volume through twenty twenty four, 185 00:09:16,160 --> 00:09:18,000 Speaker 2: and I hear the H two hundred of the same 186 00:09:18,080 --> 00:09:21,959 Speaker 2: generation is still selling to this day. Every year in 187 00:09:22,080 --> 00:09:24,679 Speaker 2: Vidia releases a new GPU, lowering the value of all 188 00:09:24,720 --> 00:09:27,240 Speaker 2: the other GPUs in the process, making it harder to 189 00:09:27,280 --> 00:09:29,520 Speaker 2: fill in the holes created by all the other GPUs 190 00:09:29,600 --> 00:09:33,200 Speaker 2: is capex and costs. This whole time, nobody appears to 191 00:09:33,240 --> 00:09:34,880 Speaker 2: have found a way to make a profit, meaning that 192 00:09:34,880 --> 00:09:37,920 Speaker 2: the hole created by these GPUs remains unfilled, all while 193 00:09:37,960 --> 00:09:43,600 Speaker 2: big tech firms by more GPUs, creating more holes to fill. 194 00:09:43,760 --> 00:09:46,160 Speaker 2: So now that you know this, there's a fairly obvious 195 00:09:46,320 --> 00:09:48,880 Speaker 2: question to ask, why in the hell are they still 196 00:09:48,920 --> 00:09:52,560 Speaker 2: buying GPUs? Well, so, where the fuck are these GPUs going? 197 00:10:05,800 --> 00:10:07,800 Speaker 2: So a few weeks ago I wrote a piece Premium 198 00:10:07,840 --> 00:10:09,600 Speaker 2: one called the Hater's Guide to Nvidia, and I asked 199 00:10:09,600 --> 00:10:12,319 Speaker 2: the basic question in there, where have all the GPUs 200 00:10:12,360 --> 00:10:15,000 Speaker 2: that Invidia has sold actually gone? In particular, the six 201 00:10:15,080 --> 00:10:18,160 Speaker 2: million Blackwell GPUs that Jensen one keeps banging on about 202 00:10:18,600 --> 00:10:20,839 Speaker 2: Now there's little evidence that these are being used in 203 00:10:20,880 --> 00:10:23,360 Speaker 2: the volume which they're sold, suggesting that they're either languishing 204 00:10:23,360 --> 00:10:26,079 Speaker 2: in the supply chain or being warehoused by hyperscalers, or 205 00:10:26,120 --> 00:10:28,960 Speaker 2: even in Nvidia themselves. Now there's the argument that this 206 00:10:29,040 --> 00:10:32,800 Speaker 2: could be and this is wanky Nvidian bullshit, this could 207 00:10:32,800 --> 00:10:36,320 Speaker 2: actually be two GPUs per GPU sold because there's two 208 00:10:36,800 --> 00:10:39,920 Speaker 2: chips on each GPU. Even if that was the case, 209 00:10:40,280 --> 00:10:44,520 Speaker 2: three billion GPUs Blackwell, specifically the brand new ones, they're 210 00:10:44,559 --> 00:10:45,320 Speaker 2: not in service. 211 00:10:45,960 --> 00:10:46,080 Speaker 1: Now. 212 00:10:46,120 --> 00:10:47,839 Speaker 2: While I'm not going to go and copy paste an 213 00:10:48,000 --> 00:10:50,840 Speaker 2: entire premium piece into this script, I am, however, going 214 00:10:50,880 --> 00:10:52,960 Speaker 2: to go into detail about what I found. And the 215 00:10:53,040 --> 00:10:56,280 Speaker 2: truth is, I can only really see in this inclodes, 216 00:10:56,320 --> 00:11:00,800 Speaker 2: looking over like bunches of data center maps, reading hundreds 217 00:11:00,800 --> 00:11:04,319 Speaker 2: of press releases, documents, earning statements. I've only been able 218 00:11:04,400 --> 00:11:09,439 Speaker 2: to find maybe a couple hundred thousand Blackwell GPUs in existence, 219 00:11:09,520 --> 00:11:12,840 Speaker 2: maybe half a million to seven hundred and fifty thousand 220 00:11:12,840 --> 00:11:15,200 Speaker 2: if you include the stuff that hasn't even been built yet. 221 00:11:15,640 --> 00:11:18,840 Speaker 2: Let's go into it. So Stargate Apilene allegedly four hundred 222 00:11:18,880 --> 00:11:22,600 Speaker 2: thousand Blackweld gps A, going there now Oracle CEO Coco, 223 00:11:22,679 --> 00:11:26,200 Speaker 2: I should say, Clay McGurk, that's probably not how you 224 00:11:26,240 --> 00:11:29,400 Speaker 2: say that. He claimed very recently there were ninety six 225 00:11:29,440 --> 00:11:33,120 Speaker 2: thousand of them in stored, so not great. There's theoretically 226 00:11:33,160 --> 00:11:35,760 Speaker 2: in one hundred and thirty one thousand Blackwell GPU cluster 227 00:11:35,840 --> 00:11:38,360 Speaker 2: owned by Oracle that they announced in March twenty twenty five, 228 00:11:39,360 --> 00:11:42,960 Speaker 2: so that should be online. Never five thousand blackweld GPUs 229 00:11:42,960 --> 00:11:46,200 Speaker 2: at the University of Texas Austin, which sound like they're online, 230 00:11:46,280 --> 00:11:49,199 Speaker 2: more than fifteen hundred in a Lambda data center in Columbus, Ohio. 231 00:11:49,280 --> 00:11:52,319 Speaker 2: Those are online. The Department of Energy is still in 232 00:11:52,360 --> 00:11:56,280 Speaker 2: development one hundred thousand GPU supercluster as well as ten 233 00:11:56,320 --> 00:11:58,840 Speaker 2: thousand in Vidia blackweld GPUs that are expected to be 234 00:11:58,880 --> 00:12:02,680 Speaker 2: available in twenty twenty six, and it's Equinox cluster. Really 235 00:12:02,720 --> 00:12:05,160 Speaker 2: can't establish how many of those are actually in operation. 236 00:12:05,760 --> 00:12:08,880 Speaker 2: Fifty thousand of these blackweld GPUs going into the still 237 00:12:08,960 --> 00:12:13,520 Speaker 2: unbuilt musk Run Colossus to supercluster, Corby's largest GB two 238 00:12:13,640 --> 00:12:16,520 Speaker 2: hundred Blackwell cluster of two four hundred and ninety six 239 00:12:16,559 --> 00:12:21,280 Speaker 2: Blackwell GPUs, tens of thousands of them deployed globally by Microsoft, 240 00:12:21,280 --> 00:12:25,120 Speaker 2: including forty six hundred Blackwell Ultra GPUs and two hundred 241 00:12:25,160 --> 00:12:28,120 Speaker 2: and sixty thousand of them, these Blackwell GPUs going into 242 00:12:28,120 --> 00:12:32,000 Speaker 2: five AI data centers for the South Korean government, and yeah, 243 00:12:32,080 --> 00:12:33,640 Speaker 2: I just want to be clear that that is also 244 00:12:33,679 --> 00:12:37,840 Speaker 2: fairly recently announced, so probably not even not even built, 245 00:12:38,440 --> 00:12:41,600 Speaker 2: let alone powered on. I goan to be honest, I'm 246 00:12:41,679 --> 00:12:46,680 Speaker 2: genuinely unable to find one million Blackwell GPUs like inexistence. Now, 247 00:12:46,960 --> 00:12:48,880 Speaker 2: some of you might say, oh, there's a bunch of 248 00:12:48,960 --> 00:12:51,800 Speaker 2: secret ones. There's a bunch of them. They don't announce 249 00:12:51,840 --> 00:12:54,600 Speaker 2: every single one. Here's the thing. Three million of these 250 00:12:54,640 --> 00:12:57,920 Speaker 2: fucking things have allegedly been shipped. I can't find a 251 00:12:57,960 --> 00:13:01,360 Speaker 2: million of them. And considering everybody always talks about their 252 00:13:01,400 --> 00:13:05,480 Speaker 2: GPU purchases, I'm kind of shocked at calm. Now. I 253 00:13:05,600 --> 00:13:07,960 Speaker 2: do not know where these six million black Weld GPUs 254 00:13:08,000 --> 00:13:10,200 Speaker 2: have gone, but they certainly haven't gone into data centers 255 00:13:10,200 --> 00:13:12,480 Speaker 2: that are powered and turned on. In fact, power has 256 00:13:12,520 --> 00:13:14,640 Speaker 2: become one of the biggest issues of building these things, 257 00:13:14,760 --> 00:13:16,679 Speaker 2: and the fact it's really difficult and maybe impossible to 258 00:13:16,679 --> 00:13:18,360 Speaker 2: get the amount of power these things need to the 259 00:13:18,400 --> 00:13:22,720 Speaker 2: goddamn data centers. In really simple terms, there isn't enough 260 00:13:22,800 --> 00:13:25,400 Speaker 2: power or built data centers for those black world GPUs 261 00:13:25,400 --> 00:13:27,680 Speaker 2: to run, in part because the data centers aren't built, 262 00:13:27,679 --> 00:13:29,560 Speaker 2: and in part because there isn't enough power for the 263 00:13:29,559 --> 00:13:33,240 Speaker 2: ones that are. Microsoft CEO Sacha Nadella recently said in 264 00:13:33,280 --> 00:13:36,160 Speaker 2: a podcast that his company and I quote didn't have 265 00:13:36,240 --> 00:13:39,679 Speaker 2: the warm shells to plug into, meaning buildings with sufficient power, 266 00:13:40,400 --> 00:13:43,320 Speaker 2: and heavily suggested that Microsoft may actually have a bunch 267 00:13:43,360 --> 00:13:47,800 Speaker 2: of chips sitting in inventory that they couldn't plug in now. 268 00:13:48,120 --> 00:13:50,160 Speaker 2: Just to give you an estimate here, even if we 269 00:13:50,200 --> 00:13:53,839 Speaker 2: say three million GPUs, even if we're going with the moonmath, 270 00:13:53,880 --> 00:13:56,640 Speaker 2: the Vinvidia, if we're going into the make believe world, 271 00:13:56,679 --> 00:14:01,120 Speaker 2: the twisted mind of Jensen Wang, still three million GPUs. 272 00:14:01,160 --> 00:14:04,120 Speaker 2: We'll look at still like five or six gigawatts of capacity. 273 00:14:04,679 --> 00:14:06,800 Speaker 2: It's not being built. I don't even think two gigawats 274 00:14:06,840 --> 00:14:08,880 Speaker 2: of data center capacity have been built. And I swear 275 00:14:08,920 --> 00:14:10,840 Speaker 2: to fucking God, if one of you emails me and said, 276 00:14:10,880 --> 00:14:14,240 Speaker 2: and Eric is built, and twenty gigawatts or something, power 277 00:14:14,240 --> 00:14:17,160 Speaker 2: can get built. Power can get built. You can build power, 278 00:14:17,600 --> 00:14:20,480 Speaker 2: getting it to the data center and actually powering the 279 00:14:20,560 --> 00:14:24,960 Speaker 2: data center correctly, as in things turn on, everything works, 280 00:14:25,200 --> 00:14:29,160 Speaker 2: nothing overloads, nothing blacks out, and the power is consistently done. 281 00:14:29,400 --> 00:14:32,960 Speaker 2: Takes what's just months of surveys and scientific stuff and 282 00:14:33,000 --> 00:14:35,760 Speaker 2: then years to just get it done. Stargate Appleine only 283 00:14:35,760 --> 00:14:38,600 Speaker 2: has two hundred megawatts. They're gonna need over one point 284 00:14:38,600 --> 00:14:41,360 Speaker 2: four gigawats just to turn the fucking thing on. I'm 285 00:14:41,400 --> 00:14:45,320 Speaker 2: so tired of He's God damn GPUs bo. With all 286 00:14:45,360 --> 00:14:48,760 Speaker 2: this said, why pray tell? Is Jensen Huang of Nvidia 287 00:14:48,960 --> 00:14:51,960 Speaker 2: saying that he has twenty million Blackwell and Vera Rubin 288 00:14:52,040 --> 00:14:54,600 Speaker 2: GPUs ordered through the end of twenty twenty six. Where 289 00:14:54,600 --> 00:14:57,480 Speaker 2: are they fucking going? Jensen? Now, I think that number 290 00:14:57,480 --> 00:15:00,680 Speaker 2: also includes the six million. And also, to be clear, 291 00:15:01,520 --> 00:15:03,600 Speaker 2: I know a lot of you aren't technical, which is awesome. 292 00:15:03,640 --> 00:15:06,000 Speaker 2: I love. I want you all to know about this. 293 00:15:06,400 --> 00:15:08,160 Speaker 2: You need to know that this is part of the 294 00:15:08,200 --> 00:15:12,680 Speaker 2: course within Nvidia in video loves schmushing accountancy things together 295 00:15:12,720 --> 00:15:15,360 Speaker 2: and coming up with random numbers. Credit the case could 296 00:15:15,360 --> 00:15:17,240 Speaker 2: go our friend of the show for telling me the story. 297 00:15:17,280 --> 00:15:20,880 Speaker 2: But during the early twenty twenties, so that there is 298 00:15:20,960 --> 00:15:24,800 Speaker 2: twenty twenty two during the Big Crypto rush. In Vidia 299 00:15:24,880 --> 00:15:28,080 Speaker 2: classified gaming GPUs that were sold to bitcoin miners as 300 00:15:28,080 --> 00:15:31,160 Speaker 2: gaming revenue. They got digged by the SEC. Wasn't fraud, 301 00:15:31,760 --> 00:15:34,320 Speaker 2: but just so you know, in video will move shit around. 302 00:15:34,600 --> 00:15:37,120 Speaker 2: And I truly do not know where these GPUs are. 303 00:15:37,280 --> 00:15:41,200 Speaker 2: I do not know even why anyone is still buying GPUs. 304 00:15:41,600 --> 00:15:41,760 Speaker 1: Now. 305 00:15:41,840 --> 00:15:45,160 Speaker 2: AI bulls will tell you that there's this insatiable demand 306 00:15:45,160 --> 00:15:47,760 Speaker 2: for AI and these massive amounts of orders are proof 307 00:15:47,800 --> 00:15:51,720 Speaker 2: of something or rather, and you know what, I'll give 308 00:15:51,760 --> 00:15:53,920 Speaker 2: them that it's proof that people are buying a lot 309 00:15:53,960 --> 00:15:57,400 Speaker 2: of GPUs. I just don't know why nobody has made 310 00:15:57,440 --> 00:15:59,880 Speaker 2: a profit from AI, and those making revenue aren't really 311 00:16:00,080 --> 00:16:02,640 Speaker 2: making that much. Let me give you an example. My 312 00:16:02,760 --> 00:16:05,360 Speaker 2: reporting on open ai from November twelfth suggests that the 313 00:16:05,360 --> 00:16:07,640 Speaker 2: company I only made four point three two nine billion 314 00:16:07,720 --> 00:16:10,720 Speaker 2: dollars in revenue for the end of September, extrapolated from 315 00:16:10,760 --> 00:16:13,520 Speaker 2: the twenty percent revenue share that Microsoft receives in the company. 316 00:16:14,080 --> 00:16:18,000 Speaker 2: And now some people who write really shit our substacks 317 00:16:18,400 --> 00:16:20,840 Speaker 2: have argued with the figures, claiming that they're either delayed 318 00:16:21,320 --> 00:16:23,240 Speaker 2: or are not inclusive of the revenue that open Ai 319 00:16:23,440 --> 00:16:26,520 Speaker 2: is paid from Microsoft as part of being's ai integration 320 00:16:26,640 --> 00:16:29,360 Speaker 2: and sales of open AI's models throughout Microsoft as ure. 321 00:16:29,600 --> 00:16:31,000 Speaker 2: So I want to be clear of two things. Some 322 00:16:31,160 --> 00:16:35,120 Speaker 2: a deeply bitter person. This is a crual accounting, meaning 323 00:16:35,160 --> 00:16:37,720 Speaker 2: that these numbers are revenue booked in the quarter I 324 00:16:37,800 --> 00:16:42,119 Speaker 2: reported them. Any comments about quarter long delays or naive approaches, 325 00:16:42,160 --> 00:16:44,400 Speaker 2: and you know who I'm fucking talking about, if you're listening, 326 00:16:45,040 --> 00:16:49,560 Speaker 2: are incorrect and a riboso. Also, Microsoft's revenue share payments 327 00:16:49,560 --> 00:16:54,080 Speaker 2: to open ai kind of pathetic, totally based on documents 328 00:16:54,120 --> 00:16:57,680 Speaker 2: reviewed by this newsletter publication whatever you call me, media 329 00:16:57,880 --> 00:17:01,400 Speaker 2: entity floating blob in the podcast, for sixty nine point 330 00:17:01,440 --> 00:17:04,080 Speaker 2: one million dollars in counting the year Q three, twenty 331 00:17:04,119 --> 00:17:07,840 Speaker 2: twenty five. And by the way, the actual number for 332 00:17:08,040 --> 00:17:11,120 Speaker 2: that three month period, including all royalties, is about four 333 00:17:11,200 --> 00:17:15,439 Speaker 2: point five two seven billion dollars of revenue. I just 334 00:17:15,480 --> 00:17:18,080 Speaker 2: want to be clear about something with open Ai. I'm 335 00:17:18,080 --> 00:17:21,080 Speaker 2: not saying they're misrepresenting their numbers to anyone. I hope 336 00:17:21,080 --> 00:17:23,960 Speaker 2: that open ai is being honest with their revenues. But 337 00:17:24,040 --> 00:17:26,400 Speaker 2: if it comes out, I'm right, if it comes out 338 00:17:26,440 --> 00:17:28,679 Speaker 2: that it turns out that they've been telling investors completely 339 00:17:28,680 --> 00:17:33,119 Speaker 2: different numbers, I'm gonna be absolutely fucking insufferable. I'm going 340 00:17:33,200 --> 00:17:35,600 Speaker 2: to bring in I'm going to be playing Tommy Trumpets 341 00:17:35,600 --> 00:17:38,160 Speaker 2: a walk around cheering that you can get five minutes 342 00:17:38,200 --> 00:17:41,200 Speaker 2: of monologue about that. Also in the same period, Open 343 00:17:41,240 --> 00:17:43,760 Speaker 2: Ai spent eight point six seven billion dollars on inference, 344 00:17:43,760 --> 00:17:45,680 Speaker 2: which is the process in which an l them creates 345 00:17:45,720 --> 00:17:49,240 Speaker 2: its output. This is the biggest company in the generative 346 00:17:49,280 --> 00:17:52,080 Speaker 2: AI space, with eight hundred million weekly active users in 347 00:17:52,119 --> 00:17:54,919 Speaker 2: the Mandate of Heaven in the eyes of the media, Anthropic, 348 00:17:54,960 --> 00:17:58,160 Speaker 2: it's largest competitor, Allegacy, will make eight hundred and thirty 349 00:17:58,160 --> 00:18:00,760 Speaker 2: three million dollars in revenue in December tween twenty five, 350 00:18:00,960 --> 00:18:03,000 Speaker 2: and based on my estimates, will end up having about 351 00:18:03,000 --> 00:18:05,119 Speaker 2: four and a half to five billion dollars if revenue 352 00:18:05,160 --> 00:18:07,520 Speaker 2: by the end of the year. Based on my reporting 353 00:18:07,560 --> 00:18:10,640 Speaker 2: from October Andthropic spent two point six six billion dollars 354 00:18:10,680 --> 00:18:13,400 Speaker 2: on Amazon Web Services through the end of September, meaning 355 00:18:13,440 --> 00:18:15,959 Speaker 2: that it, based on my own analysis of reported revenues, 356 00:18:16,000 --> 00:18:18,720 Speaker 2: spent one hundred and four percent of its revenue up 357 00:18:18,760 --> 00:18:21,679 Speaker 2: to that point just on AWS likely spent as much 358 00:18:21,720 --> 00:18:25,159 Speaker 2: on Google Cloud. Now, the reason I'm bringing up these 359 00:18:25,240 --> 00:18:28,760 Speaker 2: numbers is these are the champions, the champions of the 360 00:18:28,840 --> 00:18:33,920 Speaker 2: AI boom, yet their revenues kind of fucking stink. Wow. 361 00:18:34,320 --> 00:18:36,800 Speaker 2: Even if open ai made thirteen billion dollars this year, 362 00:18:36,840 --> 00:18:40,040 Speaker 2: even if Anthropic made five billion dollars, okay, wow, so 363 00:18:40,119 --> 00:18:45,240 Speaker 2: that's not even twenty billion dollars. That's like nineteen billion 364 00:18:45,320 --> 00:18:48,679 Speaker 2: dollars less than Microsoft spent on GPU's and other capex 365 00:18:48,880 --> 00:18:53,199 Speaker 2: in the last quarter. That's dogshit. I'm sorry, I'm just 366 00:18:53,280 --> 00:18:56,640 Speaker 2: tired of I am tired of humoring this. I'm sure 367 00:18:56,680 --> 00:18:59,159 Speaker 2: all of you are too. I find it loathsome that 368 00:18:59,200 --> 00:19:02,440 Speaker 2: we have to pretend these people are gifted somehow, they 369 00:19:02,480 --> 00:19:05,399 Speaker 2: have shit. Ask businesses that burn billions of dollars, and 370 00:19:05,440 --> 00:19:08,840 Speaker 2: you know what. Another thing I'm tired about is everybody 371 00:19:08,880 --> 00:19:13,600 Speaker 2: telling this story about Anthropic being more efficient and only 372 00:19:13,640 --> 00:19:16,920 Speaker 2: burning two point eight billion dollars this year. Now one 373 00:19:16,960 --> 00:19:19,440 Speaker 2: has to ask a question about why this company that's 374 00:19:19,480 --> 00:19:23,480 Speaker 2: allegedly reducing costs had to raise thirteen billion dollars in 375 00:19:23,520 --> 00:19:26,919 Speaker 2: September twenty twenty five, after raising three point five billion 376 00:19:26,960 --> 00:19:30,080 Speaker 2: dollars in March twenty twenty five, after raising four billion 377 00:19:30,119 --> 00:19:33,320 Speaker 2: dollars in November twenty twenty four. Am I really meant 378 00:19:33,320 --> 00:19:36,480 Speaker 2: to read stories about Anthropic hitting break even in twenty 379 00:19:36,480 --> 00:19:39,600 Speaker 2: twenty eight with a straight face, especially as other stories 380 00:19:39,640 --> 00:19:42,119 Speaker 2: say that we cash flow positive as soon as twenty 381 00:19:42,160 --> 00:19:45,640 Speaker 2: twenty seven. This company's as big a pile of shit 382 00:19:45,720 --> 00:19:48,840 Speaker 2: as open Ai. Open Ai raised eighteen point three billion 383 00:19:48,880 --> 00:19:51,960 Speaker 2: dollars this year. That's less than two billion dollars more 384 00:19:52,000 --> 00:19:54,960 Speaker 2: than Anthropic, who makes a bunch less revenue. Can't believe 385 00:19:54,960 --> 00:20:00,000 Speaker 2: I'm defending open Ai. But these companies are the two 386 00:20:00,000 --> 00:20:02,240 Speaker 2: two largest ones in the generative of AI space, and 387 00:20:02,280 --> 00:20:06,160 Speaker 2: by extension, the two largest consumers of GPU COMPUW Both 388 00:20:06,160 --> 00:20:08,960 Speaker 2: companies burn billions of dollars and require an infinite amount 389 00:20:09,000 --> 00:20:11,120 Speaker 2: of venture capital to keep them alive. At a time 390 00:20:11,160 --> 00:20:14,119 Speaker 2: when the Saudi Public Investment Fund is struggling and the 391 00:20:14,200 --> 00:20:16,720 Speaker 2: US venture capital system is set to run out of 392 00:20:16,720 --> 00:20:18,880 Speaker 2: cash in the next year and a half, The two 393 00:20:19,000 --> 00:20:22,080 Speaker 2: largest sources of actual revenue for selling AI compute are 394 00:20:22,119 --> 00:20:25,320 Speaker 2: subsidized by venture capital and debt. What happens if these 395 00:20:25,320 --> 00:20:28,240 Speaker 2: sources dry up. They're not paying out of cash flow, 396 00:20:28,600 --> 00:20:32,280 Speaker 2: and in all seriousness, who else is buying AI compute? 397 00:20:32,440 --> 00:20:35,639 Speaker 2: What are they doing with them? Hyperscalers other than Microsoft, 398 00:20:35,680 --> 00:20:37,919 Speaker 2: which chose to stop reporting its AI revenue back in 399 00:20:37,960 --> 00:20:40,200 Speaker 2: January when it claimed it made about a billion dollars 400 00:20:40,200 --> 00:20:43,720 Speaker 2: a month in revenue, don't disclose anything about their AI revenue, 401 00:20:43,840 --> 00:20:45,840 Speaker 2: which in turn means that we have no real idea 402 00:20:45,880 --> 00:20:49,439 Speaker 2: of how much real actual money is coming to justify 403 00:20:49,480 --> 00:20:53,000 Speaker 2: these GPUs core We've made one point three six billion 404 00:20:53,040 --> 00:20:55,000 Speaker 2: dollars in revenue and lost one hundred and ten million 405 00:20:55,000 --> 00:20:57,560 Speaker 2: dollars doing so in the last quarter. And if that's 406 00:20:57,560 --> 00:21:00,679 Speaker 2: indicative of the kind of actual real demand for AI compume, 407 00:21:01,000 --> 00:21:03,400 Speaker 2: I think it's time to start panicking about whether all 408 00:21:03,400 --> 00:21:06,040 Speaker 2: of this was for nothing. Corweave has a backlog of 409 00:21:06,080 --> 00:21:09,200 Speaker 2: over fifty billion dollars in compute, and twenty two billion 410 00:21:09,240 --> 00:21:11,879 Speaker 2: dollars of that is open AI company that burns billions 411 00:21:11,880 --> 00:21:14,919 Speaker 2: of dollars a year and lives on venture subsidies. Fourteen 412 00:21:14,920 --> 00:21:16,879 Speaker 2: billion dollars of that is Meta, which is yet to 413 00:21:16,880 --> 00:21:18,640 Speaker 2: work out how to make any kind of real money 414 00:21:18,680 --> 00:21:21,520 Speaker 2: from generative AI. And no, it's generative AI ads are 415 00:21:21,520 --> 00:21:23,439 Speaker 2: not the future four or four media. I love you, 416 00:21:23,480 --> 00:21:26,320 Speaker 2: but that story was bunk and the rest of it 417 00:21:26,359 --> 00:21:29,080 Speaker 2: is likely a mixture of Microsoft and Video, which agreed 418 00:21:29,080 --> 00:21:31,400 Speaker 2: to buy six point three billion dollars of any unused 419 00:21:31,400 --> 00:21:35,040 Speaker 2: compute from Corewave through twenty thirty two. Should also be clear, 420 00:21:35,240 --> 00:21:36,879 Speaker 2: I do pay and subscribe to four or four I 421 00:21:36,920 --> 00:21:39,520 Speaker 2: love it. Just the AI ads story was wank, I 422 00:21:39,520 --> 00:21:42,280 Speaker 2: love you, I love you, Joe, I love I love 423 00:21:42,359 --> 00:21:45,359 Speaker 2: I love the publication. Sorry. I also forgot Google, by 424 00:21:45,359 --> 00:21:47,600 Speaker 2: the way, which is renting capacity from Corewave to rent 425 00:21:47,640 --> 00:21:50,359 Speaker 2: to open AI, and I'm not shitting you. Oh fuck Sorry. 426 00:21:50,400 --> 00:21:53,119 Speaker 2: I also forgot to mention that Corwy's backlog problem stems 427 00:21:53,119 --> 00:21:56,879 Speaker 2: from data center construction delays. That and Corewave has fourteen 428 00:21:56,920 --> 00:22:00,200 Speaker 2: billion dollars in debt, mostly from buying GPUs, which is 429 00:22:00,440 --> 00:22:02,919 Speaker 2: able to raise by using GPUs as collateral, and then 430 00:22:02,920 --> 00:22:05,919 Speaker 2: it had contracts and customers willing to pay for it, 431 00:22:06,000 --> 00:22:09,000 Speaker 2: such as in Video, who is also selling it the GPUs. 432 00:22:09,760 --> 00:22:11,879 Speaker 2: I also left something out of this script, which is 433 00:22:11,880 --> 00:22:14,400 Speaker 2: that just the last week, core We've just raised another 434 00:22:14,440 --> 00:22:18,159 Speaker 2: two billion dollars of debt. When this all ends, I 435 00:22:18,200 --> 00:22:20,679 Speaker 2: am going to be a little insufferable. But let's just 436 00:22:20,720 --> 00:22:23,840 Speaker 2: be abundantly clear. Core Weaver has bought all those GPUs 437 00:22:23,880 --> 00:22:27,760 Speaker 2: to rent open AI, Microsoft for open Ai, Meta Google 438 00:22:27,880 --> 00:22:30,000 Speaker 2: for open Ai, and in Video, which is the company 439 00:22:30,040 --> 00:22:35,960 Speaker 2: that benefits from core Weave's continuum ability to buy GPUs? Otherwise, 440 00:22:35,960 --> 00:22:38,720 Speaker 2: where's the fucking business? Exactly? Who are the customers, who 441 00:22:38,720 --> 00:22:40,919 Speaker 2: are the people renting the GPUs? And what is the 442 00:22:40,920 --> 00:22:43,399 Speaker 2: purpose for which they're being rented? How much money is 443 00:22:43,440 --> 00:22:46,840 Speaker 2: renting those gps? Can you? Can you tell me? Can 444 00:22:46,880 --> 00:22:49,040 Speaker 2: anyone tell me? Can anyone tell me anything? You can 445 00:22:49,080 --> 00:22:51,760 Speaker 2: sit and wank and waffle on about the supposed glorious 446 00:22:51,760 --> 00:22:54,679 Speaker 2: AI revolution all you want, but where's the goddamn money? 447 00:22:55,160 --> 00:22:59,040 Speaker 2: And why exactly are we still buying GPUs? What are 448 00:22:59,040 --> 00:23:01,960 Speaker 2: they doing, who are they being rented for what purpose? 449 00:23:02,119 --> 00:23:04,160 Speaker 2: And why isn't it creating the kind of revenue that's 450 00:23:04,200 --> 00:23:07,639 Speaker 2: actually worth sharing or products that are actually worth using. 451 00:23:08,400 --> 00:23:10,560 Speaker 2: Is it because the products suck? Is it because the 452 00:23:10,600 --> 00:23:13,800 Speaker 2: revenue sucks? Is it because it's unprofitable to make the revenue? 453 00:23:14,240 --> 00:23:17,119 Speaker 2: And why at this point in history do we not 454 00:23:17,359 --> 00:23:20,480 Speaker 2: know hundreds of billions of dollars that have made in 455 00:23:20,600 --> 00:23:23,080 Speaker 2: Vidia the biggest company on the stock market, and we 456 00:23:23,119 --> 00:23:25,320 Speaker 2: still do not know why people buy these fucking things, 457 00:23:25,480 --> 00:23:28,439 Speaker 2: nor do we know what they fucking cost imagine if 458 00:23:28,480 --> 00:23:31,080 Speaker 2: we sold cars and we didn't have a milesber gallon rating. 459 00:23:32,280 --> 00:23:51,560 Speaker 2: I'm serious. That's effectively where we are. Oh God, and 460 00:23:51,640 --> 00:23:54,199 Speaker 2: Video is currently making hundreds of billions of dollars in 461 00:23:54,280 --> 00:23:56,720 Speaker 2: revenue selling GPUs to companies that either plug them in 462 00:23:56,760 --> 00:23:58,840 Speaker 2: and start losing money or I assume put them in 463 00:23:58,880 --> 00:24:02,440 Speaker 2: a warehouse for safety. And those companies increasingly a raking 464 00:24:02,520 --> 00:24:04,760 Speaker 2: up mountains of debt to do so, and billions more 465 00:24:04,760 --> 00:24:07,480 Speaker 2: in long term lease payments. And this brings me to 466 00:24:07,520 --> 00:24:13,080 Speaker 2: my core anxiety. Why exactly a company's pre ordering GPUs? 467 00:24:14,000 --> 00:24:16,760 Speaker 2: What benefit is there in doing so? Blackwell does not 468 00:24:16,800 --> 00:24:19,240 Speaker 2: appear to be more efficient in a way that actually 469 00:24:19,240 --> 00:24:22,160 Speaker 2: makes anybody a profit. And we're potentially years from seeing 470 00:24:22,160 --> 00:24:24,440 Speaker 2: these GPUs in operation in data centers at the scale 471 00:24:24,440 --> 00:24:27,359 Speaker 2: they're being shipped, So why is anyone buying more? I 472 00:24:27,400 --> 00:24:29,359 Speaker 2: just want to be really specific about something, because I 473 00:24:29,400 --> 00:24:31,159 Speaker 2: don't feel like I nailed this down two and a 474 00:24:31,200 --> 00:24:34,280 Speaker 2: half years. Fifty billion dollars per gigabot of data centers. 475 00:24:34,359 --> 00:24:36,760 Speaker 2: You may be thinking, what black Wells, You'll just shove 476 00:24:36,800 --> 00:24:38,800 Speaker 2: them in the old data centers, right, No, they use 477 00:24:38,880 --> 00:24:42,200 Speaker 2: these Oberon racks specific new racks. They take a bunch 478 00:24:42,240 --> 00:24:44,199 Speaker 2: more power, and they need a bunch of liquid cooling. 479 00:24:44,480 --> 00:24:47,840 Speaker 2: You can't just retrofit easily. You have to bulldoze shit 480 00:24:47,880 --> 00:24:51,240 Speaker 2: and rebuild well, remove all the housing and then add 481 00:24:51,400 --> 00:24:54,160 Speaker 2: HVAC stuff. It's very expensive and takes a long time. 482 00:24:55,080 --> 00:24:58,480 Speaker 2: And look, I just don't know what's happening with these GPUs, 483 00:24:58,480 --> 00:25:01,000 Speaker 2: and I'm a little bit concerned, And I doubt these 484 00:25:01,000 --> 00:25:04,760 Speaker 2: are new customers. They're likely hyperscalers, neo clouds like core Weave, 485 00:25:04,800 --> 00:25:07,760 Speaker 2: and resellers like Dell and supermicro, who also both sell 486 00:25:07,800 --> 00:25:10,639 Speaker 2: to Gorewave. Because the only companies that can actually afford 487 00:25:10,760 --> 00:25:14,159 Speaker 2: to buy GPUs are those with massive amounts of cash 488 00:25:14,320 --> 00:25:17,000 Speaker 2: or debt, to the point that even Google, Amazon, Meta, 489 00:25:17,040 --> 00:25:19,520 Speaker 2: and Oracle are taken on massive amounts of new debt 490 00:25:19,840 --> 00:25:22,159 Speaker 2: or without a plan to make a profit. Oracle is 491 00:25:22,200 --> 00:25:25,000 Speaker 2: looking potentially at fifty six billion dollars of debt. It's 492 00:25:25,040 --> 00:25:29,280 Speaker 2: completely bonkers. In Video's largest customers are increasingly unable to 493 00:25:29,280 --> 00:25:32,080 Speaker 2: afford its GPUs, which appear to be increasing in price 494 00:25:32,119 --> 00:25:36,199 Speaker 2: with every subsequent generation. In Video's GPUs are so expensive 495 00:25:36,359 --> 00:25:37,960 Speaker 2: that the only way you can buy them is by 496 00:25:38,000 --> 00:25:40,600 Speaker 2: already having billions of dollars or being able to raise 497 00:25:40,640 --> 00:25:43,280 Speaker 2: billions of dollars, which means in a very real sense 498 00:25:43,520 --> 00:25:45,840 Speaker 2: that in Video is dependent not on its customers, but 499 00:25:45,920 --> 00:25:48,960 Speaker 2: on its customers credit ratings and financial backers, and the 500 00:25:49,080 --> 00:25:53,359 Speaker 2: larger private credit institutions, which I'm eventually going to have 501 00:25:53,400 --> 00:25:56,320 Speaker 2: to do a newsletter on and a podcast an because honestly, 502 00:25:56,359 --> 00:25:59,240 Speaker 2: every time I read about the private credit situation with 503 00:25:59,320 --> 00:26:03,800 Speaker 2: blue out again, I'll begin here in the wa bit 504 00:26:03,880 --> 00:26:07,560 Speaker 2: from kill bill it's not good. And to make matters worse, 505 00:26:07,600 --> 00:26:09,600 Speaker 2: the key reason that one would buy a GPU is 506 00:26:09,640 --> 00:26:12,439 Speaker 2: to either run AI services using it, or rent it 507 00:26:12,480 --> 00:26:15,040 Speaker 2: to somebody else to run AI services, and the two 508 00:26:15,119 --> 00:26:18,000 Speaker 2: largest parties spending money on these services are open Ai 509 00:26:18,040 --> 00:26:20,800 Speaker 2: and Anthropic, both of whom lose billions of dollars and 510 00:26:20,920 --> 00:26:23,760 Speaker 2: thus are much like the people buying the GPUs depending 511 00:26:23,800 --> 00:26:27,040 Speaker 2: on venture capital and debt. Now remember open ai and 512 00:26:27,080 --> 00:26:30,040 Speaker 2: Anthropic both have lines of credit, four billion dollars for 513 00:26:30,080 --> 00:26:32,560 Speaker 2: open Ai and two and a half billy for Aanthropic. 514 00:26:33,160 --> 00:26:35,560 Speaker 2: In simple terms, in Nvidia's customers rely on debt to 515 00:26:35,560 --> 00:26:38,280 Speaker 2: buy as GPUs, and in Vidia's customers customers rely on 516 00:26:38,320 --> 00:26:41,800 Speaker 2: debt to pay to rent them. Yeah, it's not great 517 00:26:42,280 --> 00:26:45,520 Speaker 2: yet it actually gets worse from there. Who after all 518 00:26:45,560 --> 00:26:48,399 Speaker 2: of the biggest customers paying the company's renting GPUs to 519 00:26:48,440 --> 00:26:51,760 Speaker 2: sell their AI bottles. That's right, AI startups, all of 520 00:26:51,800 --> 00:26:55,560 Speaker 2: which are deeply unprofitable. Cusser. Anthropic's largest customer and now 521 00:26:55,560 --> 00:26:58,680 Speaker 2: it's biggest competitor in the AI coding sphere, raised two 522 00:26:58,760 --> 00:27:01,879 Speaker 2: point three billion dollars in November after raising nine hundred 523 00:27:01,920 --> 00:27:04,520 Speaker 2: million dollars in June. But plex D, one of the 524 00:27:04,560 --> 00:27:07,280 Speaker 2: most popular right put that in their quotes, raised two 525 00:27:07,320 --> 00:27:09,960 Speaker 2: hundred million dollars in September after raising one hundred million 526 00:27:10,000 --> 00:27:12,400 Speaker 2: dollars in July, after seeming to fail to raise half 527 00:27:12,440 --> 00:27:15,400 Speaker 2: a billion dollars in May. After raising five hundred million 528 00:27:15,480 --> 00:27:18,560 Speaker 2: dollars in December twenty twenty four, Cognition raised four hundred 529 00:27:18,560 --> 00:27:21,359 Speaker 2: million dollars in September after raising three hundred million dollars 530 00:27:21,359 --> 00:27:24,120 Speaker 2: in March, and Coher raised one hundred million dollars in September, 531 00:27:24,160 --> 00:27:26,840 Speaker 2: a month after it raised five hundred million dollars. None 532 00:27:26,880 --> 00:27:30,080 Speaker 2: of these companies are profitable, not even close. I read 533 00:27:30,119 --> 00:27:32,840 Speaker 2: a story in Newcomer by Tom Datan that said the 534 00:27:33,000 --> 00:27:35,400 Speaker 2: cursor cents one hundred percent of its revenue to Anthropic 535 00:27:35,400 --> 00:27:39,880 Speaker 2: to pay for its models. Very cool. So I really 536 00:27:39,880 --> 00:27:41,719 Speaker 2: want to lay this out for you because it's very 537 00:27:41,800 --> 00:27:45,359 Speaker 2: bad when you think about So venture capital is feeding 538 00:27:45,400 --> 00:27:49,240 Speaker 2: money to startups. They fund the startups AI startups, and 539 00:27:49,280 --> 00:27:52,800 Speaker 2: then they pay either or both open ai or Anthropic 540 00:27:52,880 --> 00:27:55,520 Speaker 2: to use their models. Now open Ai and Anthropic need 541 00:27:55,600 --> 00:27:58,399 Speaker 2: to serve those models, right, So they then raised venture 542 00:27:58,400 --> 00:28:01,159 Speaker 2: capital or there to pay hypersu scalers or neoclouds to 543 00:28:01,200 --> 00:28:05,240 Speaker 2: rent Nvidia GPUs. At that point, hyper scalers and neoclouds 544 00:28:05,240 --> 00:28:07,320 Speaker 2: then use either debt or exist in cash flow in 545 00:28:07,359 --> 00:28:10,440 Speaker 2: the case of Hyperscalers, though not for long, to buy 546 00:28:10,560 --> 00:28:14,760 Speaker 2: more in Vidia GPUs. Only one company appears to make 547 00:28:14,880 --> 00:28:17,919 Speaker 2: profit here, and it's in Vidia. Well in video and 548 00:28:17,960 --> 00:28:20,320 Speaker 2: its resellers like Dell and super Micro, which buy in 549 00:28:20,400 --> 00:28:22,840 Speaker 2: Vidia GPUs, put them in service and sell them to 550 00:28:24,440 --> 00:28:27,560 Speaker 2: neo clouds like Lamter or Core with at some point 551 00:28:27,600 --> 00:28:30,399 Speaker 2: to link in. This debt back chain breaks because very 552 00:28:30,440 --> 00:28:33,359 Speaker 2: little cash flow exists to prop it up. At some point, 553 00:28:33,440 --> 00:28:36,560 Speaker 2: venture capitalists will be forced to stop funneling money into unprofitable, 554 00:28:36,680 --> 00:28:39,800 Speaker 2: unsustainable AI companies which will make those companies unable to 555 00:28:39,840 --> 00:28:42,760 Speaker 2: funnel money into the pockets of anthropic and ope and 556 00:28:42,800 --> 00:28:45,400 Speaker 2: AI who rent the GPUs will then not be able 557 00:28:45,440 --> 00:28:48,520 Speaker 2: to funnel money into the pockets of those buying GPUs, 558 00:28:49,320 --> 00:28:51,680 Speaker 2: which will make it harder for those companies to justify 559 00:28:51,680 --> 00:28:56,120 Speaker 2: buying GPUs. And at that point some of this comes 560 00:28:56,120 --> 00:28:59,400 Speaker 2: to InVideo and Invidia doesn't make so make so much money. 561 00:29:00,000 --> 00:29:02,800 Speaker 2: I'm honest, none of Nvidia's success really makes any sense. 562 00:29:03,440 --> 00:29:05,480 Speaker 2: Who's buying so many GPUs and where are they going? 563 00:29:06,360 --> 00:29:09,560 Speaker 2: Why are in Vidia's inventories increasing? Is it really just 564 00:29:10,080 --> 00:29:13,240 Speaker 2: pre buying parts for future orders? Why are their accounts 565 00:29:13,240 --> 00:29:16,360 Speaker 2: receivable climbing? And how much product is Nvidia shipping before 566 00:29:16,360 --> 00:29:19,960 Speaker 2: it gets paid? While these are both explainable as this 567 00:29:20,000 --> 00:29:22,520 Speaker 2: is a big company and this is how companies do business, 568 00:29:22,520 --> 00:29:26,640 Speaker 2: and that's true, why do receivables not seem to be 569 00:29:26,680 --> 00:29:31,440 Speaker 2: coming down? And how long, realistically can the largest company 570 00:29:31,440 --> 00:29:34,440 Speaker 2: on the stock market continue to grow revenues selling assets 571 00:29:34,440 --> 00:29:37,120 Speaker 2: that only seem to lose its customers money and don't 572 00:29:37,120 --> 00:29:41,080 Speaker 2: seem to even be in use for years. I worry 573 00:29:41,120 --> 00:29:43,040 Speaker 2: about it in Nvidia, not because I think there's a 574 00:29:43,040 --> 00:29:45,920 Speaker 2: massive scandal, but because so much rights and its success 575 00:29:45,960 --> 00:29:48,360 Speaker 2: and its success rights on the back of dwindling amounts 576 00:29:48,360 --> 00:29:51,680 Speaker 2: of venture capital, and there because nobody is actually making 577 00:29:51,800 --> 00:29:55,040 Speaker 2: money to pay for these GPUs, let alone running them. 578 00:29:55,800 --> 00:29:58,520 Speaker 2: In fact, I'm not even saying in video goes tits up. 579 00:29:58,520 --> 00:30:00,360 Speaker 2: I want to be clear about that. I think they 580 00:30:00,360 --> 00:30:02,920 Speaker 2: may even have another good quarter or two in them. 581 00:30:03,200 --> 00:30:04,880 Speaker 2: It really just comes down to how long people are 582 00:30:04,880 --> 00:30:07,080 Speaker 2: willing to be stupid and how long Jensen Wong is 583 00:30:07,120 --> 00:30:09,440 Speaker 2: able to call up Sachin Adella and Co. At three 584 00:30:09,480 --> 00:30:12,840 Speaker 2: in the morning and say, buy one billion dollars of GPUs, 585 00:30:12,880 --> 00:30:18,360 Speaker 2: you pig Finnom style baby. But really, I think much 586 00:30:18,400 --> 00:30:20,680 Speaker 2: of the US stock market's growth is held up by 587 00:30:20,680 --> 00:30:22,760 Speaker 2: how long everybody is willing to be gas lit by 588 00:30:22,800 --> 00:30:26,000 Speaker 2: Jensen Wong into believing that they need more GPUs. At 589 00:30:26,040 --> 00:30:29,600 Speaker 2: this point, it's barely about AI anymore, as AI revenue 590 00:30:29,760 --> 00:30:33,040 Speaker 2: real cash made from selling services run on those GPUs, 591 00:30:33,200 --> 00:30:36,040 Speaker 2: doesn't even cover the costs, let alone create the cash 592 00:30:36,040 --> 00:30:39,840 Speaker 2: flow necessary to buy more seventy thousand dollars GPUs thousands 593 00:30:39,880 --> 00:30:43,400 Speaker 2: at a time. It's not like any actual innovational progress 594 00:30:43,480 --> 00:30:46,680 Speaker 2: is driving this bullshit. In any case, the market's crave 595 00:30:46,680 --> 00:30:49,480 Speaker 2: are healthy in video has so many hundreds of billions 596 00:30:49,480 --> 00:30:51,920 Speaker 2: of dollars of invidious stock sits in the hands of 597 00:30:51,960 --> 00:30:54,720 Speaker 2: retail investors and people's four O one ks, and its 598 00:30:54,800 --> 00:30:57,120 Speaker 2: endless growth has helped paper over the pallid growth of 599 00:30:57,160 --> 00:31:00,520 Speaker 2: the US stock market and by extension, the decay of 600 00:31:00,560 --> 00:31:04,200 Speaker 2: the tech industry's ability to innovate. Once this pops, and 601 00:31:04,280 --> 00:31:06,600 Speaker 2: it will pop because there's simply not enough money to 602 00:31:06,600 --> 00:31:09,160 Speaker 2: do this forever, there must be a referendum on those 603 00:31:09,160 --> 00:31:11,640 Speaker 2: that chose to ignore the naked instability of this era 604 00:31:11,840 --> 00:31:15,040 Speaker 2: and the endless lies that inflate the AI bubble. I 605 00:31:15,120 --> 00:31:17,920 Speaker 2: will be walking around with a gavel. I am going 606 00:31:17,960 --> 00:31:21,280 Speaker 2: to be taking heads. I am fucking sick of this era. 607 00:31:21,520 --> 00:31:24,040 Speaker 2: And what I'm most sick of is that so few 608 00:31:24,120 --> 00:31:27,080 Speaker 2: people are still to this day willing to admit how 609 00:31:27,120 --> 00:31:30,160 Speaker 2: bad this is. And I know in the next few 610 00:31:30,160 --> 00:31:32,680 Speaker 2: months we're going to get articles some major media outlets 611 00:31:32,720 --> 00:31:34,479 Speaker 2: that say, how could we have seen this coming? And 612 00:31:34,560 --> 00:31:36,680 Speaker 2: like I said in the previous episode, they could have 613 00:31:36,720 --> 00:31:39,520 Speaker 2: fucking looked. All of them could have looked, and they 614 00:31:39,520 --> 00:31:42,360 Speaker 2: could have looked a year ago. The incredible support I 615 00:31:42,400 --> 00:31:44,800 Speaker 2: get from all of you truly makes this show a 616 00:31:44,880 --> 00:31:46,920 Speaker 2: joy to make, even though I've done way too many 617 00:31:46,960 --> 00:31:49,800 Speaker 2: retakes on this and apologies to Mattasowski for the noises 618 00:31:49,840 --> 00:31:52,600 Speaker 2: I make. But I think in the next few months 619 00:31:52,640 --> 00:31:54,800 Speaker 2: we're all going to be validated. It's going to be 620 00:31:54,800 --> 00:31:59,720 Speaker 2: the great vindication. But until then, everybody is betting billions 621 00:31:59,720 --> 00:32:02,240 Speaker 2: on the eye idea that wily coyote won't look down. 622 00:32:02,920 --> 00:32:13,680 Speaker 2: He's gonna have to at some point, won't it. Thank 623 00:32:13,680 --> 00:32:16,400 Speaker 2: you for listening to Better Offline. The editor and composer 624 00:32:16,440 --> 00:32:19,280 Speaker 2: of the Better Offline theme song is Matasowski. You can 625 00:32:19,320 --> 00:32:21,720 Speaker 2: check out more of his music and audio projects at 626 00:32:21,720 --> 00:32:25,200 Speaker 2: Mattasowski dot com, m A T T O. S O 627 00:32:25,880 --> 00:32:29,360 Speaker 2: W s ki dot com. You can email me at 628 00:32:29,360 --> 00:32:32,200 Speaker 2: easy at Better offline dot com or visit Better Offline 629 00:32:32,200 --> 00:32:34,240 Speaker 2: dot com to find more podcast links and of course, 630 00:32:34,320 --> 00:32:37,440 Speaker 2: my newsletter. I also really recommend you go to chat 631 00:32:37,480 --> 00:32:40,120 Speaker 2: dot Where's Youreed dot at to visit the discord, and 632 00:32:40,160 --> 00:32:42,880 Speaker 2: go to our slash Better Offline to check out our reddit. 633 00:32:43,640 --> 00:32:45,000 Speaker 2: Thank you so much for listening. 634 00:32:45,840 --> 00:32:48,520 Speaker 1: Better Offline is a production of Cool Zone Media. For 635 00:32:48,640 --> 00:32:51,840 Speaker 1: more from Cool Zone Media, visit our website. Cool zonemedia 636 00:32:51,880 --> 00:32:54,720 Speaker 1: dot com, or check us out on the iHeartRadio app, 637 00:32:54,760 --> 00:33:17,800 Speaker 1: Apple Podcasts, or wherever you get your podcasts.