WEBVTT - CZM Rewind: The Case Against Generative AI (Part 2) 0:00:02.480 --> 0:00:07.040 Zone Media, Hello im ed Zetron, and this, of course 0:00:07.160 --> 0:00:21.000 is better offline what Welcome to the second part of 0:00:21.000 --> 0:00:24.120 our four part series, where I give you my most comprehensive, 0:00:24.320 --> 0:00:26.599 most up to day explanation of why we're in a 0:00:26.600 --> 0:00:29.640 bubble and what that even means. The reason why I'm 0:00:29.680 --> 0:00:32.280 taking my time to be descriptive and comprehensive is because 0:00:32.320 --> 0:00:34.280 I want this to make sense to those who listen 0:00:34.320 --> 0:00:37.159 to it. Having written hundreds of thousands of words this 0:00:37.280 --> 0:00:39.879 year about the AI bubble, so many of the arguments 0:00:39.880 --> 0:00:42.320 I've made and the secrets I've exposed are contained in 0:00:42.360 --> 0:00:45.320 their own discreet little episodes or newsletters. This is my 0:00:45.440 --> 0:00:48.319 series to consolidate all of the information I've put out 0:00:48.320 --> 0:00:51.040 there in one place, and I want to make it 0:00:51.080 --> 0:00:53.760 makes sense to anyone who listens to him. I want anyone, 0:00:53.840 --> 0:00:56.440 even who someone who doesn't even know that much about AI, 0:00:56.520 --> 0:00:58.360 to listen to the audience I've been making for the 0:00:58.400 --> 0:01:01.160 past three years, to understand why things are dire and 0:01:01.200 --> 0:01:03.360 to feel the same alarm I'm feeling, or at least 0:01:03.440 --> 0:01:05.800 understand why I'm alarmed. Because I don't like to tell 0:01:05.800 --> 0:01:08.120 you how you feel. Old school bit of FEEDBACKERD got 0:01:08.160 --> 0:01:10.560 from a listener once, and I appreciate that to this day. 0:01:11.840 --> 0:01:14.560 Now today I'll make the case the generative AI's fundamental 0:01:14.600 --> 0:01:17.000 growth story is flawed and explain why we're in the 0:01:17.040 --> 0:01:20.320 midst of an egregious bubble. This industry is sold by 0:01:20.400 --> 0:01:22.679 keeping things vague and knowing that most people don't dig 0:01:22.800 --> 0:01:25.679 much deeper than a headliner problem I simply do not have. 0:01:26.319 --> 0:01:29.600 This industry is effectively in service of two companies, open 0:01:29.640 --> 0:01:32.800 Ai and Nvidia, who pump headlines out through endless contracts 0:01:32.840 --> 0:01:35.760 between them or subsidiaries or investments to give the illusion 0:01:35.760 --> 0:01:39.040 of activity. Open Ai has now promised over four hundred 0:01:39.080 --> 0:01:42.000 billion dollars in the next four years, though honestly they 0:01:42.080 --> 0:01:44.119 might owe about a trillion dollars with all the data 0:01:44.120 --> 0:01:46.959 centers they're signed up for. All of these are egregious 0:01:47.000 --> 0:01:50.040 sums for a company that have already forecasted billions in 0:01:50.120 --> 0:01:52.720 losses with no clear explanation as to how it will 0:01:52.720 --> 0:01:54.920 afford any of this beyond we need more money and 0:01:54.960 --> 0:01:57.400 they're vague hope that there's another soft bank or Microsoft 0:01:57.400 --> 0:02:00.160 waiting in the wings to swoop in and save the day. Now, 0:02:00.160 --> 0:02:01.880 I'm going to walk you through where I see this 0:02:01.960 --> 0:02:04.240 industry today and why I see no future for it 0:02:04.360 --> 0:02:09.880 beyond a horrible, fiery car rack. While everybody reasonably harps 0:02:09.880 --> 0:02:12.840 on about hallucinations, which to remind you, is when a 0:02:12.880 --> 0:02:16.560 model authoritativety states something that isn't true, the truth of 0:02:16.960 --> 0:02:19.160 why that's bad is far more complex and actually far 0:02:19.200 --> 0:02:22.119 worse than it seems. You cannot rely on a large 0:02:22.200 --> 0:02:24.320 language model to do what you want, even though its 0:02:24.440 --> 0:02:27.320 highly tuned models on the most expensive and intricate platforms 0:02:27.320 --> 0:02:30.560 can't actually be relied upon to do exactly what you want. 0:02:30.880 --> 0:02:33.080 And I know some people might say, well, yes, they 0:02:33.080 --> 0:02:36.720 do every time, one hundred percent of the time. A 0:02:36.760 --> 0:02:40.040 hallucination isn't just when these models say something that isn't true. 0:02:40.280 --> 0:02:42.520 It's when they decide to do something wrong because it 0:02:42.560 --> 0:02:44.520 seems the most likely thing to do, or when a 0:02:44.560 --> 0:02:46.800 coding model decides to go on a wild goose chase, 0:02:47.200 --> 0:02:49.160 failing the user and burning a ton of money in 0:02:49.200 --> 0:02:53.000 the process. The advent of reasoning models those engineered to 0:02:53.240 --> 0:02:55.480 think through problems in the way reminiscent of a human. 0:02:55.520 --> 0:02:58.440 But it's not thinking. They don't think. They have no consciousness. 0:02:58.680 --> 0:03:01.480 They literally you ask them something and they break down 0:03:01.520 --> 0:03:04.560 what the prompt might mean and then choose it's not thinking, 0:03:04.919 --> 0:03:07.120 and the expansion of what people are trying to use 0:03:07.240 --> 0:03:10.320 llms for demands that the definition of an AI hallucination 0:03:10.480 --> 0:03:13.720 be widened, not merely referring to factual errors, but fundamental 0:03:13.800 --> 0:03:16.600 errors in understanding the user's request or intent or what 0:03:16.680 --> 0:03:20.080 constitutes a task, in part because these models, as I said, 0:03:20.120 --> 0:03:23.360 cannot think and do not know anything. However successful a 0:03:23.400 --> 0:03:25.760 model might be in generating something good once, it will 0:03:25.800 --> 0:03:29.120 also often generate something bad, or it will generate the 0:03:29.200 --> 0:03:32.120 right thing, but in an inefficient and over verbos fashion, 0:03:32.520 --> 0:03:34.640 you do not know what you're going to get each time, 0:03:34.720 --> 0:03:37.680 and hallucinations multiply with the complexity of the thing you're 0:03:37.720 --> 0:03:40.680 asking for or whether a task contains multiple steps, which 0:03:40.720 --> 0:03:43.520 is a fatal blow to the idea of agents. You 0:03:43.560 --> 0:03:46.080 can add as many levels of intrigue and reasoning as 0:03:46.080 --> 0:03:48.600 you want, but large language models cannot be trusted to 0:03:48.600 --> 0:03:51.840 do something correctly or even consistently, let alone every time. 0:03:53.040 --> 0:03:56.400 Model companies have successfully convinced everybody that the issue is 0:03:56.400 --> 0:03:58.880 that users are prompting the models wrong and that the 0:03:58.920 --> 0:04:01.000 people need to be trained to use AI. But what 0:04:01.040 --> 0:04:04.040 they're doing is training people to explain away the inconsistencies 0:04:04.080 --> 0:04:07.600 of large language models and to assume individual responsibility for 0:04:07.640 --> 0:04:10.720 what is an innate floor in how these fucking things work. 0:04:11.720 --> 0:04:16.040 Large language models are also uniquely expensive. Many mistakenly try 0:04:16.080 --> 0:04:18.080 and claim that this is like the dot com boom 0:04:18.160 --> 0:04:21.160 or uber, but the basic unique economics of generative AI 0:04:21.360 --> 0:04:25.600 are insane. Providers must purchase tens or hundreds of thousands 0:04:25.600 --> 0:04:29.360 of GPUs, each costing fifty thousand to seventy thousand apiece, 0:04:29.720 --> 0:04:32.279 and the hundreds of millions or billions of dollars of 0:04:32.320 --> 0:04:34.960 infrastructure that goes around them are so expensive and hard 0:04:34.960 --> 0:04:38.599 to install, and that's without mentioning things like staffing or construction, 0:04:38.720 --> 0:04:42.360 or power or water or even permitting. Then you turn 0:04:42.400 --> 0:04:45.359 them on and immediately they start losing new money. Despite 0:04:45.440 --> 0:04:48.320 hundreds of billions of GPUs sold, nobody seems to actually 0:04:48.360 --> 0:04:50.560 make any of it, other than Video of course, the 0:04:50.600 --> 0:04:53.120 company that makes them, and resellers like Dell and Supermicro 0:04:53.240 --> 0:04:55.360 who buy the GPUs put them in servers and sell 0:04:55.360 --> 0:04:58.200 them to other people. Now, if you're an eager listener, 0:04:58.800 --> 0:05:01.320 I would love to hear from you on one question, 0:05:01.400 --> 0:05:03.600 then this is just something that's been bouncing around my head. 0:05:03.760 --> 0:05:07.760 Super Micro is in Video a customer of super Micro. 0:05:07.800 --> 0:05:10.640 They're a custom in super Micro is a huge customer 0:05:10.640 --> 0:05:12.440 of Video. I've read something like seventy percent of the 0:05:12.560 --> 0:05:16.920 cost of goods sold is buying GPUs. But I read 0:05:16.960 --> 0:05:18.279 that in Video as a customer of them. But I 0:05:18.279 --> 0:05:21.360 can't find anything else. Reach out easy at better Offline 0:05:21.400 --> 0:05:24.039 dot com if you've got any thoughts there anyway, But 0:05:24.160 --> 0:05:26.560 back to those resourers. This arrangement works out great for 0:05:26.640 --> 0:05:29.000 Jensen Wang, the CEO of in Video, and terribly for 0:05:29.040 --> 0:05:32.600 everybody else. Today, I'm gonna explain the insanity of the 0:05:32.640 --> 0:05:35.080 situation we'll find ourselves in and why I continue to 0:05:35.120 --> 0:05:38.640 do this work underterred. The bubble has entered its most pornographic, 0:05:38.680 --> 0:05:41.440 aggressive and destructive stage, where the more obvious it becomes 0:05:41.440 --> 0:05:43.960 that we're all cooked here in AI land, the more 0:05:44.080 --> 0:05:47.839 ridiculous the GENERATIVAI industry will act a dark juxtaposition against 0:05:47.880 --> 0:05:50.480 every new study that says generative AI does not work, 0:05:50.720 --> 0:05:54.039 or new story about chat gpd's uncannyability to activate mental 0:05:54.080 --> 0:05:57.560 illness in people. And we're going to start looking at 0:05:57.600 --> 0:06:00.520 one company in Video which now dominates the stock market 0:06:00.600 --> 0:06:03.760 and has taken extraordinary and dangerous measures to sustain growth. 0:06:03.760 --> 0:06:07.680 That is, to any sane person, completely unsustainable and unrealistic 0:06:07.720 --> 0:06:11.279 on every level. But let's start simple. Nvidia is a 0:06:11.320 --> 0:06:14.800 hardware company that sells GPUs, including consumer GPUs that you'd 0:06:14.839 --> 0:06:17.599 see in a modern gaming PC. But when you read 0:06:17.680 --> 0:06:20.520 someone say GPU within the context of AI, they mean 0:06:20.760 --> 0:06:24.640 enterprise focused GPUs like the A one hundred, h one hundred, 0:06:24.920 --> 0:06:27.719 H two hundred, and more modern GPUs like the Blackwell 0:06:27.839 --> 0:06:31.560 Series B two hundred and GB two hundred, which combines 0:06:31.560 --> 0:06:35.680 two GPUs with an Nvidia CPU. This is all complex sounding, 0:06:35.720 --> 0:06:38.320 but I want you to have the groundwork. These GPUs 0:06:38.360 --> 0:06:41.120 cost anywhere from fifty to seventy thousand dollars and require 0:06:41.160 --> 0:06:44.640 tens of thousands of dollars more of infrastructure networking to 0:06:45.080 --> 0:06:48.000 cluster these server acts of GPUs together to provide compute 0:06:48.040 --> 0:06:50.360 and massive cooling systems to deal with the massive amounts 0:06:50.360 --> 0:06:52.960 of heat they produce, as well as servers themselves that 0:06:53.000 --> 0:06:54.920 they run on, which typically use top of the line 0:06:55.000 --> 0:06:58.160 data center CPUs and contain vast quantities of high speed 0:06:58.240 --> 0:07:02.640 memory and storage. GPUs itself is likely the most expensive 0:07:02.640 --> 0:07:05.240 single it and within an AI server the other costs. 0:07:05.279 --> 0:07:07.520 And I'm not even factoring in the actual physical building 0:07:07.600 --> 0:07:10.200 that the server lives in, or the water or electricity 0:07:10.200 --> 0:07:14.440 that uses. Well, all this crap adds up. I've mentioned 0:07:14.440 --> 0:07:17.160 in Video because it has a virtual monopoly in this space, 0:07:17.720 --> 0:07:21.200 Genera IVAI effectively requires in video GPUs, in part because 0:07:21.240 --> 0:07:23.200 it's the only company really making the kinds of high 0:07:23.200 --> 0:07:26.520 powered cards that GENERAIVAI demands, and because in Video created 0:07:26.560 --> 0:07:30.160 something called Kuda. Cuda a collection of software tools that 0:07:30.240 --> 0:07:33.040 lets programmers write software that runs on GPUs, which were 0:07:33.040 --> 0:07:36.640 traditionally used primarily for rendering graphics in games. While there 0:07:36.680 --> 0:07:39.600 are some open source alternatives as well as alternatives from 0:07:39.600 --> 0:07:42.320 Intel with its Arc GPUs and AMD, in Vidia's main 0:07:42.400 --> 0:07:45.000 rival in the consumer space, these aren't nearly as mature 0:07:45.120 --> 0:07:48.600 or feature Richkuda's been around for ten fifteen years now, 0:07:49.360 --> 0:07:50.800 they really knew what they were doing. They also have 0:07:50.840 --> 0:07:52.880 bought a company called Melanox, which did the high speed 0:07:52.960 --> 0:07:57.840 networking back in twenty nineteen, I think, for six billion dollars. Anyway, 0:07:58.360 --> 0:08:00.680 due to the complexities of AI models, one cannot just 0:08:00.720 --> 0:08:02.880 stand up a few of these GPUs either you need 0:08:02.960 --> 0:08:05.840 clusters of thousands, tens of thousands, or hundreds of thousands 0:08:05.880 --> 0:08:08.520 of them for it to be worthwhile making any investment 0:08:08.560 --> 0:08:11.120 in GPUs and the hundreds of millions or billions of dollars, 0:08:11.320 --> 0:08:14.680 especially considering they require completely different data center architecture to 0:08:14.680 --> 0:08:17.840 make them run. You've probably read a bunch of stuff 0:08:18.080 --> 0:08:22.440 about cryptominers turning into AI data center providers. These crypto 0:08:22.600 --> 0:08:24.840 data centers have to be knocked down and replaced. They 0:08:25.160 --> 0:08:27.440 you can't just put the same GPUs in isn't going 0:08:27.520 --> 0:08:29.920 to work, and with the new Blackwell ones, the brand 0:08:29.960 --> 0:08:32.360 new ones, and then the Rubens following them, same deal. 0:08:33.280 --> 0:08:35.880 A common request like asking a generative AI model to 0:08:35.880 --> 0:08:37.800 pass through thousands of lines of code and make a 0:08:37.880 --> 0:08:40.800 change or in addition, may use multiples of these fifty 0:08:40.800 --> 0:08:43.880 thousand dollars GPUs at the same time. And so if 0:08:43.920 --> 0:08:46.680 you aspire to serve thousands or millions of concurrent users, 0:08:46.720 --> 0:08:50.880 you need to spend big, really really really big. It's 0:08:50.920 --> 0:08:53.439 these factors, the vendor lock in the ecosystem and the 0:08:53.440 --> 0:08:55.680 fact that generative AI really only works when you're buying 0:08:55.760 --> 0:08:59.120 GPUs at scale that underpin the rise of Nvidia. But 0:08:59.200 --> 0:09:02.120 Beyond the economic and technical factors, there are human ones too. 0:09:03.000 --> 0:09:05.680 To understand the AI bubble is to understand why CEOs 0:09:05.720 --> 0:09:08.360 do the things they do. Because the executive job is 0:09:08.440 --> 0:09:10.720 so vague, they can telegraph the value of their labour 0:09:10.760 --> 0:09:15.480 by spending money on initiatives and partnerships and stratagem. AI 0:09:15.520 --> 0:09:18.439 gave hyperscalers the excuse to spend hundreds of billions of 0:09:18.480 --> 0:09:20.920 dollars on data centers and buy a bunch of GPUs 0:09:20.920 --> 0:09:23.000 to go in them, because that, to the markets, looks 0:09:23.040 --> 0:09:26.480 like they're doing something. By virtue of spending a lot 0:09:26.520 --> 0:09:29.840 of money in a frighteningly short amount of time, Satchnadella 0:09:29.920 --> 0:09:33.040 received multiple glossy profiles, all without having to prove that 0:09:33.080 --> 0:09:35.440 AI can really do anything, be it a job or 0:09:35.480 --> 0:09:39.280 make Microsoft money. Nevertheless, AI allowed CEOs to look busy, 0:09:39.280 --> 0:09:41.160 and once the markets and journalists had agreed on the 0:09:41.160 --> 0:09:43.719 consensus opinion that AI would be big, all that these 0:09:43.760 --> 0:09:46.360 executives had to do was buy GPUs and do AI 0:09:46.960 --> 0:09:49.959 or plug AI within their own software profitts. But really 0:09:49.960 --> 0:09:52.120 it was just jump on the big stupid asshole train. 0:10:02.800 --> 0:10:04.760 We are in the midst of one of the darkest 0:10:04.840 --> 0:10:08.840 forms of software in history, described by many as unwanted guest, 0:10:08.960 --> 0:10:12.480 invading their products, their social media feeds, their bosses, empty minds, 0:10:12.520 --> 0:10:15.680 and resting in the hands of monsters. Every story of 0:10:15.720 --> 0:10:18.840 AI's success feels bereft of any real triumph, with every 0:10:18.920 --> 0:10:22.120 literal description of its abilities involving multiple caveats about the 0:10:22.160 --> 0:10:24.840 mistakes it makes or the incredible costs of running in 0:10:25.520 --> 0:10:28.320 generator of AI really exists for two reasons, to cost 0:10:28.400 --> 0:10:31.560 money and to make executives look busy. It was meant 0:10:31.600 --> 0:10:33.640 to be the new enterprise software and the new I 0:10:33.760 --> 0:10:36.160 phone and the new Netflix all at once, a panacea 0:10:36.200 --> 0:10:38.760 where the software guys pay one hardware guy for GPUs 0:10:38.800 --> 0:10:42.320 to unlock the incredible value creation of the future. In 0:10:42.360 --> 0:10:44.360 many ways, Jenera iv AI was always set up to 0:10:44.360 --> 0:10:46.440 fail because it was meant to be Everything. Was talked 0:10:46.440 --> 0:10:49.720 about like it was everything. It's still sold like it's everything. 0:10:49.800 --> 0:10:51.920 Yet for all the fucking hype, it comes down to 0:10:52.120 --> 0:10:56.360 two companies, Open Ai and Nvidia, and in Video was 0:10:56.480 --> 0:10:59.160 for a while living high on the hog. All CEO 0:10:59.240 --> 0:11:01.319 Jensen Huang had to do every three months would say, 0:11:01.400 --> 0:11:03.840 check out these numbers and the markets and business journalist 0:11:03.840 --> 0:11:06.480 would squeer with glee, even as he said stuff like 0:11:06.679 --> 0:11:08.560 the more you buy, the more you save, in part 0:11:08.600 --> 0:11:10.880 tipping his head to the very real and sensible idea 0:11:10.920 --> 0:11:13.720 of accelerated computing, but frame within the context of the 0:11:13.720 --> 0:11:17.440 cash inferna that's generated AI, and it all seems kind 0:11:17.480 --> 0:11:21.480 of fucking ludicrous. Huang showmanship worked really well for Invidia 0:11:21.559 --> 0:11:23.960 for a while because for a while the growth was easy. 0:11:24.240 --> 0:11:28.200 Everybody was buying GPUs, Meta, Microsoft, Amazon, Google, and to 0:11:28.240 --> 0:11:30.840 a lesser extent, Apple and Tesla made up forty two 0:11:30.880 --> 0:11:33.440 percent of Invidia's revenue, creating at least for the first 0:11:33.520 --> 0:11:36.280 four a degree of shared mania where everybody justified buying 0:11:36.280 --> 0:11:38.400 tens of billions of dollars of GPUs by saying the 0:11:38.440 --> 0:11:41.240 other guy's doing it. This is one of the major 0:11:41.280 --> 0:11:44.120 reasons the AI bubble is happening, because people conflated in 0:11:44.200 --> 0:11:47.840 Vidia's incredible sales with interest in AI rather than everybody 0:11:47.840 --> 0:11:51.920 buying GPUs at once. Don't worry, I'll explain the revenue 0:11:51.920 --> 0:11:54.080 side a little bit later. We're here for the long haul. 0:11:54.120 --> 0:11:57.079 Sit down, get comfort. You're gonna need to be anyway, 0:11:57.320 --> 0:12:00.200 Invidia is now facing a big problem that the only 0:12:00.240 --> 0:12:04.079 thing that grows forever is cancer. On September ninth, twenty 0:12:04.120 --> 0:12:06.360 twenty five, The Wall Street Journal said that in Vidio's 0:12:06.400 --> 0:12:09.520 wow factor was fading, going from beating analyst estimates by 0:12:09.520 --> 0:12:12.080 nearly twenty one percent in its fiscal year Q two 0:12:12.200 --> 0:12:15.120 twenty twenty four earnings to scraping by with a pathetic 0:12:15.360 --> 0:12:17.679 measly one point five to two percent beat in its 0:12:17.679 --> 0:12:20.880 most recent earnings, something that for any other company would 0:12:20.880 --> 0:12:23.200 be a good thing because they made so much money, 0:12:23.240 --> 0:12:27.600 But framed against the delusional expectations that generative AI has inspired, well, 0:12:27.640 --> 0:12:30.120 the figure looks nothing short of ominous. I quote the 0:12:30.120 --> 0:12:33.600 Wall Street Journal. Already, in Vidia's fifty six percent annual 0:12:33.679 --> 0:12:36.240 revenue growth rate in its latest quarter was its slowest 0:12:36.240 --> 0:12:39.120 in more than two years. If analyst projections hold, growth 0:12:39.160 --> 0:12:43.280 will slow further in the current quarter. In any other scenario, 0:12:43.360 --> 0:12:45.320 fifty six percent year of a year growth would lead 0:12:45.320 --> 0:12:48.079 to an abundance of dom perignon and one signing hundreds 0:12:48.080 --> 0:12:50.200 of boobs. But this is in video, and that's just 0:12:50.240 --> 0:12:52.959 not good enough. Back in February twenty twenty four, in 0:12:53.040 --> 0:12:55.360 Video was booking two hundred and sixty five percent year 0:12:55.400 --> 0:12:57.800 of a year growth but in its February twenty twenty 0:12:57.840 --> 0:12:59.920 five earnings in Video only grew by a measly per 0:13:00.080 --> 0:13:03.679 thetic disgusting seventy eight percent year of a year. I'm 0:13:03.720 --> 0:13:06.599 being sarcastic, of course. It isn't so much that in 0:13:06.720 --> 0:13:09.520 Vidia isn't growing, but to grow year over year at 0:13:09.559 --> 0:13:12.920 the rates that people expect is insane. Life was a 0:13:12.920 --> 0:13:15.640 lot easier when in Video went from six point oh 0:13:15.679 --> 0:13:18.319 five billion dollars in revenue in Q four fiscal year 0:13:18.360 --> 0:13:21.200 twenty twenty five to twenty two billion dollars in revenue 0:13:21.200 --> 0:13:24.000 in Q four fiscal year twenty twenty four. But for 0:13:24.080 --> 0:13:26.000 it to grow even fifty five percent year of a 0:13:26.080 --> 0:13:28.360 year from Q two FY twenty twenty six, I'm just 0:13:28.400 --> 0:13:30.880 going to truncate that now, which is forty six point 0:13:30.880 --> 0:13:33.960 seven billion dollars to Q two twenty twenty seven. That 0:13:34.000 --> 0:13:36.080 would require them to make seventy two point three eight 0:13:36.160 --> 0:13:38.560 five billion dollars in revenue in the space of three months, 0:13:38.559 --> 0:13:41.000 mostly from selling GPUs, which make up about eighty eight 0:13:41.000 --> 0:13:43.400 percent of its revenue. Just want to be clear that 0:13:44.360 --> 0:13:47.080 in a year they would have to make seventy two 0:13:47.160 --> 0:13:51.080 billion dollars just selling pretty much GPUs and the associated 0:13:51.080 --> 0:13:55.320 hardware in the space of three months. It's it's insane. 0:13:55.360 --> 0:13:57.959 This is this is really that it's too much. It's 0:13:58.000 --> 0:14:00.400 too much to expect. And this, by the way, we 0:14:00.400 --> 0:14:02.400 would put in video in the ballpark of Microsoft, who 0:14:02.400 --> 0:14:04.720 made seventy six billion dollars in their last quarterly earnings, 0:14:04.760 --> 0:14:06.839 and within the neighborhood of Apple, who made ninety four 0:14:06.880 --> 0:14:09.160 billion dollars in their last quarter of earnings. And they 0:14:09.160 --> 0:14:11.760 would do this predominantly making money in an industry that 0:14:11.800 --> 0:14:13.760 a year and a half ago barely made the company 0:14:13.800 --> 0:14:17.120 six billion dollars in a quarter. And the market needs 0:14:17.160 --> 0:14:19.640 in Vidia to perform. They must, they must, as the 0:14:19.680 --> 0:14:21.760 company makes up seven to eight percent of the value 0:14:21.800 --> 0:14:24.240 of the S and P five hundred. It's not enough 0:14:24.280 --> 0:14:26.440 for Invidia to be wildly profitable, or to having a 0:14:26.480 --> 0:14:29.240 monopsony on selling GPUs, or for it to have effectively 0:14:29.280 --> 0:14:33.840 ten x their stock in a few years. No, no, no, more, more, more, 0:14:34.000 --> 0:14:36.960 always more. Number must go up. It must continue to 0:14:37.000 --> 0:14:39.920 grow at the fastest rate of anything ever, making more 0:14:39.960 --> 0:14:42.160 and more money, selling more and more of these GPUs 0:14:42.200 --> 0:14:45.920 to a small group of companies that immediately start losing money. 0:14:46.000 --> 0:14:49.400 The moment they plugged them in. It's not brilliant, is it. 0:14:49.880 --> 0:14:51.760 While a few members of the Magnificent Seven could be 0:14:51.800 --> 0:14:53.960 depended on to funnel tens of billions of dollars into 0:14:53.960 --> 0:14:56.960 a furnace each quarter, there were limits even for companies 0:14:57.000 --> 0:14:59.160 like Microsoft, which had brought over four hundred and eighty 0:14:59.160 --> 0:15:02.160 five thousand g us in twenty twenty four alone. To 0:15:02.200 --> 0:15:04.680 take a step back about how people actually make money 0:15:04.800 --> 0:15:08.440 from buying these GPUs, companies like Microsoft, Google, and Amazon 0:15:08.480 --> 0:15:11.280 make their money by either selling access to large language 0:15:11.320 --> 0:15:14.120 models that people incorporate into their products, or by renting 0:15:14.120 --> 0:15:17.080 out servers full of those GPUs to run influence the thing, 0:15:17.200 --> 0:15:20.360 to generate the output, or train AI models for companies 0:15:20.400 --> 0:15:23.600 that develop and market their models themselves, namely Anthropic and 0:15:23.640 --> 0:15:26.480 Open AI, with some smaller competitors that don't really matter 0:15:27.200 --> 0:15:30.680 that latter revenue stream. Renting out GPUs is where Jensen 0:15:30.720 --> 0:15:33.880 Wong found a solution to that horrible eternal growth problem 0:15:34.400 --> 0:15:39.760 the neo cloud, namely companies like core Weave, Lambda, and Nebius. Now, 0:15:39.800 --> 0:15:43.080 these businesses are fairly straightforward. They own or lease data 0:15:43.080 --> 0:15:45.520 centers that they then feel full of servers that are 0:15:45.560 --> 0:15:47.960 full of Nvidio GPUs, which they then rent out on 0:15:48.000 --> 0:15:51.120 an hourly basis to customers either on a per GPU basis, 0:15:51.120 --> 0:15:53.760 are in large badges for large customers who guarantee they'll 0:15:53.800 --> 0:15:55.920 use a certain amount of compute and sign up for 0:15:55.920 --> 0:15:58.640 a long term agreement for moderan an era a time 0:15:58.720 --> 0:16:03.000 couple years perhaps these larger commitments. A neocloud is a 0:16:03.040 --> 0:16:06.440 specialist cloud compute company that exists only to provide access 0:16:06.440 --> 0:16:10.560 to GPUs for AI, Unlike Amazon Web Services, Microsoft Azure, 0:16:10.640 --> 0:16:13.080 and Google Cloud, all of which to have healthy businesses 0:16:13.120 --> 0:16:15.440 selling other kinds of compute with AI, as I'll get 0:16:15.480 --> 0:16:18.120 into later, failing to provide much of a return on 0:16:18.160 --> 0:16:21.000 investment at all. It's not just the fact that these 0:16:21.000 --> 0:16:24.440 companies are more specialized than say AWS or Azure, as 0:16:24.480 --> 0:16:27.040 you've gathered from the name. These are new, young, and 0:16:27.120 --> 0:16:30.680 in almost all cases incredibly precarious businesses, each with financial 0:16:30.720 --> 0:16:34.400 circumstances that will make a Greek finance minister blush. That's 0:16:34.440 --> 0:16:37.520 because setting up a neocloud is expensive, even if the 0:16:37.520 --> 0:16:40.320 company in question already has data centers as core We've 0:16:40.320 --> 0:16:44.240 did with its cryptocurrency mining operation, AI requires, as I said, 0:16:44.280 --> 0:16:47.600 completely new data center infrastructure to run and call the GPUs, 0:16:47.960 --> 0:16:50.160 and those GPUs also need paying for. And then there's 0:16:50.200 --> 0:16:52.400 the other stuff I mentioned earlier like power, water and 0:16:52.440 --> 0:16:54.520 the other bits of the computer. CPU are the bah 0:16:54.640 --> 0:16:57.520 blah blah blah blah blah. As a result, these neoclouds 0:16:57.520 --> 0:16:59.720 are forced to raise billions of dollars in debt, which 0:16:59.720 --> 0:17:03.000 they can lateralized using the GPUs they already have, along 0:17:03.000 --> 0:17:05.520 with contracts from customers, which they then use to buy 0:17:05.640 --> 0:17:10.040 more GPUs. That's right, they buy GPUs from Nvidia. They 0:17:10.440 --> 0:17:13.000 raised dead on those GPUs and then they used that 0:17:13.040 --> 0:17:16.040 debt to my more GPUs from Nvidia. It's enough to 0:17:16.080 --> 0:17:19.200 drive a man insane. Corwave, for example, has twenty five 0:17:19.200 --> 0:17:22.040 billion dollars in debt on an estimated five point three 0:17:22.160 --> 0:17:24.520 five billion dollars of revenue in twenty twenty five, losing 0:17:24.680 --> 0:17:27.600 hundreds of billions of dollars per quarter. Now you know 0:17:27.680 --> 0:17:31.480 who also invests in these neoclouds. You'll never guess. It's 0:17:31.520 --> 0:17:35.160 in video. Invidia is also one of Corweed's largest customers, 0:17:35.160 --> 0:17:37.800 accounting for fifteen percent of its revenue in twenty twenty four, 0:17:38.000 --> 0:17:40.160 and just signed a deal to buy six point three 0:17:40.200 --> 0:17:43.240 billion dollars of any capacity that core Weave can't otherwise 0:17:43.280 --> 0:17:46.480 sell to someone else through twenty thirty two, an extension 0:17:46.480 --> 0:17:48.560 of a one point three billion dollar twenty twenty three 0:17:48.640 --> 0:17:51.440 deal reported by the Information. It was also the anchor 0:17:51.480 --> 0:17:55.480 investment in Corewey's IPO about two hundred and fifty million dollars. 0:17:55.840 --> 0:17:58.000 In Vidia is currently doing the same thing with Lambda, 0:17:58.080 --> 0:18:01.000 another neocloud that Invidia invested in, which also plans to 0:18:01.000 --> 0:18:03.520 go public next year. In Vidia is also one of 0:18:03.640 --> 0:18:06.520 Landa's largest customers, signing a deal with it this summer 0:18:06.560 --> 0:18:09.520 to rent ten thousand GPUs for one point three billion 0:18:09.520 --> 0:18:13.239 dollars over four years in the UK. And Video has 0:18:13.240 --> 0:18:16.200 also just invested seven hundred million dollars in n Scale, 0:18:16.320 --> 0:18:19.119 a former cryptomner that has never built an AI data 0:18:19.160 --> 0:18:22.840 center that that has despite having no experience, committed one 0:18:22.920 --> 0:18:26.320 billion dollars and or one hundred thousand GPUs to an 0:18:26.320 --> 0:18:30.080 open AI data center in Norway. On Thursday, September twenty fifth, 0:18:30.480 --> 0:18:32.960 Nscale announced that it closed another funding round within Video 0:18:33.040 --> 0:18:35.200 listed as the main backer. Although it's unclear how much 0:18:35.200 --> 0:18:37.320 money it put in, it would be a safe to 0:18:37.400 --> 0:18:40.520 assume it's probably at least one hundred million dollars. In 0:18:40.640 --> 0:18:44.280 Vidia also invested in Nebius, an outgrowth of Russian conglomerate Yandex, 0:18:44.280 --> 0:18:46.800 and Nebbeus provides, through their partnership with Nvidia, tens of 0:18:46.840 --> 0:18:49.399 thousands of dollars of compute credits to companies in Videa's 0:18:49.480 --> 0:18:53.800 Inception startup program. Look, in Vidia's plan is simple, fund 0:18:53.800 --> 0:18:56.680 these neoclouds. Let these neoclouds load themselves up with debt, 0:18:56.840 --> 0:18:59.399 with which point they buy bunches of GPUs from Video, 0:18:59.560 --> 0:19:01.760 which can be used as collateral for loans along with 0:19:01.800 --> 0:19:04.440 contracts and customers, allowing the neoclouds to buy even more 0:19:04.480 --> 0:19:10.320 GPUs from Nvidia. It is just that simple. It's infinite money, right, 0:19:10.440 --> 0:19:14.480 just money, me money. Now, you fund the company, the 0:19:14.520 --> 0:19:16.680 company buys from you, you fund them again. They've used 0:19:16.720 --> 0:19:18.480 the thing they bought to buy from more from you, 0:19:18.960 --> 0:19:23.560 unlimited money. Except that is for ones more problem. These 0:19:23.560 --> 0:19:27.000 companies don't They don't really appear to have that many customers, 0:19:27.000 --> 0:19:43.840 and they don't appear to be making much money. As 0:19:43.840 --> 0:19:46.760 I went into a recent premium newsletter, and Video funds 0:19:46.760 --> 0:19:49.720 and sustains neoclouds as a way of funneling revenue to itself, 0:19:49.920 --> 0:19:52.480 as well as partners like super Micro and Dell resellers 0:19:52.560 --> 0:19:55.200 that take in video GPUs like I mentioned and put 0:19:55.240 --> 0:19:57.480 them in service to sell pre built to customers. These 0:19:57.480 --> 0:19:59.920 two companies made up thirty nine percent of in video 0:20:00.119 --> 0:20:05.600 revenues last quarter. Yet when you remove hyperscalar revenue Microsoft, Amazon, Google, 0:20:05.640 --> 0:20:08.400 Open Ai, and n Video from the revenues of these neoclouds, 0:20:08.480 --> 0:20:11.240 there's barely one billion dollars in revenue can bind across 0:20:11.280 --> 0:20:15.120 core Weave, Nebius and Lambda. Corwev's five point three five 0:20:15.200 --> 0:20:17.600 billion dollars in revenue is predominantly made up with its 0:20:17.600 --> 0:20:20.919 contracts with Nvidia, Microsoft, who are offering that compute to 0:20:20.960 --> 0:20:24.320 open Ai, Google who have hired Corewave to offer compute 0:20:24.359 --> 0:20:26.680 to open Ai. And I'm not kidding, and of course 0:20:26.720 --> 0:20:29.600 open Ai itself, which has now promised core Weave twenty 0:20:29.600 --> 0:20:31.960