WEBVTT - The Enshittifinancial Crisis: Part Four 0:00:03.200 --> 0:00:07.080 Media. Hello and welcome back to Better Offline. I'm your 0:00:07.120 --> 0:00:20.880 host ed ZiT Try. This is the last part of 0:00:20.920 --> 0:00:23.640 our four part series about how the financial system and 0:00:23.680 --> 0:00:28.040 the financial markets are the last victims of the inshitification process. 0:00:28.920 --> 0:00:31.400 Over the past four episodes, in a twenty thousand or 0:00:31.400 --> 0:00:34.400 so word newsletter which this series is based on, I've 0:00:34.440 --> 0:00:37.280 explained how and in this final episode, I'm going to 0:00:37.320 --> 0:00:40.479 explain how all this was going to fall apart and 0:00:40.560 --> 0:00:43.800 make it clear where I'm so fucking alarmed. Now let's 0:00:43.800 --> 0:00:46.920 get back to those data centers. We've already got some 0:00:47.640 --> 0:00:51.159 signs of concern within the banking world around them, and 0:00:51.200 --> 0:00:53.920 I'm kind of worried, and so are they about their exposure. 0:00:54.440 --> 0:00:57.240 In November, the ft reported the Deutsche Bank, which back 0:00:57.320 --> 0:00:59.920 cor we've multiple times and several data centers, was the 0:01:00.080 --> 0:01:03.000 exploring ways the hedge it's exposure to the AI data 0:01:03.040 --> 0:01:06.520 centers after extending billions of dollars in debt, including shortening 0:01:06.560 --> 0:01:09.880 a basket of AI related stocks or buying default protection 0:01:09.920 --> 0:01:12.880 on some of its debt using synthetic risk transfers, which 0:01:12.920 --> 0:01:15.760 are when a bank sells the full or partial credit 0:01:15.800 --> 0:01:17.920 risk of a loan or loans to another bank or 0:01:18.040 --> 0:01:20.080 keeping the loans on their book paying a monthly fee 0:01:20.120 --> 0:01:23.560 to investors. This is a simplification. Please don't be mad 0:01:23.600 --> 0:01:28.320 at me. In December, Fortune reported the Morgan Stanley three 0:01:28.360 --> 0:01:31.720 times with core Weave IPI partners, Hyperions, Soft Banks Bridge 0:01:31.720 --> 0:01:35.560 Loan was also considering synthetic risk transfers on loans to 0:01:35.640 --> 0:01:39.119 businesses involved in AI infrastructure. One of what those were 0:01:39.160 --> 0:01:43.600 for now. Back in April, SMBC sold synthetic risk transfers. 0:01:44.640 --> 0:01:48.640 Back in April, SMBC sold synthetic risk transfers tied to 0:01:48.680 --> 0:01:52.280 private debt BDCs. And while this predates the larger data 0:01:52.320 --> 0:01:55.080 center deals done by Blue Al, who is a BDC themselves, 0:01:55.440 --> 0:01:59.600 smbc's oversee multiple blue ol deals in the past. In December, 0:01:59.800 --> 0:02:04.040 s MBC closed another SRT selling off risk from Australian 0:02:04.080 --> 0:02:07.720 and Asian project finance loans, though we can't confirm if 0:02:07.720 --> 0:02:10.480 any of those were data center related now. In December, 0:02:10.520 --> 0:02:13.640 Goldman Sachs pause the planned mortgage bond sale for data 0:02:13.639 --> 0:02:17.079 Center Operate A Cyrus one, with the intent to revive 0:02:17.120 --> 0:02:18.919 it in the first quarter of twenty twenty six. I 0:02:18.919 --> 0:02:22.240 have heard nothing from that since Oracle's credit risk reaches 0:02:22.280 --> 0:02:24.880 sixteen year high in the middle of December, with credit 0:02:24.919 --> 0:02:28.200 default swaps basically betting the Oracle will default on its debt, 0:02:28.200 --> 0:02:31.320 which is unlikely yet no longer impossible, climbing to their 0:02:31.400 --> 0:02:36.079 highest price since the Great Financial Crisis. While Morgan Stanley 0:02:36.120 --> 0:02:38.720 and Deutsche Banks SRTs are yet to close, it's still 0:02:38.760 --> 0:02:40.880 notable that two of the largest players in data center 0:02:40.919 --> 0:02:44.600 financing feel the need to hedge their bets. So what 0:02:44.720 --> 0:02:48.320 exactly are they hedging against? Simple that tenants won't arrive 0:02:48.400 --> 0:02:51.440 and debts won't get paid. I also believe that they're 0:02:51.440 --> 0:02:54.160 going to need bigger hedges because I don't think there's 0:02:54.240 --> 0:02:58.200 enough actual AI demand to meet the data center's being built, 0:02:58.240 --> 0:03:01.000 and I think most data center loans up being underwater 0:03:01.040 --> 0:03:03.920 within the next two years. To be clear, this isn't 0:03:03.919 --> 0:03:06.799 a speculative chance, but rather one wrong on the chain 0:03:06.840 --> 0:03:09.320 of pain that's been forged and looks like it's ready 0:03:09.360 --> 0:03:12.320 to snap. Let's start by quoting my Premium newsletter from 0:03:12.320 --> 0:03:15.920 December fifth. While many people talk about how circular the 0:03:15.960 --> 0:03:18.400 AI bubble may or may not be, the reality is 0:03:18.440 --> 0:03:20.919 that it's far more like a chain, a deeply vulnerable 0:03:20.919 --> 0:03:24.160 one held together by debt and venture capital. A company 0:03:24.160 --> 0:03:26.799 buys a GPU from Nvidia, at which point nobody is 0:03:26.840 --> 0:03:29.679 making any profit anymore. These GPUs are purchased for the 0:03:29.760 --> 0:03:33.160 most part using debt provided by banks or financial institutions, 0:03:33.520 --> 0:03:36.720 while hyperscalers like Microsoft can and do fund GPUs using 0:03:36.760 --> 0:03:39.160 cash flow even they have started to turn to debt. 0:03:40.400 --> 0:03:42.520 At that point, the company that bought the GPUs six 0:03:42.680 --> 0:03:45.160 hundreds of millions of dollars or even billions of dollars 0:03:45.240 --> 0:03:47.560 to build out a data center, and once it turns on, 0:03:48.000 --> 0:03:51.240 provides compute to a model provider such as open ai Oranthropic, 0:03:51.440 --> 0:03:54.320 or even a smaller company, which then begins losing money 0:03:54.360 --> 0:03:57.600 selling access to those GPUs, or of course training models 0:03:57.600 --> 0:04:00.920 on them, which only loses money eat sample. Both open 0:04:00.960 --> 0:04:03.560 ai and Anthropic loose billions of dollars, and both rely 0:04:03.680 --> 0:04:06.400 on venture capital to fund their ability to continue paying 0:04:06.640 --> 0:04:10.960 for accessing those GPUs. At that point, open ai and 0:04:10.960 --> 0:04:14.160 Anthropic offer either subscriptions, which cost far more to offer 0:04:14.160 --> 0:04:16.760 than the revenue they provide, or API access to their 0:04:16.800 --> 0:04:20.800 models on a per million token basis. These rates are 0:04:20.839 --> 0:04:24.760 often subsidized too. AI startups pay to access those models 0:04:24.800 --> 0:04:27.000 to run their services, which end up costing more than 0:04:27.040 --> 0:04:29.360 the revenue that they make from offering them, which means 0:04:29.400 --> 0:04:31.720 they have to raise venture capital to continue paying to 0:04:31.760 --> 0:04:35.240 access those models. Outside of hyperscalers paying in video for 0:04:35.279 --> 0:04:38.080 GPUs out of cash flow, None of the AI industry 0:04:38.120 --> 0:04:42.400 is fueled by revenue. Every single part of the industry 0:04:42.680 --> 0:04:45.800 is fueled by a kind of subsidy, debt, or equity. 0:04:46.680 --> 0:04:49.040 As a result, the AI bubble is really a stress 0:04:49.080 --> 0:04:52.640 test of the global venture capital, private equity, private credit, 0:04:52.880 --> 0:04:56.200 institutional and banking system and its willingness to fund all 0:04:56.240 --> 0:04:58.640 of this forever. Because there isn't a single generative AI 0:04:58.720 --> 0:05:02.560 company that's got a path to profit stability. You see, 0:05:02.680 --> 0:05:05.200 every little link in the chain of pain is necessary 0:05:05.200 --> 0:05:09.280 to understand things. The intitified stock market pumped not by 0:05:09.360 --> 0:05:12.760 actual cash flow or productivity, by which I mean products 0:05:12.760 --> 0:05:16.239 that are selling Google, Microsoft, Amazon matter they don't actually 0:05:16.279 --> 0:05:19.520 generate real revenue and definitely not profit from AI, but 0:05:19.560 --> 0:05:22.400 by signals read by analysts and investors trained over decades 0:05:22.400 --> 0:05:25.080 the push consumers to invest in the Magnificent seven stocks. 0:05:25.560 --> 0:05:27.800 These stocks represent as much as forty percent of the 0:05:27.839 --> 0:05:30.279 value of the S and P five hundred and their values, 0:05:30.279 --> 0:05:32.520 as I said, were pumped by analysts and the media, 0:05:32.560 --> 0:05:37.240 misleading investors into believing that revenue growth for companies like Microsoft, Meta, Amazon, 0:05:37.279 --> 0:05:39.880 and Google have anything to do with AI versus the 0:05:39.920 --> 0:05:42.520 growth of their regular services. The other part in the 0:05:42.600 --> 0:05:45.440 chain is venture capital's liquidity crisis, one peaking at a 0:05:45.440 --> 0:05:48.360 time when startups have become more capital intensive than any 0:05:48.400 --> 0:05:51.120 other point in history, and the underlying value of their 0:05:51.160 --> 0:05:54.560 investments is only valuable as long as these companies stay alive. 0:05:54.960 --> 0:05:57.599 And then, of course, the ballooning centralized data center debt 0:05:57.640 --> 0:06:00.680 bubble funded based on customer contracts, will built for demand 0:06:00.680 --> 0:06:03.800 that does not exist, finding massive data centers of GPUs 0:06:03.800 --> 0:06:06.799 that immediately become commoditized as a result of this hysteria. 0:06:07.480 --> 0:06:10.719 In really simple terms, I believe that almost every investment 0:06:10.720 --> 0:06:13.200 in a data center AI startup may go to zero. 0:06:14.000 --> 0:06:17.000 Let me explain if we assume that half of the 0:06:17.040 --> 0:06:19.120 one hundred and seventy one point five billion dollars of 0:06:19.200 --> 0:06:21.799 data center debt is in GPUs, so eighty five billion 0:06:21.920 --> 0:06:24.360 or so, that's about three point two gigawats of data 0:06:24.360 --> 0:06:26.960 set capacity. Based on my model of in Videa's approximate 0:06:27.000 --> 0:06:30.719 split of sales between different aigpus from a premium priest 0:06:30.760 --> 0:06:33.520 from a few weeks ago. As a brief explainer, I 0:06:33.560 --> 0:06:36.039 made the model by taking in videos data center revenues, 0:06:36.160 --> 0:06:38.920 estimating how many of each GPU they sell, and dividing 0:06:38.960 --> 0:06:41.400 those revenues by those numbers, and markeplying the result by 0:06:41.400 --> 0:06:43.880 the power output of the GPUs. In really simple terms, 0:06:43.960 --> 0:06:45.880 I worked out how many GPUs they sold them what 0:06:46.160 --> 0:06:48.040 the power to run them would be. There'd be about 0:06:48.040 --> 0:06:51.880 three point two gigawts. The likelihood of the majority of 0:06:51.880 --> 0:06:54.640 these products being a completed within the next year and 0:06:54.680 --> 0:06:58.440 b completed on budget is very very small. Every delay 0:06:58.520 --> 0:07:01.839 increases the likelihood off, as each of these projects is 0:07:01.880 --> 0:07:05.080 heavily debt based. I also want to add open Ai claims, 0:07:05.080 --> 0:07:07.800 and I actually think they're bullshitting here that they had 0:07:07.839 --> 0:07:11.240 two gigawatts of compute at the end of twenty twenty five, 0:07:12.200 --> 0:07:15.679 two gigawatts. They are the only company with any real 0:07:15.680 --> 0:07:19.000 consumer exposure. They're the only ones with any number of 0:07:19.040 --> 0:07:22.520 real users other than Google, who has literally changed Google 0:07:22.520 --> 0:07:25.640 a system for Gemini. I don't think there's even like 0:07:26.080 --> 0:07:29.440 five hundred megawatts of real demand outside of people training 0:07:29.480 --> 0:07:33.280 models or Elon Musk's s SAM Generator, and that's a 0:07:33.320 --> 0:07:37.280 whole different story now. In general, the customers of these 0:07:37.280 --> 0:07:40.120 projects are either hyperscalers who are only doing AI because 0:07:40.120 --> 0:07:42.240 they have no other hyper growth ideas and because Wall 0:07:42.280 --> 0:07:46.040 Street currently approves, or AI startups, all of whom are unprofitable. 0:07:47.160 --> 0:07:49.520 While there are potentially hedge funds or other small companies 0:07:49.560 --> 0:07:52.120 looking for private AI integrations. I think this is a 0:07:52.200 --> 0:07:55.400 very very small market. On top of that, AI compute 0:07:55.400 --> 0:07:58.240 itself may not be profitable. And because by my estima, 0:07:58.520 --> 0:08:01.240 everybody is spent about eighty five billion dollars on filling 0:08:01.320 --> 0:08:04.360 data centers with the same GPUs, the aggregate price of 0:08:04.400 --> 0:08:08.080 renting out GPUs will decline. Already, the average price of 0:08:08.080 --> 0:08:10.600 renting out Blackwell GPUs is declined to an average of 0:08:10.680 --> 0:08:13.200 four dollars and forty five cents an hour according to 0:08:13.240 --> 0:08:16.280 Silicon Data, and that's before the majority of black Well 0:08:16.320 --> 0:08:20.120 powered GPUs come online. Yeah, the customer base shrinks from 0:08:20.160 --> 0:08:23.760 there because the majority of AI startups aren't actually renting GPUs. 0:08:24.080 --> 0:08:26.600 They build products on top of models built by Open 0:08:26.680 --> 0:08:29.520 Ai or Anthropic, who have made it clear they're buying 0:08:29.520 --> 0:08:32.640 capacity from either hyperscalers, or, in open AI's case, getting 0:08:32.679 --> 0:08:35.280 Oracle or core Weave to build it for them. Why 0:08:35.559 --> 0:08:39.400 Because building your own model is incredibly capital intensive, both 0:08:39.440 --> 0:08:41.160 on the compute and the data you have to get, 0:08:41.520 --> 0:08:43.080 and it's hard to tell if the results will be 0:08:43.080 --> 0:08:47.440 worth it. Now, let's assume I don't actually believe it will, 0:08:47.440 --> 0:08:49.640 but let's try anyway that all of that three point 0:08:49.640 --> 0:08:55.000 two gigawatsop capacity comes online? How much of it? How much? Actually? 0:08:55.640 --> 0:08:58.319 How much compute does an AI company use? Like I said, 0:08:58.360 --> 0:09:00.920 open Ai ended twenty twenty five to two gig what's 0:09:00.960 --> 0:09:03.720 of capacity? And they apparently have nine hundred million weekly 0:09:03.760 --> 0:09:06.480 active users. I don't think there are any AI companies 0:09:06.480 --> 0:09:08.760 with even ten percent of that user base. But even 0:09:08.760 --> 0:09:10.840 if they were, open ai spent eight point six seven 0:09:10.880 --> 0:09:14.160 billion dollars on inference through the end of September. Who 0:09:14.240 --> 0:09:16.200 can afford to pay even ten percent of that a 0:09:16.240 --> 0:09:19.840 year or five percent? Yeah? In reality, open Ai is 0:09:19.920 --> 0:09:22.360 likely more indicative of the overall compute spend of the 0:09:22.520 --> 0:09:25.960 entire AI industry. As I've said, most companies are powered 0:09:25.960 --> 0:09:28.240 not by their own GPU driven models, but by renting 0:09:28.240 --> 0:09:31.920 them from model providers like open Ai. Themselves. Open Ai 0:09:31.960 --> 0:09:34.200 and and Thropic Spenter combined eleven point three to three 0:09:34.200 --> 0:09:37.760 billion dollars in compute on Zoro AWS, respectively through the 0:09:37.760 --> 0:09:41.120 first nine months of last year. And that's my own reporting, 0:09:41.160 --> 0:09:43.640 by the way, and as the two largest consumers of 0:09:43.679 --> 0:09:48.160 AI compute, Well, I'm a little worried. These things suggest 0:09:48.280 --> 0:09:51.839 two things. The market for AI compute is actually very 0:09:51.960 --> 0:09:54.800 very small. If you assume that Anthropics spent the same 0:09:54.840 --> 0:09:57.520 amount on Google Cloud as it did in AWS, that'd 0:09:57.559 --> 0:09:59.959 be about a total of five point three two billion dollars, 0:10:00.000 --> 0:10:02.680 and that Core Weaves revenue called it five billion dollars, 0:10:02.960 --> 0:10:05.839 most of which was open Ai via Microsoft or in video. 0:10:06.160 --> 0:10:09.000 That doesn't appear to be an AI compute market outside 0:10:09.040 --> 0:10:13.120 of servicing two companies. The other problem is the market 0:10:13.160 --> 0:10:15.680 for AI compute is not actually growing. In the last 0:10:15.920 --> 0:10:19.480 two years, no new major customers of AI compute have emerged. 0:10:19.880 --> 0:10:22.840 Every company that assigned a large compute deal has either 0:10:22.840 --> 0:10:26.920 been open Ai, Anthropic or a hyperscaler. Even if Cursor 0:10:27.080 --> 0:10:29.840 and AI Coding Company were to dump its entire two 0:10:29.920 --> 0:10:32.680 point three billion dollars into funding AI compute, that would 0:10:32.679 --> 0:10:35.640 still not be very much. In fact, it would take 0:10:35.720 --> 0:10:39.000 sinking every single dollar of venture capital over two hundred 0:10:39.000 --> 0:10:42.280 billion dollars every single year, and then some funneled into 0:10:42.280 --> 0:10:44.960 AI compute just to produce the revenue to justify these 0:10:45.040 --> 0:10:48.800 deals existence. In the space of a year, Microsoft Zure 0:10:48.840 --> 0:10:52.000 made seventy five billion dollars, Google Cloud forty three billion dollars, 0:10:52.000 --> 0:10:56.040 and Amazon Web Services one hundred billion dollars. Now that 0:10:56.280 --> 0:10:59.160 is just add those numbers together, it's a little over 0:10:59.160 --> 0:11:04.319 two hundred billion dollars, and that's the three largest providers 0:11:04.360 --> 0:11:06.640 of any kind of compute, not just AI compute. Most 0:11:06.679 --> 0:11:09.160 of that is not AI compute. In fact, they barely 0:11:09.160 --> 0:11:10.880 make anything on that. They won't even talk about. The 0:11:10.920 --> 0:11:24.000 AI revenue is probably because they're so stinky. Well, if 0:11:24.000 --> 0:11:26.400 you need more proof, If you still don't believe me, 0:11:26.760 --> 0:11:29.560 skip to page eighteen of Nvidia's most recent earnings and 0:11:29.600 --> 0:11:32.800 I quote multi year cloud service agreement commitments as of 0:11:32.840 --> 0:11:35.640 October twenty six, twenty twenty five, were twenty six billion dollars, 0:11:35.760 --> 0:11:38.520 for which one billion dollars, six billion dollars, six billion dollars, 0:11:38.520 --> 0:11:41.079 five billion dollars, four billion dollars, and four billion dollars 0:11:41.320 --> 0:11:43.920 will we paid in fiscal years twenty twenty six, fourth quarter, 0:11:44.120 --> 0:11:47.040 twenty twenty seven, twenty twenty eight, twenty twenty nine, twenty 0:11:47.080 --> 0:11:51.520 thirty and twenty thirty one and thereafter respectively. Now I 0:11:51.559 --> 0:11:53.880 know that was confusing, but just to line that up, 0:11:53.920 --> 0:11:57.559 that means in fiscal year twenty twenty seven, which begins Oh, 0:11:57.600 --> 0:12:01.400 I would have to look that up. Yeah, I'm literally 0:12:01.440 --> 0:12:03.720 doing this on the air, and I'm gonna shit. So 0:12:03.760 --> 0:12:07.080 their fiscal year begins in February, So basically twenty twenty 0:12:07.080 --> 0:12:11.120 seven begins this February twenty twenty six. So they're going 0:12:11.160 --> 0:12:14.320 to be spending six billion dollars on AI compute in 0:12:14.360 --> 0:12:16.320 twenty twenty seven. This is a company that does not 0:12:16.440 --> 0:12:20.680 rent out compute. If there's such incredible surging demand, why 0:12:20.720 --> 0:12:23.520 is it? Why exactly is in Vidia spending six billion 0:12:23.559 --> 0:12:25.720 fucking dollars a year in twenty twenty six and twenty 0:12:25.760 --> 0:12:28.400 twenty seven on it. In Vidia doesn't need compute, it 0:12:28.520 --> 0:12:31.720 just shut down. It's Amazon Web Services. Compared to those 0:12:31.720 --> 0:12:34.160 it's called DJX cloud. It looks far more like in 0:12:34.240 --> 0:12:37.040 video is propping up an industry with non existent demand. 0:12:37.920 --> 0:12:41.120 I'm afraid there is no secret Amazon Web Services size 0:12:41.160 --> 0:12:43.480 spend waiting in the wings for the right moment to pounce. 0:12:43.800 --> 0:12:46.360 There is no secret demand wave, nor is there any 0:12:46.360 --> 0:12:50.480 capacity crunch that is holding back incredible swarths of revenue oracles. 0:12:50.480 --> 0:12:52.880 Five hundred and twenty three billion dollars in remaining performance 0:12:52.920 --> 0:12:56.720 obligations are made up of OpenAI meta and you'll never 0:12:56.800 --> 0:13:03.560 guess who in video fucking in VideA. The AI bubble 0:13:03.679 --> 0:13:06.240 is just in Vidia being handed money and then handing 0:13:06.280 --> 0:13:08.560 it to people. They're very profitable as well. They do 0:13:08.640 --> 0:13:12.960 grame everyone else not so much. For AI data center 0:13:13.040 --> 0:13:15.280 deals to make sense, most startups would have to start 0:13:15.320 --> 0:13:18.199 becoming direct users of AI compute while also spending more 0:13:18.240 --> 0:13:22.079 on cloud compute services than they've ever spent. The largest 0:13:22.080 --> 0:13:26.640 consumers of AI compute are both unprofitable unsustainable monstrosities dependent 0:13:26.720 --> 0:13:30.880 on venture capital. Eventually, reality will dawn on one or 0:13:30.920 --> 0:13:33.680 more of these banks. Projects will get delayed thanks to 0:13:33.720 --> 0:13:36.959 weather or budgetary issues, or when customers walk away, as 0:13:37.040 --> 0:13:40.720 just happened to a data center ariit called FERMI, and 0:13:40.840 --> 0:13:44.679 loan payments will start going unpaid elsewhere. AI startups will 0:13:44.679 --> 0:13:47.440 start asking for money again and again, and for a 0:13:47.440 --> 0:13:49.960 while they'll keep raising until the evaluations get too high, 0:13:50.040 --> 0:13:52.839 or VC coffers get too low. You're probably going to 0:13:52.840 --> 0:13:54.880 say at this point that Open AI or Anthropic might 0:13:54.920 --> 0:13:57.520 go public, which will infuse capital into the system, and 0:13:57.559 --> 0:13:59.520 I want to give you a preview of what that 0:13:59.640 --> 0:14:03.480 might look like. Courtesy of AI Labs, Minimax, and Zipu, 0:14:03.520 --> 0:14:06.840 as reported by the Information which just filed to go Bubblica. 0:14:07.760 --> 0:14:09.400 I'm a little scared of These are the first real 0:14:09.480 --> 0:14:11.440 numbers about the AI BUBBLA. Hope, I'm not going to 0:14:11.440 --> 0:14:14.960 get embarrassed. Everybody's saying that behind the scenes these companies 0:14:14.960 --> 0:14:19.200 are secretly great. Let's take a look. Oh my god, 0:14:19.480 --> 0:14:22.760 these numbers aren't great at all. These numbers are terrible. 0:14:23.160 --> 0:14:25.480 In the first half of this of twenty twenty five, 0:14:25.680 --> 0:14:27.760 Zipu had a net loss of three hundred and thirty 0:14:27.760 --> 0:14:31.920 four million dollars on twenty seven million dollars of revenue. Meanwhile, 0:14:31.960 --> 0:14:34.720 Minimax made fifty three point four million dollars in revenue 0:14:34.720 --> 0:14:36.800 in the first nine months of twenty twenty five and 0:14:36.840 --> 0:14:40.080 bun two hundred and eleven million dollars to earn it. 0:14:40.080 --> 0:14:42.840 It's time to wake up. These are the real life 0:14:42.880 --> 0:14:46.359 costs of running an AI company, Open Ai and Anthropic 0:14:46.400 --> 0:14:49.560 are going to be even worse than this. This is 0:14:49.600 --> 0:14:52.600 why nobody wants to take AI company's public This is 0:14:52.640 --> 0:14:54.680 why nobody wants to talk about the actual costs of 0:14:54.760 --> 0:14:58.000 running AI. This is why nobody wants you to know 0:14:58.040 --> 0:15:00.680 the hourly cost of running a GPU, And this is 0:15:00.720 --> 0:15:04.000 why open Ai and Anthropic burn billions and billions of dollars. 0:15:04.240 --> 0:15:07.880 The margins fucking stink. Every product is unprofitable and none 0:15:07.920 --> 0:15:10.520 of these companies can afford their bills based on their 0:15:10.560 --> 0:15:14.360 actual cash flow. Generative AI is not a functional industry, 0:15:14.400 --> 0:15:18.960 and once the money works that out, everything burns. Though 0:15:19.000 --> 0:15:22.400 many AI data centers boast of having tendency agreements, remember 0:15:22.560 --> 0:15:24.720 these agreements are either with AI startups that will run 0:15:24.720 --> 0:15:27.360 out of money who really shouldn't be renting GPUs by 0:15:27.400 --> 0:15:30.680 the way, or hyperscalers with legal teams numbering in the thousands. 0:15:31.160 --> 0:15:35.080 Every single deal that Microsoft, Amazon, Meta, Googler, and Video 0:15:35.200 --> 0:15:39.640 signs is riddled withouts specifically hedging against this scenario, and 0:15:39.680 --> 0:15:41.920 there won't be a damn thing that anybody can do 0:15:42.000 --> 0:15:45.400 about it. If Hyperscalers decide to walk away before then, 0:15:45.560 --> 0:15:48.360 in Video's bubble is likely to burst. As I discussed 0:15:48.360 --> 0:15:50.400 a few weeks ago, and Vidia claims to have shipped 0:15:50.480 --> 0:15:52.960 six million Blackwell GPUs, and while it may be employing 0:15:53.080 --> 0:15:55.960 very dodgy maths claiming that each GPU is actually to 0:15:56.360 --> 0:16:01.920 chips Jensen Jensen, I've been saying. My modeling of its 0:16:02.040 --> 0:16:04.640 last three quarters suggests that in videos shipped around five 0:16:04.680 --> 0:16:07.640 point three to three gigawatts worth of GPUs, and based 0:16:07.680 --> 0:16:10.520 on reading about every single data center deal I can find, 0:16:10.800 --> 0:16:13.480 it doesn't appear that many have been built and powered on, 0:16:14.400 --> 0:16:18.640 or even started building in many cases. Worse still, in 0:16:18.720 --> 0:16:22.040 Video's diversified revenue is collapsing. In the first quarter of 0:16:22.040 --> 0:16:24.960 its fiscal year twenty twenty six, I'm sorry, you're just 0:16:25.000 --> 0:16:27.040 going to have to ride with me on this one, 0:16:27.160 --> 0:16:30.400 two customers represented sixteen percent and fourteen percent in revenue. 0:16:30.760 --> 0:16:32.880 In the second quarter of fiscal year twenty six, two 0:16:32.920 --> 0:16:36.240 customers represented twenty three percent and sixteen percent, and in 0:16:36.280 --> 0:16:40.160 the third quarter, four customers represented twenty two percent, fifteen percent, 0:16:40.240 --> 0:16:43.200 thirteen percent, and eleven percent of total revenue, with all 0:16:43.280 --> 0:16:46.080 of that money going towards GPUs or networking gear. In 0:16:46.120 --> 0:16:48.600 simpler terms, in Vidia's revenue is no longer coming from 0:16:48.640 --> 0:16:51.760 a diverse swarth of customers. In its first quarter fiscal 0:16:51.840 --> 0:16:54.000 year twenty twenty six, it had thirty point eight four 0:16:54.040 --> 0:16:57.200 billion dollars of diversified revenue. That's from a customer that 0:16:57.520 --> 0:16:59.680 from customers that do not number more than ten percent 0:16:59.680 --> 0:17:02.640 of its revenue, and in the second quarter of FY 0:17:02.720 --> 0:17:05.640 twenty six that was twenty eight point five billion Q three. 0:17:05.640 --> 0:17:09.600 It's twenty two point twenty three billion. In simpler terms, 0:17:10.320 --> 0:17:13.280 there are only a few companies that can buy GPUs, 0:17:13.359 --> 0:17:16.080 and other companies are no longer buying GPUs at the 0:17:16.080 --> 0:17:18.600 same rate. I doubt this number is going to increase. 0:17:19.640 --> 0:17:23.160 You see in video, GPUs are astronomically expensive four point 0:17:23.200 --> 0:17:25.320 five million dollars for a GB three hundred radc of 0:17:25.600 --> 0:17:28.520 seventy two B three hundred GPUs, for example, and feeling 0:17:28.600 --> 0:17:31.639 data centers full of them requires debt unless you're a hyperscaler. Well, 0:17:31.680 --> 0:17:34.320 I can't say for sure. I believe in videos, diversified 0:17:34.320 --> 0:17:36.959 revenue collapses a sign that smaller data center projects are 0:17:36.960 --> 0:17:40.399 starting to have issues getting funded and or hyperscalers are 0:17:40.400 --> 0:17:43.320 pulling back on the GPU purchases from those Taiwanese companies. 0:17:43.359 --> 0:17:46.640 I was mentioning, Now, let me look through the eyes 0:17:46.680 --> 0:17:49.720 of an AI boosted for a second. Okay, okay, everything 0:17:49.800 --> 0:17:52.440 is blue and yellow as usual. Okay. One might say 0:17:52.480 --> 0:17:55.080 that these big customers are covering the loss of revenue, 0:17:55.160 --> 0:17:57.280 but the reality is that these big projects are run 0:17:57.320 --> 0:17:59.920 on debt issued by banks are becoming increasingly worried about 0:18:00.080 --> 0:18:04.640 nobody paying them back. The mistake that every investor, commentator, analyst, 0:18:04.680 --> 0:18:06.879 and member of the media makes about Nvidia is believing 0:18:06.880 --> 0:18:09.520 that its sales are an expression of demand for AI compume, 0:18:09.880 --> 0:18:12.159 when in reality it's more of a statement about the 0:18:12.200 --> 0:18:16.000 availability of debt from banks in private credit. Similarly, the 0:18:16.000 --> 0:18:18.720 continued existence of AI startups is an expression of the 0:18:18.760 --> 0:18:21.960 desperation of venture capital and the continuing flow of massive 0:18:22.000 --> 0:18:24.360 funding rounds as assigned that they see no other revenues 0:18:24.400 --> 0:18:28.480 for growth and that their startups can't wipe their own assholes. Eventually, 0:18:28.600 --> 0:18:30.760 data centers are going to go unbuilt and data center 0:18:30.840 --> 0:18:34.240 debt packages will begin to fall apart. Remember, Oracle's thirty 0:18:34.240 --> 0:18:36.680 eight billion dollar data center deal is actually yet to close, 0:18:37.040 --> 0:18:39.960 much like Stargate New Mexico is yet to close, and 0:18:40.040 --> 0:18:43.040 much like Stargate Michigan is yet to close. These deals 0:18:43.080 --> 0:18:46.640 while seeming like they're trending positively, are all incredibly important 0:18:46.640 --> 0:18:49.040 to the future of the AI bubble, and any failure 0:18:49.160 --> 0:18:53.119 will spooken already notice mark them only one link in 0:18:53.200 --> 0:18:57.200 this chain of pain needs to break every single part 0:18:57.320 --> 0:19:01.400 of the AI bubble. This fucking charade is unprofitable save 0:19:01.480 --> 0:19:04.120 for in Video and the construction firms erecting future laser 0:19:04.200 --> 0:19:08.320 tag arenas full of negative margin GPUs tell me what 0:19:08.520 --> 0:19:11.560 happens if the debt stops flowing data centers? How will 0:19:11.600 --> 0:19:14.840 in Vidia sell their so called twenty million Blackwell and 0:19:14.960 --> 0:19:17.919 veried Ruben GPUs their claiming they'll ship by the end 0:19:17.960 --> 0:19:21.360 of twenty twenty six. What happens if venture capitalists start 0:19:21.440 --> 0:19:24.320 running low on funds and can't keep feeding hundreds of 0:19:24.359 --> 0:19:26.879 millions of dollars to AI startups so that they can 0:19:27.000 --> 0:19:30.439 feed them directly to open Ai or Anthropic. What happens 0:19:30.480 --> 0:19:33.639 to open Ai and Anthropic and they're already negative margin 0:19:33.720 --> 0:19:37.280 businesses when their customers can't pay them. What happens to 0:19:37.359 --> 0:19:40.080 Oracle or core weaves, work in progress data centers if 0:19:40.119 --> 0:19:43.359 open ai can't pay their bills? What happens to Anthropics? 0:19:43.400 --> 0:19:46.240 Twenty one billion dollars of Broadcom orders or tens of 0:19:46.280 --> 0:19:49.959 billions of dollars of Google cloud spend. And what if 0:19:50.000 --> 0:19:53.119 I'm right about everything I've said, not just in this series, 0:19:53.359 --> 0:19:57.159 but in the episodes and newsletters before him. In the 0:19:57.280 --> 0:20:00.040 last year, I estimate I've been asked the question what 0:20:00.160 --> 0:20:03.440 if You're wrong? Over twenty five times. Every single time 0:20:03.520 --> 0:20:06.680 the question comes with this undercurrent venom, the suggestion that 0:20:06.760 --> 0:20:09.280 I'm being an asshole for daring to question the wondrous 0:20:09.320 --> 0:20:12.919 AI bubble. Every single person who's asked me this has 0:20:12.960 --> 0:20:14.960 been poorly read, both in terms of my work and 0:20:15.000 --> 0:20:18.479 the surrounding economics and technological possibilities of large language models 0:20:18.640 --> 0:20:21.639 and believes they're defending technology, when in reality they're defending 0:20:21.720 --> 0:20:24.240 growth in the rock economy's growth at all cost mindset. 0:20:25.200 --> 0:20:28.280 In many cases, they are not excited about technology at all, 0:20:28.520 --> 0:20:30.480 but the prospects of being the first in line to 0:20:30.560 --> 0:20:34.240 lick an already sparkling boot. This has never ever been 0:20:34.280 --> 0:20:37.840 about progress or productivity. If it was, we'd seen actual 0:20:37.880 --> 0:20:42.560 progress or productivity boosts. So anything other than the frothiest debt, 0:20:42.960 --> 0:20:46.240 the frothiest venture capital markets, and the most annoying and 0:20:46.400 --> 0:20:51.159 incorrect headlines I've ever read in the news. Large language 0:20:51.200 --> 0:20:54.919 models do not create novel concepts. They are inconsistent and unreliable, 0:20:55.080 --> 0:20:57.320 and even the things that people like them for, like 0:20:57.480 --> 0:21:00.320 coding very wildly thanks to the dramatic very tents of 0:21:00.359 --> 0:21:05.040 a giant probability machine l alams, are not good enough 0:21:05.080 --> 0:21:07.919 for people to pay regular software prices at any scale, 0:21:08.119 --> 0:21:10.359 and the consequences of this will be that every single 0:21:10.440 --> 0:21:13.440 dollar spent on GPUs has been for exactly one point 0:21:14.280 --> 0:21:19.800 manipulating the value of stocks by making hyperscalers look good 0:21:20.080 --> 0:21:22.800 so that they can keep pretending that they create anything. 0:21:24.359 --> 0:21:27.040 AI does not have the business returns and may indeed 0:21:27.119 --> 0:21:32.760 have negative gross margins across the board. It is inconsistent, ugly, unreliable, 0:21:33.040 --> 0:21:37.639 expensive and environmentally ruinous, pissing off a large chunk of consumers, 0:21:37.880 --> 0:21:41.000 and underwhelming most of the rest, other than those convinced 0:21:41.000 --> 0:21:43.240