WEBVTT - How The AI Bubble Bursts 0:00:02.480 --> 0:00:07.440 All Zone Media. Hello, and welcome to Better Offline Live 0:00:07.560 --> 0:00:09.760 and direct, coming at you from the middle of nowhere 0:00:10.080 --> 0:00:12.840 the center of everywhere, and I'm your host ed zetron. 0:00:21.440 --> 0:00:26.760 What else it feels like forever since I talked directly 0:00:26.840 --> 0:00:29.440 to you, my dear listeners. So I'll get right to it. 0:00:29.680 --> 0:00:32.200 And as a reminder, I'll be including links to everything 0:00:32.240 --> 0:00:35.040 I'm talking about in the episode notes. And if you're 0:00:35.080 --> 0:00:37.840 just catching up, all you really need to know here 0:00:38.120 --> 0:00:41.400 is that open ai makes large language models like GPT 0:00:41.560 --> 0:00:43.920 and of course the product chat GPT and throp it 0:00:44.000 --> 0:00:47.680 makes claude a similar product, and that generative AI products 0:00:47.680 --> 0:00:50.080 have yet to really prove a use case that justifies 0:00:50.080 --> 0:00:55.720 them losing money on every single transaction. Let's begin. August second, 0:00:55.800 --> 0:00:58.560 twenty twenty four was Black Friday for the AI boom, 0:00:58.640 --> 0:01:00.680 as a week of rough earnings from big tech led 0:01:00.720 --> 0:01:04.600 to what felt like the entire media industry asking one question, 0:01:05.319 --> 0:01:08.840 is the AI bubble popping? And that's the question I'm 0:01:08.840 --> 0:01:10.880 going to try and answer for you today and in 0:01:10.920 --> 0:01:14.640 the next episode two, the Guardians sought to answer why 0:01:14.640 --> 0:01:16.919 the big seven tech companies had been hit with AI 0:01:16.959 --> 0:01:20.680 boomed outs. CNN asked has the AI bubble burst, and 0:01:20.720 --> 0:01:23.720 The Atlantic suggested several months too late that the generative 0:01:23.720 --> 0:01:27.960 AI revolution may indeed be a bubble. The Financial Times 0:01:27.959 --> 0:01:31.959 reported hedge fund Elliott Management told investors that Nvidia was 0:01:32.040 --> 0:01:35.000 a bubble, and Bloomberg reported that big tech had failed 0:01:35.040 --> 0:01:39.039 to convince Wall Street that AI was paying off. It's 0:01:39.080 --> 0:01:41.640 all part of a growing trend where people are suddenly 0:01:41.680 --> 0:01:45.720 realizing what I've been saying for months, that this unprofitable, 0:01:45.840 --> 0:01:49.560 energy hungry technology that creates mediocre outputs is not actually 0:01:49.600 --> 0:01:53.800 the future. It's just another cloud do hickey. I had 0:01:53.800 --> 0:01:55.840 not a great way of putting it, I realized, but 0:01:56.760 --> 0:01:59.480 this is my nice way of saying. Generative AI isn't 0:01:59.520 --> 0:02:03.440 completely useless, but the large language models they kind of 0:02:03.480 --> 0:02:06.480 are based on the costs. These articles were for the 0:02:06.480 --> 0:02:09.120 most part, talking about the sudden and violent declines and 0:02:09.160 --> 0:02:11.960 the share prices of companies like Microsoft, Amazon, and Google, 0:02:12.200 --> 0:02:15.440 which for the most part haven't really recovered and as 0:02:15.480 --> 0:02:18.840 of about August tenth, are down significantly over the past month. 0:02:19.720 --> 0:02:22.680 Although these articles didn't say anything particularly new or reveal 0:02:22.720 --> 0:02:25.760 any individual missteps or scandals that might have prompted the slide, 0:02:26.040 --> 0:02:28.520 they hinted to a broader awareness among Wall Street that 0:02:28.560 --> 0:02:31.600 the AI ambitions of these companies and by which I 0:02:31.639 --> 0:02:35.600 really mean generative AI, will require these massive upfront investments 0:02:36.120 --> 0:02:38.960 and the payoff might not n actually be there, and 0:02:39.040 --> 0:02:41.440 at least if it is, it's not going to be 0:02:41.480 --> 0:02:44.919 there for a while. We're talking years, if not a decade, 0:02:44.960 --> 0:02:47.600 and once the narrative gets settled, is very, very very 0:02:47.639 --> 0:02:51.680 hard to move in. While a stock market isn't always rational, 0:02:51.800 --> 0:02:54.560 check out Tesla for more. The one day sell off 0:02:54.600 --> 0:02:57.200 of these stocks and the comments from analysts and industry 0:02:57.200 --> 0:02:59.880 figures that followed, suggest that Wall Street is growing in 0:03:00.000 --> 0:03:02.680 increasingly uncomfortable with the vast amounts of money required to 0:03:02.680 --> 0:03:06.760 build and grow generative AI into whatever the fuck that's 0:03:06.800 --> 0:03:10.600 meant to be. These companies, while still making over ten 0:03:10.639 --> 0:03:13.760 billion dollars in profit, of course, referring to Microsoft, Google, Amazon, 0:03:13.800 --> 0:03:16.440 and the like in the last quarter alone, have also 0:03:16.560 --> 0:03:19.320 spent an absolute shit ton of money on infrastructure to 0:03:19.360 --> 0:03:24.840 capture the so called demand for cloud services from generative AI. However, 0:03:25.840 --> 0:03:28.280 one little problem, none of them seem to actually be 0:03:28.400 --> 0:03:32.320 making that much money from the thing they're investing in. 0:03:32.320 --> 0:03:36.120 In the last fiscal year, Microsoft's capex the capital expenditures 0:03:36.680 --> 0:03:39.200 was about fifty five point seven billion dollars, which is 0:03:39.280 --> 0:03:41.960 up seventy five percent year over year, with more than 0:03:42.000 --> 0:03:45.240 one third nineteen billion dollars spent in the last quarter 0:03:45.360 --> 0:03:48.880 and in June thirtieth, twenty twenty four. This is reportedly 0:03:48.880 --> 0:03:52.360 split fifty to fifty between infrastructure and technology, which suggests 0:03:52.360 --> 0:03:55.360 an aggressive data center build out, with Chief financial officer 0:03:55.400 --> 0:03:59.280 for Microsoft Amyhood saying that the company expects capital expenditures 0:03:59.280 --> 0:04:02.280 to increase on a sequential basis, giving cloud and the 0:04:02.400 --> 0:04:05.480 and I quote AI demand that, as I've repeatedly said, 0:04:05.600 --> 0:04:09.840 is not really there. And what exactly does increasing on 0:04:09.880 --> 0:04:14.760 a sequential basis mean in dollar terms? I have absolutely 0:04:14.800 --> 0:04:17.520 no idea. It's vague and perhaps vague enough to rattle 0:04:17.520 --> 0:04:20.320 the markets, particularly when Microsoft will be starting from an 0:04:20.360 --> 0:04:24.640 already quite high valuation. We're still, Hood added, and I 0:04:24.720 --> 0:04:29.400 quote Microsoft's earnings calles that AI related spend represented nearly 0:04:29.640 --> 0:04:32.760 all of our total capital expenditures, with roughly half for 0:04:32.839 --> 0:04:35.880 infrastructure needs that will support monetization over the next fifteen 0:04:35.920 --> 0:04:40.240 years and beyond ugh. In essence, Microsoft spent nineteen billion 0:04:40.279 --> 0:04:42.760 dollars in the last quarter on cloud and AI expenses 0:04:42.880 --> 0:04:44.800 and has made it clear that it's not done spending 0:04:44.839 --> 0:04:48.160 more money than it's ever spent before on a technology 0:04:48.200 --> 0:04:50.760 that neither makes Microsoft nor the people paying them that 0:04:50.920 --> 0:04:55.479 much money. It's very stupid and for context, Microsoft made 0:04:55.520 --> 0:04:58.760 twenty two point zero four billion dollars in profits in 0:04:58.839 --> 0:05:02.520 Q two twenty twenty five. Is this really worth sinking 0:05:02.520 --> 0:05:05.880 an entire quarters worth of profits into? Let me give 0:05:05.880 --> 0:05:08.400 it another way to look at it. Microsoft's net profit 0:05:08.440 --> 0:05:10.679 margin has dropped from thirty nine point four four percent 0:05:10.720 --> 0:05:13.000 in Q three twenty twenty three to thirty four point 0:05:13.240 --> 0:05:15.680 four percent in Q two twenty twenty four, meaning that 0:05:15.680 --> 0:05:18.800 it's taking home less money than it usually would because 0:05:18.800 --> 0:05:21.359 they're really investing in this thing that only loses money. 0:05:21.400 --> 0:05:24.080 It's very good, it's very good that this is happening. 0:05:24.400 --> 0:05:27.640 This is what you love to see. The other cloud 0:05:27.640 --> 0:05:32.040 providers aren't really doing much better. Google's capital expenditures are 0:05:32.120 --> 0:05:35.280 forecast to be fifty billion dollars in twenty twenty four, 0:05:35.440 --> 0:05:37.799 and it's spent eleven billion dollars in Q four twenty 0:05:37.800 --> 0:05:41.040 twenty three, driven by and I quote mostly by technical 0:05:41.040 --> 0:05:45.640 infrastructure meaning servers and data centers, and twelve billion dollars 0:05:45.920 --> 0:05:47.920 was what they spent in Q one twenty twenty four. 0:05:48.640 --> 0:05:51.080 The reason I'm not breaking things out much with Google 0:05:51.160 --> 0:05:55.000 is because it hasn't been extremely guarded about its AI expenses, 0:05:55.480 --> 0:05:57.960 probably because they're really high and not making any money 0:05:58.000 --> 0:06:01.400 from it. Amazon's similarly guarded, with its capex last year 0:06:01.440 --> 0:06:04.760 hanging somewhere around forty eight point four billion dollars. It's 0:06:04.800 --> 0:06:07.360 spent thirty point five billion dollars so far in twenty 0:06:07.360 --> 0:06:10.960 twenty four, an absolutely ridiculous amount considering its profit for 0:06:11.040 --> 0:06:13.040 Q two twenty twenty four was thirteen point four to 0:06:13.040 --> 0:06:16.880 eight billion. They're just sinking their profits into these things, 0:06:17.360 --> 0:06:21.520 and there's really no sign that anything changes. Every hyperscaler 0:06:21.560 --> 0:06:24.000 has said that they intend to keep spending all of 0:06:24.040 --> 0:06:26.719 this money on AI, and I haven't even mentioned companies 0:06:26.760 --> 0:06:29.240 like Oracle, which expects to spend ten billion dollars on 0:06:29.240 --> 0:06:32.200 cloud infrastructure this year, with much of that new capacity 0:06:32.200 --> 0:06:35.520 going to support Microsoft and Google Cloud. I don't know 0:06:35.600 --> 0:06:38.200 what happens if they don't need it anymore. Anyone think 0:06:38.240 --> 0:06:43.560 about anyway anyway, But no Black Friday. It was a 0:06:43.560 --> 0:06:46.159 collective realization of the scale and the cost of AI 0:06:46.520 --> 0:06:48.760 and the first signs that the markets are starting to 0:06:48.760 --> 0:06:52.360 ask those annoying little questions about whether it's actually worth it. 0:06:53.240 --> 0:06:57.240 Yet the real chaos, and one that comparatively speaking slipped 0:06:57.279 --> 0:06:59.719 under the radar, came in the form of one of 0:06:59.760 --> 0:07:03.279 my pale horses. Last month, I put out an episode 0:07:03.320 --> 0:07:05.679 called top Culture where I suggested that the first signs 0:07:05.680 --> 0:07:08.120 of the AI bubbles collapse would be in the failure 0:07:08.120 --> 0:07:10.880 of a major AI company, though one not operating at 0:07:10.880 --> 0:07:14.480 the scale of say Open Ai, and I specifically picked 0:07:14.520 --> 0:07:16.960 one out character dot Ai, which raised one hundred and 0:07:17.000 --> 0:07:19.800 fifty million dollars in funding, and the information had already 0:07:19.880 --> 0:07:23.640 hinted might sell itself to one of the big tech companies. Now. 0:07:23.640 --> 0:07:25.680 The reason I picked them is that their app which 0:07:25.680 --> 0:07:28.960 allows you to talk to AI chatbots of Elon Musk 0:07:29.000 --> 0:07:34.200 and Staragojo from jiu Jitsu Kaizen, the anime manga which 0:07:34.240 --> 0:07:37.800 we all know I'm tired of waiting for. That's not 0:07:37.880 --> 0:07:40.440 a business, by the way, and they never really had 0:07:40.440 --> 0:07:43.400 a sustainable business model or really a meaningful product. And 0:07:43.440 --> 0:07:45.400 there one hundred and fifty million dollar raise was the 0:07:45.440 --> 0:07:48.600 first big stupid capital raise in the generative AI boom. 0:07:48.720 --> 0:07:51.360 Early twenty twenty three, when nobody else could raise at all. 0:07:52.600 --> 0:07:55.720 In my opinion, their continued existence, other than being a 0:07:55.800 --> 0:07:59.600 disgusting insult to company building and startups everywhere, was proof 0:07:59.600 --> 0:08:02.800 that the bubble existed, and their death or whatever form 0:08:02.920 --> 0:08:05.240 this really is, is a sign that the tolerance for 0:08:05.320 --> 0:08:09.120 bullshit is leaving the market, as my blunt force for 0:08:09.160 --> 0:08:12.360 shadowing suggests. On Friday August second, and yes, most of 0:08:12.400 --> 0:08:15.360 this stuff happened. On that Friday, Google said that it 0:08:15.360 --> 0:08:19.120 would license Character AI's technology and hire the company's leaders, 0:08:19.160 --> 0:08:22.360 Noam Shazia and Daniel Defritas, along with their research team 0:08:22.400 --> 0:08:25.640 of thirty people, to work at Google's DeepMind AI division. 0:08:26.320 --> 0:08:28.440 The fates of the other one hundred and forty employees 0:08:28.480 --> 0:08:32.160 remain uncertain. I assume they're alive, but I imagine some 0:08:32.240 --> 0:08:34.960 will remain on as staff as the character dot ai 0:08:35.000 --> 0:08:39.160 app remains operational. It's a strange, sad end for a 0:08:39.160 --> 0:08:42.000 company that never really had any business existing, and in 0:08:42.040 --> 0:08:44.679 many ways feels like a con happening in broad daylight. 0:08:45.520 --> 0:08:48.680 Shazir and Defritas are both former Google employees, having left 0:08:48.679 --> 0:08:51.800 the company in twenty twenty one to create Character dot ai. 0:08:52.200 --> 0:08:55.640 It's unclear whether given their previous employment, they'll be required 0:08:55.640 --> 0:08:58.080 to wear the new GLO propeller hat that you get 0:08:58.320 --> 0:09:01.120 if you're new to Google. I actually don't know if 0:09:01.160 --> 0:09:03.400 they still do that. Email me is easy at better 0:09:03.480 --> 0:09:05.760 offline dot com. That's the letter E than the letter 0:09:05.840 --> 0:09:09.360 Z or Z for my British and Canadian fans. If 0:09:09.400 --> 0:09:11.959 you know about this propeller or indeed can get me one, 0:09:12.000 --> 0:09:14.920 I would very much like one anyway. While this is 0:09:14.960 --> 0:09:18.319 being framed as a typical licensing employee poaching deal where 0:09:18.600 --> 0:09:22.400 nothing acquisition adjacent has happened, it is actually an acquisition. 0:09:22.840 --> 0:09:26.520 Google's paying two point five billion dollars to investors. Employee 0:09:26.520 --> 0:09:29.200 stock options will vest, and that means when you get 0:09:29.200 --> 0:09:31.120 stock in a company, it usually takes time for you 0:09:31.240 --> 0:09:33.440 to actually earn it, so that you can't just take 0:09:33.480 --> 0:09:36.160 it and leave it immediately, as I would, And that 0:09:36.240 --> 0:09:39.080 will keep vesting until July twenty twenty six at the 0:09:39.120 --> 0:09:41.520 acquisition rate of eighty eight dollars a share paid for 0:09:41.679 --> 0:09:43.800 by the money from the licensing deal. That is not 0:09:43.880 --> 0:09:47.120 an acquisition from Google. But after that point your shit 0:09:47.200 --> 0:09:49.800 out of luck. It's no longer guaranteed. This is a 0:09:50.120 --> 0:09:53.959 terrible situation. All around for everyone other than the earliest investors. 0:09:53.960 --> 0:09:57.600 It's another great situation where the bad guys win. In 0:09:57.720 --> 0:10:00.360 essence anyone who has options that aren't fully there. By 0:10:00.440 --> 0:10:02.880 July twenty twenty six, maybe shit out of luck and 0:10:02.960 --> 0:10:06.480 character AI's original funding valued them a billion dollars, making 0:10:06.480 --> 0:10:09.200 this a situation where and Resent Horowitz aw a one 0:10:09.240 --> 0:10:11.480 point five x return on their investment in a company 0:10:11.480 --> 0:10:15.319 that never really did anything. Employees kind of got screwed. 0:10:15.360 --> 0:10:19.400 Maybe they didn't, It's unclear, But what is clear is 0:10:19.400 --> 0:10:22.840 that as of now, character ai is effectively dead. The 0:10:22.880 --> 0:10:25.640 company will shift from using their own models their literal 0:10:25.720 --> 0:10:28.960 only thing that they did, to publicly available open source 0:10:29.000 --> 0:10:32.600 large language models like metas Lama, and the original engineering 0:10:32.679 --> 0:10:35.719 team is effectively gone. The company will likely shamble a 0:10:35.760 --> 0:10:38.160 long life lessly until it curls up in the corner 0:10:38.200 --> 0:10:40.439 like an old cat, Except an old cat has more 0:10:40.480 --> 0:10:44.640 purpose and meaning in the universe than the bullshit chatbot company. 0:10:45.800 --> 0:10:49.440 So why does this matter? On one level, the death 0:10:49.480 --> 0:10:52.440 of character dot ai is an indicator of the unsustainability 0:10:52.440 --> 0:10:55.800 of many of these generative AI applications and companies. Even 0:10:55.840 --> 0:10:58.160 with one hundred and fifty million dollars in funding, which 0:10:58.200 --> 0:11:00.880 is a decent amount of money for a normal company. 0:11:01.440 --> 0:11:04.240 Character likely couldn't keep the lights on for very long, 0:11:04.600 --> 0:11:06.440 which kind of made it necessary for them to be 0:11:06.520 --> 0:11:10.240 absorbed or acquired before they shut down and embarrassed everyone involved. 0:11:11.280 --> 0:11:15.400 It's also just like a Charlatan's Olympics. It's this company sucked. 0:11:15.440 --> 0:11:19.000 It never made money, it didn't really do anything that different, 0:11:19.960 --> 0:11:23.000 but they just get bailed out. These fucking start up people. 0:11:23.000 --> 0:11:25.439 They always go on about the meritocracy. They're always going 0:11:25.480 --> 0:11:28.559 on about, Oh, it's about working hard, It's not about 0:11:28.559 --> 0:11:32.960 your circumstances. These jack holes, these fucking idiots. They got 0:11:32.960 --> 0:11:35.640 to sell their bullshit company to Google, a company they 0:11:35.720 --> 0:11:38.680 left the found it. It's just so annoying. And this 0:11:38.800 --> 0:11:42.560 agreement also strongly resembles an earlier one between Inflection AI 0:11:42.880 --> 0:11:46.760 and Microsoft, where Microsoft bought Inflections technology and team without 0:11:47.000 --> 0:11:50.520 actually acquiring its equity, so the actual stock in the company. 0:11:51.080 --> 0:11:54.480 Founded by DeepMind co founder mister Fassuliman and LinkedIn co 0:11:54.600 --> 0:11:57.480 founder Reid Hoffman in twenty twenty two, Inflection was, in 0:11:57.600 --> 0:12:00.959 Layman's terms, intended to be a more emotion aware version 0:12:00.960 --> 0:12:04.360 of CHATGBT just one year after its creation, they'd raised 0:12:04.400 --> 0:12:07.360 one point three billion dollars at evaluation of four billion. 0:12:07.920 --> 0:12:11.400 As I noted previously, I believe that Microsoft's acquisition of 0:12:11.400 --> 0:12:14.800 Inflection was an attempt to avoid regulatory scrutiny by structuring 0:12:14.800 --> 0:12:17.600 an acquisition as a transfer of tech and talent, rather 0:12:17.640 --> 0:12:21.480 than buying the company where the tech and the talent live. 0:12:22.040 --> 0:12:24.599 And yet I'm also convinced that Inflection would struggle to 0:12:24.679 --> 0:12:27.440 last out on its own, even with its smaller user 0:12:27.480 --> 0:12:29.960 base and with a one point three billion dollar war chest, 0:12:29.960 --> 0:12:32.640 and eventually it would have had to get acquired or 0:12:32.760 --> 0:12:37.679 just die. Regardless, I'm beginning to see it patten AI's 0:12:37.720 --> 0:12:42.280 point of failure is shifting, or more accurately centralizing. At 0:12:42.280 --> 0:12:46.120 one point, the burden of this entire generative AI fast 0:12:46.240 --> 0:12:48.840 was shouldered by a large and disparate group of startups 0:12:48.840 --> 0:12:51.840 and investors, and it's now moving to a few shoulders, 0:12:51.880 --> 0:12:55.000 those of giants like Microsoft, Google, Amazon, Meta and to 0:12:55.120 --> 0:12:59.080 a much lesser extent, Apple. The two most prevalent large 0:12:59.160 --> 0:13:02.840 language models outside of course of Metaslama open AI's GPT 0:13:02.960 --> 0:13:06.880 and Anthropics claud are effectively big tech welfare recipients receiving 0:13:06.920 --> 0:13:09.400 billions of dollars in cloud credits to run their extremely 0:13:09.440 --> 0:13:12.839 expensive models without having to build their own infrastructure. It's 0:13:12.840 --> 0:13:15.960 also kind of a con because that's not real money, 0:13:16.240 --> 0:13:19.679 that's chuckie geese tokens, and how are those valued. Ten 0:13:19.720 --> 0:13:22.559 billion dollars goes to open Ai, mostly in cloud credits, 0:13:22.640 --> 0:13:25.480 and they get equity in return. That doesn't make it. 0:13:25.520 --> 0:13:27.800 I can't invest in companies and air miles. Why can 0:13:27.840 --> 0:13:32.280 Microsoft do it? Anyway? Microsoft invested billions of dollars in 0:13:32.320 --> 0:13:35.360 open Ai, and Google and Amazon are propping Upanthropic too, 0:13:35.400 --> 0:13:37.800 with the latter completing its four billion dollar investment in 0:13:37.800 --> 0:13:40.880 the ai company earlier this year. I need to be 0:13:41.000 --> 0:13:43.760 clear that open Ai and Anthropic are wholly reliant on 0:13:43.840 --> 0:13:46.480 the mercies of a handful of trillion dollar companies who 0:13:46.600 --> 0:13:49.280 are themselves at the mercy of the rot economy in 0:13:49.320 --> 0:13:52.440 the public markets, and especially in the case of open 0:13:52.480 --> 0:13:55.520 AI's relationship with Microsoft, they're not just dependent on big 0:13:55.559 --> 0:13:58.240 tech firms financial support, but also on their ability to 0:13:58.280 --> 0:14:01.679 procure and build the actual infratre structure of these products. 0:14:02.320 --> 0:14:05.640 Open Ai does not have their own service, neither does Anthropic. 0:14:06.000 --> 0:14:09.160 Perhaps they have some a few, but the majority of 0:14:09.240 --> 0:14:12.760 their actual processing is done by big tech firms that 0:14:13.600 --> 0:14:16.840 paid them in Chuck e GE's tokens in cloud credits 0:14:16.840 --> 0:14:19.920 to run their services. Who knows how you even amortize that. 0:14:20.120 --> 0:14:23.600 It's completely insane. It drives me insane that they're able 0:14:23.680 --> 0:14:27.080 to do this every time they find a horrible little 0:14:27.080 --> 0:14:30.800 thing to do. But me, I'm at the mercy of 0:14:30.840 --> 0:14:32.840 one of the fine products that you'll hear after this 0:14:32.920 --> 0:14:36.280 ad break, one that will no doubt echo my exact thoughts, feelings, 0:14:36.280 --> 0:14:39.560 and desires. Buy their things or don't. That's up to you, 0:14:49.760 --> 0:14:54.720 and we're back. As I was previously saying, Friday, August second, 0:14:54.840 --> 0:14:56.960 twenty twenty four is a rough day for AI, and 0:14:57.000 --> 0:15:01.720 it somehow got worse. That's SAE. The Information reported that 0:15:01.840 --> 0:15:05.840 Nvidia has reportedly told Microsoft and another unnamed cloud customer 0:15:06.040 --> 0:15:09.560 that its next generation Blackwell chips, which are designed primarily 0:15:09.560 --> 0:15:12.760 for accelerating AI compute tasks, will be delayed by three 0:15:12.800 --> 0:15:16.400 months due to an unspecified design floor identified by its 0:15:16.400 --> 0:15:21.800 contract foundry TSMC. It's entirely possible that this floor takes 0:15:21.840 --> 0:15:25.640 a little more than three months to resolve. Semiconductor manufacturing 0:15:25.720 --> 0:15:28.880 is really, really, really really hard and issues that initially 0:15:28.880 --> 0:15:32.120 see minor can rapidly spiral out of control. On top 0:15:32.160 --> 0:15:34.840 of that, you can't rush them in the same way 0:15:34.880 --> 0:15:37.600 you can with software. You can't just throw a bunch 0:15:37.680 --> 0:15:41.320 more money. TSMC has plenty. But on top of that, 0:15:42.240 --> 0:15:45.360 this isn't a money problem. This is one of those 0:15:45.440 --> 0:15:48.960 scientific ones that you can't just skip over. And with 0:15:49.000 --> 0:15:52.000 black Whill representing a major leap forward in capabilities both 0:15:52.000 --> 0:15:54.720 in the terms of power efficiency and sheer compute power, 0:15:55.280 --> 0:15:57.440 this is a huge blow to open ai and it's 0:15:57.640 --> 0:16:01.360 many many competitors, by which I mean it's other competitor, 0:16:01.400 --> 0:16:05.280 which is also bankrolled by a big tech sugar daddy. 0:16:05.320 --> 0:16:07.840 It's also likely to delay some of open AI's grander 0:16:07.840 --> 0:16:10.800 ambitions for an indeterminate amount of time, which in turn 0:16:11.280 --> 0:16:14.200 kind of dampens its appeal to potential investors. And to 0:16:14.240 --> 0:16:16.760 be clear, open ai basically has to raise in the 0:16:16.800 --> 0:16:19.920 next six months, as the Information reported that open ai 0:16:20.000 --> 0:16:23.440 could lose as much as five billion dollars this year alone. 0:16:23.640 --> 0:16:26.320 I will get to that later, but just to be clear, 0:16:27.000 --> 0:16:30.200 that's not five billion in losses plus revenue. That is, 0:16:31.560 --> 0:16:34.960 after you take the money out that they made. It's 0:16:35.040 --> 0:16:38.200 really bad it's really really bad. Though. It's all so 0:16:38.480 --> 0:16:41.400 very bad. I'm not sure why more people aren't freaking out. 0:16:42.240 --> 0:16:45.240 I'm mostly chilling just because they don't really care if 0:16:45.240 --> 0:16:47.920 they live or die. I feel bad for the people 0:16:47.960 --> 0:16:50.680 with the jobs, but I'm rambling. We'll get back to it, 0:16:51.440 --> 0:16:53.520 and this is a problem that expands to Google and 0:16:53.560 --> 0:16:56.760 Microsoft too. In Nvidious Blackwell chips was supposed to be 0:16:56.760 --> 0:17:00.320 a major technological leap for generative AI, designed with both 0:17:00.320 --> 0:17:03.160 influence the way that large language models generate answers in 0:17:03.200 --> 0:17:06.240 training in mind, and thus providing vastly more compute parerent 0:17:06.400 --> 0:17:10.480 energy efficiency necessary to make this shit move. But as 0:17:10.520 --> 0:17:13.280 of now, these chips have been delayed, with their availability 0:17:13.359 --> 0:17:15.800 not expected into the first quarter of twenty twenty five 0:17:16.040 --> 0:17:19.200 at the earliest, and the Information reports that open Ai 0:17:19.280 --> 0:17:21.920 probably won't get access to them until March twenty twenty four, 0:17:22.760 --> 0:17:25.320 and Nvidia might be forced to do new test runs 0:17:25.359 --> 0:17:27.879 with its foundry partner before scaling up to mass production. 0:17:29.000 --> 0:17:32.040 One particularly worrying quote, and I kind of just hinted 0:17:32.040 --> 0:17:35.000 in it there, was that Microsoft managers had planned to 0:17:35.000 --> 0:17:38.040 make Blackwell powered servers available to open Ai by January, 0:17:38.080 --> 0:17:40.320 but like I said, may need to plan for March 0:17:40.359 --> 0:17:42.520 or early spring, and that was said by a person 0:17:42.560 --> 0:17:47.439 with knowledge of the situation. Beautiful name the reason that's bad, Well, 0:17:47.920 --> 0:17:52.359 it's not really just one. First, open ai desperately needs 0:17:52.400 --> 0:17:55.320 something new. They need something that will show both investors 0:17:55.320 --> 0:17:58.359 and the media that open ai is building something meaningful. 0:17:58.560 --> 0:18:00.720 And while it's possible it will be able to deliver 0:18:00.800 --> 0:18:03.240 GPT five, which is their next model, in the near 0:18:03.320 --> 0:18:06.640 term future, even potential customers don't believe the jump will 0:18:06.680 --> 0:18:10.800 be significant enough from GPT four. Second, for reasons I've 0:18:10.840 --> 0:18:13.600 alluded to at the start of this episode, investors are 0:18:13.640 --> 0:18:16.439 getting a touch itchy with the AI hype boom, and 0:18:16.560 --> 0:18:19.639 more specifically with the giant tech companies that are bankrolling it. 0:18:20.200 --> 0:18:22.840 The street needs to see results, or at the very 0:18:22.920 --> 0:18:26.280 least ap plausible case for the future profitability and marketability 0:18:26.440 --> 0:18:29.520 of generative AI, a thing that it's genuinely bad at. 0:18:30.320 --> 0:18:33.040 In vidious Blackwell chips could help open ai models run 0:18:33.119 --> 0:18:36.880 faster and ingest training data faster, as well as, even 0:18:36.960 --> 0:18:40.320 if it's not possible, fooling investors into believing that open 0:18:40.359 --> 0:18:42.520 ai had the latest tech and thus could build the 0:18:42.960 --> 0:18:46.119 average general intelligence that sam Ortmand's been lying about and 0:18:46.320 --> 0:18:51.600 allow him to continue his con regardless, the article from 0:18:51.600 --> 0:18:54.720 the Information put it very plainly open AI only gets 0:18:54.760 --> 0:18:57.960 access to the latest technology as fast as Microsoft permits, 0:18:58.080 --> 0:19:00.199 or as fast as Microsoft is able to if I 0:19:00.240 --> 0:19:03.480 did on their own, even one of these events would 0:19:03.480 --> 0:19:06.040 be a deeply worrying sign of the bubble popping, and 0:19:06.080 --> 0:19:08.080 together they threatened to begin a collapse that I've been 0:19:08.080 --> 0:19:11.000 predicting since March, where I said that AI companies had 0:19:11.000 --> 0:19:14.320 about three quarters to prove themselves and I quote savaging 0:19:14.320 --> 0:19:17.000 the revenues of the biggest companies in tech when things 0:19:17.040 --> 0:19:20.960 don't work out. But I've realized now that it isn't 0:19:20.960 --> 0:19:24.000 really super useful to attach things to time. Though I 0:19:24.160 --> 0:19:26.439 stand by bud prediction, and thus I think it's more 0:19:26.480 --> 0:19:29.159 illustrative to suggest what the terms of the bubble popping 0:19:29.160 --> 0:19:33.880 actually are, So let's define them. What do I mean 0:19:33.880 --> 0:19:37.120 when I talk about the bubble popping? For the sake 0:19:37.160 --> 0:19:40.080 of clarity, I'm defining it as the major cloud companies 0:19:40.119 --> 0:19:43.159 reducing capital expenditures related to generative AI in a public 0:19:43.200 --> 0:19:46.520 and significant manner, as in an actual statement from satching 0:19:46.560 --> 0:19:49.760 Adelaer at Microsoft or Sun Duppishi at Google or one 0:19:49.760 --> 0:19:52.640 of the major large language model companies Anthropic or open 0:19:52.680 --> 0:19:56.880 AI collapsing in some way. No cheap thrills, no half measures. 0:19:57.280 --> 0:20:00.760 We're not calling this fucking thing until it actually pops, 0:20:01.000 --> 0:20:03.080 and it will not pop into one or both of 0:20:03.119 --> 0:20:06.119 these companies go tits up. To be clear, when I 0:20:06.160 --> 0:20:10.040 talk about Anthropic and Open Air collapsing, I don't actually 0:20:10.119 --> 0:20:13.760 mean in the Enron style way. These companies are probably 0:20:13.800 --> 0:20:18.240 not going to fall apart, shut down, close their doors suddenly. 0:20:19.320 --> 0:20:22.439 It could be that they're absorbed into another company on 0:20:22.520 --> 0:20:26.800 unfavorable terms, you know, like character AI and inflection, like 0:20:26.880 --> 0:20:29.800 the things that just happened. And in that case, I 0:20:29.800 --> 0:20:32.840 think investors are going to take a massive loss. Open 0:20:32.880 --> 0:20:36.040 AI or Anthropic might be forced to radically limit their operations, 0:20:36.080 --> 0:20:39.359 either by shutting down their free chatbots or limiting access 0:20:39.359 --> 0:20:42.520 to enterprise customers, or by slowing down development massively, which 0:20:42.640 --> 0:20:46.200 by the way, will be death's embrace. It will be over. 0:20:47.000 --> 0:20:49.000 We could see massive layoffs, or we could see these 0:20:49.000 --> 0:20:51.800 companies pivot to a less capital intensive business model, such 0:20:51.800 --> 0:20:55.800 as licensing their patterns and IP to other technologies or companies, 0:20:56.200 --> 0:20:58.960 but not really offering an actual tangible product to consumers 0:20:59.040 --> 0:21:01.400 or business customers, as opposed to what they're doing today, 0:21:01.400 --> 0:21:04.800 which is offering a tangible thing that doesn't really doom 0:21:04.840 --> 0:21:08.640 that much. I'm rambling again. Yeah, I can imagine all 0:21:08.680 --> 0:21:11.280 sorts of terminal endings for Generative AI is big two, 0:21:11.280 --> 0:21:13.840 and some of them aren't even legally actionable. But I 0:21:13.880 --> 0:21:16.840 believe that the collapse or absorption of these companies is 0:21:16.880 --> 0:21:20.160 the one critical sign to look for. Open ai began 0:21:20.240 --> 0:21:23.320 this hype cycle and Sam Altman is definitely the pt 0:21:23.480 --> 0:21:26.440 barnum of the large language model circus, and open ai 0:21:26.520 --> 0:21:29.000 has absorbed more money and attention than any other startup 0:21:29.040 --> 0:21:30.879 of the last few years, and outside of Uber, I 0:21:30.920 --> 0:21:34.520 think they might be getting the most all time. God, 0:21:34.560 --> 0:21:37.000 I hate it. It's symbolic of the excess in the 0:21:37.000 --> 0:21:39.199 waste of the generative AI boom and it's death or 0:21:39.320 --> 0:21:42.960 as I mentioned, some alternative collapse is the sign that 0:21:42.960 --> 0:21:45.240 we're done here in the same way that FTX signaled 0:21:45.240 --> 0:21:48.600 the end of the cryptocurrency boom. And if open AI's 0:21:48.680 --> 0:21:52.000 collapse marks the apocalypse for generative AI, I believe it 0:21:52.040 --> 0:21:54.800 will be accompanied by one or several of the following 0:21:54.840 --> 0:21:58.560 pale horses, real tangible signs that things are falling apart, 0:21:58.760 --> 0:22:01.160 things you can send to your friends, family and laugh at. 0:22:02.040 --> 0:22:06.399 Let's take a look, shall we. Number one, Any significant 0:22:06.400 --> 0:22:09.000 price changes by any large language model company are a 0:22:09.000 --> 0:22:12.920 bad sign. Desperate times require desperate measures, and any oscillation 0:22:13.040 --> 0:22:15.119