WEBVTT - Monologue: OpenAI's Albatross 0:00:01.880 --> 0:00:05.480 Al Zone Media, give me that total next shirt, a 0:00:05.480 --> 0:00:08.000 tim pan apple, a German shepherd, a wristbandstand, and a 0:00:08.080 --> 0:00:10.680 lurching red bird and more brave than the Turning Network. 0:00:10.760 --> 0:00:13.039 This is your weekly Better Offline monologue, and I'm your 0:00:13.039 --> 0:00:24.000 host ed zetron. Now, before we go any further, I 0:00:24.079 --> 0:00:27.280 need your help. I look Better Offline is up for 0:00:27.320 --> 0:00:28.960 a webby and I really need you to vote for 0:00:29.000 --> 0:00:31.040 best episode in the business category. It's the man who 0:00:31.120 --> 0:00:34.360 killed Google Search. It's Propagart Ragavan. Let's get him. I 0:00:34.440 --> 0:00:36.520 realize it's a huge pain in the ass to sign 0:00:36.600 --> 0:00:38.720 up for something and vote, but I've never won an 0:00:38.760 --> 0:00:41.160 award in my life and I'd really appreciate it. Link 0:00:41.240 --> 0:00:42.960 is going to be in the episode notes, and while 0:00:42.960 --> 0:00:46.000 you're there, also vote for the wonderful Mollyconger's Weird Little Guys, 0:00:46.000 --> 0:00:48.440 which I'll also have a link to. I know signing 0:00:48.520 --> 0:00:51.360 up to stuff is annoying. I'm asking a lot from you, 0:00:51.440 --> 0:00:56.040 but there you go. I'm doing it anyway. To the monologue, 0:00:56.280 --> 0:00:58.720 I feel like we're approaching a choke point in the 0:00:58.720 --> 0:01:01.280 whole General v Ai bubble, the culmination of over a 0:01:01.360 --> 0:01:04.119 year of different narratives and pressures that I believe will 0:01:04.440 --> 0:01:07.839 lead to an ultimate collapse. Last week, open Ai released 0:01:07.840 --> 0:01:10.880 an image generator with GPT four to zero, which quickly 0:01:10.920 --> 0:01:13.440 gained massive attention for its ability to create images in 0:01:13.480 --> 0:01:17.440 the style of famed Japanese animation company Studio Ghibli. And 0:01:17.480 --> 0:01:19.600 to be clear, I think these images are an abomination 0:01:19.680 --> 0:01:21.880 and everyone involved in launching this tool has committed a 0:01:21.920 --> 0:01:27.800 mortal sin anyway. Nevertheless, creating these disgusting, disgraceful images comes 0:01:27.800 --> 0:01:30.679 at in incredibly high cost, and for the last week, 0:01:30.760 --> 0:01:33.479 open Ai CEO Sam Ortman has been complaining about their 0:01:33.480 --> 0:01:36.960 GPUs melting, leading to open ai having to limit free 0:01:37.080 --> 0:01:40.200 users to only three image generations a day, along with 0:01:40.280 --> 0:01:43.400 longer wait times than capacity issues with video generator Sora. 0:01:44.120 --> 0:01:46.960 To make matters worse, Ortman also announced that and I 0:01:47.200 --> 0:01:50.080 quote by the way, that users should expect new releases 0:01:50.120 --> 0:01:52.720 from open Ai to be delayed, stuff to break, and 0:01:52.760 --> 0:01:55.200 for services to sometimes be slow as we deal with 0:01:55.200 --> 0:02:00.800 capacity challenges. This led me to ask a very simple question. 0:02:00.960 --> 0:02:03.480 I think everybody in the tech media really should be asking, 0:02:04.040 --> 0:02:08.600 why can't Sam Waltman ask Microsoft for more GPUs. The 0:02:08.639 --> 0:02:10.720 answer is, as you may have guessed from my last 0:02:10.760 --> 0:02:13.359 monologue is that there may not actually be capacity for 0:02:13.440 --> 0:02:16.920 them to do so. Open AI's relationship with Redmond has 0:02:16.960 --> 0:02:20.160 grown kind of chilly over the past year. I'd speculate 0:02:20.160 --> 0:02:23.120 that Microsoft has refused to provide additional immediate capacity or 0:02:23.120 --> 0:02:25.400 has refused to provide capacity on the chummy terms that 0:02:25.400 --> 0:02:28.920 open Ai previously enjoyed, receiving a significant discount on the 0:02:28.960 --> 0:02:32.240 usual ticket prices in the past. We know that Microsoft 0:02:32.240 --> 0:02:34.680 has both walked away from two gigawats of future compute 0:02:34.680 --> 0:02:37.560 capacity and declined the option to spend another twelve billion 0:02:37.600 --> 0:02:39.960 dollars on core Weave's compute and core Weave if you 0:02:40.000 --> 0:02:44.040 don't remember there that the publicly traded data centered AI 0:02:44.200 --> 0:02:48.880 company a whole dog's dinner onto itself, and analyst house 0:02:48.919 --> 0:02:51.560 TD Cohen suggested in that this is a sign that 0:02:51.639 --> 0:02:54.799 Microsoft is no longer willing to shoulder the immense financial 0:02:54.800 --> 0:02:58.440 burden of supporting open Ai, even though open ai picked 0:02:58.440 --> 0:03:00.600 that option up, which by which I mean they took 0:03:00.639 --> 0:03:03.239 the twelve billion dollars of compute. It isn't clear if 0:03:03.400 --> 0:03:06.840 Corwave can actually build the capacity they need, and definitely 0:03:06.960 --> 0:03:08.200 don't think they're going to be able to do it 0:03:08.240 --> 0:03:11.520 in the time they need it. Microsoft allegedly walked away 0:03:11.520 --> 0:03:14.440 from Corewave due to its failure to deliver and that 0:03:14.760 --> 0:03:17.120 deliver the services they asked for, and indeed probably the 0:03:17.120 --> 0:03:19.880 compute as well. If that's true, it's unclear what has 0:03:20.000 --> 0:03:22.800 changed to make core Weave magically able to support open Ai, 0:03:23.440 --> 0:03:25.800 or even how a company that's drowning in high interest 0:03:25.840 --> 0:03:28.720 debt can finance the creation of several billion dollars worth 0:03:28.760 --> 0:03:32.239 of new data centers. Also, it's not quite as simple 0:03:32.280 --> 0:03:34.480 as open ai calling up a data center company with 0:03:34.520 --> 0:03:37.120 a bunch of GPUs and running chat, GPT, dot ex. 0:03:37.920 --> 0:03:41.000 Open Ai likely has reams of different requirements, and the 0:03:41.040 --> 0:03:43.720 amount of GPUs they will need will likely vary based 0:03:43.760 --> 0:03:46.520 on demand, putting them in a problematic situation where they 0:03:46.520 --> 0:03:48.320 could be commuting to a bunch of compute that they 0:03:48.400 --> 0:03:51.760 don't need if demand slows down. I've heard that companies 0:03:51.800 --> 0:03:55.000 generally want a six to twelve month commitment for GPUs 0:03:55.040 --> 0:03:57.240 two the cost is fixed no matter how much they 0:03:57.280 --> 0:04:00.560 get used, or at least there's a minimum commitment. But 0:04:00.800 --> 0:04:04.120 let's assume for a second that demand for chat GPT 0:04:04.240 --> 0:04:08.200 continues to rise. How does OpenAI actually get that compute 0:04:08.280 --> 0:04:11.440 if Microsoft isn't handing it over, and the Information reports 0:04:11.480 --> 0:04:14.040 that open ai still projects to spend about thirteen billion 0:04:14.080 --> 0:04:16.680 dollars on as your Cloud in twenty twenty five, there 0:04:16.720 --> 0:04:19.200 aren't really a ton of other options, especially for a 0:04:19.200 --> 0:04:23.360 company with such gigantic requirements, meaning that whatever infrastructure open 0:04:23.360 --> 0:04:26.160 ai is building is a patchwork between smaller players, and 0:04:26.279 --> 0:04:31.320 using so many smaller providers likely creates unavoidable inefficiencies and overhead. 0:04:31.839 --> 0:04:35.280 I'm naming another pale horse of the AI apocalypse by 0:04:35.320 --> 0:04:40.039 the way limits to service and service degradation across chat GPT. 0:04:40.720 --> 0:04:43.200 Open ai is running out of compute capacity. They've talked 0:04:43.200 --> 0:04:45.919 about it since October of last year, and chat GPT's 0:04:45.960 --> 0:04:49.120 new image generation is a significant drain on their resources, 0:04:49.360 --> 0:04:52.320 meaning that to continue providing their services, they're going to 0:04:52.400 --> 0:04:56.880 need to expand capacity or reduce access to services otherwise. 0:04:57.600 --> 0:05:01.360 The problem is that expanding is extremely different. Data centers 0:05:01.440 --> 0:05:03.920 take three to six years to build, and open ai 0:05:03.960 --> 0:05:06.800 has planned Stargate data Center won't have anything ready before 0:05:06.800 --> 0:05:09.640 twenty twenty six at the earliest, which means we're approaching 0:05:09.640 --> 0:05:12.120 a point where there simply might not be enough data 0:05:12.160 --> 0:05:16.240 centers or GPUs to burn, while open ai could theoretically 0:05:16.279 --> 0:05:19.000 go to Google or Amazon. Both of those companies are 0:05:19.000 --> 0:05:21.520 invested in anthropic and have little incentive to align with 0:05:21.600 --> 0:05:24.839 open Ai. Meta is building their own chet GPT competitor, 0:05:24.920 --> 0:05:29.160 and Elon must despises Sam Mortman real shithead versus fuckwad 0:05:29.279 --> 0:05:33.000 situation there. While I can't say for certain, I can't 0:05:33.000 --> 0:05:36.279 work out where open ai will get the capacity to continue, 0:05:36.360 --> 0:05:38.240 And I just don't know how they're going to expand 0:05:38.240 --> 0:05:41.600 their services if Microsoft isn't providing capacity. Yes, there's a 0:05:41.600 --> 0:05:44.039 Oracle which open ai has a partnership with, but they're 0:05:44.120 --> 0:05:48.720 relatively small in this space. Chat GPT's immage generation has 0:05:48.760 --> 0:05:51.240 become this massive burden on the company right at the 0:05:51.240 --> 0:05:54.280 point where it's introducing some of its most expensive models ever, 0:05:54.520 --> 0:05:57.880 and the products themselves are extremely expensive to run. Deep 0:05:57.920 --> 0:06:00.400 research is perhaps the best example using O open ai 0:06:00.480 --> 0:06:03.359 is extremely expensive O three model, which can cost in 0:06:03.400 --> 0:06:05.720 some cases as much as one thousand dollars per query. 0:06:05.960 --> 0:06:10.040 Deeper search is probably cheaper, but not that much cheaper, probably, 0:06:10.360 --> 0:06:12.680 I would. I've heard rumors, and this is a rumor. 0:06:13.279 --> 0:06:15.400 It's a rumor. I've heard like a dollar or two 0:06:15.400 --> 0:06:18.479 per query. If that's the case, that's fucking insane. Anyway. 0:06:18.720 --> 0:06:22.200 While open Ai could absorb the remaining capacity at say Crusoe, 0:06:22.279 --> 0:06:25.200 Lambda and core Wave, this creates a systemic risk where 0:06:25.200 --> 0:06:28.200 every GPU provider is reliant on open AI's money, and 0:06:28.240 --> 0:06:30.880 this assumes that they'll actually have enough to begin with. 0:06:31.680 --> 0:06:35.000 Open Ai also just close the largest private funding round 0:06:35.040 --> 0:06:38.880 in history, forty billion theoretical dollars, valuing the company at 0:06:38.880 --> 0:06:42.159 a ridiculous three hundred billion dollars raised from he gets 0:06:42.240 --> 0:06:45.120 the soft Bank and other investors. That's good news, right, 0:06:46.120 --> 0:06:49.520 Not really? In truth, open Ai really only raised ten 0:06:49.560 --> 0:06:51.560 billion dollars, with seven and a half billion of those 0:06:51.600 --> 0:06:53.880 dollars coming from soft Bank and another two point five 0:06:53.920 --> 0:06:57.680 billion dollars coming from other investors, including Thrive Capital and Microsoft. 0:06:58.080 --> 0:07:00.600 The remaining thirty billion dollars off, where which soft Bank 0:07:00.680 --> 0:07:02.719 is on the hook for twenty billion dollars off, will 0:07:02.800 --> 0:07:06.320 arrive at the end of the year. That's all we've gone. 0:07:06.480 --> 0:07:09.160 But open Ai will only get ten billion dollars from 0:07:09.200 --> 0:07:11.920 soft Bank, so bringing it down to a thirty billion 0:07:11.960 --> 0:07:14.720 dollar round total. If open ai fails to convert from 0:07:14.760 --> 0:07:17.120 a nonprofit to a for profit company by the end 0:07:17.160 --> 0:07:21.240 of twenty twenty five, a massive acceleration there. As a reminder, 0:07:21.440 --> 0:07:24.080 open ai is a weirdly structured nonprofit with a for 0:07:24.200 --> 0:07:26.720 profit arm, and their last round of funding from October 0:07:26.720 --> 0:07:29.480 twenty twenty four had another caveat that if open ai 0:07:29.640 --> 0:07:32.119 failed to become a for profit company by October twenty 0:07:32.120 --> 0:07:35.720 twenty six, all investment dollars would convert into debt. I've 0:07:35.760 --> 0:07:38.160 also read that they would have to hand the money back. 0:07:38.560 --> 0:07:41.600 I'm not sure whether that's the case. Debt is the 0:07:41.600 --> 0:07:44.920 one that's been reported the most. Furthermore, open Ai loses 0:07:44.920 --> 0:07:48.000 money on every single prompt on Chat GPT, even from 0:07:48.040 --> 0:07:51.200 their two hundred dollars a month chet GPT Pro subscribers. 0:07:51.600 --> 0:07:53.840 The burdens some interest payments would make it even harder 0:07:53.880 --> 0:07:56.960 for open ai to reach break even, which right now 0:07:57.000 --> 0:07:59.760 it doesn't even seem like they can do anyway. As 0:07:59.800 --> 0:08:02.600 an another reminder, soft Bank is a company that has 0:08:02.640 --> 0:08:05.720 now invested in two different fraudulent schemes, Wirecard and Green 0:08:05.760 --> 0:08:08.240 Silk Capital, the latter of which helped put the nail 0:08:08.240 --> 0:08:10.640 in the coffin of credit sweee back in twenty twenty 0:08:10.680 --> 0:08:15.040 three and put sixteen billion dollars into we work. It 0:08:15.080 --> 0:08:19.360 will be incredibly some might say impossibly difficult, and I'll 0:08:19.400 --> 0:08:21.480 cover this in the future episode to convert open ai 0:08:21.560 --> 0:08:23.760 into a for profit company, and the fact that soft 0:08:23.840 --> 0:08:26.640 Bank is putting this caveat on their investment heavily suggests 0:08:26.800 --> 0:08:29.000 that they have doubts it will happen. And I must 0:08:29.080 --> 0:08:31.960 be clear. When the monopoly man is getting nervous, you 0:08:32.000 --> 0:08:35.800 should get nervous too. The fact OpenAI accepted these terms 0:08:35.800 --> 0:08:38.760 also suggest they're desper and I don't blame them. They've 0:08:38.760 --> 0:08:41.720 committed eighteen billion dollars to the Stargate Data Center project, 0:08:41.800 --> 0:08:44.440 will spend thirteen billion dollars on Microsoft Computer alone in 0:08:44.440 --> 0:08:47.400 twenty twenty five, according to the information, and they've now 0:08:47.440 --> 0:08:51.640 created an incredibly popular product that will guarantee people come 0:08:51.679 --> 0:08:54.040 and use it like twice and then never use it again. 0:08:55.360 --> 0:08:57.960 Now keep a keen eye on any restrictions that open 0:08:57.960 --> 0:09:00.400 ai makes on chat GPT in the coming month. I 0:09:00.440 --> 0:09:02.440 do not see how this company survives, nor do I 0:09:02.480 --> 0:09:05.680 see how they expand their capacity much further. Price increases, 0:09:05.800 --> 0:09:08.000 rate limits and other ways of slowing down the pressure 0:09:08.040 --> 0:09:11.240 on their servers will likely suggest that open eye is 0:09:11.320 --> 0:09:13.400 up against the wall, both in their ability to support 0:09:13.400 --> 0:09:15.600 the services they provide and the costs they must bear 0:09:15.640 --> 0:09:18.680 to provide them. We are entering the hysterical era of 0:09:18.679 --> 0:09:20.960 the bubble, time when the craziest stuff will happen as 0:09:21.000 --> 0:09:23.240 the money does everything it can to keep the dream alive. 0:09:23.880 --> 0:09:25.720 I look forward to telling you what happens next.