WEBVTT - Exclusive: Here's How Much Anthropic Spends on AWS 0:00:02.480 --> 0:00:07.400 Alz Media. Hello and welcome to a special exclusive episode 0:00:07.400 --> 0:00:21.640 of Better Offline. I'm Your host ed Zitron. As a 0:00:21.720 --> 0:00:24.479 result of discussions with sources and documents viewed of the 0:00:24.480 --> 0:00:27.240 amounts built on Amazon Web Services, I am for the 0:00:27.240 --> 0:00:29.680 first time in history able to disclose how much AI 0:00:29.720 --> 0:00:33.720 firms are spending on AWS, specifically Anthropic and a coding 0:00:33.720 --> 0:00:37.680 company Cursor, its largest customer for API services. I can 0:00:37.720 --> 0:00:40.920 exclusively reveal today how much Anthropics spent on AWS for 0:00:40.920 --> 0:00:42.960 the years twenty twenty four and from the beginning of 0:00:42.960 --> 0:00:45.520 twenty twenty five through the end of September twenty twenty five, 0:00:45.720 --> 0:00:48.120 and from what I can see, their compute spend may 0:00:48.200 --> 0:00:52.239 vastly exceed what has previously been reported. Furthermore, I can 0:00:52.240 --> 0:00:54.800 confirm that through the end of September twenty twenty five, 0:00:54.960 --> 0:00:57.320 Anthropic is spent around one hundred percent of their revenue 0:00:57.320 --> 0:01:00.319 in twenty twenty five on Amazon Web Services, spending two 0:01:00.320 --> 0:01:03.400 point sixty six billion dollars on compute on an estimated 0:01:03.440 --> 0:01:06.840 two point five five billion dollars in revenue. Go to 0:01:06.840 --> 0:01:09.080 the newsletter. I source the whole goddamn thing, and if 0:01:09.120 --> 0:01:11.399 I'm honest, this piece is the culmination of several months 0:01:11.400 --> 0:01:14.679 of articles about how Anthropics business tactics have made be 0:01:14.760 --> 0:01:18.280 turned the screws on their biggest customer. I can exclusively 0:01:18.319 --> 0:01:21.080 reveal today, as well as many other numbers in the newsletter, 0:01:21.319 --> 0:01:24.000 the curs of Amazon Web services bills doubled from six 0:01:24.040 --> 0:01:26.440 point two million dollars in May twenty twenty five to 0:01:26.480 --> 0:01:28.960 twelve point six million dollars in June twenty twenty five, 0:01:29.360 --> 0:01:31.959 and have stayed inflated since Anthropic increased the costs with 0:01:32.000 --> 0:01:35.440 the launch of priority servers, tiers and aggressive rent seeking measure. 0:01:35.959 --> 0:01:38.240 I need to be clear I cannot one hundred percent 0:01:38.319 --> 0:01:40.160 guarantee that's what did it. I'm going to hedge my 0:01:40.200 --> 0:01:43.039 bets very hard on that, but it certainly Bloody Wealth 0:01:43.080 --> 0:01:45.560 seems that way. It's my guard instinct. I'm not going 0:01:45.560 --> 0:01:47.640 to say it declaratively, but I'm going to show you 0:01:47.680 --> 0:01:50.400 why I believe this, And I admit I struggled with 0:01:50.440 --> 0:01:52.560 how to turn this into an episode because the newsletter, 0:01:52.600 --> 0:01:54.760 which is on my free feed, is a series of 0:01:54.840 --> 0:01:57.080 numbers and analyzes that if I just read them aloud, 0:01:57.080 --> 0:01:59.320 would sound extremely dull and at times be quite hard 0:01:59.320 --> 0:02:02.120 to follow. It's not something that naturally plays well for radio. 0:02:03.320 --> 0:02:05.600 So instead of giving you the audible version, I'm going 0:02:05.600 --> 0:02:07.040 to give you the cliff notes and speak to a 0:02:07.080 --> 0:02:10.120 degree of vindication I feel on reading these costs. So 0:02:10.240 --> 0:02:12.800 let's start with a number. One point two two five 0:02:12.880 --> 0:02:15.880 billion dollars. That's how much Anthropic spent on Amazon Web 0:02:15.919 --> 0:02:18.680 Services in the third quarter of twenty twenty five. They 0:02:18.680 --> 0:02:20.880 spent eight hundred and twenty nine point seven million in 0:02:20.960 --> 0:02:23.400 Q two twenty twenty five, and six hundred and ten 0:02:23.400 --> 0:02:25.960 million dollars in Q one twenty twenty five. Oh and 0:02:26.000 --> 0:02:28.680 one other number, they spent one point three five billion 0:02:28.680 --> 0:02:32.360 dollars on AWS and twenty twenty four. So yeah, just 0:02:32.600 --> 0:02:34.840 in another way, talking of their twenty twenty five numbers, 0:02:34.880 --> 0:02:38.160 anthropics spend on AWS doubled over the course of three quarters. 0:02:38.840 --> 0:02:42.440 Now a little backstory about Anthropic that's necessary to understand 0:02:42.440 --> 0:02:45.920 this fully. Anthropic was originally invested in by both Google 0:02:45.960 --> 0:02:48.880 and Amazon. According to The New York Times, Google owns 0:02:48.919 --> 0:02:51.720 around fourteen percent of the company. An analyst, yes Tomate, 0:02:51.800 --> 0:02:55.240 Amazon owned somewhere between fifteen and eighteen percent, and both have, 0:02:55.520 --> 0:02:58.440 in not so many words, said that they're the main 0:02:58.600 --> 0:03:02.680 or primary compute part for Aanthropic. It's unclear how much 0:03:02.720 --> 0:03:05.799 Anthropic spends on Google Cloud, but semi analysis believes they're 0:03:05.800 --> 0:03:08.920 a big client, and that's about as much detail as 0:03:08.960 --> 0:03:11.760 I can get from anywhere I've really looked. In any case, 0:03:11.840 --> 0:03:14.840 Anthropic is spending effectively every dollar they make on Amazon 0:03:14.840 --> 0:03:17.440 Web Services, and Amazon has appears to be booking this 0:03:17.480 --> 0:03:20.400 as revenue, though I can't directly confirm that, though I 0:03:20.440 --> 0:03:23.840 do know these numbers are cash, they're after credits. Though 0:03:23.880 --> 0:03:26.919 in the recent months, Anthropic has lowered the amount of 0:03:26.960 --> 0:03:29.359 revenue they're spending on it to eighty six point two 0:03:29.360 --> 0:03:31.640 percent in Q three twenty twenty five, which is an 0:03:31.680 --> 0:03:33.760 improvement from Q two to twenty twenty five, where they 0:03:33.760 --> 0:03:36.000 spent one hundred and six percent of their revenue and 0:03:36.120 --> 0:03:38.360 Q one, where they spent one hundred and seventy five 0:03:38.360 --> 0:03:41.480 percent of what they made on Amazon Web Services. It's 0:03:41.600 --> 0:03:45.040 quite horrifying when you say it out loud. Now, if 0:03:45.040 --> 0:03:47.520 you're thinking that, because these numbers are quite close, that 0:03:47.560 --> 0:03:51.480 this might suggest that Anthropics costs are improving, think again. Anthropics. 0:03:51.520 --> 0:03:54.640 Amazon Web services costs of a habit of massively spiking. 0:03:54.720 --> 0:03:57.720 For example, their AWS bill led from three hundred and 0:03:57.800 --> 0:04:00.600 eighty three point seven million dollars in August twenty twenty 0:04:00.640 --> 0:04:03.480 five to five hundred and eighteen point nine million dollars 0:04:03.480 --> 0:04:06.720 in September twenty twenty five. That's one hundred and thirty 0:04:06.720 --> 0:04:09.440 five million goddamn dollars. And my hunch is it's because 0:04:09.480 --> 0:04:11.800 they have a massive problem where clawed code users are 0:04:11.920 --> 0:04:14.480 each costing them thousands of dollars despite only paying one 0:04:14.520 --> 0:04:17.279 hundred or two hundred dollars a month. There's also the 0:04:17.360 --> 0:04:20.840 nasty matter of Google Cloud. Anthropics Amazon Web Services bill 0:04:20.920 --> 0:04:23.000 is two point sixty six billion dollars from January through 0:04:23.040 --> 0:04:25.080 the end of September, as I said, and that is 0:04:25.120 --> 0:04:27.320 pretty close to two point five to five billion dollars 0:04:27.360 --> 0:04:30.520 in revenue. But if Anthropics spend on Google Cloud was 0:04:30.560 --> 0:04:33.039 only twenty five percent of what they spent on AWS, 0:04:33.200 --> 0:04:35.440 its compute cost would jumped to three point three three 0:04:35.440 --> 0:04:37.960 billion dollars through the end of September, way more than 0:04:37.960 --> 0:04:40.800 it brings in. If it's half of what they spent 0:04:41.320 --> 0:04:43.960 on Amazon Web Services, this becomes a three point nine 0:04:43.960 --> 0:04:46.560 to nine billion dollar compute bill, and if they spend 0:04:46.600 --> 0:04:49.240 the same amount, the bill becomes five point three billion dollars, 0:04:49.279 --> 0:04:52.520 and again that's just through the end of September. Another note, 0:04:52.680 --> 0:04:55.880 cursors spend on Amazon Web services is comparatively small, but 0:04:55.960 --> 0:04:59.080 includes some spend on Anthropics models because Amazon is allowed 0:04:59.080 --> 0:05:01.479 to sell them. And I believe that the reason that 0:05:01.520 --> 0:05:04.680 they do this, because they do directly pay Anthropic like 0:05:04.680 --> 0:05:07.560 they actually send money directly to them, is because Amazon 0:05:07.600 --> 0:05:10.719 offers significant discounts in some cases for running models through 0:05:10.760 --> 0:05:13.479 their service. I think it's their bedrock service, and my 0:05:13.560 --> 0:05:15.520 source confirmed that this was the case, though I could 0:05:15.560 --> 0:05:19.000 not get granular data on what exactly CURSES spend was 0:05:19.040 --> 0:05:21.599 on Amazon, like I can't say, oh, they use this 0:05:21.720 --> 0:05:24.600 model or that model. Now, Cursor spends most of their 0:05:24.640 --> 0:05:27.680 compute money directly with Anthropic, as well as every other 0:05:27.720 --> 0:05:30.760 model developer whose models they use. AWS is a small 0:05:30.760 --> 0:05:34.240 piece of the puzzle, and while small, it's spending data 0:05:34.279 --> 0:05:37.080 provides evidence of how much this shit actually costs, though 0:05:37.120 --> 0:05:39.560 I also concede that some of the money CURSES spends 0:05:39.560 --> 0:05:41.960 with AWS likely goes to the non AI part of 0:05:41.960 --> 0:05:46.240 the business, like file hosting and other tech infrastructure. Nevertheless, 0:05:46.240 --> 0:05:48.920 the timing of the massive jumps in CURSES AWS bill 0:05:48.960 --> 0:05:51.040 from six point two million dollars in May to twelve 0:05:51.040 --> 0:05:53.880 point six million dollars in June directly correlate with the 0:05:53.920 --> 0:05:56.719 massive changes made to their product, increasing the costs on 0:05:56.800 --> 0:05:58.760 any users that wanted to use Cursor in the way 0:05:58.800 --> 0:06:00.960 they had in the past by making them face the 0:06:01.000 --> 0:06:04.080 actual costs of serving models on a per million token basis. 0:06:04.240 --> 0:06:07.120 I've written about this a lot, By the way, it's 0:06:07.160 --> 0:06:09.960 hard to describe it in detail because it's going to 0:06:10.000 --> 0:06:13.200 take forever. But around mid June, Cursor had to change 0:06:13.240 --> 0:06:17.080 everything because mysteriously they had to stop spending so much 0:06:17.120 --> 0:06:19.599 money with their customers. Their customers would burning a hole 0:06:19.640 --> 0:06:21.640 in their pocket, and I think we can kind of 0:06:21.680 --> 0:06:25.120 see why. Curses costs have also never come down, spiking 0:06:25.160 --> 0:06:27.520 to a higher fifteen point five million dollars in June, 0:06:27.560 --> 0:06:30.200 dropping to a still high nine point six million dollars 0:06:30.440 --> 0:06:32.640 in August. I need to spike again to twelve point 0:06:32.720 --> 0:06:37.040 nine million dollars in September, though I cannot declaratively state 0:06:37.080 --> 0:06:40.720 that this is exactly what happened. Curses costs doubled immediately 0:06:40.720 --> 0:06:43.760 following the addition of anthropic service tiers in late May 0:06:43.800 --> 0:06:47.080 twenty twenty five, which require an upfront commitment of token 0:06:47.120 --> 0:06:50.320 spend and token throughput, and when Curser announced the launch 0:06:50.320 --> 0:06:53.160 of their two hundred dollars a month ultraplan amidst massive 0:06:53.160 --> 0:06:56.040 product changes, they cited how it was and I quote 0:06:56.320 --> 0:06:59.880 made possible by multi year partnerships from open Ai, Anthropic, 0:07:00.160 --> 0:07:03.320 Google and Xai, and that their support was instrumental in 0:07:03.360 --> 0:07:17.760 offering this volume of computer the predictable price. Now, really, 0:07:18.240 --> 0:07:21.080 I'm being as fair as I can. Another factor might 0:07:21.120 --> 0:07:24.520 be that the new Claude four models was significantly more expensive. 0:07:24.880 --> 0:07:27.640 It's entirely possible that all of these things are true. 0:07:27.840 --> 0:07:29.640 I just want to make sure I cover my basis 0:07:29.640 --> 0:07:32.640 because I do not know for sure. But the timing, 0:07:33.120 --> 0:07:36.840 the timing man and another thing, you know what. Anthropic 0:07:36.960 --> 0:07:40.440 also launched a week before service Tears a competing product, 0:07:40.440 --> 0:07:43.000 a Cursor called claud Code one that they could run 0:07:43.040 --> 0:07:45.080 with as little restraint as they'd like to drain as 0:07:45.120 --> 0:07:48.200 many monthly customers away from Cursor, who is also their 0:07:48.280 --> 0:07:52.840 largest customer, buying Anthropic models through their API. Real fucking mystery, right, 0:07:52.920 --> 0:07:54.640 If it quacks like a dark where's a T shirt 0:07:54.640 --> 0:07:57.080 that says dark? And Claude tells you you're absolutely right, 0:07:57.120 --> 0:07:59.240 that's a duck. When you upload a picture of it, 0:07:59.240 --> 0:08:01.960 it's probably a fucking duck. But I obviously can't say 0:08:01.960 --> 0:08:05.400 for sure. I need to be explicit here with what happened, though. 0:08:05.920 --> 0:08:09.320 Anthropics supplied access to their models to a company Cursor, 0:08:09.560 --> 0:08:12.200 and then released a product claud Code that did exactly 0:08:12.240 --> 0:08:15.080 the same thing as that company Cursor, turning it both 0:08:15.080 --> 0:08:18.240 into a customer and a competitor in the process, creating 0:08:18.280 --> 0:08:21.360 a massive conflict of interest, as not only did Anthropic 0:08:21.400 --> 0:08:24.320 have an incentive for that customer or competitor to fail, 0:08:24.360 --> 0:08:26.560 though they also needed their compute revenue, which is kind 0:08:26.560 --> 0:08:29.200 of a bugger. Anthropic also had the means to make 0:08:29.240 --> 0:08:31.560 this failure happen in the most painful and expensive way 0:08:31.640 --> 0:08:34.760 possible by worsening the terms in which that competitor required 0:08:34.760 --> 0:08:38.320 the compute it needed to function. Could be a coincidence, 0:08:38.360 --> 0:08:41.040 I guess. And when I say compute, I mean tokens. 0:08:41.240 --> 0:08:44.439 Just I'm reading a script. Okay, where us going to 0:08:44.480 --> 0:08:46.960 eam Anyway, I'm not going to turn this into a massive, 0:08:47.040 --> 0:08:49.400 sprawling episode about this company. I wanted to give you 0:08:49.480 --> 0:08:52.079 the raw information so you can go and read the 0:08:52.120 --> 0:08:55.240 detailed analysis. I did. It's free, by the way, don't worry. 0:08:55.880 --> 0:08:57.400 But now I want to talk about how all this 0:08:57.559 --> 0:08:59.880 made me feel, because that's what makes this show unique. 0:09:00.000 --> 0:09:01.800 I think is the appropriate way of coming out this. 0:09:02.360 --> 0:09:05.120 I'm going to be honest, I find what it looks like, 0:09:05.200 --> 0:09:08.240 and I'm hedging my beds again. Anthropic did to Cursor 0:09:08.600 --> 0:09:12.800 truly disgusting. Cursor hit five hundred million dollars in annualized 0:09:12.840 --> 0:09:15.360 revenue in the same month that they then saw their 0:09:15.440 --> 0:09:19.080 cost double, dramatically reducing the value of their subscription product 0:09:19.120 --> 0:09:22.320 at the apex of their success. Yes, Cursor is an 0:09:22.360 --> 0:09:25.480 unsustainable AI company I know, and like all of these companies, 0:09:25.520 --> 0:09:28.520 has no part of the profitability. Anthropic should have always 0:09:28.640 --> 0:09:31.240 charged sustainable rates, even if it meant that it wasn't 0:09:31.240 --> 0:09:34.400 possible to build a big company based on their models. Sadly, 0:09:34.480 --> 0:09:36.960 we don't live in that universe. And while you could 0:09:37.000 --> 0:09:39.240 make the case that startups like Uber didn't at first 0:09:39.320 --> 0:09:41.679 charge sustainable rates, I'd argue that the reason why its 0:09:41.720 --> 0:09:44.560 initial rates weren't successful was because of the steep upfront 0:09:44.600 --> 0:09:46.840 cost of customer acquisition, which is the problem that could 0:09:46.840 --> 0:09:49.000 be solved through the lifetime of the customer and Uber 0:09:49.000 --> 0:09:50.920 had the means to gradually ratchet up the costs of 0:09:51.040 --> 0:09:53.839 rides or mores. Italy reduced the cut that they pay 0:09:53.880 --> 0:09:57.439 to drivers in a way that wouldn't be immediately paying for. Furthermore, 0:09:57.720 --> 0:10:00.440 Uber never had a fuel problem. What Anthropic has as 0:10:00.480 --> 0:10:02.640 a fuel problem, they have a compute problem for the 0:10:02.720 --> 0:10:06.040 amount that they're paying to run their goddamn services. Cursor 0:10:06.120 --> 0:10:09.160 is also Anthropic's largest customer, and the timing of priority 0:10:09.200 --> 0:10:11.280 tiers to coincide at the moment when they were growing 0:10:11.400 --> 0:10:15.360 fastest is a suspicious and potentially disgraceful move. While you 0:10:15.400 --> 0:10:17.760 could describe it as a necessary step in the direction 0:10:17.840 --> 0:10:21.280 of sustainability, that plausible excuse is undercut by the overall 0:10:21.360 --> 0:10:24.319 timing of the move. One cannot ignore how close the 0:10:24.400 --> 0:10:26.600 launch of these tiers were to the launch of anthropics 0:10:26.679 --> 0:10:29.680 clawed Code, a product that lacks curses, flashy front end, 0:10:29.840 --> 0:10:33.319 but performs similar functions, all subsidized by anthropics massive hordes 0:10:33.360 --> 0:10:36.880 of venture capital, and its chummy relationships with hyperscalers like 0:10:37.000 --> 0:10:40.040 Amazon and Google. The thing is, even with these moves, 0:10:40.080 --> 0:10:42.840 Anthropics still spent a dollar and four cents on Amazon 0:10:42.880 --> 0:10:45.240 Web services for every dollar they made through the end 0:10:45.280 --> 0:10:48.160 of September twenty twenty five, and that's for just twenty 0:10:48.240 --> 0:10:51.040 twenty five. By the way, their costs increase nearly with 0:10:51.160 --> 0:10:53.679 their revenue. And while they've improved, when they spent a 0:10:53.720 --> 0:10:56.200 remarkable two hundred and twenty seven percent of their revenue 0:10:56.200 --> 0:10:59.000 on AWS in January, they still spent eighty eight point 0:10:59.080 --> 0:11:02.520 nine percent of it on a western September. Now, if 0:11:02.559 --> 0:11:05.120 you're worried hearing how close these numbers before, like I said, 0:11:05.400 --> 0:11:09.920 means there's somehow approaching profitability. Good lord, No, I'm repeating myself. 0:11:09.960 --> 0:11:11.680 I realized, But I really need you to come away 0:11:11.720 --> 0:11:14.760 with this with reality in your brain. These digital mister 0:11:14.840 --> 0:11:18.320 beans very likely spend comparable sums on Google Cloud unlikely 0:11:18.360 --> 0:11:20.760 another billion or two one salaries data And I don't 0:11:20.800 --> 0:11:23.080 know that one point five billion dollar settlement with all 0:11:23.080 --> 0:11:26.000 the authors that they just agreed to. This company absolutely 0:11:26.040 --> 0:11:28.680 fucking sucks. I don't care if you like Claude Sonnet 0:11:28.760 --> 0:11:30.880 or claud Opus. I don't give a fuck. Claude Opus 0:11:30.920 --> 0:11:33.280 and Claudes on It are not worth burning billions of 0:11:33.360 --> 0:11:36.640 dollars a year in cloud costs, fueling an environmentally destructive 0:11:36.720 --> 0:11:39.719 plagiarism charge pseudo company that would roll over and die 0:11:39.800 --> 0:11:42.319 within months if it didn't constantly get fed billions of 0:11:42.360 --> 0:11:44.800 dollars a year. What are you gonna tell me they're 0:11:44.800 --> 0:11:47.000 gonna turn this ship around. They're gonna make some sort 0:11:47.040 --> 0:11:50.680 of autonomous AI coder. You know that's bullshit. Every goddamn 0:11:50.760 --> 0:11:53.600 one of you boosters knows that total bullshit. I'm sure 0:11:53.720 --> 0:11:56.360 son at four point five is somewhat better than Sonnet four, 0:11:56.559 --> 0:12:00.400 but what does that actually mean? Anthropic raised twenty billion 0:12:00.400 --> 0:12:03.000 dollars this year? Do we give them more next year? 0:12:03.840 --> 0:12:06.680 I've heard reports that they're actually targeting twenty billion dollars 0:12:06.720 --> 0:12:09.280 in anialized revenue, so one point six to seven billion 0:12:09.320 --> 0:12:12.679 dollars a month in revenue by the end of next year. 0:12:12.679 --> 0:12:16.040 It's an absolute fucking joke. But the only thing funnier 0:12:16.120 --> 0:12:18.160 than that joke is that it will likely cost them 0:12:18.160 --> 0:12:20.880 twenty five billion dollars to make that fictional money. And 0:12:20.960 --> 0:12:25.040 where preytell is that coming from? And why? Why? What 0:12:25.240 --> 0:12:27.360 is so remarkable about this company that gives them a 0:12:27.400 --> 0:12:29.760 free paster burn two point sixty six billion dollars in 0:12:29.800 --> 0:12:33.600 AWS in fucking nine months. I'm not talking about your 0:12:33.640 --> 0:12:36.640 cynical oh, Amazon is booking at is revenue crony capitalisms? 0:12:36.679 --> 0:12:39.080 Here an'tswer it. I'm not, I'm not. I'm talking about 0:12:39.120 --> 0:12:42.920 the scientific or technological reasoning for keeping Anthropic alive. And yes, 0:12:43.040 --> 0:12:46.280 I feel exactly the same way about Open AI. What 0:12:46.520 --> 0:12:50.080 possible achievement does Anthropic have that warrants this needless, endless, 0:12:50.200 --> 0:12:54.120 sprawling financial destruction. Why are we rewarding a company with 0:12:54.240 --> 0:12:57.079 bad business practices for making a product that loses more 0:12:57.120 --> 0:13:00.240 money the more money it makes. I'll even try and 0:13:00.360 --> 0:13:03.160 see this through the eyes of an AI booster. Damn, 0:13:03.200 --> 0:13:05.520 all I'm seeing is blue and yellow anyway, And even 0:13:05.640 --> 0:13:08.439 from here, the only reason to keep Anthropic alive is 0:13:08.480 --> 0:13:11.560 because you see these companies as sports teams. You see 0:13:11.640 --> 0:13:15.760 Dario Amiday as the equivalent of Dan Campbell or Greg Popovich. 0:13:15.800 --> 0:13:17.800 You root for them and their causes because you think 0:13:17.880 --> 0:13:20.760 that if they win, you as a fan will be rewarded. 0:13:21.120 --> 0:13:22.839 You don't think too hard about what it is that 0:13:22.920 --> 0:13:25.959 Claude Sonnet or Claude Opus do, and you find enough 0:13:26.000 --> 0:13:28.640 ways that this is somewhat kind of useful to you, 0:13:28.880 --> 0:13:31.400 and you use those reasons to justify the proliferation of 0:13:31.440 --> 0:13:36.840 a wasteful and destructive technology. What exactly happens here? Anthropics 0:13:36.880 --> 0:13:40.240 AWS bills are not really going down. They've normalized in 0:13:40.280 --> 0:13:42.439 an eighty eight to ninety five percent range and they're 0:13:42.480 --> 0:13:44.679 clearly going to stay there. And if your argument is 0:13:45.000 --> 0:13:47.760 they'll go down, your argument is quite literally no, eh. 0:13:48.480 --> 0:13:51.000 Go read semi analysis for seventeen hours and come up 0:13:51.040 --> 0:13:55.160 with some demented GPU based argument about inference smax scores, 0:13:55.360 --> 0:13:57.400 pretend like you give a shit, come up with a 0:13:57.480 --> 0:14:00.200 real argument against mine, because I am working hard, are 0:14:00.240 --> 0:14:02.240 at this than you are, And if you believe otherwise, 0:14:02.360 --> 0:14:04.320 you should ask yourself why the guy who said Sam 0:14:04.400 --> 0:14:07.439 Altman's no it loads refused cash dump in a premium 0:14:07.480 --> 0:14:11.079 newsletter got this scoop and you did not. But that 0:14:11.160 --> 0:14:14.120 actually leads me to a key question. How long do 0:14:14.280 --> 0:14:17.760 we hand Anthropic and by extension, open AI billions of dollars? 0:14:18.400 --> 0:14:21.080 And for the first time in your goddamn life, it's 0:14:21.160 --> 0:14:24.400 time to ask, what if I'm right? What if these 0:14:24.480 --> 0:14:28.160 companies are incapable of becoming profitable? What if there really 0:14:28.280 --> 0:14:31.960 is no massive demand for generative AI? Do you really 0:14:32.000 --> 0:14:34.720 think Anthropic will make one point six billion dollars a month? 0:14:34.800 --> 0:14:37.840 Sometime in twenty twenty six, do you really think that? 0:14:38.520 --> 0:14:42.120 And even for Amazon, it's kind of shit wow to 0:14:42.840 --> 0:14:45.880 do a couple billion on one hundred and five billion 0:14:45.920 --> 0:14:47.760 dollars of capex. I might have even said this later 0:14:47.800 --> 0:14:50.760 in the script, but just thinking about it makes me 0:14:50.840 --> 0:14:54.240 feel a little crazy. And look, I get there's a 0:14:54.320 --> 0:14:56.840 middle ground here where people say that there's some sort 0:14:56.920 --> 0:14:59.320 of use case that sort of works for AI, where 0:14:59.360 --> 0:15:01.440 you hit it hard enough to write good enough prompts 0:15:01.480 --> 0:15:03.320 or whatever that you like it for search, that your 0:15:03.360 --> 0:15:05.680 brainstorm with it, they helped you pick out a hat 0:15:05.760 --> 0:15:07.840 that you used it to solve some sort of problem. Once, 0:15:08.200 --> 0:15:10.680 I just want to ask you how much of those 0:15:10.720 --> 0:15:14.240 anecdotes really worth to you? How impressed with these things 0:15:14.360 --> 0:15:18.360 are you? Would you pay double, triple, quadruple? Would you 0:15:18.440 --> 0:15:20.600 pay on a meted basis where those little flights of 0:15:20.680 --> 0:15:23.000 fancy cost you a few cents, then ten cents, then 0:15:23.040 --> 0:15:25.600 a dollar, because that's how much it costs to provide 0:15:25.640 --> 0:15:27.320 these services, and at some point you're going to be 0:15:27.360 --> 0:15:43.280 made to pay for it one way or another. Advertising 0:15:43.320 --> 0:15:45.840 won't be the answer. By the way, the literal only 0:15:45.920 --> 0:15:49.080 company to try advertising in large language models is an 0:15:49.120 --> 0:15:52.840 AI search engine company called Perplexity, and they just paused 0:15:52.880 --> 0:15:56.280 accepting new advertisers too, and I quote ad week rethink 0:15:56.320 --> 0:16:00.520 how ads fit into its AI search experience. They made 0:16:00.600 --> 0:16:04.160 twenty thousand dollars in twenty twenty four in advertising revenue. 0:16:04.560 --> 0:16:07.520 Are we supposed to be impressed that Perplexity made enough 0:16:07.600 --> 0:16:11.080 revenue to buy a second hand Toyota Corolla. There are 0:16:11.120 --> 0:16:15.000 people making more money than that's slinging fucking Herbalife. And 0:16:15.120 --> 0:16:18.680 this is literally the exact company that should have succeeded 0:16:19.000 --> 0:16:22.080 based on any kind of ads will fix everything argument. 0:16:22.480 --> 0:16:25.440 And they couldn't even buy courtside tickets at the NBA playoffs. 0:16:26.480 --> 0:16:29.720 The costs are increasing linearly with revenue, and I've fucking 0:16:29.880 --> 0:16:32.640 proved it. I am open to any compelling arguments that 0:16:32.720 --> 0:16:35.640 can explain how this ever changes. And my god, if 0:16:35.720 --> 0:16:38.920 you say trainium, I will absolutely lose my shit. Chips 0:16:38.920 --> 0:16:43.400 aren't fixing this. By the way, if your answer is 0:16:43.440 --> 0:16:46.400 the anthropic will make some sort of theoretical ultrapowerful large 0:16:46.440 --> 0:16:49.080 language model or invent agi, you are a goddamn mark. 0:16:49.160 --> 0:16:55.480 You are being conned. Look join me. I'm serious. There's 0:16:55.560 --> 0:16:58.320 no harm in being wrong I've been wrong tons of 0:16:58.360 --> 0:17:01.400 times in my life. Being wrong and admitting you're wrong 0:17:01.520 --> 0:17:04.000 is an act of bravery. Shit, actually kind of get it. 0:17:04.480 --> 0:17:06.640 This stuff feels if you let it, like it's doing 0:17:06.760 --> 0:17:09.119 something for you, even though interacting with it is actually 0:17:09.200 --> 0:17:11.560 draining you because you're constantly having to find ways to 0:17:11.640 --> 0:17:13.240 make it do what you want it to do, to 0:17:13.359 --> 0:17:15.159 the point that when it actually does something for the 0:17:15.240 --> 0:17:19.120 first time, it almost feels magical. You feel very powerful, 0:17:19.560 --> 0:17:21.800 despite the fact that you have been put to work 0:17:22.119 --> 0:17:25.320 to make automation work. That's not how automation is meant 0:17:25.400 --> 0:17:29.119 to work. And sure, there are software engineers out there 0:17:29.119 --> 0:17:31.600 who have, like any good software engineer, found a way 0:17:31.600 --> 0:17:33.800 to take the useful parts of llms and use them to, 0:17:34.000 --> 0:17:36.239 to quote Carl Brown of the Internet at bugs, make 0:17:36.320 --> 0:17:39.000 the easy things easier. Then there are the ones that 0:17:39.040 --> 0:17:41.679 are spending more time than they were building software, prompting 0:17:41.760 --> 0:17:44.800 l elms and rewriting claude dot md files and thinking 0:17:44.840 --> 0:17:47.320 that because things sort of worked off they hit enter 0:17:47.359 --> 0:17:50.639 that they're privy to a great becoming. And there are 0:17:50.680 --> 0:17:54.119 the victims, of course, of vibe coding startups companies that 0:17:54.240 --> 0:17:56.840 sell the outright lie that somebody who cannot read or 0:17:56.840 --> 0:18:01.760 write software can write secure, effective and functional software. Look, 0:18:02.320 --> 0:18:04.960 I'm serious, join me. If you're an AI booster, I 0:18:05.080 --> 0:18:09.280 don't care. Everybody is welcome. In reality, I don't care 0:18:09.320 --> 0:18:10.879 who you are. I don't care if I've called you 0:18:10.960 --> 0:18:13.639 a booster and given you a verbal swirly one hundred times. 0:18:14.119 --> 0:18:16.200 Now is the time to accept that this software is 0:18:16.240 --> 0:18:19.960 too expensive, too destructive, and too wasteful to continue backing it. 0:18:20.359 --> 0:18:22.240 I'm not even saying you have to say fuck AI 0:18:22.440 --> 0:18:25.119 or shun chat GPT like you're an armish teenager that 0:18:25.200 --> 0:18:28.000 looked at porno. But it's time to be loud and 0:18:28.119 --> 0:18:30.440 direct that these products are not worth the egregious and 0:18:30.440 --> 0:18:34.879 perpetual annihilation of billions of dollars every fucking year. I 0:18:34.960 --> 0:18:36.560 don't even know if this means you have to stop 0:18:36.680 --> 0:18:39.440 using them. I don't want you to, but I don't like. 0:18:39.840 --> 0:18:42.520 What are we gonna do. These things are not gonna 0:18:42.560 --> 0:18:45.119 go away because you stopped using Claude. They're gonna go 0:18:45.200 --> 0:18:48.359 away because you stop talking about them. They're gonna go 0:18:48.440 --> 0:18:51.760 away because they cost too much in their pay pigs 0:18:51.800 --> 0:18:55.840 stop paying them. What I am advocating for is for 0:18:56.000 --> 0:18:58.639 everybody to openly discuss that the amount of money it 0:18:58.720 --> 0:19:01.040 costs to run these companies is that odds with what 0:19:01.160 --> 0:19:03.600 they have built, are building and will build in the future. 0:19:04.280 --> 0:19:07.439 Nothing they are building is moving towards superintelligence or AGI. 0:19:07.920 --> 0:19:11.199 No combination of Amazon Tranium or Google TPUs is going 0:19:11.280 --> 0:19:14.159 to usher in the birth of the machine guard. The 0:19:14.280 --> 0:19:17.640 products they make are, at best, and in inconsistent moments, 0:19:17.880 --> 0:19:21.600 kind of cool, but one hundred times more often mediocre, unreliable, 0:19:21.640 --> 0:19:26.600 and outright ridiculous, even if you really get a lot 0:19:26.800 --> 0:19:29.560 out of these models. Do you think that these companies 0:19:29.600 --> 0:19:31.920 should be allowed to burn billions of dollars every year? 0:19:32.760 --> 0:19:34.560 How much do you think they should be allowed to burn? 0:19:35.240 --> 0:19:38.520 And how much is too much for you? It's time 0:19:38.600 --> 0:19:42.280 to start having this conversation and having it publicly, especially 0:19:42.320 --> 0:19:45.200 as Clammy Sam Mortman Bloviate's about building two hundred and 0:19:45.240 --> 0:19:48.879 fifty gigawatts of data centers in seven goddamn years at 0:19:48.880 --> 0:19:51.680 the cost of one third of America's entire fucking economic 0:19:51.760 --> 0:19:54.680 output in twenty twenty four. Anyway, this has been a 0:19:54.720 --> 0:19:56.280 big day for me, so I'm going to leave it there. 0:19:56.480 --> 0:19:59.280 It's a huge scoop. I'm grateful that I get to 0:19:59.320 --> 0:20:01.879 do this every day, grateful for you listening, and grateful 0:20:02.000 --> 0:20:05.159 for your reading. I hope you've enjoyed this episode and 0:20:05.440 --> 0:20:15.680 thank you as eva for supporting my work. Thank you 0:20:15.720 --> 0:20:18.440 for listening to Better Offline. The editor and composer of 0:20:18.480 --> 0:20:21.440 the Better Offline theme song is Matasowski. You can check 0:20:21.480 --> 0:20:24.400 out more of his music and audio projects at Matasowski 0:20:24.440 --> 0:20:27.879 dot com, M A T T O. S O w 0:20:28.280 --> 0:20:31.800 Ski dot com. You can email me at easy at 0:20:31.840 --> 0:20:34.439 Better Offline dot com or visit Better Offline dot com 0:20:34.520 --> 0:20:36.920 to find more podcast links and of course, my newsletter. 0:20:37.359 --> 0:20:39.840