Subprime AI Crisis, Mr. Beast’s Guide For Excellence, Amazon’s Culture Problems

0:00Is AI in a subprime crisis as spending mounts without returns? The iPhone 16 goes on sale today. Are we supposed to be excited? Mr. Beast has an internal guide to excellence that we'll talk about. And is Amazon's culture in trouble? All that and more is coming up right after this.

0:20You're used to hearing my voice on the world bringing you interviews from around the globe. And you hear me reporting environment and climate news. I'm Carolyn Buehler. And I'm Marco Werman. We're now with you hosting the world together. More global journalism with a fresh new sound. Listen to the world on your local public radio station and wherever you find your podcasts.

0:47Welcome to Big Technology Podcast Friday edition, where we break down the news in our traditional cool-headed and nuanced format. So much is going on. OpenAI is fundraising. Apple's releasing new phones. Mr. Beast's large internal guide to how to be great at his company has leaked and it's fascinating. Meanwhile, Amazon has employees going back to work five days a week. And Andy Jassy has some very revealing comments about the culture potentially going south. So we're going to break that all down. Joining me as always is Ranjan Roy of Margins. Ranjan, welcome. It's good to see you back in New York as well.

1:22I see your Brooklyn apartment right behind you. And I think after all of our summer travels and podcasting technical complexity, it's good to both be back in New York right now. I've never been happier to have exposed brick behind me, good Wi-Fi around me, and a solid mic in front of me. This is it. This is nirvana for podcasting right here. This is all you need in life. Exactly. Exposed brick, a sure microphone, and good Wi-Fi. and what a week to talk about artificial intelligence because there's been this thing that we've been talking about for a while which is that some of the math doesn't seem to add up when you have the amount of money that open ai is losing and the amount of money it's trying to invest and still dubious sort of application of the technology within companies and a piece from Ed Zitron put it all together this week.

2:21Now, let me be the first to say that I don't agree with this post completely. I think that it is wrong on a number of facts, but it's interesting to discuss because I think it very clearly articulates the bear case on AI and it's worth picking apart. Yeah, I think the Ed Zitron piece, and I agree that it's definitely in a number of his pieces have been very, very aggressive against open AI, but I think they echo a number of the things that we've been saying. And one of the most interesting parts of this piece was it actually brought me back to that Goldman Sachs report from I think three months ago that we were saying could be the canary in the coal mine for generative AI.

3:02Because the biggest question is going to be when people actually have to pay for things. When companies, you know, they invest the budget upfront, but when they actually have to start renewing, when they have to start actually extending the contracts beyond just proof of concepts and small innovation group type contracts. And I think we're starting to see it. He lays out a lot of really good detail around what the cost structure looks for open AI in companies like that in that category, and then what the economics look like from the client side. And they just don't match up. They don't match up in any way.

3:40Absolutely. So let's go through some of the numbers because he does do a very good aggregation of lots of numbers that have been reported. And I feel like it's valuable to just speak them out loud and we can talk about them. So he says OpenAI is paying Microsoft an estimated$4 billion in 2024 to power ChatGPT and its underlying models. And that's with Microsoft giving OpenAI a$1.30 per GPU hour cost, as opposed to the regular$3.40 or$4 that other customers pay. So basically what he's saying is that OpenAI is basically paying Microsoft a third of what you would pay typically. And he said training costs are currently$3 billion for the year and certain to increase.

4:28The information has reported that OpenAI's revenue, meanwhile, is$3.5 to$4.5 billion a year. And this is sort of the... Oh, here's another point. The Microsoft... So if you think about the AI providing value to enterprises, this was a pretty stunning stat that he shared. With Microsoft's 365 suite, which has Word, Excel, PowerPoint, Outlook, he says that companies are barely adopting the AI-powered co-pilot, with somewhere between 0.1 % and 1 % of its 440 million seats doing so. What do you think about when you hear all these stats? Yeah, I think, so there's three parts, as you kind of outlined there.

5:17There's the economics for OpenAI, the fundraising for OpenAI, and then the adoption within the enterprise. Let's start with the adoption with the enterprise. So that stat between only 0.1 % to 1 % of Microsoft Office 365 subscribers have actually signed up for Copilot, which is an additional$30 to$50 a person. I think one of the biggest miscalculations that Microsoft and Google are making is starting by anchoring the price at a per seat basis like this before they have shown genuine value for these products, for these add-ons essentially, and made them so good and so, you know, where people are addicted to it, where every user is like, I can't live without this.

6:05But everyone I know who's tried out these products, some of them work pretty well. Some of them don't. And it's not the magic that you're just going to go in Excel and they just release something where you can actually, you know, there's like a Python compiler that you can actually integrate within Excel. But that was just released as a demo. And that could work pretty well at some point, I'm sure. But in the meantime, we will all just upload our CSVs into Claude and ask it to do basic analyses for us. like they came out so strong and I get the business logic of anchoring the price at a high enough point because you see at some point the value will get there but when it's not there and suddenly you ask people companies that have 20 to 50 ,000 seats imagine you're a CIO and you have like 50 ,000 seats with Microsoft Office 365 and they're asking you to add$30 to every single seat.

7:01And I'm sure you'll get a little bit of a discount when only a few of your people even know how to use generative AI. So to me, the adoption part, I think is going to be very rough for on the enterprise side. And again, we've talked about it, that it's that first renewal period. That's kind of the, where the rubber meets the road and we're getting there right now. Before you move on, I would just want to note that like they sold$30 a seat licenses back when the hype was at its height. And you basically would get a yes to say, we want to pilot AI in any organization, in any environment. And if you didn't pilot AI, or if you said, yeah, we're not going to use this stuff, your job was at risk.

7:41And now they've tried it. And the number of people getting value from it is so low. And you're right. It's the first renewal that comes up. Okay. So that is the enterprise application. now let's talk about open ai's economics as you said as the numbers are laid out it starts to get so much weirder and actually i think the most valuable part of this piece again as you said it took all these different bits of reporting and really laid out a very strong case the fact that open ai is getting a significantly discounted rate from microsoft that means that its economics aren't even being truly reflected in whatever it's showing to investors.

8:22And we'll get into the funding. Then that$3.5 to$4.5 billion a year in July number that was reported, we talked about this. It was a bit fishy because it was actually even from a few months before. And then again, it was extrapolating a run rate out for the entire 12 months, which is never a good sign for any kind of revenue projection when you're actually talking about where is this going to go, especially when generative AI's hype has already been at its peak and is starting to wane a little bit. And then there was additional reporting from the New York Times that said it's actually only above$2 billion, that it's going to be at the lower end of the year to date.

9:01So it's definitely going to be at the lower end of the estimate. So the company is going to definitely keep burning money. And there's two ways to approach this business strategy. take the models you have as of today and try to get more people using and paying for them and i don't think that's an unreasonable business strategy because chat gpt plus is pretty good chat using 4.0 4.0 mini but instead they're going all in i'm sure they're investing heavily in sora i'm sure they're i mean uh they've released oh one they're trying to say reasoning is the next level agentic AI, et cetera, et cetera.

9:40So rather than just focusing on the actual adoption and conversion of users, they're still going out in this all out AI arms race mentality that we just have to build the biggest models possible and the most advanced models possible before the actual business case is proven in the near term. So I think it is a reminder on the actual OpenAI's intrinsic economics, not separate from their ability to fundraise, they are so far from making any possible sense right now. And again, we don't have great numbers or reporting on a clod and how Anthropic is doing with it, or even perplexity or any of the other competitors in the space.

10:23But it seems to me that any company that's going AI arms race, builds the next generation of models rather than let's just actually build out this business is is at least not on a good trajectory right now i think ranjan this might be why we haven't had any open ai guests on the show but i'd rather speak to these issues truthfully and not tell not sugarcoat them in the hope that sam altman might come on the podcast like we're here to tell the truth to listeners and this is the truth and we're going to talk about it and if opening ai doesn't want to come on and face tough questions and speak about this stuff, then that's fine.

11:01And I do think, though, this is a good place for nuance, which is that there's two things here that we're discussing, right? So obviously, we talked about the low adoption of Microsoft. That might just be of the Microsoft co-pilot, and that might just be a business situation. But we can both agree that the underlying product, generative AI, is actually extremely valuable. And we've both found super valuable use cases in our day to day using chat GPT using Claude. I had a crazy thing that I started this week where, you know, you would have needed an app for this previously. But I like told Claude, Claude, I'm going to go on a whole foods diet, I'm going to update you with all the meals that I'm eating.

11:43And you give me a whole food score and give me coaching on how I'm going to get to more whole foods type meals. And I'm just like, typing in taking pictures of stuff, It reads the labels. It gives me feedback. It gives me whole food score. It's like an unbelievable nutritional coach. And of course, like, does it hallucinate a little bit? Probably, but you can still see through that. And it's just been, it's unbelievable. Like this, this is a better application, quote unquote, than any fitness app, you know, I would have used previously. And it's all built with Enclod and I'm paying$20 a month for it.

12:16So wait, sorry. What is a whole foods diet? Whole foods diet is you want to eat basically food that's not processed. And going back to my interview with Johan Hari about I was just gonna I was just gonna say I listened to your Ozempic grocery episode. And I, I much more than that Netflix documentary, that interview that you had with on the Ozempic topic, I think almost radicalized me against processed food more than yes. Yeah, reading the book, speaking with Johan left a lasting impact on me. And yeah, so I'm trying to reduce processed foods and I'm using Claude to do it. And I'm paying$20 a month for Claude, the premium version.

12:58And that's just one use case that I have with it. So obviously it's a great deal for me. And obviously there's great value in this technology. Like people who are like this technology is worthless and just a party trick are absolutely wrong. But I think the question that we're getting at here is, are the economics feasible in the long term? And I think that really is the deeper question. No, that's a good point. I mean, any longtime listener knows very well that I'm actually incredibly bullish on generative AI. And I think it's the most transformative technology in the world. And I think before the big technology podcast becomes a health and nutritional podcast, and we start selling supplements, this is what we focus on the actual business and nuance of the, like how this is playing out rather than just getting too hyped up in the actual technology.

13:48And I agree. And my question there is, is$20 the actual optimal price for Anthropic to charge you for using Claude? Or is it still underpricing its product in order to increase adoption? And they're actually losing money based on the compute that you're requiring. These are the kind of things that we don't know yet. And it seems like, at least from the OpenAI numbers we have, that the business model side of it, these companies are not paying attention to. I know, like, OpenAI, for myself, I've had, even paying a reasonable amount for API access, essentially, on the enterprise side, they don't have built-out customer success teams in the way that Microsoft or Google do.

14:41They're not building out the business the way a traditional enterprise company would. Even a SaaS startup that's scaling and growing that's worth like$50 million will have a more built out infrastructure traditionally to serve the enterprise. So the way they're building out these businesses, it's different. It's still more research lab-ish, which is a good segue into the funding side of it. But before we get there, I think that this is an important part to stop on, which is that this could be the subprime issue, which is that you have AI companies building on top of this software, not just AI companies, everyday companies.

15:21and I mean I'll shout out some of the companies that we think have like pretty uh constructive approaches on AI companies like ServiceNow and Box and are they just going to get so used to these low prices that when and build their products around them that when the rubber meets meets the road and these companies need to justify themselves financially the prices go up and it becomes challenging for the business models and the companies that have gotten used to building with AI and the clients that have gotten used to using the products built within the system. I mean, that's basically the idea of subprime here.

15:55I'm just going to read it from Ed's piece. My concern is that I believe we're in the midst of a subprime AI crisis where thousands of companies have integrated generative AI at prices that are far from stable and even further from profitable. Almost every AI-powered startup that uses LLM features is based on some combination of GPT or Claude. The models are built by two companies that are deeply unprofitable. Anthropic is on course to lose$2.7 billion this year and have the pricing designated to get more customers rather than make any kind of profit. Actually, okay, yeah, this is a good point to stop on.

16:33I agree because, again, having worked in trading at a big bank during the 2000s and sitting front row for the subprime crisis that drove the global financial crisis. That is what I thought was the most powerful part of this piece. And I think the phrase subprime AI for readers, get used to it, because I think you're going to be hearing it for listeners. You're going to be hearing it a lot, lot more. Because, yeah, it's this idea that there's been no true price discovery around the businesses being built on generative AI. and when everything is subsidized, anything that goes wrong. OpenAI, if they don't raise this next funding round, suddenly they have to jack up prices.

17:16The downstream effects of that will put any number of startups out of business. It'll completely transform what a box or whoever else has to charge different clients. And the downstream effects will be so significant that, again, it can completely derail the type of progress that we're at least all hoping for. And yeah, I think that's my biggest regret about this. And as someone who is very proud of the tech or very excited about the technology is this idea that because people are ignoring the actual business fundamentals around it and simply just trying to grow at all costs and assume it'll be subsidized, that we're not going to build the proper business case for generative AI.

18:02And you know how we like always talk about how this stuff is kind of weird and doesn't fully make sense when you think about the structures and the goals and the mission and what's happening. I just want to read this one thing that I found was very interesting from this person on Twitter, Rune, who apparently works for OpenAI. I'm just going to read it because I think this does really get the mindset within the company, right? He says, unfortunately, I don't think building nice AI products today or making them widely available matters very much. Minor improvements in daily active users or usability especially doesn't matter.

18:35Close to 100 % of the fruits of AI are in the future from self-improving intelligence. Every model until then is a minor demo, pitch deck, to hopefully raise capital for even larger data centers. People need to look at the accelerating arc of recent progress and remember that core algorithmic and step change progress towards self-improvement is what matters. One argument that has been that products are at a steady path towards generality, general intelligence, not sure that's true. So basically, this is what's going on in the minds of the research houses. Basically, that they're making money right now is an accident.

19:15Let me actually put a finer point on that from Ethan Mollick from Wharton. So he posts this photo of the tweets from Rune on LinkedIn and comments. this is an important thing to understand the major ai labs including open ai are very much focused on racing for the future it is almost accidental that the early products are making billions of dollars their goal is explicitly agi artificial general intelligence not your use case they are not spending much time commercializing because most of their effort is going to be is going toward building more powerful AIs. So this is one of those cases where I thought I was making a potentially an original point by talking about them ignoring the business case and actually building out business infrastructure.

20:07And then someone just tweeted it out. They said, we actually don't care about the business side of it. And your daily active users and profitability are not even worth spending time on because it's just about AGI. And yeah, I think Ethan Mollick summed it up well. That's exactly it. And it's interesting because they were nonprofit, research-oriented companies to start. So they just never converted. So at least even if you start to convert some of it and you start to get a different cap table, in the end, the core culture of the company is not, you're not Oracle. You're not like Salesforce. you are still at research at your core man this is just so fascinating it's the big think about this this is the biggest technological development of the decade it is being driven by folks who don't really care i mean if taking this at face value i think there's truth to it who don't care as much at about the current moment as you might imagine who are working towards this mythical goal of AGI and the economics don't make sense.

21:16I mean, that's, that's one way to look. The other way to look at it is of course, they're all about economics and you know, they're, they're using the AGI thing as hype, which we heard about last week from, um, Parmi from Bloomberg. But I mean, if they were just kind of ruthless business people, you'd think they'd be better at commercialization. So where, where do you think this is going, man? Ah, I wish I knew. I mean, well, I know where this is going. We're going to trough, trough of disillusionment. It's going to be there. It's already kind of starting to get there. There has to be a shakeout.

21:48There has to be some event that forces companies to be more realistic, forces an open AI or an anthropic to say, actually, it's not about AGI. Let's actually increase our daily active users and try to understand how to commercialize this technology. There has to be something. But as long as the money keeps flowing in, it's going to keep pushing that back. And the money does certainly appear to be flowing in given OpenAI's latest fundraising. There's reporting yesterday that it's oversubscribed and people are clamoring to get into it. and they're going to get to choose who gets to actually allocate capital into the round.

22:31So it appears that the interest is still there. Right. But there's also another interesting detail, which is that, first of all, they've already gotten a billion from Thrive Capital, which is Kushner's VC fund, the FT reports. But Andreessen Horowitz and Sequoia, not part of the new round. Now, the FT says they're sitting out. so I'm curious this is potentially interesting from two ways right okay they're trying to open AI is trying to raise it a 150 billion dollar valuation so do these companies uh sorry do these VC firms think that it's just not going to work from a business point of view and that's why they're sitting out I mean think about Andreessen which is so public relations oriented deciding that they don't want to be in on the next round of open AI like that's a pretty interesting decision or is open AI telling, you know, picking and choosing and being like, yeah, you know what?

23:24We're not going to go with you this time. Yeah, actually. So in the funding, some of the news that came out, like again, Sequoia, Andreessen Horowitz on the sidelines, maybe you can make the case that the more generous case would be, they have so many competing investments now that they don't want to come at this late stage with an open AI, generous interpretation. But then to me, the more interesting part is, so you have strategic investors, Oracle, Apple, NVIDIA, Tiger Global has shown up again. Even the United Arab Emirates MGX fund is supposedly one of the big investors for this latest round.

24:04So suddenly you're getting back into this world where the most on the ground investors are stepping back. And then it's all the outside money, whether it's a strategic corporate, whether it's a sovereign wealth fund, whether it's Tiger Global, who's become quite famous for late stage momentum style investing. And I've written about numerous times. So we are definitely at that moment where it's gone from pure tech investment, like promise on the technology to let's just try to keep growing this thing by just funneling more money into it if this goes up in flames how do you think it happens i think something falls through in this round and actually one thing that you've seen in the way every open ai funding open ai for as we've been saying a research house well one thing they're very good at is fundraising pr because and this is something have you tried search gpt it's like google but with open ai And therefore, the valuation should be trillions.

25:08And here is O1, which is reasoning and agentic because everyone is saying agentic AI. So I actually think, again, maybe when we said ruthless business people, the one thing they are is ruthless marketers because they are two investors, two investors. Because the one thing they're very good at is one, as you just said, kind of dangling the next big phase of growth in front of investors. But I honestly think the other thing is these leaks around fundraising. Because again, in the good old days when an actual fundraising round meant that every investor tried to avoid talking to journalists about it, and it would only maybe come out after the funding was done through a press release, and they wouldn't even disclose the valuation.

25:54Now, every day, there's more news that makes it seem positive. Suddenly, it's oversubscribed, and this information comes out, and everyone's clamoring to get into it. So I think on both counts, they're very good at spreading the word around their fundraising needs and kind of pushing that narrative. But, yeah, I think that's if something goes wrong this round, I think this is going to be a huge problem in the industry. Some indication that the money is not infinite would completely change the entire course of where generative AI is going. Fascinating. Okay, so let me ask you this, because I hate starting a segment with a question and then not answering it.

26:35Is there a subprime AI crisis going on right now? Yes. 100 %? I 100 % believe it, because we kind of walked through this entire thing that the economics don't make sense. No one at these companies is actually trying to make them make sense. And they don't make sense even from the client side in terms of pricing what a Microsoft is charging for Copilot. So I'm not just putting OpenAI and Anthropic and the likes in this. And then the downstream effects, no one knows what the real economics of the products they're building are. Maybe your Suno AI music creation app actually needs to cost$70 a month instead of$10 to make it make sense for the company.

27:19There's so many downstream apps and products and like services that are being created right now that could just completely get upended the moment someone at the foundation model companies has to actually turn a profit. Okay. And I'm going to say maybe not, because the one argument that you can make on behalf of the research houses is that the product that they've delivered has just gotten so much better at a point where we expected it to slow down. and if you think about like early chat gpt compared to where the models are now early chat gpt compared to where the latest claude model is it's crazy things like artifacts like you've talked about here on the show where you can build you know different games or even like have it write up documents for you and then alter the documents based off of prompts i think this stuff is especially powerful and i think that yes we're getting a good deal on it but the prices have also been coming down, at least from the APIs, and there's probably a happy middle that they can live in.

28:22So that would be the optimistic take. Although I don't, having walked through the subprime argument, I don't fully rule it out. It's possible that Ed is right. All right. I'll admit it. I'll take that. I'm sticking with 100%. Yes. And where I'm going from here, the place that could really save artificial intelligence, I'm going to go online to the Apple store and see whether I can get my hands on an iPhone 16. No, I'm just kidding. Speaking of overhyped artificial intelligence, let's move on to the iPhone. I will not be online. In fact, it doesn't even seem like there are lines for this phone.