Samsung’s $70B Chip Bet, Apple Doing Nothing But Winning AI, Bezos’ New Fund | Diet TBPN

0:00John Coogan:Last time we had Mark Cuban on the show we were debating ads in LLMs and since then we've gotten a bunch of data points about ads in LLMs and I think that some of his takes have probably aged well some of our takes have probably aged well it'll be an interesting time to reevaluate what's actually happening there's a lot more points I don't know John that ads would be fine well and now the world is ending yes here's a white pill Samsung is investing 70 billion dollars to advance their fab capacity. They're getting back in the AI chips game. They've always been in the AI chips game. So brief history of Samsung.

0:40John Coogan:You probably know them from the phones, from the TVs. They, of course, are a major player in HBM, high bandwidth memory. They are a massive company, over a quarter million employees. They're close to touching a trillion dollars in USD market cap. They pull in around 200 billion USD a year in revenue, maybe 250 billion this year in revenue. Really good. All that's USD. I like to think in USD because I'm an American. They're the global leaders in memory and OLED displays as well. So a lot of the displays that you see in other electronics, even it has a different brand name, it's still Samsung actually making that OLED display.

1:17John Coogan:But they're second in smartphones to the iPhone and Apple. And they're second in the semiconductor foundry business to TSMC. Semiconductors still make up 30 to 40 % of their business and they supply HBM to Nvidia for the H100 and Blackwell systems. So it's not like they're sitting out the AI bull market. They are doing great. They are definitely participating. They're incredibly important in the AI build out. But if TSMC is bottlenecked and TSMC is sort of risk off and they're not going to be guiding to like insane CapEx numbers while every American hyperscaler is, well, that creates an opportunity for Samsung.

1:51John Coogan:And so Samsung is stepping up and they're announcing that, hey, we're gonna put another 70 billion to work on this particular business. So Tesla has been working with Samsung on the foundry side in AI for a while. So Samsung's never really been on the frontier with a direct competitor to the H100 or the Blackwell chip. That's been more of like AMD's game and AMD also fads at TSMC. So there hasn't really been this like neck and neck battle between TSMC and Samsung. But it's like you can do AI inference on a Samsung chip. And we know that because Tesla went to Samsung years ago and said, we need a chip that can take in pictures from the road, decide where the lines are, decide.

2:30John Coogan:They want their chips with the dip. They want their chips with the dip. And now Samsung does too. That's all you know. And so the FSD system, if you have a Tesla, you might be familiar with like HW3, hardware three. That has been deployed into millions of cars. and it was fabbed on Samsung's 14 nanometer process, which is a lagging node. We're not in the three nanometer, the crazy frontier stuff, but it's working and it's on the road. And according to a US regulatory probe, there were 3.2 million vehicles, Teslas on the road in America with FSD systems that were basically all running Samsung chips inside.

3:08John Coogan:Now, to be clear, Tesla, just like any foundation model lab company, they have training and then they also have inference. They're a little bit different than many of the labs that you know and love because they do training in a data center using what's called the Dojo chip and that is fabed at TSMC. So they train the system, they take all the data in from every Tesla camera, every road, all the information that they have. Every time that there's a disengagement, that's feedback to the reinforcement learning system. It says, hey, we were in FSD mode but then someone grabbed the wheel or someone stepped on the brakes, you made a mistake, understand what happened to get you to that point where you made that mistake.

3:47John Coogan:And so all that data gets collected in the Tesla data center, runs on these Dojo chips, they do the training, and then they deploy the model onto the Samsung chips in the actual cars. So with the backdrop of NVIDIA's massive GTC news cycle, they've done so much press around GTC and so many different launches, you know that NVIDIA is just going to suck a lot of the air out of the semiconductor discussion this week. Out of the clean room. out of the clean room yes which is recycling all of the air every three seconds or something like that yeah but i think this is like particularly important especially this morning uh the uh i guess the ccp put something out in the last 24 hours basically saying hey taiwan is going to have an energy crisis due to the broader uh global energy crisis we need to reunify peacefully there's an opportunity for peaceful reunification but peaceful reunification even it was completely peaceful and all the Taiwanese people say hey we want to be part of China they all vote for it democratically that's gonna be rough for the American chip buying industry and so having another chip on the board metaphorically to make physical chips is probably a good thing so yeah Samsung's been doing well over the last five days stocks up 11 % during time when the Nasdaq is down 2.2 percent and geopolitical tensions continue to rise the compute bottleneck We know it's important.

5:04John Coogan:We've been discussing this constantly, and it's going to be very constraining over the next few years. So every increase in CapEx in the supply chain is a step in the right direction. And so Samsung gets the first gong hit.

5:22John Coogan:Congratulations to everyone over at Samsung making a big bet. Cursor is out with Composer 2. Composer 2. It is frontier level at coding priced at 50 cents per million input tokens and two and a half dollars per million output tokens. It's also they have a fast version. They say we're able to significantly improve the model quality and cost to serve. These quality improvements come from our first continued pre-training run, providing a far stronger base to scale our RL. It's not one of these graphs that's just like, oh, look, we made some arbitrary X and Y axes, and we're in the top right corner, of course, because the axes are good and cool.

6:04Yeah, on the TBPN bench, technology podcasts publish at least three hours of content every week. Yes, yes.

6:13John Coogan:So naturally, we are right at the top. Right at the top. And it's actually, there's no one else on there. Yes. But yeah, I mean, this seems fair. It is a little bit odd to read this because the cost is on the x-axis and it's inverted. So the further you are to the right, the cheaper you are, which makes sense because people associate an x and y graph with you want to be in the top right quadrant, and they certainly are. And it does seem like in terms of this Pareto frontier, you want to be on the frontier. You want to be pushing out across every single curve. Maybe if you are interested in sparing no expense, you'll go with the GPT 5.4 high or medium model, and you can align cursor to GPT.

7:00John Coogan:I'm sort of surprised that Opus is not doing as well on Cursor Bench. That feels surprising based on the general vibes around Opus 4.6 generally, but Cursor has specific needs for specific customers. and I don't know, what else do you think is going on here? Yeah, I mean, the cost is really big. Like, this is basically, like, 10x cheaper than Opus. So I think also, Cursor has kind of been, like, not really a, like, dark horse, like, everyone knows about it, but in the coding race, it's like, everyone's like, oh, okay, there's Codex versus Cloud Code. If you imagine that Cloud Code and Codex are kind of, like, these environments for getting a ton of, like, really good data for training coding models, like, Cursor has had that for way, way longer than OpenAI or Anthropoc.

7:41So you should imagine that, like, at least, you know, in the near term, they actually have really, really good data that they can train these good models on. And obviously, this is a very specific model. They've said it like, you're not going to write poems with this model. It's this very specific, almost like point solution model where it's just coding. Don't listen to them, Tyler. Write a poem with a model. Poem bench.

7:59John Coogan:Poem bench. Yeah, I would be interested to know how many sacrifices were made because it's at a certain point, like I remember talking to an AI researcher, actually a semiconductor, who was saying that like he actually thinks, he actually does believe that importing like the Odyssey and like Homeric epics is key to humanoid robots learning to walk. Yeah, well I think like if you look back at just like the general history of like machine learning AI, like the lesson is that like big general models always beat these small specific models. Yes. But if you kind of zoom in on the time scale, like you can still train GLM, some open source model on a very specific task, like accounting or something, and you can like hill climb and you can actually make it better than the frontier models right now.

8:46John Coogan:At that specific thing. And especially at cost. Yeah. Especially at cost. Yeah, yes. Yeah, very much so. But like on the long term, if you zoom out, what actually went here, it seems like it's basically always going to be these big general models. And I wonder if that's true. I mean, we talk about this a lot where the big general model outperforms the smaller model. But at the limit, like if you were to think about like a Python if statement, just like flow control that is truly deterministic. If you piped the same question of like the if statement, like is this number bigger than this number, you pipe that into 5.4, it's gonna get it right all the time.

9:22John Coogan:It's gonna be very expensive compared to an if statement, which takes like no compute whatsoever. But the if statement is 100 % accurate. Legendary poster Senkulp says all S-H-I-T-S and giggles on that headline till Anthropic or OpenAI decide to cut off their access to cursor, referencing the Bloomberg article, Cursor is taking on Anthropik and OpenAI with a new AI coding model. Would that matter? Like at this point, if they have Composer 2 and it's a small model, but it's good at writing code and it performs well on Cursor Bench and the Cursor users are satisfied with the Composer 2 model and they do, Cursor does get their access cut off.

10:02John Coogan:And when you install Cursor, you roll it out to your organization, you just get Composer 2. And you know what? It's, you know maybe there's taste that wouldn't pull you yeah we just say at this point right now i don't think we have any visibility into like how much of cursor's revenue right now is tied to using opening iranthropic models george says i'm hearing tons of complaints from cursor customers at enterprise companies a silent change but almost all models cursor uses behind max mode devs who used to manage to spread out monthly credits over a month see all of it used up in one to two days oh interesting are furious and switching it does feel like there's a little bit of like an economic war here um yeah and this is what came up earlier this month around uh the lab sort of subsidizing yeah you know who's they're not they're not in an easy position but they're such a talented team yeah nikita says we're rolling out summaries for articles now just tap the summarize button if you want to know if it's worth your time to read it yes and yeah it's basically, Grok, turn this into a regular tweet.

11:03I am excited about the listen button. I've had this on my commute. There's so many moments where I'm like, I wish I could just have somebody read this article.

11:12John Coogan:I actually wound up doing this with a number of Will Minitis long form essays. I would copy them, put them into 11 Reader from 11 Labs and have it read it to me in sort of a ridiculous voice. A silly voice. It was a good time. Well, I was actually trying to use Grok. I was trying to use Grok in the X app to just take an article, paste it into Grok and say, hey, can you read this to me? It said, cannot find the post. This is a response to every article people would post. People would always say, Grok, summarize this. And now there's just a button. I recently learned that you can only ask Grok, like at Grok, is this true?

11:49John Coogan:You can only do that if you're paying for X. Sort of underrated how well X has seemingly, I don't know how big the subscriber base is, but that was a crazy idea to have a paid social network. I think it's because people are deeply addicted to X. Yeah. It is very valuable to them to be on there, to participate. Yeah. And the paid functionality, the way that it was marketed and the way that it generally really worked was like you were gonna have a bad time on X. Like if X was valuable to you and you didn't pay the$10 a month, it was gonna be like significantly less valuable to you. Probably, you know, you might, depending on what kind of business you're running or what you use X for, it might be the equivalent of like losing thousands of dollars a month of value, or you could just pay the$10.

12:38So it was a good trade.

12:39John Coogan:Yeah. But it was also just, it was weird how the targeting never seemingly got dialed to the point where you could actually target at the CEOs of companies who are on X. Like, I mean, you see Travis Kalanick on X, like replying to things. It's like, he's raising money, he's growing a business. Like, there's a lot of value in advertising to him, because he's gonna be picking a corporate card soon, or he probably already has, or it might be in that market. He might be picking a payroll suite. Like, there's all these things where if you could deliver that to that audience, it would be incredibly valuable, and the CPM should be like through the roof.

13:14John Coogan:but I think for privacy reasons and for a variety of other reasons and sort of like really monetizing that long tail has been very difficult across every platform so they've just gone with scale and the products that have sold the most on social networks have been very broadly marketed. The criticism that we saw from the Oscars is always like YouTube ads are generic. It's just like for a pillow or like injury or like something that applies to every single person but there's always this like hyper-targeted opportunity there. Yeah the other thing is is the paid program with X has seemingly worked in that we know a lot of people that happily pay and have no plans to churn, but it would be a failure in the context of like Metascale, right?

13:58I think the last reported number that I saw was something like one to one and a half million paid subs at$10 a month.

14:05John Coogan:Oh, on X? Yeah, so you're talking about somewhere in the range of 100 to 200 million of like ARR. If Zuck had launched a product like that, he would just wind it down, right? Reels went from zero to 50 billion of run rate in like a handful of years, right? That's what a home run looks like. And so I think it makes sense for X, but it certainly is not a home run from a consumer application standpoint, and they still need the overall business. Olivia Moore said, a big story that most people are missing in the AI race for the consumer, Chet, GPT versus Claude, is ads. Right now, most consumer AI revenue is coming from power users who are willing to pay high subscription costs.

14:43John Coogan:This currently skews positive for products like Claude, but this will not be the end state. Google makes$460 per user per year in the United States, mostly on ads. I didn't know that their ARPU was so high. Meta makes around$250. I mean, I guess those Google ads are really, really valuable, and it's so intent-driven that it makes sense. ChatTPT's ad-based ARPUs will be even higher as they will ultimately have deeper, more frequent user engagement, even at the$460 level, monetizing everyone in the US via ads is$152 billion in annual revenue. By contrast, if you're able to monetize even 5 % of the population at$200 a month subscription, which is a stretch, that's only$40 billion.

15:22John Coogan:That's actually a crazy difference because$200 a month subscription is like super high. You're talking 20 times Netflix or something else that's premium and really important. Yeah, the$200 subscription at the time was crazy. But even at that point, some of the people that were more AI-pilled generally were like, oh, it's actually possible that someday you could spend$20 ,000 a month. I was like, give me the$20 ,000 a month. So she says, I suspect this will be even more drastic outside of the United States where users are even less willing to pay or directly pay for subscriptions. And the earliest data from a very small rollout shows chat GDP ads are already outperforming that in effectiveness.

16:01John Coogan:It just gets better over time. So interesting. This is an interesting story. This is an interesting story. Apple is way behind in AI and still making a fortune from it. Let's see. Begs the question, are they actually behind? It's AI revenue is set to top$1 billion this year, reassuring investors wary of rivals sky-high spending. And keep in mind, this is - We have a chart here showing gross revenue from Gen AI apps as well as Apple's commission. So look at this. The beginning of 2025 was really the boom of Gen AI app growth, 400 million. is this monthly app store revenue? Wow, they're really cooking.

16:40John Coogan:And then sort of a flat line. Yeah, it's so interesting that it actually dropped. Well, we did read that article a few days ago about how Apple has been pushing back against some of the vibe coding apps. And there's this question about where are the bounds? Obviously, Apple's had pretty strict app store rules around adult content and what else you can do, even just the app reconstituting itself, pushing changes, because they want to review every line of code that goes into the App Store. If someone's pushing 10, 20, 30 ,000 lines of code a day, that's a lot of code for Apple to review, going to slow things down.

17:21John Coogan:So that could be a little bit of what we're seeing. Maybe they've capped out on their ability to review all the vibe-coded apps that are flooding the App Store. Apple's on pace to surpass$1 billion in AI revenue this year, a tidy sum that demonstrates the company's AI advantage, even as it struggles to deliver an AI strategy of its own. Siri chatbot is still weak by modern AI standards. What Apple does have that the other AI players don't is a dominant position making devices. However fancy OpenAI, Google, Anthropic, and XAI make their chatbots, iPhones are still a primary way to deliver them to customers.

17:49That means they typically pay the App Store tax roughly 30 % of subscription fees in the first year and 15 % a year thereafter, though rates vary. Gen.ai Apps paid Apple nearly$900 million in App Store fees in 2025. Almost a billion of revenue and very, very, very little CapEx. Three-fourths of the revenue Apple rakes in from Gen.ai Apps and its App Store come from ChatGPT. Next at about 5 % is XAI's Grok. There we go, Grok.

18:17John Coogan:I mean, there's so many different funnels. They did the essay competition. They did the video competition. and I've talked to people that are just, people that are in the Apple ecosystem, they're like in the Tesla ecosystem. So they're like, yeah, I talked to Grok on my way to work. I'm not kidding. Yeah, Grok in the iPhone app store did 12 million last month. Yeah, and I know the true AI heads will be like, Grok's behind on this benchmark model or whatever. Tyler, is that a correct characterization? Yeah, Grok did more revenue last month than Claude in the iPhone app store. I've started having conversations with, I mean, I'm using ChatGPT, but I wanted to just, I wanted to get up to speed on Taiwan and just the, like, what was the reason for the original civil war and stuff.

19:07John Coogan:And so I was just having a conversation back and forth. At no point was I like, oh, it really needs to be like, you know, GPT 5.4 Pro. It's like, these are things that exist just like with one search to Wikipedia or one search to any, and it's probably baked into the weights of 3.5. So like if I'm just going to be like chatting with someone who's like reasonably smart, like I would say Grok is there. And so what do you think? Yeah, but like you could be talking to someone who's really, really. No, like not if you're asking like basic, basic knowledge retrieval questions that like, they're like any model is going to one shot and just be.

19:43John Coogan:Yeah, but you're just describing stuff that you could just like actually Google. But I can't Google via voice in my car on the drive. And for someone who's driving a Tesla and has a Grok integration right there, they're just like, sure, this is great. Okay, yeah, that's fair. But I don't think those people have actually tried like GPT 5.4 Pro. It's so good. It is good, but it's slow. And truthfully, you can fire off the exact same query to 5.4 Pro and 5.4 fast. Fast. And if the query is simple enough, the answer will be exactly the same. Because if I ask 5.4 extended thinking, like, what is the capital of California?

20:29John Coogan:And it thinks for 10 minutes, and it just tells me, say, Sacramento. See? There you go. That's why you need to think. A lot of people, a lot of smart people might say. I told you, I run my life on GPT-2. I hallucinate a lot. People have said I have the mind of GPT-2. It's true. It's true. Anyway, let's continue. Apple's revenue from generative AI apps rose from about$35 million in January to a high of$100 million in August. Do nothing win. Do nothing win. Sales have fallen from their peak, partly because Chachapiti downloads have declined, according to the data. as a proportion of Apple's total sales,$1 billion is small, yet Gen AI apps are the growth driver for Apple's services business, which investors have focused on in recent years because it has grown faster than device sales and boasts higher profit margins.

21:23John Coogan:Apple's dominant share at the top of the smartphone market affords it another luxury, time to get its own AI strategy right. So they're making money while they figure everything else out. Apple's AI plan runs counter to strategies of competitors that are spending hundreds of billions of dollars on chips and data centers to build frontier language models. Apple is spending a fraction of that, aiming instead to use all of the personal information people store on their iPhones together with the chips that it designs itself to power an on-device AI strategy. If they can act as a toll road for providers of AI, then they'll probably end up looking good long term for not having the big capex overhang.

22:00John Coogan:I have to imagine that Apple is not capturing any revenue from enterprises, developers, cloud code, codex, any of those developers. They're probably not, even if they are winding up using like a ChatGPT subscription in codex, they're probably setting that new subscription up on desktop. It's not a toll road on the actual side. Yeah, but it's a toll road on consumer, which is consumer sales. All the more reason to get to the ads, honestly, because Apple does not tax those. And AI is exciting for Apple because they need a new product that they can just randomly bill you like$2.99 anytime they need some additional cash.

22:42What are you talking about?

22:42John Coogan:$2.99? Like, don't you get just random bills from Apple? Like, here and there. $2.99? Like$2.99? Yeah, like, I feel like every time I check my email, it's like, Apple has charged you some random amount for some subscription. In other news, Rolls-Royce has scrap plans to go all electric by 2030 as, quote, drivers prefer V12 engines. Would you look at that? I mean, and this is just a total shock. Total shock. Yeah? Total shock. Yeah. Drivers totally had to experience, you know, being forced, EVs forced upon them for the last few years to know that they preferred combustion engines after all. Of course, I'm kidding.

23:28John Coogan:Elon has been saying the Roadster reveal will blow your mind. If it has a V12, people are going to be going crazy. If he drops a V12, that would completely break the internet. It would be incredible. Let's talk about this Tesla that you were following yesterday. Oh, yes. Did you drop this in the chat already? I sent it to you. We shouldn't share the actual picture. I saw a Tesla that was a very funny mix of it. It had the anti-Elon club on it, but also an 1199 license plate. And it was a plaid, and it just mixed every possible political ideology. And it had a vanity plate that was very sci-fi.