Chip Stocks and the A.I. Explosion

Data is everywhere. When orchestrated properly, it sings. At Morningstar, we analyze and enrich data, making it actionable and powerful for you. Morningstar, where data speaks.

But the hardware that NVIDIA is providing is getting so much better that the algorithms will improve exponentially, which means you have a use case explosion. And with that use case explosion, you need way more chips.

Hello, you, and welcome to— it's so dumb. I think we're going to try a different one. I thought that was going to sound better than it did. Hey, you, it's been a while. Hello and welcome to the Barron Streetwise Podcast. I'm Jack Howe, and the voice you just heard is Daniel Pilling. He's a semiconductor analyst at Sands Capital. That's a Washington, D.C.-based investment company managing more than $50 billion with an emphasis on innovation and technology. In a moment, Daniel will explain how artificial intelligence will shape future demand for chips and which stocks will shine. Listening in is our audio producer, Jackson. Hi, Jackson. Hello, you. Welcome back from your bike trip in Japan. Good time? Great time. Great people, yummy food. One flat tire. Nice. Do you want to talk about some Paramount drama before we get into our chip chat? Believe it or not, I watched 8 episodes of Yellowstone, the whole first season, on the plane back home. You're definitely getting a cowboy hat. That's how it starts. The real Paramount drama was on Wall Street this past week. I mean, first of all, there was an earnings call on Tuesday and the stock went down 7%. That on its own is not really that big of a deal for Paramount. It's something of an earnings day disappointer lately. That, that 7% drop is only the fourth worst earnings day stock decline in the past 4 years. In this case, I don't think it was the financial results. They were okay. CEO Bob Bakish— former CEO, I should say— is out. Replacing him is what you might call an executive triumvirate, but I call it a trium-invertebrate because it's three managers who appear to be chosen for their unlikeliness to stand up to Shari Redstone. She's the one who controls the votes. If you're not familiar with the Paramount story, I describe it in a recent column like a three-season streaming drama. Season 1 introduced Sumner Redstone. He's the son of a Boston linoleum peddler. He turned a family investment in a drive-in movie theater into a media empire that at its peak was valued at more than $80 billion. He was known as a tough negotiator and a frequent litigator. He's the guy who brought Viacom and CBS together for the first time. He meddled into the business well into his 80s. He spent his 90s incapacitated with girlfriends fighting over his fortune. Season 2 started with CBS CEO Leslie Moonves being ousted in 2018 after more than a dozen women accused him of sexual misconduct. Sumner died 2 years later. His daughter Shari, who had won control of the business by that point, brought Viacom and CBS together for the second time. As Paramount Global, led by Bob Bakish. Then you have the current season. The TV business is struggling almost all over. For most companies, it's a balance between managing these declining but still profitable legacy assets and this growing streaming side that isn't yet profitable, and it's unclear how profitable it will become. That's certainly the situation for Paramount, only maybe a little bit worse. Morgan Stanley wrote this past week that the problem for Paramount is that it, quote, approaches this optimization challenge with assets that are both highly exposed to the declining linear TV profit pool and arguably subscale in streaming. Okay, so in the current and surely final season, season 3, Paramount begins with a junk rating by S&P, a lot of debt, and there's a $26 billion buyout offer from Apollo Global, and it's pretty straightforward, it's all cash. And Paramount rebuffs that offer and instead enters exclusive talks with a relative Hollywood newcomer called Skydance. That company is backed by private equity and by the profoundly wealthy Larry Ellison, founder of Oracle, and Skydance is run by Larry's son David.— and its proposal for Paramount is anything but simple. But it's quite lucrative for Shari Redstone. She controls these special super-voting shares. Skydance will buy her out at a handsome premium. Afterward, Paramount would buy Skydance for $4 to $5 billion. And that's it. Shari gets a lucrative buyout. Ordinary shareholders get shares in the new company. Mind you, Paramount stock was recently down 73% over the past 5 years, so unclear how this new business combination would turn things around. Shareholders weren't loving it. Bakish was trying to find another way, which is one of the reasons he's out. A few directors have also been replaced. When you control the votes, you can do what you want. We should point out that Paramount has appointed a committee to review the terms of any deal and make sure that it's fair to shareholders. 3 last things to know. Skydance recently came up with a sweetener for its deal. It said that one of its backers would infuse $3 billion into Paramount to help it pay down debt. Okay. Also, Apollo is back with its original $26 billion offer, and this time it's partnering with Sony. That's important because one of the reasons Paramount gave for pushing aside the Apollo deal was it wasn't sure how Apollo would come up with the cash. Having a big partner makes that more clear. The last thing is that Paramount has a deal with Charter Communications, a standard cable deal to distribute its networks over Charter's footprint, and Paramount gets money for that. Well, that deal just expired. And standard practice is that the network owner asks for more and the cable company says, we can't do it. And the network owner goes to the public and says, hey, the greedy cable company is trying to take your favorite shows away. And in the end, the cable company buckles and comes up with more money, and life moves on. But that's not what happened last fall when there was a similar standoff between Disney and Charter. Disney said, we want more money. Charter said, we will literally walk away from the cable television business entirely because it's terrible, subscribers are leaving, we'd rather just do broadband and leave television alone. And Disney, I would imagine, thought, hey, wait a second, We really need the cash flows from this legacy television business to fund our streaming service while we're building that service out. So we better come to a deal here. And Disney basically caved. It agreed to share distribution of its streaming services with Charter to cut Charter in on the business. Now, that deal that Disney gave in to, that's basically a ceiling, I'm guessing, for how good Paramount's deal is going to be in its Charter negotiations. It's going to have to figure out something soon. And if the deal ends up being particularly bad for Paramount, that could complicate its talks over selling the company. Anyhow, that's it. I don't know how the drama will end. I think it's too late for a happy ending for shareholders. They've taken a bath, but it's not too late for a fair deal for shareholders. We'll see. Jackson, do we need to say anything else about television, or you think I have it covered? I'm definitely not going to get in, just so you know, to the NBA and negotiations over those rights. Those are 10-year deals and one's just coming due, and Warner Brothers and Disney are the two big rights holders. But there are new bidders on the scene, Amazon and NBC, and the rumor is that the price is going to double or more, and the NBA is going to make great money, and Warner Brothers might take a beating because its TNT network is all in on basketball, and it's paying around $1.2 billion a year on average for something that NBC is willing to pay $2.5 billion a year. So best-case scenario, Warner pays a lot more for something it already has. Worst-case scenario, it loses its deal for NBA games. It's got that show, you know, with, um, Shaquille O'Neal and Charles Barkley. You know, it just signed a 10-year deal It's like a 9-figure, 10-year deal for Barkley. He's only 2 years into the deal. So they're really on the hook for a lot of basketball money. And now it's unclear whether they'll keep their games. Anyhow, I'm definitely not going to talk about any of that stuff. I won't mention that Barkley built in an escape clause into his contract. So if TNT doesn't keep its NBA rights, he gets to become a free agent, probably get even more money from some other company. He talked this past week about building that escape clause into his contract ahead of the NBA rights negotiation for this exact reason. He said— and keep your finger on the bleep button there, Jackson, because you got some profanity coming your way— he said, I wanted to cover my ass. I mean, he didn't say anything really that bad. He just said, yeah, we got it. Jackson, should we take a quick break before we get to our conversation about chips? Give folks a chance to hear from a sponsor. Stretch the hammies, the quads, the deltoids, the lats. Name, name, name 2 or 3 muscle groups quickly, quickly. Uh, glutes. No, forget it. You took a lot of them. Triceps. That's the first one I thought. Listeners don't need you talking about their glutes. We'll be back right after this. Data is everywhere, but is it ready for consumption? Morningstar developed the language of global investment data so you have the right ingredients to help you shine. Morningstar, where data speaks.

Harvard Business School Executive Education delivers breakthrough learning for leaders, stimulating classes led by faculty at the forefront of their fields. Topics that will define the future of business, discussions that transform perspectives and ways of thinking, and access to the brightest business minds on the planet. Learn more at hbs.me/breakthrough. That's hbs.me/breakthrough.

Welcome back. Let's hear from Daniel Pilling. He's a semiconductor analyst at Sands Capital. I spoke with him recently about artificial intelligence and self-driving cars and Taiwan and which stocks to favor.

So this is the first time in the semiconductor industry that we're going from humans being the biggest demand driver, i.e., you and I and everyone else buying smartphones, PC, et cetera, et cetera, to actually having this technology which is not bottlenecked by the number of people on planet Earth. So the world talks a lot about AI models and the training piece. And we agree, right? This is super important, great topic. I think the other edge side of that coin is that when these models get bigger and better, that also means we're going to have a lot of edge AI use cases. So we're going to have AI on the smartphone. We're going to have AI within vehicles. We're going to have it within robots, within the PC. And so the better the center becomes, the models within the clouds, the better the edge becomes, whereas the edge hasn't even started growing yet. So we don't even have an AI smartphone yet today from Apple. Sooner or later, we probably will. And what that might mean is that we start buying more smartphones in faster times. Today, I might buy a smartphone every 3 years. Maybe I'll buy it again every 2 years. Same for the PC, same for cars, et cetera, et cetera. So not only do you have this massive avalanche of demand within the center, which, and the center to me is the hyperscalers and AI training and inference within the center. But also you're going to have all this demand on the edge, which hasn't even started yet.

Explain that concept for people who don't know what it means or why it's important. What is edge computing? Why is it important? We used to need to have the computing power close to us, and then the computing power moved to central locations somewhere on the internet. Now we can have dumb devices that access that computing power. And now it's once again becoming important to have computing power close to us.

Why? So I think the crux of it is latency. And maybe I'll give you an example from our own lives, right? So when I buy something on Amazon and the Amazon app doesn't work, I might abandon it within 10 seconds. We all have this thing, sort of everything needs to be fast nowadays. Now, if you look at the smartphone, for example, there will be issues running AI inference on the smartphone within the cloud. Why? We may be in an area of poor connectivity. 5G is not fast enough or not good enough either, et cetera, et cetera. So it's going to be slow. So if you're Apple and you want to grow your servicing business again, you need to make sure that inference works on your smartphone as fast as possible because we're all used to zero latency. Whereas actually on the PC, there's a second reason. If you're Microsoft and you sell us, sell you and me a Copilot, it's much better for the consumer to pay for the inference, right? So if I buy the PC and the inference runs on my computer, it's much better because Microsoft doesn't have to pay for it. So maybe there's a little bit of a sort of business reason as well. So those two. And then three, there's a privacy reason as well. So maybe one day we have really models that can memorize because today these models cannot memorize anything. If you go on ChatGPT, it doesn't really know who you are when you come back. In the future, we will probably want to have our own models that know us. And then maybe that model just resides on your phone for privacy reasons. So you don't want to share that on the cloud. So maybe that's another reason.

What did you mean a few minutes ago when you said that demand is no longer bottlenecked by the number of people on Earth? Is it that we're not directly buying the devices that have the chips one for one? Or what do you mean by that?

Yeah, so in the past, right, when you look at smartphones, for example, you can say, okay, there's X number of people in the world and they're going to require a certain number of smartphones. They're going to buy them every 2 or 3 years. And that's sort of the market size. And that's it, right? So that's sort of the max you can get. For AI, it doesn't really matter how many people you have in the world. For AI, what matters more is like, how many computing problems do I have that need to be solved? So ultimately what you're going to have here is you probably will have machines talking to machines driving AI use cases. You'll have all the enterprise companies in the world. You have the hyperscalers and various venture finance startups. And that means that to me, that's incredibly interesting because the incremental driver here is not a human that has a certain budget, but it's actually the gross margin line of a Fortune 2000 company. And there's a lot of space to grow there, right? Obviously, the average person, including all of us, are kind of budget constrained in terms of how much we can spend on a smartphone.

When you start talking about machines talking to machines and figuring out new use cases for AI, it just gets me wondering, what will they come up with? Like, how will the world change? What do you come up with 10 years from now for ways that ordinary people might feel the change in their lives from AI?

Maybe I'll give you a couple examples, but I will preface this with one thing. Just put some numbers around it if you like. So as you may know, NVIDIA had a developer's technology conference conference recently, and then Broadcom, another big AI chip company, had one as well. The crux of it was like this. Broadcom was talking about the cluster size of these AI chips growing by 30 times. And then if you say the models, the GPUs improved by 5 to 10 times on top, that means you're going to have something 150 to 300 times more compute for these models, like enormous numbers of incremental compute. So when you throw more compute at these models, they get much better. That's been happening for 4 or 5 years now. And then the conclusion of that is then to say, when the models get better, the use case explodes, which then effectively means the IQ of these models will go up tremendously. And you can draw sort of a curve of where we are today to the future and put new use cases on them, right? We met NASA recently, and NASA is using AI to design their spaceships or certain parts of them. So if you're a Star Trek fan, you see that some spaceships there look very different than ours today. That's what they're doing. So the AI is AI is like 90% better than a human or something like that, right? It's significantly better.

Does anything look like the Starship Enterprise? Were they like— did they get it right early with the shape of that ship, or is that still fiction?

So it looked more like, um, it looked a bit more like insectoids, you know, like more like insects today. So it wasn't as sort of geometric as we humans like to do things, but it was sort of optimized in a very odd-looking fashion.

Space insects. All right, I'm intrigued.

What else? So the second use case I found intriguing for us that might actually impact our own lives is sort of within the medical space. The idea is to say that you, I, every one of us is going to have a sort of AI replica somewhere within the cloud. And then if you're unwell, the doctors will actually use medicines on that replica of you. To see how you react. That's an enormously positive use case. And I mean, I think we're far away from that. But then again, the way that GenAI works, it can take you and create 1,000 scenarios of you and apply different uses of drugs and see which one was the best one for whatever ailment one may have. And maybe the third one I'm going to mention to you. So when you look at what happened at Meta, Meta sort of had issues with Apple and privacy concerns. That was a big problem, right? So the stock price had The company had issues because of that. Lo and behold, they used GenAI to get around those issues. And what they're doing now is using GenAI to improve the ad targeting. And it's with phenomenal success. I think what they may do in the future, and instead of showing you a, let's say, a Nike commercial, they might show you a Nike commercial with you wearing the shoes because they know that maybe you like Nike, maybe you like Air Jordans in particular, and maybe you like that color in particular. So the ad targeting is going to become much, much, much, much better. And then maybe lastly, I mean, I hope for some sort of Copilot type of GenAI thing that knows my preferences really well and might help me in my day-to-day sort of life choices. That would be nice.

Is NVIDIA still, I mean, I guess you remain big believers in NVIDIA, right? Its products are pricey. There are companies that want to compete in that space. What's going to keep NVIDIA safe?

So yes, we're firm believers in the competitive differentiation of NVIDIA. And in our opinion, there's a few angles to that, if I may highlight those. But first and foremost, designing these chips requires very specialized mathematicians and scientists. Our research indicates that most of those work at NVIDIA. And NVIDIA is able to pay them multiples more than their competitors. And the reason why is because they just have much more scale. I think the second interesting piece is that AI chips are becoming more of a system problem. So you may think of kind of like Apple. Apple optimized the software and the hardware. The same thing is happening here. So whereas the problem is moving from designing a single chip to an entire rack with thousands of different parts, and then on top of that, you have to optimize the software that talks to the chips. NVIDIA has done all of that, yeah? And you may say it's kind of like a Swiss watch that you have to update each year, and the parts just have to be perfectly working together. Nobody else has done that even close to. And then 3, I think it's also interesting that NVIDIA has moved from a sort of 2-year, 2.5-year cadence of putting out new chips to a 1-year cadence. I think the others will follow, but they will struggle to follow in that type of speed. So I think NVIDIA for now is running ahead because of people, because the systemization of the problem, which is both a hardware and a software problem. And as mentioned, sorry, I'm blanking on that third point right now. But I mean, those 3 basically.

That's why I never start on a chip thing by saying, "Okay, 3 things," because I always forget the third thing. That's most helpful. Thank you. But Nvidia, the stock valuation isn't crazy, but it is ambitious. That doesn't turn you off at all. You think it'll grow into that valuation and then some.

Yeah. Nvidia, I'm going to quote consensus numbers here today. I think trades on, if we take calendar year '25 and about 27 times earnings. Their biggest customers, the hyperscalers, have spent about 3 times more on data center CapEx in the past 6, 7 years. So if you compare 2018 to this year, it's about 3 times more. And as you've seen in the last few earnings calls, all of them are spending much more again. And maybe, can I give you like a mental model on this, if I may? Sure. So I think this all goes back to the App Store and Apple. So if you look at what happened at the App Store in 2011, you had ringtones, right? Because the hardware wasn't good enough. If you look at it today, you have millions of apps on the App Store. Why? Because the hardware got really good. The same thing is happening here. Today we have ChatGPT-4, et cetera, et cetera. But the hardware that NVIDIA is providing is getting so much better that the algorithms will improve exponentially, which means you have a use case explosion. And with that use case explosion, you need way more chips.

What are some other stocks that you like? Tell me about a couple of your favorites.

Yeah, so within the semiconductor space, I think you could bucket it in— I'll do it again this time, I'll get it right— in 3.

Okay, I'm going to hold you to it. I'm counting.

Number 1, go ahead. So number 1 is sort of the semi-capex companies. So that's the semi-equipment companies. So those are companies that manufacture the equipment that's required to produce the chips. These are companies like ASML, Lam Research. There's a business called Entegra, a smaller chemical and filtration type of company. And I can talk about all of those, but the second aspect then is sort of the guys who buy the semi-capex equipment, which is the foundries of the world. And really, as of today, there's one dominant foundry in the world, which is Taiwan Semi. And then on top of that, you have sort of the chip designers and Nvidia is one of them. Within that space, we only own Nvidia. But then you can say, well, okay, so at the bottom, the first two layers, they will benefit from anything that happens within Moore's Law and anything that happens in the semiconductor space. So if I need more chips, I need more semi-capacitor equipment, and I need more foundries, so they will do really well. And then at the top, I think you want to be a little bit more cautious in terms of picking your areas of interest, right? So for example, I wouldn't necessarily look at smartphone chip designers because smartphones aren't growing that much anymore, whereas AI is obviously a different question that's growing a lot. So you want to pick your sort of chip designer effectively.

And ASML is— that's the Dutch company that does the EUV. I'm not going to try to come up with it. Am I— extreme ultraviolet lithography? Have I got it?

That's right. That's right.

That's the new thing you need to make the new chips for this sort of leading-edge miniaturization of these circuits, and they're the only one who's got them. If ASML has the latest and greatest machines and they're the only one that has them, why do you also like a company like Lam Research?

So ASML, you may think of sort of like, it's called lithography. So you basically take a light and you expose a pattern to light and that helps it then to pattern whatever chip you want on the wafer. The other aspects you need is so-called etching and deposition. So imagine, If you are an artist and you're trying to chisel a figure out of clay, you need a chisel to do that. And LAMM is that. So when you build like transistors, you need to create, you need to etch holes and then you need to deposit chemicals on top of these holes. And they look like 3D architectures, right? If you zoom in, you can maybe think of, so it looks a little bit like Manhattan or something like that. So those are two sides of the coin effectively, right? So you need litho, but you also need etch and deposition. I think within semiconductor chip design, we're moving sort of from a flat-ish city, right? So when you go to London, for example, London is more of a flatter city, right? So there are skyscrapers, but far and few between. And you're moving from London to Manhattan, where it's like tons of skyscrapers everywhere. And so chip design is the same, right? You're moving from from planar to 3D and sort of these enormous skyscrapers. And companies like Lam are required to do that.

And with Taiwan Semi, does the fact that it's a company in Taiwan add complexity? I mean, it must be hard to figure out. You have to not just understand chips, you have to understand politics. I mean, Taiwan is, as China calls it, China. And, you know, so it's in this weird sort of state. Politically, and how do you think about that?

Yeah, we look at that a lot, and maybe I'll give you a mental model. Now, I will all say these are people involved with a lot of power, so the logic may not matter that much, but I think the logic that we have hopefully makes sense. I'll give you a few factoids. First of all, I think there's this idea of if China were to Taiwan, that they would fix their problems within the semiconductor space. That's not true. If China were to control Taiwan tomorrow, then the ASMLs of the world would stop delivering spare parts, which would mean that all of the Taiwanese equipment would stop working within 2 or 3 weeks. So you would not get any chips. The second impact would be probably that if that were to happen, that China would get completely cut off of of leading-edge chips, at least in the parts that haven't been cut off yet, which then means that China would also not be able to develop these big AI models. And then arguably, if that's the most important technology of our time and the most exponential technology of our time, you would lose that access to that technology if you were to do something with Taiwan. So my sense is that at least the logical case would be you wouldn't do that because you're not getting anything out of it. You would literally be cut off of the most important technology, and, uh, you would not be getting any semiconductors in Taiwan at all because you don't have the ASMLs of the world.

I think they call the position of US policymakers towards Taiwan a strategic ambiguity. So it sounds like, sounds like strategic ambiguity might be the best we could do for a while, and maybe, you know, that it might keep working for Taiwan Semi. Let me ask you one last thing. I want to circle back to use cases for AI. Do you do any thinking about self-driving cars? Do you pay attention to that market? Because that's something for ordinary people where they can get their head around that in a big way. And we already see some of the early features in cars that are on the road. There are some people out there say, hey, end of the decade, we'll be there. And there are some people who say not even close. Maybe it might take a decade or more beyond that. Where do you think we're headed in terms of the cars driving themselves?

Yeah, Jack, no, we do a lot of thinking about that. So maybe a few points. Is this going to happen? I cannot imagine it doesn't happen. And I'll tell you why. Because ultimately, all of these things are actually the same problem. It's just a question of how much compute and how much data. That's it. And we know that the compute amounts are going to go up just tremendously. We know that. In terms of semiconductor content, we're talking about thousands of dollars here, potentially. And that's nowhere in the numbers today, nowhere. So if and when that happens, you're going to have a nice little incremental growth opportunity within semis. And then the third thing I think that one might think of is synthetic data. Now, I'm not entirely sure how this will impact, but NVIDIA talks about using synthetic data to train these models. And obviously Tesla has a different opinion here. But then again, if I say, well, so the data availability for this problem is going to shoot up tremendously because we can simulate these things. And we're gonna have massively amounts of rising compute. And again, leads me to believe we're gonna have a solution sooner or later. Now again, all these things are exponential and if and when they happen, we should have a great sort of investment opportunity ahead of us, right? Because when it happens, there will be disbelief and all of a sudden a lot of growth, just like similar to Nvidia, yeah? But on the timing part, I shall graciously pass in terms of we're not sure.

Hey Daniel, I like talking with you and I learned plenty. Thanks for joining us.

Thank you, Jack. Thanks for having me.

Thanks to all of you for listening. If you have a question you'd like answered on the podcast, Jax, go ahead and sing the question solicitation song quickly. 1 and 2. Open your voice memo app and send it to jack.how@barons.com. Unless you have an Android, then there's other recording apps that I assume can also send to recordings to the same email. Okay, save some, save some for the concert tour. Jackson Cantrell is our producer. You can follow the podcast on Apple Podcasts, Spotify, YouTube— I still don't think that's real— or wherever you listen to podcasts. See you next week.

Data is everywhere.

When orchestrated properly, it sings. At Morningstar, we analyze and enrich data, making it actionable and powerful for you.

Morningstar.

Where data speaks.