Transcript of THESE Stocks might Get SCREWED
Meet Kevin
0:00Meta Muse and agents are taking over.0:01So, does this mean that CPUs are0:04realistically going to take over? And0:08how many of them do we really need? And0:10this is where there's some estimates0:12that we're going to refer to and a0:14little bit of a discussion that we0:15should have around this because recently0:18we have seen an explosion0:21in stocks related anything to the CPU0:25trade. For example, we have seen this0:27explosion in AMD. In fairness, we did0:31call for a breakout around the mid-4s0:33over here. It happened, but it has gone0:37kooky-dooky. We went all the way up to0:38645 on AMD. It is overbought on the0:41daily. It is overbought on the weekly.0:43And like, hey0:45I'm I'm um I'm you know, what what0:47what's the what's the phrase when you're0:48not the kind of person that's like You0:50know, some people get pissed when0:51somebody else wins and then you're not0:53part of it cuz they think life is like a0:54zero-sum game, right?0:56Uh what's the opposite of that? Well,0:57whatever the opposite of that is is what0:59I am. It's like I think everybody can1:01win and there's no like I got no shorts,1:03you know what I mean? I'm not bringing1:04this up to be biased. Uh I just think1:08some of this got a little ahead of1:09itself because I don't think we need1:11that many CPUs. And I'm going to show1:13you why. So, first of all, look at the1:15jump here in some of the CPU plays. So,1:18specifically Intel is up 3X this year,1:23214%1:25year-to-date. AMD is up 183%, which is1:28also almost a 3X this year. Which is I1:32mean, very impressive. It's phenomenal.1:34Uh but some of this might be a little1:36bit overblown on the latest1:39uh1:40CPU enthusiasm regarding Meta Muse. Why?1:44Because there are some folks chatting1:46with Meta Muse and they're like, "Hey,1:47so what do you use? Like what's your1:49hardware? What do you use?" And what1:51people are saying is that these CPUs or1:54the CPUs that are running for like a1:55Meta Muse are basically two virtual1:58CPUs. They're really part of a giant2:00basket of CPUs and RAM and SSD, but then2:03everybody kind of gets their slice of2:05it. And so specifically, your Meta Muse2:08or your agent might use two virtual CPUs2:11in a in a virtual machine.2:13Uh so it's basically a virtual CPU, a V2:15CPU. Uh and then 8 GB of RAM, and then2:18100 GB of SSD. So a lot of people are2:21extrapolating this going, "Oh my gosh, I2:23could get so many CPUs2:27that are needed or or the world is going2:29to need so many CPUs to run an agent for2:31everybody."2:32But let's do some math here. And uh2:34Vikram,2:36who posts on Substack, I think he had a2:37a good piece over here, Vikram uh Sekar.2:40He ended up writing that an AMD Venice2:43chip itself, which is a 256 core CPU,2:48could potentially support about 8962:52agents2:53uh agentic virtual machines. So think2:56about that. One AMD chip could support2:59896, almost 900 of these agent3:02instances. Assuming that CPU time is3:06shared, given that agents do sit idle a3:08lot and you kind of share them in3:09between, not assuming that everybody's3:11running at the same time. So we are3:13making the assumption that there's a3:14shared CPU-ness here, which totally3:16makes sense. Like I don't use Meta Muse3:19all the freaking time. Like somebody3:20else may as well be using my shared CPU3:23while I'm not using it, right? Duh.3:25That it would be ridiculous to assume3:28otherwise.3:30Now,3:31if you guess that if you take current3:33estimates that let's say by the end of3:352026 there were going to be 10 million3:38active agents, and I'm not saying like3:40you spun it up and then you forgot about3:41it and you never went back to it.3:4310 million active agents by the end of3:452026.3:46And you assume that we were going to3:48grow not 10x but 100x to get to 13:53billion agents by the end of 2030. 13:57billion agents, that's basically one per3:58every seventh person in the world, or4:00maybe that's like 10 for 100 million4:03people, right? You really need more than4:0510 agents? I don't know, probably not.4:07May- Yeah, I'm going to get all the4:08comments like, "I have 50 agents." Yeah,4:10and you probably don't check all but4:12like one of them, right?4:13>> [laughter]4:13>> So, there's there's some efficiency to4:15work out there, but you know, probably4:17the odds are people who are watching you4:20know, a video topic about CPUs and4:22agents, the odds are I'm talking to a4:24group of people who are more likely to4:26use CPUs uh and agents than than the4:29average. So, we have to keep that in4:31mind.4:32You know, it's kind of funny, too, with4:33like people4:35people leave me these comments like,4:36"But Kevin,4:38you pitched Kalshi.4:42You pitched a prediction market. You're4:46trying to get kids to use prediction4:49markets." I'm like, "Bro,4:51kids don't watch my channel, okay? I4:54look at my analytics and it's like, how4:56many 13- to 18-year-olds watch? Zero.4:59How many like 18- to 24-year-olds watch?5:02Most videos? Zero.5:05>> [laughter]5:05>> Okay? It's like older people watch my5:07stuff. So, if you want to sign up for5:08Kalshi, you can do that, all right? It's5:11sponsored. But like, I think people use5:13prediction markets for intelligence.5:15That's what I like to do. I like to look5:16at them and go, "Okay, this is this is5:18where sentiment is right now. Okay, this5:20is where people's heads are right now.5:21Okay, interesting." And then that to me5:24informs, "Wow, people are really5:26believing that we're going to get five5:27rate hikes." Or people really believe5:29that we're going to get, you know, uh no5:31deal with Iran. And I'm like, "Oh man,5:33well, I'll take the opposite side of5:34that bet." And then, you know, I'll go5:36buy the dip.5:37The stock market. Uh so, I I actually5:40think they're great, but I I think it's5:41funny when people are like, "But Kevin."5:43So, we kind of have to know who's5:44watching. And so, the reason I bring5:45that up is5:47you know, when you think about this,5:48don't think about how many agents you5:50use. Think about how many agents on5:52average. Like your mom, how many agents5:55is she going to use? You know, how many5:57agents is your father-in-law going to5:59use, right? That's kind of where your6:01wife, I don't know if you're watching,6:02maybe maybe your husband. I I don't6:04know.6:05Anyway.6:06So with this math, you would have 16:09billion agents by 2030.6:13On a CPU basis, you would then therefore6:16need about 1.2 million CPUs to service6:19this given that we can put 900 agents on6:22these virtual machines.6:24That would basically generate revenue of6:27about half a billion dollars in 20276:30or 5.5 billion dollars between 2027 and6:342030.6:36Let me just be6:37ridiculously6:39clear about that.6:41That is nothing.6:43That is literally nothing. It is a6:45complete freaking rounding error. Okay?6:50Like6:51Nvidia does for data centers like 90 to6:54100 billion dollars per quarter. So I6:58want you to think about that, where the7:00money is going, okay? So 100 billion7:03dollars per quarter is basically a7:04billion dollars per day. I don't even7:07need a calculator, okay? In five and a7:10half days, Nvidia does what agents what7:141 billion agents would need between 20277:17and 2030, which is7:19three to four years depending on if7:20you're counting the last year or not.7:22Which means agents7:25needing CPUs is not a good trade.7:29Anthropic and OpenAI continuing to7:31survive because they go public and they7:33keep getting funding and they keep7:35sucking money in and Elon Musk keeps7:37sucking money in and they keep throwing7:38it at Nvidia and AMD for frontier chips,7:42that is a good trade because7:46in fairness, you know, they have their7:47missions, but it is a little loony how7:49much money they are spending on Frontier7:51chips and Nvidia and AMD. Jensen and7:54Lisa Su are salivating over the fact7:56that they don't even have to manufacture7:58this crap. They just design it and then8:00Elon comes in hook, line, and sinker and8:02goes8:08it to Jensen."8:10I also don't know why Elon all of a8:11sudden sounds like that, but whatever,8:12we're going with it. So, so the point is8:15a agentic AI isn't a CPU boom that it8:19seems it is, but it's not just this8:21because maybe you assume oh, well, it's8:22going to be double that, it's going to8:23be triple that. Fine. Then it's going to8:25take 10 days or 15 days to match8:29Nvidia's8:30sales. So, the Frontier is where the8:32money is in this trade. But, the next8:34thing that you have to realize is8:35there are products like this.8:37You could just8:39I I I you can even get 10% off8:40apparently. I don't even know. This8:41sounds like it's a sponsorship. It's8:43not, okay? I have no I honestly, they8:45probably couldn't even afford a8:46sponsorship. So, these mini PCs, which8:50is like uh you know, it's driven by an8:52AMD R7, 16 GB of RAM, 512 GB of SSD.8:58Okay, you buy this for $417.9:01You can buy it and you can do what a9:0310-year-old does and it's play Minecraft9:06on it. You could go clip the $40 coupon9:09and for $380, you could have yourself an9:13agentic CPU.9:15That's all you need. In fact, you could9:17even be more extreme and say you don't9:20even need that. So, I was having this9:22discussion, I guess you would call it.9:24Uh that uh9:26seeking anything but Alpha, which is9:29which is funny because, you know, I have9:31the Alpha report. I guess I guess he9:32doesn't like the Alpha report. Although,9:33I think he follows. And I'm assuming9:35it's a he. I don't know. It's just a9:36picture of the sky. Maybe they're a9:38pilot. I don't know what they are. They9:40follow. They do follow. All right. And9:42then their ad is Ivan. So maybe this is9:43Ivan. Let's assume it's a he. Let's9:45assume the person's name is Ivan. So9:48Ivan says you could just use low-end9:49cell phones. And And I didn't even think9:51about it, but I when I read it, I'm9:53like, yeah, that's actually a really9:54fair point. So uh we have racks of these9:57little CPUs that we use because they're9:59dirt cheap and we get a lot of value out10:01of them. But anyway, for consumers, not10:03businesses, frankly a phone or an iPad10:05is probably all you need. The memory10:07bandwidth on this little mini PC works10:09out to 45 GB per second. The iPhone 1810:13Pro Max does 150 GB per second. So10:17literally the iPhone does 3x what this10:19little thing does, and the little thing10:21is already good enough. That's the10:23point. Like this10:25is already good enough. And the latest10:28iPhone does 3x good enough.10:30Uh GB of RAM, you got 16 over here. On10:32the iPhone you got 12. And there's some10:34limitations based on the OS, obviously.10:36You're not going to dedicate all of your10:37available memory to10:39uh you know, like an agent on your10:41phone. So hence maybe that matters less.10:44Uh but uh you know, whereas over here I10:46can install Linux and dedicate the whole10:48damn thing to an agent. You're not going10:50to do that with a phone cuz it still has10:51to get a phone call and a text message,10:53right? It still has background tasks.10:56But anyway, if you just look at it from10:58an LLM speed point of view, these little11:01mini PCs can run you eight tokens per11:03second, whereas the iPhone Pro Max, if11:04you had it fully dedicated, which again11:06you won't, uh it could probably run11:07about 30 tokens per second. You know,11:09that's11:10that's actually really wild when you11:13think about it that way.11:15You might not need for agentic AI that11:19much power. Like Ivan goes on to say11:23that you could essentially have a $30011:26phone. And if you annualize it to a11:283-year time span, you're literally11:29talking about less than a penny per day11:33to run these agents.11:36And maybe you could even batch them and11:37and serve more agents than you need. So,11:39the CPU trade, in my opinion,11:42is overblown. Now, Vic goes on to argue11:46that memory is the play.11:49That's really a topic for a different11:50video. Suggest that it's not CPUs that11:53are the limiting factor, it's memory11:54that's the limiting factor. And then we11:56could get into, you know, 3D DRAM, high11:59bandwidth memory. Really a topic for a12:01different video.12:02But this idea here that12:05we need so many more frontier CPUs12:08because people want to spin up an agent12:10to save money on, uh, you know,12:14some subscription that they signed up12:16for, like to go negotiate their Verizon12:18bill.12:19>> [laughter]12:19>> And now you probably have like agents12:21competing with each other, like, "Oh,12:23no, we're at our lowest." Well, then I'm12:24going to cancel. You probably have that12:26kind of stuff going on. It's honestly12:28pretty wild.12:29But it doesn't need that much compute.12:33The frontier is where the computer is.12:35That's where the profit is. That's why12:37the canaries of this freaking coal mine12:39are Anthropic and OpenAI. And then of12:41course the labor market. So, it's really12:42those two were the frontier, in my12:44opinion. And then you got the labor12:45market. One of those rolls over,12:48it's like wax on, wax off. One of those12:50rolls over, we're screwed.12:51But outside of that, I don't know, man.12:54This the the CPU thing, just my initial12:56reaction, and I'm I'm not trying to be12:58bearish on CPUs. I mean,13:00>> video does the CPUs, too. Like, they're13:01all Vera Rubins. Probably not as good as13:04as the AMD ones, and I think AMD's a13:06great company. Uh,13:08that said,13:09it might be a little overblown. Now,13:11everything might go up13:13>> [music]13:14>> with an Iran deal that comes. Or if we13:16don't get an Iran deal, everything might13:18go down. So, in the short term, who13:19knows? But longer term,13:22we're probably a little overblown on13:23CPUs. If you like that video, check this13:25one out. I think you're going to love13:27it.13:27>> Why not advertise these things that you13:28told us here? I feel like nobody else13:30knows about this.13:31>> We'll we'll try a little advertising and13:32see how it goes.13:33>> Congratulations, man. You have done so13:34much. People love you. People look up to13:36you. [music]13:37>> Kevin Paffrath here, financial analyst13:39and YouTuber Meet Kevin. Always great to13:41get your take.
2,516 words · 381 lines
