[00:00] Metamuse and agents are taking over, so does this mean that CPUs are realistically going to take over? And how many of them do we really need? And this is where there's some estimates that we're going to refer to
[00:14] and a little bit of a discussion that we should have around this because recently we have seen an explosion in stocks related anything to the CPU trade.
[00:26] for example we have seen this explosion in AMD in fairness we did call for a breakout around the mid-fours over here it happened but it has gone kooky dooky we went all the way up to 645 on AMD
[00:40] it is overbought on the daily it is overbought on the weekly and like hey I'm um I'm you know what what's the what's the phrase when you're not the kind of person that's like you know some people get pissed when somebody else wins and then you're not part of it because they think life is like a
[00:55] zero-sum game, right? What's the opposite of that? Whatever the opposite of that is, is what I am. It's like, I think everybody can win, and I got no shorts, and I'm not bringing this up to be biased.
[01:07] I just think some of this got a little ahead of itself, because I don't think we need that many CPUs. And I'm going to show you why. So first of all, look at the jump here in some of the CPU plays. So specifically, Intel is up 3x this year,
[01:23] 214% year-to-date. AMD is up 183%, which is also almost a 3x this year. Which is, I mean, very impressive. It's phenomenal.
[01:35] But some of this might be a little bit overblown on the latest CPU enthusiasm regarding Metamuse. Why? Because there are some folks chatting with Metamuse,
[01:47] and they're like, hey, so what do you use? Like, what's your hardware? What do you use? And what people are saying is that These CPUs or the CPUs that are running for like a Metamuse are basically two virtual CPUs.
[01:59] They're really part of a giant basket of CPUs and RAM and SSD. But then everybody kind of gets their slice of it. And so specifically, your Metamuse or your agent might use two virtual CPUs in a virtual machine.
[02:14] So it's basically a virtual CPU, a V CPU. And then 8 gigabytes of RAM and then 100 gigabytes of SSD. So a lot of people are extrapolating this and going, oh my gosh, I could get so many CPUs that are needed.
[02:28] Or the world is going to need so many CPUs to run an agent for everybody. But let's do some math here. And Vikram, who posts on Substack, I think he had a good piece over here of Vikram Seeker,
[02:40] He ended up writing that an AMD Venice chip itself, which is a 256-core CPU, could potentially support about 896 agents, agentic virtual machines.
[02:56] So think about that. One AMD chip could support 896, almost 900 of these agent instances. Assuming that CPU time is shared, given that agents do sit idle a lot,
[03:09] you kind of share them in between, not assuming that everybody's running at the same time. So we are making the assumption that there's a shared CPU miss here, which totally makes sense. Like, I don't use Metamuse all the freaking time. Like somebody else may as well be using my shared CPU while I not using it right Duh That it would be ridiculous to assume otherwise Now if you guess that if you take current estimates that let say by the end of 2026
[03:36] there are going to be 10 million active agents. And I'm not saying like you spun it up and then you forgot about it and you never went back to it. 10 million active agents by the end of 2026. and you assumed that we were going to grow not 10x, but 100x
[03:52] to get to 1 billion agents by the end of 2030. 1 billion agents. So basically one per every seventh person in the world. Or maybe that's like 10 for 100 million people, right?
[04:05] You really need more than 10 agents? I don't know, probably not. I'm going to get all the contracts. I've got 50 agents. Yeah, and you probably don't check all but like one of them, right? so there's some efficiency to work out there
[04:17] but you know probably the odds are people who are watching a video topic about CPUs and agents the odds are I'm talking to a group of people who are more likely to use CPUs and agents
[04:29] than the average so we have to keep that in mind you know it's kind of funny too people believe you use but Kevin you pitch Cassie
[04:42] You pitch a prediction market. You're trying to get kids to use prediction markets. I'm like, bro, kids don't watch my channel, okay?
[04:54] I look at my analytics, and it's like, how many 13 to 18-year-olds watch? Zero. How many, like, 18 to 24-year-olds watch most videos? Zero. It's like, older people watch my stuff.
[05:08] So if you want to sign up for county, you can do that. but like I think people use prediction markets for intelligence that's what I like to do I like to look at them and go okay this is this is where sentiment is right now okay this is what people's
[05:21] heads are right now okay interesting and then that to me informs wow people are really believing that we're going to get five great hikes so people really believe that we're going to get you know uh no deal with it one I'm like oh man I'll take the opposite side of that bed and then
[05:35] you know, I'll go buy the dip in the stock market. So I actually think it's great, but I think it's funny when people are like, but Kevin! So we kind of have to know who's watching. So the reason I bring that up is, you know, when you think about this,
[05:49] don't think about how many agents you use. Think about how many agents on average. Like, your mom. How many agents is she going to use? You know, how many agents is your father-in-law going to use, right? That's kind of where your wife, I don't know if you're watching,
[06:03] maybe your husband. I don't know. Anyway, so with this math, you would have 1 billion agents by 2030. On a CPU basis, you would then therefore need about 1.2 million CPUs to service this,
[06:20] given that we could put 900 agents on these virtual machines. That would basically generate a revenue of about half a billion dollars in 2027, or $5.5 billion between 27 and 2030.
[06:36] Let me just be ridiculously clear about that. That is nothing. That is literally nothing. It is a complete fracking rounding error Okay Like NVIDIA does for data centers like to billion per quarter
[06:59] So I want you to think about that, where the money is going, okay? So $100 billion per quarter is basically $1 billion per day. I don't even need to check that out, okay? In five and a half days,
[07:11] NVIDIA does what agents, what 1 billion agents would need between 2027 and 2030, which is three to four years, depending on if you're counting the last year or not.
[07:23] Which means agents needing CPUs is not a good trade. Anthropic and OpenAI continue to survive because they go public, and they keep getting funding,
[07:35] and they keep sucking money, and Elon Musk keeps sucking money, and they keep throwing it at NVIDIA and AMD for frontier chips. That is a good trade. because, in fairness, you know, they have their mission,
[07:48] but it is a little loony how much money they are spending on Frontier Chips and Nginia and A&D. Jensen and Lisa Sue are salivating over the fact that they don't even have to manufacture this crap. They just design it.
[08:00] And then Elon comes in hook, line, and sinker and goes, You guys are the best. Vive, go race, all the money in the world, and give it to Jensen. I also don't know why Elon all of a sudden sounds like that,
[08:12] but whatever, we'll go with it. So, the point is, agentic AI isn't a CPU boom that it seems it is. But it's not just this, because maybe you assume, oh, well, it's going to be double that, it's going to be triple that.
[08:25] Fine. Then it's going to take 10 days or 15 days to match NVIDIA's sales. The frontier is where the money is in this trade. But the next thing that you have to realize is there are products like this.
[08:38] You can just, and you can even get 10% off, apparently. I don't even know. This sounds like it's a sponsorship. if it's not, okay? I have no... Honestly, they probably couldn't even afford a sponsorship. So, these mini PCs,
[08:50] which is like a... You know, it's driven by an AMD R7, 16 gigabytes of RAM, 512 gigabytes of SSD. Can you buy this for $417? You could buy it,
[09:02] and you could do what a 10-year-old does, and it's play Minecraft on it. You could go clip the $40 coupon, and for $380, you can have yourself an agentic CPU.
[09:16] That's all you need. In fact, you could even be more extreme and say, you don't even need that. So I was having this discussion, I guess you would call it, that seeking anything but alpha,
[09:29] which is funny because, you know, I have the alpha report. I guess he doesn't like the alpha report. Although I think he follows. And I'm assuming it's a he. I don't know. It's the picture of the sky. Maybe they're a pilot. I don't know what they are. They follow.
[09:41] They do follow. All right. And then their ad is Ivan. So maybe this is Ivan. Let's assume it's a heap. Let's assume the person's name is Ivan. So Ivan says you can just use low-end cell phones. And I didn't even think about it. But when I read it, I'm like, yeah, that's actually a really fair point.
[09:55] So we have racks of these little CPUs that we use because they're dirt cheap and we get a lot of value out of them. But anyway, for consumers, not businesses, frankly, a phone or an iPad is probably all you need.
[10:07] The memory bandwidth on this little mini PC works out to 45 gigabytes per second. The iPhone 18 Pro Max does 150 gigabytes per second So literally the iPhone does 3x what this little thing does and the little thing is already good enough That the point Like this
[10:25] is already good enough, and the latest iPhone does 3x good enough. Gigabytes of RAM, you have 16 over here, on the iPhone you have 12, and there's some limitations based on the OS, obviously. You're not going to dedicate all of your available memory to, you know,
[10:40] like an agent on your phone, so hence, maybe that matters less. But, you know, whereas over here, I can install Linux and dedicate the whole damn thing to an agent. You're not going to do that with a phone, because it still has to get a phone call and a text message, right?
[10:54] It still has background tasks. But anyway, if you just look at it from an LLM speed point of view, these little main PCs can run you eight tokens per second, whereas the iPhone Pro Max, if you had it fully dedicated, which again, you won't,
[11:07] could probably run about 30 tokens per second. You know, that's actually really wild when you think about it that way. You might not need for agentic AI that much power.
[11:22] Like, Ivan goes on to say that you could essentially have a $300 phone, and if you annualized it to a three-year time span, you're literally talking about less than a penny per day to run these agents.
[11:34] and make new things and batch them and serve more agents than he needs. So the CPU trade, in my opinion, is overblown. Now, Vick goes on to argue that memory is the play.
[11:49] That's really a topic for a different video. Suggest that. If that CPU is not the limiting factor, it's memory that's the limiting factor. And then we can get into, you know, 3D, DRAM, high bandwidth memory.
[12:01] Really a topic for a different video. But this idea here that we need so many more frontier CPUs because people want to spin up an agent to save money on, you know, some subscription that they signed up for.
[12:17] Like, to go negotiate their Verizon bill. And now you probably have, like, agents competing with each other. Like, oh, no, we're at our lowest. Well, then I'm going to cancel. You probably have that kind of stuff going on. It's honestly pretty wild.
[12:30] But it doesn't need that much compute. The frontier is where the compute is. That's where the profit is. That's why the canaries of this frickin' coal mine are Anthropic and OpenAI. And then, of course, the labor market.
[12:42] So it's really those two are the frontier, in my opinion. And then you get the labor market. One of those rolls over, like, wax on, wax off. One of those rolls over was true. But outside of that, I don't know, man.
[12:54] The CC new thing, just my initial reaction, I'm not trying to be bearish on CPUs. I mean, NVIDIA does CPUs too. They're all Vera Rubin. Probably not as good as the AMD one.
[13:06] And I think AMD's a great company. As I said, it might be a little overblown. Now, everything might go up with an Iran deal that comes. Or if we don't get an Iran deal, everything might go down.
[13:19] So in the short term, who knows? But longer term, we're probably a little overblown on CPUs. If you like that video, check this one out. I think you're going to love it. Why not advertise these things each month here? I can't ask. Nobody else knows about this.
[13:31] We'll try a little advertising and see how it goes. Congratulations, man. You have done so much. People love you. People look up to you. Kevin, pass us this. My name's Lamas. A new fever. Meet Kevin. Always great to get your take.