00:24:56 · Jaimin Rangwalla
[00:24:56] S0: So historically, you think about it, memory was tied to smartphones, memory is tied to computers. Everyone has a smartphone, everyone has a computer, and so you can just say over time, as people started adding more technology devices, individual might have eight gigs of RAM on a computer, you have eight to 16 gigs of RAM on a phone. And so you've seen like a natural growth of that just over time as you've seen penetration of phones, penetration of the computers, and then the density within each device also increases over time. But now if you think about it, if you or me in a year or two, we're going have a claw that's running hundreds of agents at all time. I'm going to have a similar impact at work. I'm going to have hundreds, maybe thousands of agents running at all time. All those agents require CPUs, GPUs, memory. Now it's persistent memory because it's not running twenty four hours a day. Some will run for some amount of time, some are less, you get the kind of a general direction of where this is going, which is your footprint, your like semiconductor or technology footprint as an individual is about to expand by like 10x because you used to have a couple of devices. Maybe you're someone who has an iPad too, you have x, a little more, but like three or four devices is about to go to thousands of virtual devices on top of, and the virtual devices are the agents, and each agent requires, you know, semiconductor content. So it is hard to like fathom a world where like we're all having this like big semiconductor footprint, but that's what's happening underneath the hood, right? And if we're all gonna be users and we're all gonna be using hundreds and thousands of agents at a time, that is where we're going. Like, on our call, we had Boris, Cherny, who's former CodeTwo alum, but also the founder of Claude Code, And he was saying that, you know, he runs, you know, some number of agents during the day, but when he goes home, he's running or on on his way out, he's has a thousands of agents running to kinda do work all through the night before he comes back in the morning. So just imagine, like if we're all going be doing things like that, where we're all going to be having periods where we're running hundreds of agents, and it can be simple tasks.
00:26:55 · Jaimin Rangwalla, Molly O'Shea
It doesn't have to be like, you know, a very complex programming task, but just the footprint of each individual in terms of their semiconductor content, their power content, right? It doesn't just apply to semiconductors, applies to how much power each person consumes, is all going to increase significantly. [00:27:12] S1: Because of those behavioral changes, how is that affecting the architecture of AI?
00:27:17 · Jaimin Rangwalla
[00:27:17] S0: I think that we're now moving to a world where the chatbots, you know, we joke around internally, the chatbots have amnesia, right? Because every time you might give it some files, you give it some medical records, it does a great job of going through and analyzing it. You go back the next day and you give it the same record. You ask a question about those records, it doesn't remember. I don't if you ever watched that movie, fifty First Dates, Bob Sandler, right? Where basically he has to re explain every day what the situation is. It's very similar to that. But what's changing is now you're getting persistent memory with all of the the next evolution of, accelerators. And so I think that you're seeing memory content is increasing. The agent is not gonna have amnesia anymore. The agent is gonna be able to remember. And now because the agents are launching agents and they're doing very complex tasks but also simple tasks, they need offload from CPUs too. Historically, CPUs have done something called serial processing, which is like you just give it a set of instructions and it does one by one by one by one critical instructions, but one in row. And I think that as an agent has some element, which is like, make a reservation, that might just be very serial tasks. It might have something which is like, okay, program something that's super complex and might require parallel tasks. And so I think that that is just creating a massive need for more CPUs and more memory, and we're seeing that. We're actually just at the beginning of that. I think agents is really what is driving that.
00:28:38 · Molly O'Shea
[00:28:38] S1: There was a really interesting slide on that where you're showing the shift of GPUs to CPUs, and it was like a four to one conversion. So how did you kind of model that out, or how are you estimating it?
00:28:49 · Jaimin Rangwalla
[00:28:49] S0: Yeah, so for some amount of the last, the greater part of the last three or four years, it was very much a one CPU to even eight GPUs. That was kind of the, and it had gone up to 16 GPUs, which was all of the math and all of the computation is happening at the GPUs, and then they just kick it out to the CPU when it's time to execute the final step. But now the ratio has been actually moving to one CPU to four GPUs, and so it's improved kind of by 2x already. And we think it actually has a chance to flip the opposite direction, which is one GPU to four CPUs. And some people, very aggressive, say that it could go to one GPU to eight CPUs, right? And so I think that it's just more a function of how many of those tasks are being completed and how many of those new tasks are being done that don't require the parallel compute, but will require more just serial task completion. And I think that, again, it's just another way that I think about it is there's 7,000,000,000 people or so, plus or minus. Everyone has their technology needs. But if every person has, you know, a thousand, right, the 7,000,000,000 multiplies by a thousand. Right? And so just think about how many more people effectively are on the earth than there are today. I think that all of those people are going to require serial processing, parallel processing, memory, like all of the things that every human would have. Think that it's just another way to think about just almost population expansion. If we expanded the population by a thousand, and this is a digital population that we're expanding, just all of the resources that are required for that digital population.
00:30:21 · Molly O'Shea, Jaimin Rangwalla
[00:30:21] S1: Who are the main players within this stack? [00:30:24] S0: On the CPU side, it's Intel, AMD, and ARM. Amazon has done a great job. They actually have one of the most efficient CPUs, but they've always kind of kept it internally. They bought a private company that helped them do this, and it has helped, I think, drive AWS success, especially as we move to CPUs. I think it really helps be a tailwind to their business. But Intel for years has been almost like the forgotten child in semiconductors. They had a lot of missteps along the way, whether it was a technology issue, product issue, They've had a lot of leadership changes, but they have an amazing CEO now. Sometimes we try to keep things simple, which is you have a four to one going to one to four, which is mathematically a 16x improvement in market size. And then you have Elon Musk giving a stamp of approval to Intel and increasing the pace of innovation. And so I think that to me, it's just some of the best thesis are the simple thesis, and you don't want to overthink it. And I think that just the combination of that for a stock that, yes, it's had a big move this year, but if you think about every semiconductor company over the last four years in this heyday of AI, they're all up five, seven, 10x. It's still a laggard over any period beyond twelve months. So to me, there's still a lot of catch up for that company. [00:31:40] S1: Is that really exciting for you? Because you started your career in the semiconductor space, and now you're like, I mean, that was gaming back then, and now