The Tech Stack Powering the Intelligent Edge | Cristiano Amon | USA House

Thank you, Pat. How are you?

Oh, that was a great event last night. It seemed like a who's who of people coming in. It was great to see Michael Katzios talking a little bit about diffusion and Secretary Lutnik as well. I'm not too sure what time he went home, but he had a great time at the party. It was great to see him.

No, it was good.

Yes. Look, I will say we started talking about AI at the edge before it was popular, but maybe I'll answer your question going a little bit back in history because I think it's interesting. I remember when you talk about the beginning of personal computers and you have some of the biggest companies in the world at the time. They thought everything is going to be a mainframe. They didn't really understand that you could actually do a personal computer. You saw what happens with Microsoft and MS-DOS. That was the beginning of it.

All right. Then you fast forward. Microsoft was one of the largest company in the world and the smartphone revolution happened and you could put a computer in pocket and there's some skepticism about that and I think that you saw Android came as a result. So I think the interesting thing is it's never about cloud or edge. It's not about what is more important. The software will always find its most efficient computer and it's going to be designed. It's almost like this. I have in our phone business we have seen so many discussions over the years even when for example like webOS came and web apps like everything's going to be you don't need computing and what's actually happened was the opposite the amount of computing you have on mobile devices today is incredible that continue so when you fast forward to what's actually happening with computing AI is just a new way to do software and computing this is going to propagate to everything and it was just fascinating I think when you and I were waiting to come over here you saw the panel with David Sax and Satya Nadella the workstation it's exciting again because you can run some very large language models into a workstation and so I think what's happening we are seeing this transition that a lot of the computing everywhere is running AI and I think that's kind of what we see some of the opportunities for a company. Phones it's the edge, PCs are the edge, cars which is now computing surface by the way autonomous driving is basically AI running on the edge you have robotics you have the transformation of industrial and at the end the way this is going to play out you're going to see the edge and the cloud that is going to work as one. It's interesting the phone is the most cloud connected device ever. If you put it on airplane mode, you're not going to use it. You're going to be very frustrated. But it has a lot of compute. And I think that's it's not going to be any different for AI. And I'm excited because 2026 appears to be when the realization happens and it's going to be the year that the edge is going to become very relevant for AI.

No, very good. And maybe I'll use the opportunity Pat to talk a little bit technically about what happens with computing systems and I think this is going to be with AI when you go to a process of disaggregation. So I am going to take just a few minutes to talk about what happened with mobile phones because this is going to be an important context to talk about what we're doing in AI. So for those that understand how PCs were run it was very CPU centric architecture you run a lot of things on the CPU and then you have also a GPU that was doing a lot of rendering for games and things like that. So, if you have, for example, to play an MP3 file, I think the ones that probably remember when you're running iTunes on your PC, the CPU is going to do the decode, right? The CPU will decode the MP3 file and play the music, but it burns so much power.

So what happened in contrast when you look at a PC and this is before we entered the PC space but when you look at a PC and look at a phone you look at the architecture is very very different. In the phone you have an interesting environment that you need a lot of compute but you have a very scarce resource which is power consumption. So you have a battery then you need to figure out how to get the most efficient compute because the battery has to last all day. So what you do you start disaggregating the compute you started to say I'm going to have the separate compute that does MP3 file have the second one that does JPEG decode when you take a picture this is the image signal processor and you have the silicon compute because you disaggregate and that is interesting because it's happening on the data center right now actually if you look this is the curve of growth in compute capacity is the energy so people say how do I get this efficiency you see for example When you move from training to inference, the data centers disaggregate and they say what's my compute for prefill? What's my compute for decode? And interesting, you go to the same thought process that you have to go when you have to become very efficient and deal with what's the most efficient compute to run that task. So that goes back to your question about Qualcomm. We have developed we have most people don't know we ship a large number of GPUs every year because all the phones have GPUs we have our GPUs we have our own CPUs and we also have we develop our own NPU and we believe that for certain things for certain tasks you want to go find the most efficient compute and we decided to scale that architecture up. The first step in scaling that architecture when we went to PC and I think it was great to see Satya talking about all of the five silicon models of Microsoft running on our NPU and PCs. Then we took it to cars and you'll see a lot of stacks of autonomous driving running on our processor and now we're entering into the data center space. So we're architecting this for what we think is going to be the post-GPU architecture for inference and I think it creates interesting opportunities. We're very excited about this and I think that's the unique thing about our company. We are uniquely situated because we go from sub five watts when we do an earbud that has a model that does running on the earbud that does voice detect and send it to the agent sub five watts all the way to 500 watts and we have that architecture that scales. So we're excited about what the future holds for us.

It validates this point.