Francis Zco here with Tius Research, and I'm sitting with Ziad Asghar, who is the new GM for the XR business at Qualcomm. Ziad, we've been working together for a while now, especially with your previous role in AI. Now that we're here in Hawaii again another year, I think it'd be a good time to talk a little bit about your new role. First of all, congratulations on the new role. It's a big, I think with where AI is going, it's a really great spot for somebody like you. Maybe it's just been a few months, maybe tell us a little bit about how it's going.
I think there is a very interesting inflection point coming, and I think it's coming in the form of just like you said, AI and XR coming together. They make for some amazing use cases. They make for the technology that allows us to be able to actually leverage the real benefits of AI in a very, very strong way. Just think about it, if you have for example your smartphone, typically the smartphone camera is not exactly looking where you're looking at. But in the case of a smart glass, if you're wearing that, it sees what you see, it hears what you hear, and that part actually can really change the generative AI experience for you. We can talk a little bit more about it, but I think it's that cusp of AI and XR where I think there's going to be a lot of innovation. And really, the XR challenges that we have coming up from a perception perspective, from an AI perspective, and power perspective, these are really Qualcomm-sized problems. These are the problems that Qualcomm really excels at doing. This connects extremely well with my previous role as well, and that's why I'm super excited of being over here with XR.
Absolutely. I mean, there's so many directions you can take XR for Qualcomm. Maybe we could talk a little bit about, in your view, what you think. Let's double-click a little bit on that. Within the overall Qualcomm strategy, there's obviously the larger strategy that Cristiano has put out with diversifying your revenue channels and so forth in the areas that you're in, and we're seeing that play out. But also, I think there's the AI strategy as well. So maybe let's double-click about that. In your view, where do you see that XR/AR role within that strategy?
I think we take that really high-level look over here. Where are we coming from? We started with personal computing, right? Everybody has their laptops and desktops. And slowly we transitioned to mobile computing, and everybody has a smartphone in addition to their laptop. And what I envision is that as a next step, you go to spatial computing. And with spatial computing, now everybody will have these devices such as smart glasses or mixed reality glasses, along with all the other variables. I almost want to call it the personal network around the human. I think Cristiano put it very well that Qualcomm operates where the human is at, which is really on the body. So I'm talking about devices like smart glasses and smart watches and earbuds, and all of those coming together to be able to really change the way we interact with the devices. The human-machine interface completely changing. For example, if you have a smartphone, you can typically type, there's a keyboard. But if you have a smart glass, you're not touching anything, right? There is no keyboard at this point in time. You actually have to interface with it using your voice. So this is absolutely best positioned for agent AI. This is the agent, this is the assistant that sits on your device, and you tell it to complete a task, and it's able to tap into multiple different applications, in fact, tap into multiple different devices around you, and then complete the task for you. And I think that's really the story. It really allows us to stitch together all of the different products that we have from Qualcomm in a very meaningful fashion. And then you can, right, the AI that lives on the smartphone is still very usable for this experience. So the way I'm driving the roadmap on the AR side, for example, we have the capability to run about a billion parameter model on the smart glass directly. Exactly, right. So now you can imagine, you go for a run, you go for a bike ride, you don't need to take your phone, you can still have that same seamless experience with your assistant. It's just a smaller model that's running at the same time. If you look at an MR-like device, that actually has exactly or very similar IP as to what we have in our smartphones. I can actually run even a 7 billion parameter model on the MR device. And then you extend it into a PC, we can run a 13 or 15 billion parameter model. You take it to the smartphone, you can do 7 to 10 billion parameter model. You take it to an automotive, and you can do a 30 billion model, right? Because you have so much more AI capability, but yet you still have that great 5G link to be able to go to the cloud and then to be able to do much, much larger models like 100 billion parameter models and even more.
And we were just talking, I just came out of a session with Durga, and one of the things that I took out of there is the fact that I think he mentioned the 8 billion parameter models of today are just as powerful and just as accurate as the 170 billion parameters of two years ago. And I think that trend's just going to continue because AI is going to keep learning how to learn, and it's just going to get better.
In fact, to even add to that, one thing that happens is when a new modality comes in AI, the model is large. Then the data gets curated, curated, curated, people basically make the model much smaller but still very performant. But I think what is even more interesting to me is what people call domain-specific models. So let's say you're creating an application for medical needs or you're creating an application for the hospitality industry, you can actually have a model that's trained on that material. You reduce the chances of hallucination anyways. But now with the full orchestration that's running on there, you augment that with other pieces of information that you have on all the devices that I talked about. And now you have an experience that's far in excess of what anybody can have just in the cloud.
I think what's interesting, especially as you guys do your full strategy, is a lot of things that you're doing are all kind of building blocks towards that vision. Like all the multimodal input capability, that's perfect for like in an AR glasses type situation, right? Because you're talking via voice, but then you can refer to something you're seeing that the camera is seeing, and that brings in as an input to what's going on. So that's amazing.
And that's exactly right. So like the exact use case you're talking, you could be standing in your kitchen and you say, hey, I want to cook based on these ingredients. Now I can either take my phone and make it point towards that, or it can be a very seamless experience where the camera on your smart glasses takes that image to possibly the phone, processes it over there, brings it back to the ear, and essentially you just hear it back. I think that's the kind of experiences that you can light up with a smart glass. And especially as I move it towards the place where we can do smaller models on the smart glass, it can be a completely, wholly, fully contained device that actually is able to do those use cases very well.
Okay, so it looks like models are getting there. You got other things like, I was just at one of the demo rooms down there, put on the Meta Quest 3 and got to see the full city, which was amazing. It was the first time I actually really, I mean between the video passthrough and how well that's working, and also being able to interact with your hands on the digital interfaces. To me, the end goal here is to have that kind of capability that we have now with XR, to put it into regular glasses, where you know, fashionable, the Ray-Ban Meta type stuff. What do you think it's going to take for us to get there?
That's a great question. So today, just like the case you talked about, we have amazing video see-through technology. Number one, we have been investing in these technologies for the longest time. We acquired multiple teams. We have all of what we call the XR perception technologies in-house. So these are like hand tracking, some of the experiences that you took through in the demo. But we also have things like SLAM and 3D reconstruction and plane detection and all of those technologies. But what it will take now is from the video see-through world to change it to the optical see-through world. So with smart glasses, you can have optical see-through. You put in simple displays now, which by the way is another big factor that was still lagging. So number one step, you are able to basically now bring in the displays. Displays are being produced with waveguide-like technology at pretty high volumes and at a reasonable price point. So you get glasses that have displays now. You actually use optical see-through. You still have all the perception technologies. You interact with the world with your hands, and basically the hand tracking and all those investments that we have made, they flow through absolutely easily into this realm. The third thing that you really have to do is the smaller models. So like our engagement with Meta, for example, the Llama 3.2, we have a 1 billion parameter model. The 1 billion parameter model can run directly on the smart glass. Now you're opening up a whole gamut of new use cases that were just not there before. That's just the third step. But fourth step is the products that we have out there in the market today, the march of the transistor improvement continues. So now I can use the same use case but with much lesser power. That's the IP level improvement that I can have. The transistor level improvement is the second one. And then the third one is basically architecture level improvement. So in all products, the way I have been driving the roadmap has been that we have the active version of the AP, but we also are able to run it in what we call always-on subsystem. This is our low-power island and all. So now you can imagine that I can be doing a coarse level of correction with my smart glasses in that low-power island in the very always-on low-power domain, and then when an event happens, I bring up the larger system. So these architectural improvements, the transistor improvements, and IP improvements are going to get us to a power point where we can actually use these glasses all day long. So that's how it's all coming together.
