So I'm here with Akash Palkhiwala, the CFO and for the last couple of years I think the COO of Qualcomm. Now, we all know Qualcomm as the company that makes the chipsets in our handsets, the modems, but obviously with AI, particularly just an explosion of chips everywhere, and Qualcomm's no different. You're moving into all kinds of new devices, all kinds of new use cases. So, a few years ago as a tech reporter, if I was talking about the future of AI, I wouldn't be talking about chips. But right now, I think the world wants to understand what's happening in the chip industry because it really is the foundation, the center of so much of what's going on. So let's start with ChatGPT. We are three years on from what in many ways was the foundational moment for AI. Generative AI was going on many years before that. We could all see it building up but then suddenly just exploded into public consciousness and into the minds of every business user because of that one moment. Three years on, we are just at that point where we're saying, well, what is this thing going to do to the world? It hasn't done it yet despite a lot of claims. I've been hearing that this would be the era of AI agents for instance. We're still waiting for that to happen. So what does it look like from where you're sitting at Qualcomm? Is it AI delayed or is it happening? What are you seeing on the ground?
It's interesting. First of all, thanks for having me here. This is an incredible time to be in the technology industry, to be in the semiconductor industry, and for Qualcomm to be at this intersection of the transformation that's happening around the world with AI and to be able to take the technology we create in phones and apply it to a whole other set of industries like automotive, like PCs, like wearable devices, and now data centers and robotics. Super excited about what we can do with the technology portfolio that Qualcomm has created. In the short term, as we think about the last three years, we are incredibly excited because you can see how the momentum is building. We started with the ChatGPT moment and since then we have had a lot of small explosions of innovations and you've seen a bunch of different companies start working on models in parallel as well. The performance of these models continues to get better. We've gone from text models to multimodal models. We've gone from early inferencing to reasoning to mixture of experts. All these new technologies coming in and each one of these things makes the models a lot better. And while today there isn't as much you can hold on to in terms of the value that they're adding, especially in large corporations, we can see the clear path. You can see the dots and the dots are connecting and they're leading to an incredibly positive place in our minds. So I'll give you a few examples of what we're doing at Qualcomm. Qualcomm obviously is very much an engineering-driven company and we write a tremendous amount of software using AI models to be able to talk like a human being to the model and have it write software, have it check your software, have it write test cases and things like that. We're already experimenting at scale. It is a part of some of the development that we're doing today and deploying in our products and we see the path to really continuing to scale this very significantly over the next couple years. So that's one area.
I've heard questions about the quality of code that's coming out of these coding agents. Can you tell what contribution they're going to make? Are they going to be as good as the professional coders?
So we think about it in two parts. First, think of it as something that's assisting you to write the code. Typically in a large organization, you'll have a new engineer come in. The new engineer will write the code but you're going to have someone more senior maybe check the code before you submit it into the product. Think of AI as replacing a new engineer, a new graduate. So it is writing that quality of code. A lot of that also goes on the quality of the prompt that you're asking. If you ask a very specific question, you get a pretty specific and good quality answer.
So for the best coders, they get the biggest productivity boost.
They do, and that's why prompt engineering has become such a big thing. It's really the skill of asking the question the right way.
Yeah. Can we all look forward to that in our professional lives? Do you think that this is going to be something that really makes the most accomplished people that much more productive?
When you take the generalization that far, I'm not sure that's necessarily going to be true. But I do think that there is a skill to asking the question the right way to an AI model. The more context, the more information that you can add to the question, the better quality of the answer that you'll receive.
Yeah. And so it is one of the skills. I mean, my son is an aspiring computer science graduate and one of the things that I always keep telling him is you have to be really good at asking the question the right way. And this generation is growing up in the world of ChatGPT. It comes very naturally to them.
And they're probably going to see a lot more benefit. This is not necessarily to age, but we're seeing new graduates pick this up much faster than people who've been in the workplace for a longer time. So you're seeing coding is one real benefit. Let me give you a couple more. Second is we design a lot of chips at Qualcomm. One of the things you have to do is you design the core technology but then you use those Lego blocks to make the design of a chip. We think there is a role for AI to be played where AI will help us design the layout of the chips and it will accelerate how quickly we can get products to market. The value of that to us is tremendous because we are all in this highly competitive industry and it's a race to launch new products. To the extent you can use tools like AI to develop the products faster and get them to market.
So this is a time you just measure this by time.
Time to market is a clear benefit. There's engineering cost savings as well, but for us the primary benefit is time to market.
Yeah. So these are real benefits now or these are things you're experimenting with?
These are real benefits but we are in the early part of the deployment and we have clear line of sight where this will scale very significantly. I'll give you one more example that we think is very significant is with customer service. When we deploy our chips to our customers and they're doing development on top of it, we get a lot of questions back with how do I make this change? How do I use this piece of technology? And we have a lot of documentation that they could refer to. But with AI, it's much easier for them to ask a question, get an answer right away, sometimes without any intervention of an engineer from our side, and then they can go along their merry way and continue the development process.
Yeah. So, one thing you mentioned just now was that we are seeing these models develop. We all talked about hallucinations two or three years ago. We still talk about it but not quite as much because it seems that there are more techniques for grounding the models to make them not veer off course quite as often. So to what extent does what we currently know about these AI models fulfill this longer term view that you're talking about, this longer term vision of how much it's going to be helpful? Is it just an engineering challenge that we've got to work out how to implement it or do we need big breakthroughs? Do we need more big breakthroughs to make this live up to its promise?
Yeah. I think you almost have to break the problem statement into two parts. The two parts that you outlined. First is how do the models continue to develop and let's assume that we get incremental improvements for the next couple years. Even with that, we think the models are already in a very significant place where they can add a lot of value to things that companies do on a day-to-day basis. And I gave you a few examples. There are a few other examples on the supply chain, finance, business development side. The second part of the conversation is how do you take these models and actually implement something at scale in your organization and there's a separate process to that. That involves changing processes, that involves getting buy-in from people and changing people's behaviors. Sometimes that takes longer than actually the technology piece of it. So we're running both in parallel.
People seem to be always surprised by that second thing that you mentioned there, that this stuff just takes longer. Is that what's happening at the moment three years after ChatGPT? It's just longer than people thought?
That's a portion of it. I think changing people is a tough business to be in. It takes longer than you think. This is about a very large working population absorbing new technology and changing as a result of it. But we don't think in terms of months or a couple years. We think in terms of a few years. What is the opportunity in five years out, I think is extremely clear to most of us. I think we're going to be able to deploy these technologies at scale and incrementally transform things that companies do.
But in two or three years, not five years. When is it going to happen?
I'm very confident that in two to three years Qualcomm is going to have very significant benefit both in terms of time to market and cost as we deploy these technologies.
Well, now one thing obviously that you're known for here is mobile phones, handsets, devices of all kinds. You have low power chips that work at what they call the edge of the networks. And we've seen a type of network, the mobile networks that you thrived in. What we're moving into is going to be completely different. We spent a lot of time talking about the data center. For the last year or two, I think the world has learned much more about data centers than anybody ever thought they were going to. It feels to me like we're going to be talking a lot more about the edge the next year or two because this intelligence has to get closer to people. It has to get closer to the world to be of use. But it's taking time. There are both technology bottlenecks and use case bottlenecks. We've seen Apple for instance take longer with Apple Intelligence on the handset than they initially thought 18 months ago. So tell me a little about this. When is intelligence coming to somewhere useful close to the edge where people are going to get real benefit from it and how are we going to notice it? How is it first kind of?
That's a great question and one of our favorite topics at Qualcomm. We have this firm belief that where AI is going, of course it started in the data center and there's a lot more that'll happen in the data center and we want to participate in that. But the next step is really taking those models and taking smaller models and deploying them in edge devices and having a hybrid AI deployment. Just like computing today, some computing happens on the phone, a lot of computing happens in the cloud, and you split computing between the two. You're going to have the same thing happen in AI. There'll be some AI use cases and processing that happens on the device and some will happen in the cloud and there will be a hybrid between the two scenarios. So that's our vision going forward and we're very confident that if you move the equation of time for a second, the absolute outcome is very clear to us and I think to the rest of the industry. In terms of specific use cases on the edge, let me give you a couple things that are already happening at scale on the edge. The best place to think about AI deployed at the edge at scale is autonomous driving. When you get into your Tesla or any of the other automotives, we're seeing several ADAS deployments happening where AI is used to control the action of the car, whether it's staying in the lane, keeping the distance with the car in front of it. The sensors provide data to our AI chip and then our AI chip drives the car effectively, telling it what to do. That is the reason why you do it on the edge. There's no time to go to the cloud. Your car is driving at an incredible speed. You cannot tolerate the latency of going to the cloud and coming back. You also cannot always rely on a wireless network being available to go and run an inference on the cloud. So that's a great example of a device where it sits on the vehicle, it doesn't go to the cloud and it runs all the time. So that's one case of AI deployment at the edge. The second case, and you're seeing this emerge slowly, we have this vision around personal AI devices and Meta Ray-Ban glasses as an example is a personal AI device. You can wear it on your face. It can see what you can see. It can hear what you can hear. And it is something you can talk to. So you could look at ask your glasses, what phone is that? Who is this? Please make a payment at this QR code. And because the glasses can see what you can see, it has a camera in it. All these use cases are real. They are happening now. Remind me to take medicine at 4 p.m. And it becomes a personal assistant to you that you can just talk to and you ask questions and it'll give you answers to all those questions. I think that is happening now. Over the next two years, you're going to see this significant expansion of personal AI devices and how people interact with their devices.
Yeah. On this question of why intelligence is going to move into these new places, it strikes me one, we hear about privacy for instance, you want your data to be on your device, so you want the intelligence to come to you, not the other way around. It strikes me one thing we don't hear a lot about is the economics of all of this. But given that inference costs are becoming a huge factor in AI for any company with AI-based applications, they're facing incredible... well, we don't really know yet. The more it gets used, the higher their costs are going to be in just running these applications. So if the inference is happening on my phone, that's going to be my chip, my cost, not the application provider. How big a factor is this in the economics of AI at the edge?
I think it's a very significant factor and it becomes a bigger factor as it gets deployed broader, as it goes into emerging markets where the affordability is even lower. You should think of it as to run something on the edge, there are multiple reasons to do it. I gave you an example of latency. You don't have the time to go to the cloud and come back. Privacy and security are massive reasons. There's certain data that you don't want to send to the cloud and you can run AI on it on the device. The third reason is the device has a lot more context about you. It knows where you are, what is around you, and so it can run use cases that are different than the cloud. And then the final one is cost. Once you purchase the device, running an extra inference event is free versus using the cloud where obviously the resources are extremely expensive. So there are several good reasons to run inference and AI on the device.
Yeah, it's just a question of persuading people there is a reason to upgrade their handset to get that more powerful chip and so on.
Yeah, I think the way it works in technology is relatively straightforward. I think once the use cases come through, we've seen consumers adopt these use cases at incredible pace. This has been the history of the phone industry where we've gone from a very simple device to an incredibly capable device today and the consumers have followed it because you're adding so much value to their lives, to everything they do on a daily basis, that adoption is not really an issue once you get the right use cases in place and the technologies in place, which usually takes time.
