Good morning. Welcome back to COMPUTEX CEO Keynote. As AI moves beyond the cloud and into the real world, technologies at the edge are becoming more important than ever. From automotive to industrial systems, from smart devices to intelligent infrastructure, NXP is helping bring AI into the real world. Today's speaker has spent 30 years shaping this industry, from Motorola and Intel to Broadcom, and now leading NXP as President and CEO. He is the architect of NXP's Edge AI strategy and one of the most respected voices in the semiconductor world today. Ladies and gentlemen, please welcome Rafael Sotomayor. Welcome back, Rafael.
Thank you, James. Thank you. Well, let's start. Eight days. Eight days from today, the best footballers on the planet are going to get together, and they're going to play in this tournament that mesmerizes nations. When their team plays, countries stop. Billions of people watching. Years of preparation. The players on the field are elite. What makes them elite compared to the rest of the players? It's not fitness, because they're all fit, and it's not knowledge of the game, because all of them know the game. It's the mastering of the task. It's executing at the highest level with the highest level of pressure. That is what elite looks like. Here in COMPUTEX, and I walk the halls, there are plenty of examples on physical AI. They're everywhere.
I ponder the following question: If we're going to bring all these intelligence systems and devices into our lives, what does elite look like for them? What is the equivalent of a world-class athlete for a machine? I'm going to need you to follow me here, because that's a question we're going to answer today. If you are a fan of football, you've probably seen Messi play. Now, if you're not a fan, watch Messi play and you become one. He's very smooth, the way he handles the ball so effortlessly. When he was about to kick the ball, it seems that waits at the last minute to whether he's going to go left, he's going to go right, depending on what the opposing goalie is doing.
In that gap, that tiny, infinitesimal gap between trigger and response, is probably one of the most sophisticated feats of intelligence on the planet, and yet he doesn't think about it. His reflexes take over. We all know excellence is not about thinking harder. It's about mastering the task so deeply that thinking becomes unnecessary. Let me give you an example here with you. What's happening with you, and I see a lot of you here. What's happening with your body? Some of you guys are already adjusting your posture. You're fidgeting with your hands, your feet, and you're doing this without thinking. It turns out that most of our daily tasks, over 95% of the things that we do on a daily basis, we do it without any conscious thought, without any effort, with very little energy expenditure. Think about that. That is brilliant.
What we spend most of the day doing costs us very little. If reflexes dominate our daily life, then I would contend that reflexes are foundational for robotics, are foundational for physical AI. That's the key. Reflexes are hard, that brings us to Moravec's paradox. This concept, Moravec was the first one to highlight at the time, something that seemed very counterintuitive. What is really, really hard for human beings, reasoning, solving complex problems, playing chess, those are things that are really easy for machines. What is easy for humans, like walking, like folding clothes, that's not so easy for machines. Reflexes, not language, not reasoning, but reflexes are the hardest thing in robotics.
I want you to hold this notion of reflexes being difficult for a moment, because we're going to get back to it, because that's key to unlock the full potential of physical AI. Right now is fascinating times. It feels it here in COMPUTEX. All of you guys are here. We're living through the fastest technology evolution ever. I know you feel it. Daily, forget about daily, almost hourly, something amazing gets disclosed. Something amazing is happening, and it gets released, whether it's new stuff in agentic AI or new developments in software-defined vehicles or machines that speak any language and solve all sorts of complex problems. It's amazing, and a lot of that is coming from the innovation that is happening of AI at the cloud. Now AI is available and moving to the real world, and it's touching and transforming everything.
The real world is what we call at NXP, the Edge. The Edge, that's where we shine. The Edge products is what we do. These are fit-for-purpose products. They have real-time performance needs, that they are low latency, low power, high levels of security, high levels of safety, products that go into software-defined vehicles, smart factories, smart infrastructure. Those are our products. As we move intelligence from the cloud to the Edge, what is so exciting is now all these devices get intelligence. You get to create new use cases. You have the ability now to create devices that are making decisions. They're independent, maybe even autonomous. In order to actually reach the full potential, we still need to solve the riddle, the Moravec's paradox.
Now I'm going to talk to you about how NXP is very well positioned to provide a solution to it. In order to get clues on how to solve Moravec's paradox, we need to start in a place that already solved the problem of reflexes, which is us. Now, if I were to ask you a question today, which the following question is, where does intelligence reside in the human body? I would guess that over 95% of you would tell me the human brain. You will be partially correct, which means you're partially incorrect. Let's start with the greatest physical AI ever built, the human brain. The cerebrum is the largest part, representing 80% of the mass of the brain. This is where reasoning, conscious thinking, learning, decision-making, this all works here. This is not where the fastest responses get resolved.
Let me tell you what happened to me this morning, and this is actually real. The NXP office is really a few blocks away from here. Listen, I'm a little jet lagged, I was a little distracted, I crossed the street, I was this close of getting run over by a scooter. It was this close, and it woke me up. I'm looking at this 300 milliseconds of response time by the cerebrum. That's an eternity, okay? What happened to me this morning, and I jumped like a cat. The cerebrum didn't help me at all this morning. Now, the good thing about is that the cerebrum doesn't work alone. It has a companion. Underneath the cerebrum, it sits the cerebellum. Think about the cerebellum as the co-processor for motion, for motor control, for position, balance.
It's still faster than cerebrum, but not fast enough. Not for me to avoid the scooter this morning, tell you that. We got to go a level deeper. We get to the spinal cord. This, life-saving reflexes don't originate in conscious thought. They don't even originate in the skull. They live in the spinal cord. You can see the stats here. 40 milliseconds. It's lightning fast. This is nature's decision and design by choice. Because in this case, latency is more important than intelligence. This is what saved me this morning. The spinal cord processes, decides, and acts independently. Let me give you an example to make it more concrete other than me trying to avoid the motorcycle this morning, and I know you will relate to this. You touch something hot. The stimulus reaches your hand and sends a signal to the spinal cord.
The spinal cord gets the signal and says, whoop, and sends a muscle command control that says, remove the hand. Now the hand is away and safe. Meanwhile, the brain is still processing. Then eventually catches up, and you realize, ooh, that was hot. Ouch, that hurt. Now you can see the spinal cord allows us to act before we think. The spinal cord keeps us safe. Now, before I move on, I want to highlight another design principle. I want to highlight it. Nature, on its infinite wisdom, decided to put reflexes where the action is. Okay? Essentially, the spinal cord is a stack processing center that runs along your back, placing, sensing, and processing to the nearest possible point. Again, brilliant. Closer means faster means safer, and also means that it's the lowest possible energy use.
What evolution is showing us here is billion years of optimization. Okay, where are we? We started with this conversation of biology, looking for clues on how to solve Moravec's riddle. Okay, what are the takeaways? The takeaway says that you don't scale intelligence by making the brain bigger and bigger. You scale intelligence by putting intelligence at the right place. There are three fundamental requirements of intelligence in real life. You need to have ultra-low latency responses to actually be able to live in the real world. You have distributed control, so there's no single points of failure, and you have energy efficiency, extreme energy efficiency. You actually manage a strict budget to be able to expand life. This is the playbook. This is the blueprint. We contend that this is what is needed in physical AI.
We now take this principle and apply it to robotics. This is what we call at NXP, the neural axis architecture. Intelligence has three layers: the reasoning layer, the coordination layer, and the reflection layer. Three layers, independent but highly coordinated. Now, let's see it in practice. We have three different form factors here. All of them, which NXP participates. Three different implementations. Let's look at first one, the drones. Drones, you can see also the neural axis maps very nicely here. You have the reasoning layer doing all the flight planning, path optimization, and then you have the cerebrum managing flight balance and performance. Both of them need to be operating independently, concurrently. This is not a nice to have, okay? This decision of having this independence is what makes this device trustworthy, safe.
The reflex layer, which is now these nodes at the outer edge, they're connected to motor controls and actuation, and these are fast, reflexive. They don't wait. Microseconds. At NXP, we deliver this full system. All the KPIs, all the things that are needed to build a world-class drone. See, I give you an example. We tracked a metric that we call glass-to-glass latency, which if you think about it, camera captures, processes, transmits, the controller responds, then drone reacts. All that loop, 20 milliseconds. Miss it, drift, instability, potentially even a crash. Hit it, you're in control. The drone feels alive, safe. This is NXP's implementation of the neural axis in a drone. Maybe this one we're a little bit more familiar since we have a lot of customers in automotive, the self-driving vehicle is also a great example of the neural axis.
You have the reasoning, obviously ADAS, navigation, what is the vehicle going to do next? You have coordination, which is this is a separate compute that manages vehicle dynamics, ensuring the car always has stable motion. This is an area where NXP has created a leadership position with a central compute 5-nm S32N family of processors. Processors and products that are like any on the market. Now, on the reflexive layers, we have the zonal, which are managing mission critical functions of the vehicle like braking, like suspension. This is another area where NXP has created a leadership position in the zonal architectures with S32K family of products. Now, this separation, both logical and physically of the three layers is what makes this car reliable. In a vehicle with lives at stake, you don't have margin for error.
The reason NXP has created a leadership position in SDV is because we created a leadership position in the neural axis for SDV. Humanoids, which is the most complex implementation of all. I contend, because this is a nascent market, and you hear a lot of different topics and different conversations about the architecture of humanoid, but our belief is still very firm. If you're going to deploy, if this neural axis architecture is good enough for life, it's good enough for humanoids. Okay? I don't think this is a conversation. Let me maybe bring this alive with an example. Imagine a robot walking in a warehouse, and it's carrying a very valuable and very fragile package. Halfway down the hall, somebody drops a pallet, and it kind of brushes up the robot, and it kind of loses a bit of control.
Now think what happens now. The robot needs to recover balance, make sure that, confirms it's got still the package, adjust the grip, understand where he is in location, recover, and continue walking. All that needs to happen within 40 milliseconds. No cloud call, no waiting for the model to respond. Okay. Intelligence at the body, at the joints, at the hands, at the feet. They must own the moment. You see, the neural axis architecture is not a nice to have. That is existential. Today, you see usually humanoids with a separation of the reasoning part, with the separation of the coordination because robots continue to learn and continue to learn new use cases, but you don't want to disturb motion of the robot. Motion needs to be stable.
Instead of a spine, you have all these distributor controls at the end, at the nodes, and these processors own their function locally. They own it. Hands know how to grab, ankles know how to balance. Not waiting, no asking for permission, acting, making sure the humanoid survives. We just walked through three examples, three different form factors, three different levels of complexity, one system. One architecture, one blueprint, three layers of intelligence. What's the takeaway? Intelligence is not about a bigger brain. Stop thinking about a bigger brain. Think about a Neural Axis. Three layers of intelligence. Okay? Now, we talk a lot about motion. Now I'm going to introduce you kind of a new wrinkle. Motion and movement is not the same as understanding. The fact that a robot is able to execute a perfect reflex doesn't mean they understand why.
I would say that not understanding why in a world as complex as the one we live in, that's not okay. The question is: how do we teach a robot not only how to move, but how to understand? Let me go there next. We're not born, even though some of us may believe so, that we were born walking. We were not born knowing how to kick a ball or grab a glass of milk without spilling it. We learn, slowly through mistakes. This process of trial and error didn't only build a skill for us, it created a physical model of the world in our head. Okay? We learn the concept of gravity because we fall and it hurts, so we don't want to do that again. We learn the concept of heat because we touched something hot and we got burned.
Obviously the way we learn, that doesn't work for robots. You can't just let them in the wild and say, okay, go ahead and learn. Come back. It will be not safe, probably reckless. We have to take a different approach. Now, the one thing robots do very well, and I tell you, they see remarkably well. Let me use this example of my back here with a robot looking at the cough syrup. Probably looks and identifies the bottle, understands that it's got liquid inside, probably reads the label. Does it understand what it's holding? Does it understand that the bottle is full? Does it understand if you tilt it a little too much, you are going to spill the liquid? If you don't understand inertia, it's really hard to estimate force.
If you don't understand friction, you don't know how that syrup is going to pour. Without that understanding, the robot won't act in a safe manner. You need a bridge between perception and understanding. Perception tells you what's in front of you. The understanding of the world tells you what is going to happen when you interact with that object in front of you. We need a bridge. The question here is: how do we teach robots physics? Today, we do it ourselves. I really mean it. Individually, we do it ourselves. People teach the robots how to pick objects, how to move. They mimic, the robot adapts, learns. It works, but it has limitations. It's slow. It's very expensive. This is where world models come in.
The ability now to teach a robot to really estimate outcomes without going through the process of experiencing it. It's like you literally inject knowledge into the robot. Think about The Matrix, right? Inject knowledge into the robot. You think about it. You inject knowledge, you inject experience without having to go through the physical act of acquiring that experience. This is a fascinating field. World models are feeding VLAs. VLAs, I'm getting really, really, really techy here. VLAs stands for Vision-Language-Action models. It is the bridge between what the robot sees, what it's told, and how it moves. This is the bridge between these three different layers. It creates the bridge that it needs to be able to see and perceive, to understand. This is one of the most fascinating research areas right now in both academia and industry, VLAs.
VLAs are happening, and they're getting better and better and better. The question right now is, okay, so how do we take these VLAs that were created in the cloud with essentially unlimited resources, and you move into the edge because the edge is a constrained environment. It's a constrained environment with power, with latency, with memory, with processing. A VLA model in a research notebook, it's not very helpful for a robot in a warehouse, right? You need to bring those models into the real processors at the edge. Our answer to that challenge, to that problem is right here, our toolkit, our eIQ Toolkit. This is a software that allows to grab those models, imports them, quantize and prune them, compile them. Essentially, you grab the model and you tune it for the target hardware and for the target use case.
Our goal is to remove objections from a customer. All these things are complex, so as we look at the challenges, we want to make sure that we remove objections, we remove friction for anything that prevents physical AI from being deployed. Where are we? We talked about the neural axis as the architecture, and we talk of VLA as the understanding, and all these two things come together and make the robot useful. Is that enough? Is useful enough? Useful and trustworthy are not the same. For physical AI to get really adopted in the market, trust needs to be solved. Trust for a human being, you earn it with time. You invest in your reputation and your rapport and your track record, months, years. You can't have a machine trying to develop rapport with an operator for a year, okay?
We don't have the time. You can't design trust. You can't earn, the machine cannot earn trust. It has to be designed in from the beginning. The machine needs to be trustworthy from moment zero. Trust, it's important. Nobody likes to think that things are going to go wrong, but they go wrong. When I look at these pictures, and you could probably cringe when you look at them, I can't help myself to think about two laws, Murphy's Law and Finagle's Law. Murphy's Law, anything that can go wrong, will go wrong. Anything can go wrong in Finagle's Law will go wrong at the wrong possible time. That is the moment where trust gets defined. Trust doesn't get defined when everything is going great. Trust gets defined at the worst possible moment. How we look at trust.
Trust can actually be compromised through external issues, through external triggers, or through internal triggers. External triggers could be changes in the environment. It could be malicious people trying to attack the system. Internally, you have design flaws, model limitations, sometimes wear and tear creates a different behavior in electronics. These are important issues because as Physical AI gets deployed, more and more mission-critical functions are going to rely on them. Dealing with this is massively important. Now, we at NXP don't come to Physical AI from the cloud. We come from the edge. These are the things that we're already working on. This is part of what we do today, and now it's just, for us, it's more valuable. Some of the assets that we have at NXP to take care of trust are even more important with Physical AI.
Let me tell you how we, as a high level, discuss trust, because we have a framework today that just right now is getting enhanced. It's a very simple one at a high level. We contain. This means that we isolate the problem. We create redundancy, and we don't allow a single point of failure. We protect. We have hardware injected directly. We have security injected directly into the hardware. We protect execution of the code. We protect tampering. We protect credentials. Guess what. We're also post-quantum crypto ready because the threats today are not the same as the threats five years from now. This future-proofs our devices. The next primitive is verify. Some of the certifications that we have on SIL and ASIL, this is not just paperwork. This is how trust gets measured.
Our SafeAssure program allows a hospital or a car OEM to get the one thing that is really, really important for them. Assurances of how the device will behave when things don't go right. Obviously adapt. Nothing stays still in this market. Threats continue to evolve, environments continue to evolve. We need to evolve. We need to have the mechanism to constantly update these devices. With these devices at the edge, they stay at the edge for a decade. You need to make sure that these devices are constantly integral, secure for years and years to come. We take trust very, very seriously at NXP. Mistakes and errors in physical and at the edge, they're not digital. You cannot do a software patch for a broken bone. You cannot do a system update to deal with a collision.
I know we cannot design systems where nothing goes wrong. I know that. We can design systems where something goes wrong, still goes right. Trust is at the core of our DNA. The reason why is because we know this, because the real world has no undo button. We get to a part where I start landing now what's happening with physical AI today. Our job right now is not really to push physical AI before our customers are ready for it. Our job is to actually making sure that physical AI has a place to be. There's a reason why physical AI gets adopted. The title here is Earning the Right to be Useful. I get a lot of questions right now from customers who are trying to adopt AI.
One of the questions I get often is, hey, Rafael, how smart can physical AI become? I just ponder on that question because I don't think it's the right question. I think the right question is, where can physical AI be deployed today, that it can be done in a safely secure manner where they create value? I think the answer to that question is unraveling right now. Let me give you some examples. In factory automation, even today, in factories where physical AI-enabled robots are being deployed, already having 40% productivity enhancements over regular automation. Regular automation is pretty high already. This is on top of it, 40%. Obviously, we don't work alone. We work with our customers. We're very proud of working with one of our customers here.
I'm showing you Boston Dynamics, where we combine our hardware or software with the phenomenal robotics platform they have to make sure that robots and humans can cooperate in a productive manner in the factory floor. Now, healthcare, and this is where safety and precision are super important. They define outcomes. Already in 2025, diagnostics and lab robots increased sales by 610%. This is not a trend. This is a signal that the healthcare says physical AI, if deployed responsibly, will save lives. We're doing and very proud of the engagement that we have with GE HealthCare, where they're actually deploying intelligence systems in anesthesia, where precision safety really makes a difference. In their products, NXP is at the core of it. Now this is the time that I always do when I come to Taiwan, because it's truly amazing, the ecosystem in Taiwan.
We know we can't do this alone. NXP, without our customer and our partners, we're too small. The reason NXP punches above our weight is because of you. We can't do this on innovation alone. We do collaboration with co-creation and delivering the market. The people here on this list are not just customers and partners, are the lifeblood what makes physical AI deployment into the market and makes it a reality. For the companies who are here and the companies in this list, thank you, because we do this together. I started with Messi, and I'm going to end up there, too. What makes this player so amazing and so elite and this body does these incredible things, it's not the things that you see. It's all the things that you don't see.
Put aside his incredible commitment to his craft, but all the things that you see, the invisible architecture in his body that we call the neural axis, which is the same for physical AI and robotics, just engineered by man. We started with a question, what is elite for a machine? I think we answered that question. Elite for a machine is a machine that executes reliably in high performance in the real world and the real conditions, and when things go wrong, this machine keeps assets and people safe. Machines and systems, these elite systems have three non-negotiables: low latency, low power, and high, high level of trust. Remember, you don't scale physical AI with intelligence alone. You scale it by placing intelligence at the right place using the neural axis as its backbone. For the push of physical AI into the market, we at NXP, we're ready.
Thank you.