Dumont, it's great to have you here on The Road to Autonomy. Physical AI is exploding. It's everywhere. It's real. And ARM is at the center of it because you have a low-power architecture. How is ARM approaching the advancement and the commercialization of physical AI?
Hey Grayson, first off, thanks for having us on. I'm a big fan of your pod, and I'm just listening to the one on Robotaxi, so maybe we'll get a chance to talk about that a little later. ARM's approach here is that physical AI is where AI is coming into machines. It's embodying AI in machines — machines that sense, that can reason, that can act in the physical world. So, very simply, that is two major markets for us: that's the automotive market and the robotics market.
And those markets share some common sort of compute challenges when you sort of look into them, right? When you sort of double-click into what is similar about the automotive market and the robotics market. And I think I've heard you say that autonomous vehicles are just robots on wheels. If I'm not mistaken, I think that came from you. And I think there's a lot of truth to that, right? So, when you get into the different levels of compute which are required for robotics and autonomous vehicles, when you're talking about perception and planning, when you're talking about interaction, actuation and sensing, there's a lot of common technologies there. So, our goal is, and we see that ARM compute is at the heart of every layer of those, right? Up to and including what's happening at the moment, which is the cloud, right? Where you're seeing fleet management, additional learning, and training happening in the cloud on ARM-based compute. So, I feel like our vision is that ARM is at every layer of the compute stack when you look at a physical AI.
Us personally, we use your architecture because it's efficient. It's really good architecture and it's massively efficient. And I'm just a user, I'm a customer. But overall, in your opinion, you're sitting there, you're developing this, you're working with it — what does the ARM architecture unlock for physical AI that your competitors cannot unlock?
So, what we bring is we bring two forms of heritage, for want of a better term, right? And the first is that we've been in the automotive space for 30 years, right? We've been building IP that's been consumed by people building chips for the automotive space for 30 years. That has allowed us to learn about that market, it has allowed us to build in special features, safety features into our products that make it much more straightforward if you're building a chip based on that IP stack. You're building a chip that you have to qualify — the person taking that chip has to build their product and have to qualify it in the market. All of that stack, if you have to do that based on IP that is not qualified, it's a much harder challenge for you as a system provider. And most people don't want to get into that. They just want to trust and believe that the IP stack that they've built upon is that, right? It's trustworthy, right? And the reason you buy IP and license IP is for that trust. So, we've spent a lot of time and energy learning about the market and learning about the needs and building the right IP stack.
The other piece of heritage that we bring is what comes from the mobile market, right? And probably where we're most famous, actually. I don't know that everybody knows we're in almost every vehicle on the road, but I'm sure most people who know about ARM think of us as the mobile company. And that heritage is just low power, right? Everything is done — as an ex-engineer myself, inserting clock gates at every level in the clock tree, inserting power gating in every piece of the implementation, building an SOC that consumes as close to zero power when it's not in use, right? Like we have all of those technology facets built into the IP stack. So you bring those together, now you've got a really low-power, power-efficient solution, plus you've got the automotive-enabled piece. And the other thing, actually, which as we're talking about it, is like we don't just build CPUs, right? We're not just a CPU company. We build CPU IP. We build GPU IP. We build image signal processing IP. And then we build computer subsystems where we pull it all together. And what that allows you to do is go, if I'm building a vision pipeline to take all of the sensor data that's coming that I'm ingesting, whether that's into a robot or into a vehicle, that computer vision pipeline is all automotive. We call it AE, right? Our automotive-enabled. It's all automotive-enabled grade. So you can go, okay, everything from the sensor to the compute and back out to the actuator is already a proven safe pipeline. That's the value kind of we bring there because we're presenting it as a complete stack. And I think that would just make it easier if you're a chip maker to bring your own differentiation but to build it on a foundation of all those IPs.
Do you imagine this hypothetical scenario where ARM wasn't around with your low-power architecture and physical AI tried to scale on high power? To me it would just be a failure. I mean, to me, low power truly unlocks physical AI and what's possible with it.
If you look at most robotic solutions, they're energy sensitive, right? And that's — how much energy you're consuming drives how big a battery you need to have, right? How big your cooling solution needs to be? Can you do it with air cooling or do you need liquid cooling? And then how much does it weigh, right? If I'm building something that has liquid cooling, now all of a sudden I've got to make a decision. I've got to put a liquid cooling solution in, which is probably taking away some of my battery capacity, which is therefore taking away my day-in-the-life experience for this robot. I've seen some really cool examples on my travels of robots where people are putting them into operation for 8, 12 hours at a time, but they're always looking for more, right? The longer you have that robot in operation, the less charging time, the more productive that robot is, right? So yeah, I mean, it may not answer your question, but I just think that the robotic space is super sensitive to power, and I think power efficiency just unlocks robotics. And without it, I think you end up with some pretty poor experiences, you know.
Historically, robotics have always been, 'Oh, that's a science fair kid. Oh, that lives in this little small ecosystem.' And now it's on its way into the mainstream. When Jensen gave the speech and declared physical AI, all of a sudden the attention turned to it, but unfortunately it's fizzled and a lot of the attention is still on LLMs. So, here's the big 50,000-foot question: When does physical AI have its ChatGPT moment, would you estimate?
It's a great question. So, first, you have to give hats off to all the people who've been in robotics from the start, right? And who've been believing that this is going to happen. And I talk to these guys all the time and it's just awesome when you meet with these guys because they're just so passionate, right? And they're so open and so sharing. So, it's a fantastic community. I love engaging with that community because they're definitely none of them, even though they've been in it for many of them since high school, no sign of being jaded by it. It's really cool, right? So, but to answer the question on how do we scale physical AI, right? And I think what we have to look to is some form of standardization, right? Some form of common language, some form of way to describe robots that those building the parts for robots, those assembling the robots, those deploying the robots, those exposed to the robots can understand, right? And if I draw the analogy back to ADAS standards, most people, certainly tech people, but even outside of that in the industry, will understand what ADAS level two means, what ADAS level three means, what ADAS level four means, right? And that has allowed that ecosystem to scale. Right now, it's still at the early phases of it, but it's scaling because you now can say, 'Well, that's an L3 vehicle.' And I understand what that means. And a regulator — you can have a conversation with a regulator, whether it's in Austin, Texas, where I live, or somewhere else in the country, and you can say, 'Well, this is an L3 vehicle.' And they can go, 'Oh, what does that mean?' And there's a lookup table for what that means, and all the meaning behind it, and all the work that's been done to certify it. We don't have that yet in the robotic space. And that's something that, when we announced a robotics capability framework, one of the things we were looking to do was to start that conversation.
Right? And I very much emphasize it's the start of a conversation with those in the robotics ecosystem who are building and deploying robots. We as ARM, we don't pretend to have all the answers, and we don't pretend to show up with a perfectly defined robotics capability framework — that would be amazing hubris from our side. But what's been interesting about it is from when we've announced it, the interest in working on it is wide in the industry. So, we're getting a lot of interest on the robotics capability framework. So, I think having a standard language, and that's what we're doing with the robotics capability framework, is really important. Anybody who's listening to this, who's in the mix and working on robotics, please reach out — I'd absolutely love to connect with you. We've recently hired a chief robotics technologist to come in and work in this space. So, that's really, I think, a wide-open invitation.
The other thing that we've been talking about, when we talked about scaling physical AI, is how do you make it easier to build, assemble, procure parts, prove out software? How do you get some of the — whether it's VLAs or LLMs or whatever models — running on it? Like essentially, if you look at any other ecosystem that has scaled, there's communities, right? And there's an ecosystem of players. And there's a place where they can have common discussions and common conversations. We think with the Arm Total Design for Physical AI program that we've announced, we can help with that, right? We think we can help with that. And we put that together in just over a month, right? And this will tell you the kind of interest that we've seen. We scaled right from introducing this to some of our partners and some of our customers to having over 80 partners signed up in a month, right? Yeah, I was very surprised. I set the target for my team, 'Hey, can we get to 25? Can we get to 25 before we announce this?' And we're at 80 and there's still more coming in. And I don't think that's — I mean, that's a reflection of my team. I think they did a great job, but I think it's also a reflection on the fact that everybody in the industry feels like this market, this physical AI market, has yet to scale. And they're looking for ways and opportunities to drive that scaling. And I think when somebody says, 'Hey,' particularly with Arm's reputation for building ecosystems, says, 'Hey, we're going to help with this and we're going to participate and bring people together,' people go, 'Yeah, that's the kind of community that can make that kind of change happen, you know?'
You have the community there. You also have the resources and the technical know-how to accelerate this. Well, it sounds to me there's two threads here. One, it sounds that the framework is very similar to the SAE levels of automation, and then it sounds that the ecosystem is, let's call it homebrew — a grown-up version of homebrew, people can come together and build the future. Am I reading both of those correctly?
Yeah, well, let's start with the framework. So, just like the SAE levels framework, there's level zero to five, and it's everything from just reactive robots, you know, to robots with a plan, deliberative robots, all the way up to self-improving robots, robots that can sort of self-train, right? And as you look at those, you start thinking, okay, what are the important factors of those robots? So, they've got different capabilities. I need to understand their behaviors. I need to understand what kind of outcomes that they're expected to drive, and ultimately, what capability is beneath that level, right? It might be fixed rules, right? So, very directly programmed, old-school kind of robots. Robots with rule-based logic, self-optimizing robots, and across that spectrum. So, the robotics capability framework, as it stands, and like I said, it's something that we're inviting the community to help us shape, tries to set the boundaries of what is important there. But even in some of the early conversations we're having with partners, we're getting feedback on, 'Hey, there's a vector here that we may need to think about.' So, it's great, right? It started a conversation, and that's what we wanted, right? We wanted it to start a conversation. So, that's it on the robotics capability framework.
On the Arm Total Design ecosystem, yeah, I love that reframe as sort of maker grown-up, right? It is interesting when you're at Arm and just having sort of this front-row view into the industry and the ability to talk to all of these companies. You sit in a lot of rooms with people who've been in the automotive industry or at the early stages of the physical AI industry. And quite often what they bring up is that there's a fragmentation issue. Right? There's a fragmentation concern. And unless we can come together in a way where the guy building the low-level OS and the guy building the board and the guy writing the software and the guy training the model that runs on the software and the guy actually building the robot at the end of the day have a place to gather and to work together, they're all operating in isolated silos, right? And I think what we know from history is when you build sort of modular solutions, you buffer on every level of that modularity, right? So, you create interfaces that are standardized to make it easy for each other to talk to each other. But by doing that, you kind of remove the opportunity to optimize across those layers of, you know, whether that's computer software. So, I think it is important for the people who are building the software and the models to be talking very closely to the guys who are building the actuation system so that they can go, 'Hey, how can we make this better?' Because we're still at a stage where a humanoid robot spends an awful lot of its energy just on standing up. Right? And that's just one example, right? Where you think, 'Okay, if you could optimize the way the actuators talk to the models, maybe you can improve that.' Right? And it's not just, 'Hey, I've got a fixed actuation execution plane, and I've got to make my software work on it.' Maybe the software guy can now communicate with the actuation guy and go, 'Hey, if you built it a little bit differently, I could have a better solution.' And together, they would work better together. So, that level of communication and interaction between teams and between companies is kind of what we're trying to foster there.
I'll just name them. I mean, years ago — we're probably going back a decade ago, maybe a little bit longer — I used to attend this event. Joshua Schachter had an event up at Willow Run in Northern California called Self-Racing Cars. And it was just a bunch of us got together, and you would build cars, and you would run them fully autonomously around a racetrack with a safety driver because that was the requirement of that. But there was one company that was there: Nvidia. This is before Nvidia became Nvidia. They were there saying, 'Hey, we can offer you chip support. We can offer you this.' And then, lo and behold, fast forward today, Nvidia's pretty much in every single robo-taxi and autonomous truck out there. But it started with those local do-it-yourself events, those homebrew events. Could we see a similar thing with ARM where some of these smaller events, where you become the glue, and then you fast forward X amount of years, and all of a sudden, ARM is at the center of the physical AI economy?
Well, so let me start out by saying Nvidia is a very close and wonderful partner of ours. And I love talking to people when they go, 'Oh, we use Nvidia.' And you go, 'You do realize that's ARM?' And they go, 'Oh, no, I didn't realize that was ARM.' So, it's kind of interesting. It's been the secret technology behind some of the most successful stories. But yeah, to answer your question on how this plays out. So, ARM has always done well when our partners do well, right? And that's a key part of our company DNA is the collaborative mindset and sort of the we're enabling and we're underlying technology that allows the most advanced solutions in the world to be built, right? And our take on this physical AI space is these are going to be the most advanced computing solutions. They're going to be some of the most challenging compute solutions. I spent 10 years in the cloud space. In this space, you've got computer vision, you've got sensing, you've got the processing of that vision data, you've got the coalescing of that vision data with real-time models — you've got to have a real-time compute environment, you've got to have safety criticality, you've got to have high-performance compute, but you've got to do it all in a power footprint of less than 100 watts so that you can put it on the back of a robot. There aren't too many spaces in any ecosystem where you have those competing constraints. And the competing constraints — by the way, that's what makes this kind of cool, right? And I think it's the kind of fun part of the physical AI space is your power footprint, you're now trying to figure out what can I get out of it, right? How can I achieve things that people would not expect in that power footprint. And I think by enabling that, ARM will do just fine, right? As a company and as an ecosystem, and that's our goal. Our goal is we build ecosystems to help scale technologies that benefit both those in the ecosystem and, sure, it benefits ARM as well.
There's one common denominator across physical AI. Not everything's the same and it's not a one-size-fits-all solution. Because that's how it's going. Is that one of the benefits of your IP and your architecture, that it can be adopted and adapted for different solutions because it's not just this one solution? There are multiple solutions, and who knows in the future what other solutions or form factors are going to emerge?
That's absolutely the case, right? So, whether you're using, say, a Rockchip chip in your AMR, or you're using an Nvidia processor in your humanoid or your autonomous vehicle, and all of the others in between, and XPST for the little micros that are used in some of the dexterous hands — it's a common architecture, right? It's a common architecture that is pervasive across all of the layers of compute, but also all of the different kinds of robots. And what's useful and helpful about that is it allows the software challenge, which for many people is the challenge they really want to spend time solving, to be fairly consistent, right? So, you can build a solution that scales. If you're a company who's building robots, you don't want to have to switch architectures between your AMR and your quadruped and your humanoid. And you also — so that's a great advantage. You also don't want to have to switch architectures, if you can avoid it, going from the actuation on your humanoid up to the perception and planning and data ingestion side of your robot. So, the ability to go broad across different types of devices as well as go north-south in terms of the technology that's within them — I think that's what Arm enables people to do, and it just makes that problem go away, you know, right? It makes it much easier for them to do that.
I would summarize it in a very simple way: Arm enables companies that use your architecture to scale. I mean, you have some very public customers who are scaling rapidly using your IP and your technology on the physical AI side. If you had to estimate — and this is going to be a broad estimation — how many physical AI devices or iterations out there would you say are running Arm architecture today?
That's a great question. So, I know from last year's data that over 2 billion devices shipped based on Arm last year. So, that's in 2025. Now, if you go back, you start accumulating that over the years, it's obviously a massive number, right? But what's amazing about that is just how different those can be, right? From whether it's a Cortex-M-based micro that's managing the digits of a device, all the way up to the highly advanced compute. So, yeah, it's very pervasive.
If you look at your public filings that you've made with the SEC, you can see the growth numbers. And so, I encourage the audience to go read Arm's SEC filings so you can see all the financial data. You can see the actual numbers of how the company's growing. So, we have your data filed with the SEC, which is showing really remarkable things from a growth perspective. And then recently the company made a public announcement that they're standing up a physical AI unit, which you're heading. What is the company seeing in the market to make a dedicated unit just for physical AI?
Physical AI?
Yeah, so let me just — I'll just clarify one thing because my boss might listen to this. He actually likes this video part too. I'm not heading it. My boss is heading it, but I am running the go-to-market on it. Yeah, so we see that physical AI is on the cusp of scale, right? We see an ecosystem that is ready to transform.
And having worked in the cloud ecosystem, you know, in the last 10 to 15 years, I watched that transform, and Arm was sort of a key catalyst to help that transform, you know, with Amazon and Google and Microsoft and others, Meta adopting our technology. So we see a lot of similarities in this ecosystem in terms of the challenges. And the challenges we think, you know, we can bring solutions to. So that's why we put the physical AI business unit together — was to bring a coherent response from inside Arm to the ecosystem and then bring the ecosystem together. So yeah, we're all in on the physical AI space. Not only is it, you know, do we think it's ready to transform? It's a super exciting technology space. It's a great place to engage with the developers of those solutions, to work with people who are doing creative things, you know, all manner of whether it's robots or autonomous vehicles or autonomous trucking or autonomous, you know, diggers or what have you. It's just — it is, you know, I meet with a lot of people and a lot of companies, but there's a huge amount of intellectual investment going into this physical AI ecosystem right now, and we believe that to unlock that needs a concentration of an ecosystem and focus, which is why we put the physical AI business unit together.
