AVs and Humanoids Are More Alike Than You Think (w/ EVP of Physical AI @ Arm Drew Henry)

**Sophia Tung** (0:10)
Hey there, you're listening to the Ride AI Podcast, where we talk about news and intelligence at the intersection of technology and mobility. I'm your host, Sophia Tung. Now, I know that's a mouthful for an opening, but I find it particularly true today because we have the executive vice president of ARMS, yes, that ARM, new physical AI business unit, Drew Henry, here with us today. Drew previously spent 11 years at NVIDIA as general manager of GeForce, which is still probably the largest GPU brand in the world. Three years at SanDisk, providing the world with storage, and is going on nine years at ARM, leading Cloud Infra and now, physical AI.
That is quite the resume. Drew, welcome to the show.

**Drew Henry** (0:52)
Thank you, Sophia. I'm looking forward to the conversation.

**Sophia Tung** (0:55)
As am I. First of all, what do you, I guess, what does ARM define as physical AI? Because I think that's, I mean, physical AI, that term has been thrown around so much, but what does that actually mean?

**Drew Henry** (1:08)
Yeah, for us, we're trying to make it simple. For us, it's where AI is embodied into a machine, and that machine is then sensing, deciding, acting, safely in the real world, right? So this is, you know, robots and autonomous cars and traditional cars that are becoming much more autonomous, drones, things like that. That's kind of how we think about it.
There's a crazy key metric that we use to define it, which is this latency, the time between when a photon hits a sensor and an actuator in one of these machines actually acts. So that's kind of the way that we think about it in the big market, and then we think about it from an engineering standpoint.

**Sophia Tung** (1:53)
Very interesting. I want to talk about that latency piece, and I think that's pretty important. I mean, you know, I hear Arm talk about the divide between how distributed sensors should be, or if it should all just be done on one compute unit. But we'll talk about that later. But I understand Arm just carved out sort of like physical AI as its own business unit, alongside the cloud and edge compute stuff that Arm also does.
Why did automotive and robotics need to be like one business unit instead of separate units? What's the tell that they're actually the same problem?

**Drew Henry** (2:30)
What's the tell? Yeah, I like the question.
It's funny, we've been in this space for a really long time, a number of decades providing computing technology to transportation systems, robotic systems, and the like, kind of in the traditional sense of those markets. We've been doing this for a long time. Matter of fact, in the last 12 months, we've shipped 2 billion arm devices just into those market segments, just in the last 12 months. So it's big scale for us. And the reason why we kind of combined all that together in one place is because whether you're looking at what Tesla's trying to do with the vehicles that they build, or what BMW or Mercedes-Benz or the China manufacturers are trying to do with theirs, everyone's really moving towards this world of autonomy. And so when you kind of put that up to the highest level, autonomous cars, traditional cars becoming autonomous, more and more and more autonomous, autonomous cars being built, you know, robotaxes like those first principles or robotics platforms, all of this is just now under that umbrella. So it's more that the market has evolved into that definition more so than we've kind of cramming it into just one definition.

**Sophia Tung** (3:46)
Do you think, so we're sort of on the leading edge of hardware that can act autonomously. A large part of it is the hardware, but also a large part of it is the software. And I know you've used the line before that AI models, right, the things, the software that actually drives the intelligence, that actually drives these robots, they're not really the hard part anymore, right? So it's more like what happens when the model leaves the cloud and has to run on device.
Can you sort of explain that line of thinking?

**Drew Henry** (4:24)
Yeah, it's interesting when you kind of go back a little bit in time and go back to the early days of people trying to drive vehicles autonomously, and you go back to the DARPA challenges that existed, and kind of fast forward to where we are today. To a large extent, the autonomous driving problem kind of, from a technical standpoint, is a solved problem.

**Sophia Tung** (4:47)
Yeah, we've got CEOs of all these companies, they're saying, yeah, I believe it was one of the co-CEOs of Waymo who said that, yeah, we're not really thinking about... The driving part itself is solved.