Physical AI: A deep dive into autonomous vehicles and robotics

“The Angle” Music

Cold Open
Lee Sandquist
We will need a lot more actuation, more gears, more vision. And so I'm pretty excited and curious, if we see the datacenter-driven industrial strength kind of prolong into a physical, real-world AI cycle.

Jennifer Martin
Welcome to The Angle from T. Rowe Price, a podcast for curious investors. Just a reminder that outside of the U.S. and Australia, this podcast is for investment professionals only.

I'm your host, Jennifer Martin, a portfolio specialist at T. Rowe Price Associates here in Baltimore, Maryland.

The next frontier AI systems will do more than answer questions or complete digital tasks. They may drive cars, move through warehouses, assist in factories, and eventually work alongside humans in physical environments.

That shift changes the investment equation.

Physical AI depends not only on better models and more compute, but also on sensors, connectivity, safety, and regulation.

Tony Wang and Lee Sandquist are here to help us understand this systems integration story.

Tony manages a technology portfolio and as an analyst he covered semiconductors and autonomous vehicles.

Jennifer Martin
Lee Sandquist is an industrial tech analyst and associate portfolio manager.

Tony and Lee, welcome to The Angle.

Tony Wang
Great. Thanks for having us.

Lee Sandquist
Happy to be here.

Jennifer Martin
Great.

So, Tony, you covered NVIDIA for many years and even shared a crab feast, a Baltimore delicacy, with CEO Jensen Huang on the rooftop deck of our offices. I guess that was last September. I was there too. And Jensen, many people appreciate, often has his finger on the pulse of what's coming. At NVIDIA’s annual AI conference this year, he pronounced: “Physical AI has arrived. Every industrial company will become a robotics company.”

How can physical AI transform traditional automation and robots over the long term?

Tony Wang
Yeah, absolutely. Well, the potential is enormous. So, and I think it's an extension of what's going on in the AI market. If you think about ChatGPT was first, like thinking, and then now we have agents like coordinating and doing work. And then third phase will be AI coming out of the data center and physical AI. So it's going to be acting. And so I think it's definitely really promising. It's early, obviously, but you're seeing a lot of really good green shoots.

I think that there's a ton of potential here, right? When you think about, just high level—take the economy, right? It's like labor times productivity. AI is the productivity and then physical AI is the labor that needs to be uncapped as well. And so I think there's…obviously, the total addressable market is massive. And so huge implications in terms of what's the cost of goods. I think AI and physical AI is going to be really deflationary. So that's good for long-term growth and the cost of goods. And so I'm encouraged by what this could mean.

And, in addition, like in terms of like, how we're going to get there, I think that the models are probably the leading indicator, in terms of what the models can do, and then for the physical AI to really take off, I think that's like AI coming out of the data center. So like the models, what we're seeing is that like right now, they’re reasoning, thinking and the next phase is like time on task, right? So can these models continue to do work for longer than like a few hours? Can they do work for a week or a month? And then, like, can you string them together into a team that can, like, figure out things and work through problems together?

So I think it's going to be really interesting. It's early but definitely something that we're watching here.

Jennifer Martin
All right. It sounds like we're going to have to have a little bit of imagination as we move from agentic AI to physical AI. And so, Lee, let's bring you into the conversation and explore one of the first use cases for physical AI and that’s driverless vehicles, which seem to have finally arrived.

I used to see driverless vehicles, whether it was in Silicon Valley or San Francisco, and I felt like it was sort of, you know, a time warp to the future. Now, just last week, I was in London, and there's Waymos, which I think is actually pretty impressive—those streets are very narrow. What are some of the key lessons from their long road to commercialization?

Lee Sandquist
Autonomous driving has really highlighted the friction that can emerge when you bring an imperfect automated solution or system into situations that require perfection, right?

Since ChatGPT was launched in November of 2022, consumer LLMs have made super-fast progress. We take the functionality for granted, but the speed of progress has been remarkable. In large part, that's because you already had a wealth of data, and I think it's likely that we'll probably exhaust the entirety of internet data and some portion of printed text for pre-training in less than five years, which is crazy to think about.

But beyond data, the safety threshold was also pretty low. These were not life and death situations. And finally, the physical infrastructure and distribution, meaning the internet and mobile apps, were already in place. So if you were an internet company [and] you had the capital and the AI talent, you could train the model, build a viable product, and get it to consumers pretty efficiently.

But as we move into the, into the physical world, the safety bar becomes much, much higher and the number of edge cases grows. Not only do we have a much harder problem, but we also have to build all the tools from scratch.

In many cases, we don't actually have the pre-training data. We don't have the infrastructure. We don't have the supply chain. We don't have the distribution. These will all be solved. But as we saw with autonomous driving, it does take some time. Waymo opened to the public about six years ago, and they only have 3,000 cars on the road today. Tesla launched their robotaxi service in Austin a year ago, and it has 20 to 30 unsupervised vehicles active in that city today.

So, you know, why has that progress been slower than, than most would have expected? Well, one, it's a really hard problem, right?

Think about what's actually happening.

 A Tesla is driving down the road, and eight cameras are on the exterior of the car taking in video input. The video input then flows down into an onboard inference chip and into an end-to-end neural net, where that model is voting real time on what the cameras are seeing and what actions to take. The output is the decision to turn right, left, accelerate, or decelerate. The external environment is constantly changing, and there is zero room for error.

Many, many of the AV companies outside of Tesla also don't even have the pre-training dataset that's required. And finally, there are there are several other real-world points of friction, right? You need state-by-state regulatory approval. You need a charging infrastructure. Consumers need to adopt the service and gain comfort with the, with the technology. And fleet utilization serves as a headwind for unit economics in the early days.

So real barriers exist today. The progress has been slower than most probably would have expected, but I see substantial potential over the next five to 10 years.

Jennifer Martin
Really clear. So both of you highlighting pace of AI innovation has been rapid, but it hits a little bit of a different speed when you move into the physical world. These are not small challenges. What are the similarities and differences in how Tesla and Waymo overcame them? Because you were kind of getting into that. So let's unpack that a little.

Lee Sandquist
Yeah. Big differences in how they're attacking the TAM. But just to level.

Jennifer Martin
And TAM means…

Lee Sandquist
Total addressable market.

Jennifer Martin
We love jargon.

Lee Sandquist
But just to level-set the conversation. This is a big opportunity. Even in the, even just in the US, there are about 3 trillion vehicle miles traveled each year. And in the coming decades, I think we need to consider what percent of those could be autonomous. Is it 10%? Is it 50%?

Let's just say it's one-third penetration. That's a trillion AV, or autonomous vehicle, miles annually just in the U.S. And then the next question is, okay, what do consumers pay for each of those miles?

Let's just keep the math simple and say those, those, those miles are monetized at a dollar per mile. Even at that discount, we're still looking at a USD 1 trillion U.S. autonomous vehicle addressable market.

So it's, it's a big opportunity. You can see why these companies are motivated, but it's interesting to see each company attack the problem slightly differently. So Waymo's, Waymo has been using dozens of sensors, including LiDAR and a hybrid model architecture. So they're using neural networks but also some hard-coded rules of the road. Tesla is only using cameras, and they're only using an end-to -end neural network.

Tesla's approach has, has clear advantages in terms of cost and scalability, if they get it right. But many would argue that the sensor redundancies that Waymo has, they offer much safer outcomes. Today, Waymo is just focused on L4, which means eyes-off, hands-off robotaxis—what you saw in London. Tesla is offering L2 subscriptions, meaning you or I could buy a Tesla tomorrow and, and use their technology. But last year they also released their own L4 robotaxi offering, similar to Waymo.

Jennifer Martin
Clearly bringing driverless vehicles to, you know, to the market has been much harder, a lot slower, as you outlined, and much more capital intensive and with a regulatory overview than many investors expected. And so, Tony, does progress on autonomous vehicles help to accelerate the development and commercial timeline for general-purpose robots, or is robotics a very different problem?

Tony Wang
I mean, there's definitely similarities with the camera vision, right? And the AI models like picking up, like learning from the vision data and all the driving data. But I think it's kind of a different problem. I mean, I think that you're going to have a lot of like the actuators, the sensors, the motors, like there's going to be new bottlenecks and supply chain of like how we build this.

Like, you make a great point, Lee, in terms of like scaling an LLM is a lot easier than scaling 1,000 robots, right? Like they're out in the real world doing things like high risk. So I think it's going to be pretty different.

And one of the, the frameworks that is just like okay, which one do you want, right? Do you want the kind of low-cost, like, camera-only LLM-forward, or do you want the high-cost, not-scalable version that costs $300,000 per car? How are you thinking about that in robotics, right? Like is this same trade-off?

Lee Sandquist
I think not all physical AI is the same. And the competitive landscapes are very different. The competitive moats and differentiation is very, very different. So with Tesla, if you think about it in, in robotaxi or full self-driving, they are, at least a couple of years ago, they were the only company that had the camera-only, end-to-end neural network approach.

They had the lowest-cost hardware, and they were the only company with the dataset available to train their individual models. And that competitive advantage, especially the data, compounds on itself. The more cars you have out on the road, the more data you're collecting, the better the model can be.

Ultimately, as you switch to robotics and you think about those competitive advantages, you know, everyone will be using the lowest-cost hardware, everyone will be using vision-only, end-to-end neural nets.

And perhaps, most importantly, no one has the data. That's the biggest difference. So if you think about just the wealth of text data on the internet today—in manufacturing settings, almost 0% of it has been recorded in a way that can be used to train real-world models.

Jennifer Martin
So Tony, let's peel back the curtain on autonomous vehicles and AI-powered robots and talk about the supporting cast. How does the rise of physical AI affect the infrastructure layer? Where do you see the value accruing, and where might bottlenecks occur?

Tony Wang
Yeah, I think about that as probably the most important question in terms of stock picking and, like, where good investments are. Because you want to go places where they become more mission-critical, there's more content increase. And the, you're going from, like, abundance to scarcity in that framework.

And so, look, I think that AI infrastructure is seeing a lot of bottlenecks right now. And physical AI is going to be an extension of that. So I do think that we're going to need a lot more memory. You know, the HBM trade ratio to commodity DRAM is going up. You know, it was two time; now it's four times, in terms of the number of wafers that will get, you know, eaten up by HBM. And so I think that’s a big theme, continues to be.

And then we're also going to need more storage. Like, think about all these, you know, robots are going to be generating a ton of data. You got to store them. So that'll help, help with like NAND storage as well as HDDs. In addition, I think the networking is going to be really important to make sure that you have AI infrastructure that is passing data back and forth really fast.

So I would say, like, those are, continue to be good themes. In addition, I think that you're going to have a lot of bottlenecks in the manufacturing and, you know, sensor technology that we're going to have to also develop further.

Jennifer Martin
So what I heard is as agentic AI moves into physical systems, it's going to put even more pressure in areas like token consumption, inference intensity—really affecting some of the areas that you mentioned.

Tony Wang
Totally, absolutely.

Jennifer Martin

And the other thing that, I smile, is AI does not create more DRAM. That's, you know, AI can't do that. We have to—that's a physical constraint. And I think that's something that you also highlighted, too.

Tony Wang

Well said.

Jennifer Martin
When you think about the next, the next paradigm—where are the real-world constraints emerging in industrials as the datacenter build-out continues and the physical AI gathers steam? We were just in London at the industrial buy-side conference. You know, these were industrial companies, and all we did was talk about AI and datacenters and electricity.

Lee Sandquist
Yeah. And it's remarkable to see the fundamental improvement and the profit generation out of some of these legacy industrial companies who touch into the datacenter world. And if you think about what's driving that: It's connectivity, it's power, thermal management, it's compute for training, it's compute for inference—all of those things you need in humanoid robotics and physical AI.

You know, the number might be different, but ultimately you need all that. And I think it's important to maybe take a step back and just recognize how hard these humanoid robotic problems are that they're trying to solve. You know, the physics is very different for an autonomous vehicle versus a humanoid robot. In the car, the system is, just, as I mentioned, turning right, left, braking, or accelerating. And that's one decision kind of for the entire system.

Even in just one Optimus—Tesla's robot—even in one Optimus hand, there are 22 degrees of freedom plus three more in the wrist. So that means dozens of separate actuators, all having to make that individual physical decision. And there are also new sensor modalities required in order to assess pressure, force, balance.

And so as I think about what's driving the industrial strength, and what I mentioned, the connectivity, the power, the thermal management—all of that will still be relevant. But we will also need new hardware to enable the physical motion that I just discussed. Right? So we will need a lot more actuation, more gears, more vision. And so I'm pretty excited and curious, if we see the datacenter-driven industrial strength kind of prolong into a physical, real-world AI cycle.

Jennifer Martin
Good. I love the curiosity, so.

Tony Wang
Hey, can I ask Lee one question, actually?

Jennifer Martin
 Yes. Of course.

Tony Wang
Something, kind of—you know, I'm just like on Instagram, and I see all these robots in China doing kung fu and, like, you know, it's like, whoa! Like, you know, that seems like they're pretty much there. Is that just a parlor trick like, you know, is it like, what are they doing? Are they ahead of the U.S. robotics? Or is it just like they're just hard-coded program versus we're trying to solve a grander challenge here?

Lee Sandquist
So there are hundreds of AI-first robotics companies, many of them humanoid-focused, in China today. I think it's difficult to paint a, you know, one brush for all of them. But I would say it's important to remember that China is just the global leader in manufacturing. The supply chains over there, the work ethic is incredible. And the manufacturing expertise is, is, is top tier.

Each separate region is also very incentivized on a local level to start these companies and become domain experts. I would say that there is a big difference in in terms of model quality and…but what I think that they are doing well is not perfecting the humanoid form factor. In the U.S., Tesla [and] Figure are very focused, for better or for worse, on the humanoid form factor.

My concern is that they will ultimately spend a lot of time perfecting that form factor, and then you fall behind in the physical deployment, data collection, and model improvement flywheel that I think is super important. And so, in China, if you're deploying all of these even slightly subpar robots into the field, you're still collecting a ton of valuable data, and you can use that to improve the models.

And so we'll see if the patient, kind of perfect approach in the U.S. works longer term. But my concern is that if no one has a data advantage and everyone's using similar model architect approaches, does this ultimately become a hardware and manufacturing challenge? And if so, is China just better equipped and well ahead of the Western companies?

Tony Wang
I think that's a great point. Like when I was covering Tesla—and I think we were talking about this—it’s like, oh my God, they just ramped the Shanghai fab like to crazy numbers so fast. And then they must be able to do it in Berlin and like [the] U.S. And it was, it was partially, you know, Shanghai and China that supply chain is so robust—[they] learned so much from like the Apple ecosystem and building autos, right? And so I think that's a great point. It is a manufacturing scaling thing that like you know that's a big advantage. So, appreciate that.

Jennifer Martin
And that's a good way to end because I think AI—we talked about it in a previous episode—where there is a Nash equilibrium between sovereigns. You know, every country—U.S. [and] China, [for] example—needs to maintain spending on AI. And I think what you just made the case for, Lee, is it also extends into the physical world. And then, obviously, the reshoring efforts of U.S. highlighting that necessity to have some manufacturing prowess, too, if we are going into physical AI. So great, great point to end on.

Jennifer Martin

And let’s end on a fun note. What's your favorite robot, fictional or real, and what would you have it do for you?

Tony Wang
Oh boy. Probably my laundry. You know, like I would say that first, but I haven't really thought about my favorite robot. The only one that comes to mind is Terminator, which is not my favorite robot.

Jennifer Martin
It's a scary robot.

Tony Wang

Yeah.

Lee Sandquist
So you can buy Unitree G1 on Amazon. And so, I'm making a push internally to get one for the office. And it could go grab us coffees and go to meetings, take notes. But we'll see if I get that past approval.

Jennifer Martin
As a parent of teenagers, I just need someone to clean things off the floor. I don’t know what that is…maybe it’s a Unitree robot, I don’t know, but that would be very exciting.

So there you have it—thank you very much, Tony and Lee, for your time today.

Lee Sandquist
Thank you so much.

Tony Wang

Great. Thanks so much.

Jennifer Martin
If I had to summarize our conversation today, I would say that physical AI has enormous potential, but it involves safety, regulation, and human trust. Those constraints make it different from software-only AI adoption. Again, I'm Jennifer Martin. Thank you for listening to The Angle. We look forward to your company on future episodes. You can find more information about this and other topics on our website.

Please rate and subscribe wherever you get your podcasts. The Angle: better questions, better insights—only from T. Rowe Price.

DISCLOSURE

This podcast episode was recorded in June of 2026 and is for general information and educational purposes only. Outside of the United States and Australia it is for investment professional use only. It is not intended to be used by persons and jurisdictions which prohibit or restrict distribution of the material. This podcast does not give advice or recommendations of any nature or constitute an offer or solicitation to buy or sell any security in any jurisdiction. Prospective investors should seek independent legal, financial, and tax advice before making any investment decision. Past performance is not a guarantee or a reliable indicator of future performance. All investments are subject to risk, including the possible loss of principal. Discussions relating to specific securities are informational only and are not recommendations and may or may not have been held in any T. Rowe Price portfolio.

There should be no assumptions that the securities were or will be profitable. T. Rowe Price is not affiliated with any companies discussed. The views contained herein are of the speakers as of the date of the recording and are subject to change without notice. These views may differ from those of other T. Rowe Price associates and/or affiliates. Information is from sources deemed reliable but not guaranteed.

Please visit http://www.troweprice.com/theanglepodcast for full global issuer disclosures. This podcast is copyright by T. Rowe Price, 2026.

CON01525641

202607-5642534