AI, R2 and the Future of Everyday Driving | Rivian CEO RJ Scaringe

By 2030, it’ll be inconceivable to buy a car and not expect it to drive itself. Every single one of our cars, we want to have the ability to operate at very high levels of autonomy.

Radars are extremely cheap. LiDARs are very cheap, but the really expensive part of the system is actually the onboard inference.

My view is EV adoption in the United States is a reflection of the lack of choice. As consumers, we need lots of choices. We need to have variety. We self-identify with the thing we drive. The world doesn’t need another Model Y. The world needs another choice.

Rivian is already an incredibly cool company. How did you decide it was going to become an autonomy company? When did that happen?

From the beginning, we thought of it as a transportation and mobility company. In fact, even before Rivian became Rivian, when I was thinking about what the first products would be, it was unclear what kind of car it would be—or even if it was a car—but it was always clear we wanted to be at the front edge of helping to redefine what it means to have access to personal transportation.

Autonomy has always been part of the strategy, but it’s now fully coming to life with the technology that we’re building.

And when you think about the function of Rivian, there’s transportation, and there’s also the experience. How long ago did you guys start investing in the autonomy strategy here?

We launched R1 at the very end of 2021.

Mm-hm.

We used what I’ll broadly characterize as a 1.0 approach to autonomy. We had a perception platform. We used a third-party, front-facing camera that was essentially a third-party solution, which then plugged into an overall framework that we built, but it was all rules-based.

The camera fed a rules-based planner. The planner would then make a bunch of decisions around the feeds from the perception. The moment we launched, we knew it was the wrong approach, but it was the thing we had started working on well before the launch.

At the end of 2021, beginning of 2022, we made the decision to completely reset the platform.

Was that hard as a decision?

No, because it was so clear when we made it. When you’re building something like this, you recognize you’re going to spend many, many billions of dollars creating it. We knew that at the core of transportation is driving, and at the core of that is a shift to having the vehicle be capable of driving itself.

We made the decision to redo it as a clean sheet, with no legacy of what we had built in Gen 1. That first launched from a hardware point of view in the middle of 2024. That was with our Gen 2 vehicles.

There wasn’t a single line of shared code or a single piece of common hardware on the perception or compute side. Then we had to build the actual data flywheel. We had to grow the car park to build enough of a data flywheel to then start to train the model.

What we showed at our Autonomy Day late last year, in late 2025, was the beginning of a series of really exciting steps in how this is going to grow and expand.

I say this all the time: I think, not just for Rivian but for the auto industry in general, the last 3 years compared to the next 3 years are going to look very different. The rate of progress that we saw in autonomy between, let’s say, 2021 and 2025, and what we’re going to see between today and, let’s say, 2029 or 2030, are completely different slopes.

That really comes back to entirely new architectures now being used to develop self-driving—truly AI architectures. Before, these were not AI architectures in the true sense. They were using machine vision, but really rules-based environments that we defined as humans. We codified them, which is very different from how AI works today.

You might actually have perfect timing here. I got to be part of investing in the first wave of independent autonomy bets that were working with the OEMs at my last investing firm.

I would say 8 to 10 years ago.

Yeah. As you mentioned, there have been several architectural revolutions since then. For companies to make that shift from separate perception and planning systems to more end-to-end neural networks, I asked because I felt it was actually quite a hard decision for people when choosing their partners, from a technical perspective.

I think you can see it. If you go back to the very beginning of the idea of self-driving, a lot of effort and a lot of spend happened for companies to build these rules-based environments and more classic systems.

When transformer-based encoding came along just a couple of years ago, it shifted very rapidly. It was clear that the future state was going to be neural-network-based.

It was hard because if you’re a company that built all these systems, you had to ask, “Do I keep investing in what I had? What do I do with all this work that was built before?” The reality is that a lot of it—the vast majority of it—is going to be pure throwaway. It wasn’t a gradual shift. It was a complete rethink of how things are architected.

How did you decide that this was going to be an in-house effort versus a partner effort, given that most people who made cars said, “We’re going to partner or buy something here”?

On things that are really important, we’ve taken the approach of vertically integrating them. So, electronics, our software, and all the high-voltage systems in the vehicle—things like motors, inverters, and all the power electronics—are things we develop and build in-house.

In a few cases, we had to start with something that was either off-the-shelf or partially off-the-shelf. But today, all of that is completely in-house.

In the case of self-driving, we knew that long-term it needed to be something that was developed internally. We started, as I said, with a Mobileye-centric solution, which a lot of folks did.

Particularly in that 2015 to 2021 time frame.

When you really look at what’s necessary to be successful in a neural-network-based approach, there’s a core set of ingredients that very few people have, and I think we uniquely have them.

First and foremost, you need to have complete control of a perception platform. You need to have control of everything that the system is capable of observing, whether that’s cameras, radars, LiDAR, or some combination of all 3. There can’t be an intermediary company processing some of the information. That’s powerful because you can then feed raw signals into your system.

The system needs to be capable of triggering unique, interesting, or noteworthy events that you can then use to train. Those triggered moments need to be captured and saved on the vehicle, and then, when the time arises—when you have Wi-Fi, ideally—sent up.

The reason I say Wi-Fi is that this is a large amount of data. You could, of course, do it over LTE, but it’s expensive. You have to have a really robust data architecture on the vehicle, and then you need to be able to send it off-board and use that with a lot of training and a lot of GPUs to train a model.

Companies that are either developing independent solutions or are not a car company typically don’t have access to the type of mileage that we do—the huge amount of data that our vehicles generate. If you’re developing this from a sensor-set point of view, you typically don’t have the vehicle architecture and the vehicle car park.

We just came to the view that we have all these ingredients to do it really well.

It’s not an optional thing. The companies that do this well will exist. The companies that don’t do this well—I feel really strongly about this—they will not exist. They will shrink to nothing, asymptotically approaching zero.

Do you think it can only be delivered in a really vertically integrated way?

No. I think there are more than 1 and fewer than 5 companies outside of China that have the necessary ingredients to do this: the capital, the GPUs, and the car park with enough vehicles to generate enough data.

I’d say more than 1, fewer than 3, maybe 4. There’s a very small number of companies that can do this.

I think the unique spot we’re in right now is that the 1.0—

Can I ask explicitly, then? It’s you, it’s Tesla, it’s Waymo. Is that the 3?

I would include all 3 of those. There are maybe 1 or 2 others in the mix.

The challenge is that you have to look not just at the moment in time for performance, where we are today. Do you have the ingredients to continue making progress at a very high rate over the next 4 or 5 years?

A lot of the solutions that are more 1.0-based and are stuck in that framework have, truly, a 0% chance of progressing to be competitive with a neural-network-based approach.

The neural-network-based approach does take a lot. You have to build a ton of inference on the vehicle. You have to either buy it or build it. We decided to build it, so we built an in-house chip to do this. You need to have a car park this large—

You just mean enough onboard compute to actually run the models in the car?

Of course, NVIDIA makes those. But you need to be able to do that at scale and have it in every car. And so, we took the decision to make our chip in-house.

Is that more a capability decision or a cost decision?

It's a cost. We want to have it on everything. Every single one of our cars, we want to have the ability for it to operate at very high levels of autonomy. And so, we design, spec, and build the cameras.

Radars are extremely cheap. LiDARs are now very, very cheap. And so that's an order of magnitude more expensive than any of the perception stack. I think people focus on the perception because it's the thing we can visualize, right?

But the brain is actually the most expensive part. And so we brought that in-house as a way to remove cost from the system so that we can easily deploy this on every car.

You are taking a sort of step-by-step approach to levels of autonomy. At Rivian, how do you think about how quickly you approach Level 4, or the safety case around each of these things? How fast does your team go at this?

Yeah, this question is unique because just a few years ago—2019, 2021 even—there were very clearly delineated ways to approach autonomy. There was a Level 2 approach, which was camera-heavy, maybe with a few radars, and then there was a Level 4 approach, which of course had cameras but had a lot of LiDARs. It was sort of inconceivable to think of the Level 2 system becoming a Level 4, and similarly, the Level 4 system was way overbuilt to even conceivably think about putting that on every consumer vehicle.

Well, you didn't want all these parts—the tens of thousands of dollars of perception.

What's happened is those 2 worlds have just started to very clearly merge, where the delineation between a Level 2, a Level 3, and a Level 4, in terms of perception and in terms of compute, has started to fade. It's now essentially just about how capable the system is at addressing all these corner cases.

This is what's hard for a consumer to recognize. If you're driving a Level 2 system, a Level 3 system, or a Level 4 system, for 99.9999% of the time—like 3 or 4 nines—it feels identical, right? The difference is the 5th, 6th, or 7th nine. Those are these extreme corner cases.

I think it's actually led to a lot of confusion, where you'll be in a Level 2 system—the car could drive itself—and you're like, "Yes, it can," under most road conditions, except these very unique corner cases. And so, to your point on safety cases, the question then becomes: How confident are we in the system's capability in covering these really obscure, unlikely, rare events? Of course, if they're not covered well, it can lead to a really terrible outcome—the vehicle in a bad collision.

That's where the neural-net-based approach has just changed things a lot. The capabilities are so much stronger, and the ability now, I think, for us to deploy on a lot more vehicles and have a very large car parc is significant. We went from, a few years ago, when the state of the art was a test-and-development fleet of maybe a few hundred vehicles, maybe high hundreds of vehicles, to now thousands and thousands. Every single car on the road is part of your data fleet that's identifying these unique corner cases and then running the model against them to test.

Now, of course, we're simulating those unique cases, and we can do a lot there. The whole nature of it has changed so dramatically that I think by 2030 it'll be inconceivable to buy a car and not expect it to drive itself. Maybe that's sooner. Maybe we hope it's sooner; we're targeting a little sooner than that. But certainly, in the very, very near future, that will become a must-have in a car.

It's hard to imagine buying a car today without airbags or buying a car today without air conditioning. These things, at a moment in time, were optional. I think in not too much time—a couple of years—it'll be hard to conceive of buying a car that can't drop you at the airport or pick up your kids from school.

I would argue that right now, most of the biggest carmakers do not have the ingredients that you described to make this a reality. So do you think that's going to play out in the market where autonomy will be so important as a driving feature, a core feature of the car, that there's just going to be a big market-share shift to those who can figure it out? I know you're biased here, but—

No, no, no. I think it's a hard question to answer. I always characterize it like this: I think it's inconceivable for a car company to continue to operate at scale, like mass-market. Very niche enthusiast realms, sure, but at scale, without a software-defined architecture. Even before you get to autonomy, just: Can you do OTAs? Do you have control of a—

Sorry, can you define software-defined architecture?

Yeah, that's before we even get to autonomous—these are basics. The way car electronic systems have been designed and built, and have evolved—with the exception of Tesla and Rivian—every car on the road has what's called a domain-based architecture. You could also call it a function-based architecture.

All the functions across the vehicle—let's say chassis control, door-system control, infotainment, your air-conditioning system—all have little computers associated with them, right?

What we call ECUs, electronic control units.

In a modern car, you might have 100 to 150 of these. Each of these runs its own little island of software. That little island of software is written by a supplier, more likely a supplier to the supplier. So you go to a Tier 1, and they hire a Tier 2 who writes the codebase to run your HVAC.

That's why it's impossible to debug a software system, and it's also why it's really hard to do an update. Imagine you have 100 different islands of software written by 100 different teams that all have to coordinate. If you want a feature—something that manifests as a feature often involves combining functions from different domains.

A simple one to visualize is when you walk up to your car to get into it: You want it to automatically unlock. You want the HVAC to go to your preset. You want your seats to adjust. You want it to make an audible noise on the outside. You want the lights to do something. You probably want the audio system to do something.

Those are all different little ECUs in a traditional car. The coordination cost is really high. It's very unlikely that a car company will make a change to that sequence because it involves coordinating amongst maybe 10 different players.

In contrast, on an approach where you build a zonal architecture, where you have a very small number of computers—ideally 1, 2, maybe 3, depending on the size of the car—that are running 1 operating system that controls everything, it's very easy. You could make updates to that sequence in a matter of minutes, maybe an hour. You could change the whole sequence of what happens when you walk up to the car, issue an over-the-air update, and it's very straightforward.

How often does Rivian update?

We do about 1 a month. Typically, we add a couple of new features and refinements to existing features. We're listening to what customers are saying and asking for, but every month the car gets notably better. It's created this really amazing dynamic where customers are excited for the update. They're like, "When's the next OTA going to drop?"

The irony of all this is these domain-based architectures. How did we arrive at this? It actually goes back to fuel-injection systems. Up until the early 1960s, every car on the road was completely analog. There were no computers at all in the cars—100% analog—and the first computers were there to drive the fuel-injection systems.

Car companies said, "This isn't a core competency. Let's push that little computer to run the fuel-injection system to a supplier, and the supplier will make that." This is where you saw things like the Bosch fuel-injection systems and whatnot. It's sort of like a field of weeds.

Then over the next 60 or 70 years, everything that became computer-controlled to any degree suddenly started to have a little ECU, a little computer associated with it. It just grew into this absolute disastrous mess that is, today, the network architecture that's in truly every car on the road, with the exception of 2 companies.

What I just described is what underpins a large software-licensing deal. We did a $5.8 billion deal with Volkswagen Group, the 2nd-largest car company in the world, to essentially leverage our network architecture and ECU topology for all their various brands.

So it's an interesting final point there on your first question, which is: What happens to market share? I think it's inconceivable that a car company, if it's to be at scale, doesn't have a software-defined architecture that allows its features to become better and better, particularly thinking about how AI starts to integrate into the features. That's number 1.

Secondly, it's inconceivable to think about a car company existing at scale without the vehicles having very high levels of autonomy.

And so car companies have a choice on both of those. They can either accept that they're going to shrink. That's choice 1. Choice 2 is to go build it themselves, which is really hard because they don't typically have these skill sets. They're not software and electronics companies in terms of their organizational DNA. Or they can find a 3rd party to source it from. In both cases, there aren't great 3rd parties to go to.

In the case of autonomy, most of the 3rd parties that did emerge over the last 10 to 15 years tend to be very much classic, rules-based, what we call ADAS or autonomous vehicle 1.0 solutions. Those work pretty well for the business construct of selling a sensor and a function. But that structure is really flawed when you want to have a large data flywheel that's constantly learning and evolving, and you're issuing updates constantly. It's really hard to imagine that with an arms-length transaction. I think the vertically integrated stacks are naturally going to have some big advantages.

So this might be an irrelevant question, but I'm curious. Do you think the autonomy models developed by maybe 3, maybe 1, maybe 5 companies are fundamentally different over time? Because I spent a lot of time in the AI ecosystem, and the—let's say—the language-oriented foundation models feel like they're converging at this moment in time.

I look at a Rivian and I'm like, I don't know. People adventure in that thing. Do you actually want it to do different things, have different styles or capabilities, or is it really just as much autonomy as possible, with safety as the case?

Well, first, this is a great question. I want my car to drive.

In the LLM world, a lot of it has converged because the training datasets are nearly the same. We're taking the breadth of knowledge that's contained on the internet and training models off of that. In the case of driving a vehicle, there is no internet of driving data. So you need both a robust sensor set to be able to capture the data, and you need a car parc that has enough vehicles in it. Of course, Tesla has the largest car parc of vehicles by far.

Our approach to this is that we have a higher level of capability on our perception stacks. We have better cameras, we have radar, and of course, with R2, we'll have LiDAR as well. A huge part of that strategy is that those not only cover corner cases better, but the cameras also have incredible low-light and bright-light performance. The dynamic range of the cameras is stronger. We have more cameras and a lot more megapixels. We have radar, which is great for object detection, and LiDAR, which is a very powerful tool for training the models.

Imagine that 800 feet in front of us there's a little speck in a camera. It's hard to figure out what that is. Historically, what we would do to train that is have a LiDAR sitting on the vehicle, on a ground-truth fleet, to help train the cameras. Putting that on every single one of our cars turns our entire fleet into this amazing training platform, this data-acquisition machine.

That was a core part of how we thought about our strategy. We're going to go not as heavy as, let's say, Waymo on perception, but heavier than, let's say, Tesla, to build a really robust data platform on a vehicle-by-vehicle basis, and then with a car parc that's going to grow significantly with the expansion of R2.

Yeah. So I think, first and foremost, there is no common internet data. The datasets that we're going to be picking up, though, are going to be very similar.

But you have to go acquire it.

But there are still different decisions about what data you care about acquiring.

Yeah.

Well, I think this is what—how does a car feel? Ultimately, it needs to be safe, and the differences in the way it drives or feels are going to be more about what's the UI, the user interface, of it. Even you've just updated some of your features. You have 3 settings for how the vehicle drives: mild, medium, and spicy.

Spicy is the highest one. Yeah. This is a little bit more aggressive over time, and we've spent time thinking about this. I think this will start to become part of a key decision: How does the vehicle behave? There's work we're doing to think about how the vehicle can behave in a way that, against a set of heuristics,

drives like you.

So overall, the model is trained on how to perform in a safe way, but it actually learns some of your driving preferences and creates a model around you. Of course, in a world where you never drive the car because it's always driving for you, there's a way for you to set preferences: "I'd like it to aggressively change lanes. I'd like it to reside in the right-hand lane." Those kinds of decisions are less around the technology and more about what's the product, or the UI, if you like.

Right. The ability to collect those preferences.

Yeah. Preference-based. And I think we will see that—

And that'll be a decision that Tesla makes that may be different from how Rivian makes it. It's hard to say today.

Can we talk about what the R2 means for the company and some of the key design decisions here? I was just talking to Jonathan, one of your lead designers, about the constraints and aiming for more of a mass market and more volume here.

You said it. R1 is a flagship product. Its average selling price is around $90,000. The R1S is the best-selling premium electric SUV in the country. That's electric SUVs over $70,000, and we're the best-selling premium SUV, electric or non-electric, in the state of California.

So it sells really well. It outsells everything in its class, like the Tesla Model X, by about 2 to 1. But because of the price, it's limiting in terms of how much volume we can achieve with that platform.

R2 is our first truly mass-market product, with pricing that, as we've said, is going to start at $45,000. That allows people in the average price range for a new car in the United States, which is $50,000—in that $45,000 to $55,000 price range—to have a really great choice. To date, there haven't been a lot of great choices there.

I'd say there's a singular set of great choices with the Model 3 and Model Y. Of course, that's shown through the extreme market-share capture of roughly 50%. Market share goes up or down, but around that—call it half—the EV market is the Model 3 or Model Y.

There's such an untapped opportunity to pull customers out of ICE vehicles, out of internal-combustion vehicles, with a choice that has characteristics that are different and unique relative to a Tesla.

These are too substantive to be rapid-fire questions, but they're important for me to ask you. Do Americans want EVs? Why haven't they adopted them faster?

What? Yeah, I think, to the last question, causality is always hard to really understand, but let's zoom out here.

The overall adoption rate in the United States of EVs is around 8%. The vast majority of vehicle buyers are buying vehicles that are under $70,000, with the average sale price at about $50,000. If you look at the number of vehicle choices you have at a price point under $70,000—depending on the year, because this changes year to year—there are well in excess of 300 different vehicle-model-line choices. That's putting aside trims and performance packages, but just in terms of overall vehicle types.

You can buy hatchbacks, minivans, SUVs, 2-seaters, convertibles. There's a whole array of different things you can buy. In the EV space, I think there's more than 1 and fewer than 3 great choices. I'd say Tesla, with the Model 3 and Model Y, is absolutely 1 of those. But there are so few choices that if you're looking for a form factor that's not a Tesla—

So you think it's just a missing product set that people want? Yeah.

An extreme lack of choice, is how you put it. A shocking lack of choice. And this gets into interesting corporate psychology, but because of the success of the Model Y in particular, the EV choices that do exist outside of Tesla are often very similar to a Model Y.

So if you were to draw an outline, if you looked at the side-view profile of a lot of its alternatives and put that profile next to a Model Y—

They're all basically the same. It's like, if you want a Model Y, buy a Model Y versus getting—

You want something different.

Yes. You have all these companies trying to create their own version of the Model Y. It's unfortunate because they didn't say, "What can we do that's unique and different?"

For us, we think the Model Y is a great car. I've owned one. Many folks on our team have owned one. But the world doesn't need another Model Y. The world needs another choice.

I think this is a reframing of just how we look at transportation. It's such a big space. It's such an area of personal expression that, as consumers, we need lots of choices. We need to have variety. We self-identify with the thing we drive. We just haven't had it.

I think EV adoption in the United States is a reflection of the lack of choice. There's 1 set of really great choices with the Model 3 and Model Y.

I think there needs to be many more. Even looking at our partnership with Volkswagen Group, a big motivator for that, which ties to our mission, was: can we take our technology platform and allow that to be expressed through a variety of really interesting and very storied brands, different form factors, different price points, and, of course, different segments?

I think the more choices we have, the more it's going to lead to broader-based adoption of electric vehicles, which creates, I think, a very positive level of momentum around the space. It's worth noting that, when we look at how we develop a car like R2, we don't think of it as, "This is someone who's going to buy an EV; let's make it good." We think of it as, "Let's make the best possible vehicle we can imagine."

So incredible performance, great range, great dynamics, tons of storage. The person buying it will be drawn into electrification because the car is just the best choice they have.

And we took that same view with R1. On R1, the vast majority of our customers are first-time EV owners, and their first EV is a Rivian, which is really good. If all we were doing was moving customers between 1 or 2 brands, it wouldn't be accomplishing the goal. We have to create new EV customers with products that are so compelling that they just draw people in.

So that leads into my very last question here. I grew up thinking a car is a huge part of my identity. I love cars. I drew them. I still think they're pretty cool. As they become more like utilitarian services, with the rise of robotaxis as a concept serving some of the functions your car served before, how do you think our relationship with cars, or vehicles, changes over time?

I do think we're going to see a shift. It's an interesting philosophical question: why are cars such a part of our society, and why do we have this affinity for them in a way that we don't have that feeling for other things in our life that are really important? I don't look at my refrigerator and think, "I really love that," in the same way that I do with a car.

Part of it is that a car enables personal freedom. It allows you to explore. It's something that you not only ride in, but it becomes part of an expression of self. I think that's probably going to continue to some degree, but it is going to evolve.

The way we look at it with our products, and even how we've laid out and contemplated the purpose of the brand, is that the vehicles and products we make need to both enable people to go do the kinds of things that they would hope to have memories of years to come. We often say, the kinds of things you'd want to take photographs of.

But more than just enabling it—which is a functional requirement, like: can it drive there? Can it fit the stuff—your pets, your gear, your friends, all of your stuff?—more than just enabling it, can it inspire it? Can the brand, the way we present what we're building, and the way we make design decisions inspire you to go do the things you want to remember for years to come?

There are little design decisions we take that link to that. So a flashlight in the door

is an invitation to explore. It's an invitation to go look at things at night,

or the treehouse.

Yeah, exactly. There are all these little decisions you made throughout the whole car that are just designed to engage that element of inspiring people to imagine the life they want to have.

Awesome. Thank you so much, RJ. Congrats on the R2 and on the autonomy program.

Thank you.