Listen to this post:

Good morning,

Today’s Stratechery Interview is with Rivian founder and CEO RJ Scaringe. Last week Rivian held their Autonomy and AI Day, where the company unveiled its plans for a fully integrated approach to self-driving. Rivian is building everything from its own chips to its own sensors — including video, LiDAR, and radar — and if all goes well, the company will supply a multitude of companies, particularly Volkswagen.

In this interview we cover all aspects of Rivian, including the long path to starting the company, production challenges, and why partnerships with Amazon and Volkswagen are so important, and point to relationships in the future. We also dive into autonomy, and why Rivian is taking a different path than Tesla, plus I ask why CarPlay isn’t available on Rivian vehicles, and what that reveals about their nature.

As an aside to podcast listeners: due to a mind-boggling mistake by me, the first 20 minutes of this podcast are considerably lower audio quality. I forgot to hit ‘Record’, so the segment that remains is what the Rivian PR represenative captured on her phone. I’m incredible grateful for the save.

As a reminder, all Stratechery content, including interviews, is available as a podcast; click the link at the top of this email to add Stratechery to your podcast player.

On to the Interview:

An Interview with Rivian CEO RJ Scaringe About Building a Car Company and Autonomy

This interview is lightly edited for content and clarity.

Car Company Dreams

RJ Scaringe, welcome to Stratechery.

RJS: Happy to be here.

We are here to talk about Rivian and your recent Autonomy and AI Day. Before we get to that, however, I want to learn more about you and your background, and how you ended up sitting with me today. You were, as I understand it, into cars at a very early age.

RJS: I’ve been around cars as long as I could remember. As a kid I was restoring and working on cars. I spent time in restoration shops helping and slowly learning how to do more than “help”, but actually really help. And then around the age of, I guess 10-ish, decided I wanted to start a car company.

Oh, okay. So there was no like, “Oh, I got a computer and started typing out BASIC”, this is straight cars all the way.

RJS: Yeah, I knew I wanted to start a car company. And at that point when you’re a kid, you have no idea what it entails, you have no idea what the business is going to be, but I just knew that it was something I wanted to do and I sort of started charting out my future path with that as the end state goal, with that as the context. So I went and worked as a machinist, I ultimately went to school for engineering, I did a degree in mechanical engineering, then I went and did a master’s and a PhD focused on automotive. And then the day after I finished my PhD, I officially started Rivian.

So why did you think it was necessary to do that level of education? Not just a bachelor’s or not just gain experience, but to go all the way through to the PhD?

RJS: It was actually pretty intentional. I knew that to start a car company, I was intellectually honest with myself that it would take a lot of money and I knew that I didn’t have any money. So that meant for me to do this and be successful, I would need to get other people to invest money into the idea and typically in the tech space, you could start something with not a lot of capital that you can make a very crude version of your first product…

Because it’s software not hardware

RSJ: Exactly, and I also knew that I didn’t want to go work for 25 or 30 years to accumulate experience that would make me credible. So I was like, “What’s the fastest path to credibility?”, and my thought was it would be a PhD. I said, “If I get a PhD from a top school” — I went to MIT — I thought that would be some earned credibility that would make it more likely that investors would want to get into the company.

I didn’t grow up around venture capital, I had no idea around any of these things, that was my hypothesis. And amazingly, it proved to be a key element in Rivian’s journey, because one of our early investors, one of our earliest large investors, I should say, was someone that I was introduced to through MIT and was an alumni of the school and I was connected with them through the provost, and that was ultimately what led to some of the really critical early capital into the business.

So is it the end state that the PhD was totally worth it, but the actual academics was completely incidental to this introduction?

RJS: Yeah, and I think as is the case I think with all higher education, the biggest takeaway is to learn how to learn, and to learn how to solve complex problems. I think undergrad, you’re learning how to learn. Graduate school, particularly for technical degrees, you identify a problem and you work really hard to solve that problem, and you have broad responsibility and broad scope on doing that activity, and you build confidence and you build skillsets around problem solving. But the problems are going to change in the course of a life or in the course of your career, the things I was working on 20 years ago have nothing to do with Rivian whatsoever.

Well, I’m actually kind of curious about that. What were the things that you were working on 20 years ago, and why aren’t they applicable?

RJS: Well, in the case of automotive research, in 2005 the work that was funded, which is the kind of work that you do as a PhD student, so you get sponsored by companies or by grants, was to look at making engines, internal combustion engines more efficient, that was primarily the focus. And so I was working on something called a homogeneous charge compression ignition engine, which is a different type of combustion. We’d compression ignite a pre-mixed fuel air mixture, very hard to get-

Like diesel?

RJS: It’s like diesel efficiency with gasoline-like cleanliness is the idea. Obviously, it’s not a technology that has any runway and makes any sense in the future.

(laughing) I’m hearing about if for the first time right now.

RJS: Yeah, so it was an interesting project. It was really a study in software controls, because that was the challenge of this project. But I didn’t take a single piece of that and use it in starting Rivian.

Now, that’s different, some folks turn their PhD into the foundation of a business and start a business off the back of it, I had the benefit of just being in the automotive lab, I had the benefit of working closely with car companies. Big, large car companies were funding a lot of the work and it further solidified my view that I didn’t want to go work at one of those companies, and I thought the likelihood of me learning the necessary skills is lower working at one of those places than me learning by doing, by just going and starting a company as a 26-year-old PhD graduate.

Founding Rivian

Right. So if you start out with cars as a child and you’re coming all the way up, at what point did you know that the car — you wanted to start a car company, at what point did you know it was going to be electric and not internal combustion?

RJS: That was far less clear. I wanted to make cars very efficient and I wanted to design cars that would essentially help define what the future of state would look like. But when you’re 10, you have no idea what that means. When you’re 20 — and at this point, this was early 2000s, it still wasn’t that clear — and so it didn’t really become clear until I started the business.

Even before starting the business, one of the concepts that was competing for what Rivian ultimately would become was this idea I had for a pedal-powered car, which at the time I was thinking could be a hybrid-electric, except the hybrid, it was human-plus-electric drive and amazingly, full circle, that happens with e-bikes. E-bike is the most popular electric—

Oh right, yeah.

RJS: But 25 years ago, 20-plus years ago, that wasn’t clear that e-bikes were going to be an explosive success as they’ve been. And then I created within Rivian, a skunkworks team that’s now spun out into a new company to actually focus on this pedal and pedal-hybrid electric vehicles. So we have a quadricycle we’re doing with Amazon as a first big customer, but the name of this company is Also, and the idea of this spin-out from Rivian is that if you want to electrify the world, you need to electrify vehicles, but you also need to electrify everything else, and so Also is doing everything else.

So Tesla started in 2003. Was there any inspiration or connection there, or is it just incidental that it ended up being kind of around the same time?

RJS: Yeah, so they started in ’03, I started in ’09. Of course I was aware of Tesla, but Tesla launched their first car, the Roadster, before I even started Rivian and so they launched the Roadster and then they were working on the Model S, which it doesn’t get talked about a lot, but there was a time when the Model S was considering using a series hybrid architecture as well. Ultimately, they went pure EV but that was in like 2008, 2009, and I started the company just as, my view is there’s going to need to be a lot of successful choices, and I’d been on that mission for a while — what I didn’t expect is just the process of raising capital is really hard.

So is that actually where they did help you a lot, just because eventually once they got over the hump and it was a successful venture, did that make it easier for you in the long run?

RJS: I think so, and I think Tesla was the existence proof that I’d say more than raising capital, what Tesla did is they showed that electric cars could be cool.

Right.

RJS: And they did that with the Roadster. So they launched this Roadster, they took a Lotus Elise, they re-engineered it, they made it electric. It was super fast, it was really cool, this was way before anybody had thought about electric vehicles as something that could be fun or fast accelerating. Now it’s hard to believe that 20 years ago this was the case, but at the time they really took electric cars from this perspective like golf carts, to like, “Oh, this can be a highly capable performance machine”, and that just shifted mindset, and that was important.

So you start Rivian in 2009, I believe the first vehicle comes out in 2021. That’s a long time period, what was going on for those 12 years? Are these painful memories?

RJS: No, no, no, no, they’re useful memories. In the beginning you have no capital, so you can’t realistically make progress on building something like a car unless you have some level of capital. So if you’re spending $1 million a year, you need to spend 5,000 times more than that to ultimately launch a car, something like that, maybe more, and so you’re not able to actually make real progress, you’re just working on demos and proof of concepts. And we didn’t even start with $1 million, we started with zero. So the first financing was I refinanced the house that I owned, which is comical when I think back now that my level of conviction and optimism.

No, it’s awesome.

RJS: I thought, “I’ll refinance my house, take the $100,000 to get out of it and use that to start a car company”, so that was what we did. But it’s very hard to then hire, we’re just getting a semblance of traction to have some capital, some amount of money that we could actually make real progress on a product.

Right. And this was all — you still didn’t really know for sure what you were going to build, right?

RJS: That’s what I was going to say, it’s actually really helpful that in those years we didn’t have capital, because we could have started building the wrong thing, so it provided me this few year period where I was learning how to run a company. I’d never run a company before, I was learning how to lead teams, I was learning how to hire, I was learning how to have hard conversations, I was learning how to raise capital, I was learning about strategy and design and brand and all these things. And so it was a really wonderful period of time for me because we were iterating so dramatically, so significantly on the strategy, the product, the type of company we’re building, the skillsets we want to accumulate and build in-house, in ways that we couldn’t today. I couldn’t walk in the door to Rivian, say, “All right, everybody, we’re going to do a completely different set of products, get ready”.

You were on an e-bike back then, you could sort of go where you wanted to. The bigger you get, the more locked in you are.

RJS: Yeah, the whole team could fit in one little room with one little table, investor management was very straightforward, it just gave us the freedom to be very iterative. And I look back and I’m so thankful that happened because this squiggly path led us to what Rivian ultimately became, this idea of building a really strong brand around enabling and inspiring adventure that scales across different price points and form factors.

We came up with the concept for R1 as the flagship product, then we would follow that with R2, which is now about to launch, and then R3, which is going to launch shortly thereafter, and things took a lot longer. Once we even got all that defined, we still had to raise a lot more capital. We then raised a lot of capital and we’re on the path of execution, and there’s some big unplanned events. COVID was very, very, very challenging and maybe the worst possible time you could imagine it.

Right. You’re just about to launch.

RJS: Yeah, so trying to build a plant starting in 2020, which is sort of wild, and then turn on a supply chain with a bunch of suppliers that didn’t want to work with us, we had to pay extremely high prices to them just to get them to provide us components. Just a bunch of things that when you’re planning it years before, you don’t think, “Well, there’s going to be a supply chain crisis, there’s going to be a pandemic”, and there’s going to be all these externalities that make it really hard to start in that moment.

You mentioned getting to this adventure brand identity, the R1T, R1S being your initial products, what was the process of honing in on that? Why did you decide that was the way to go? I mean, from the outside, you view obviously, Rivian, you’re always going to be compared to Tesla in a certain respect. They have this futuristic car looking very aerodynamic, and Rivian comes along, it’s like, “Yes, thank God, a pickup truck”, not a Cybertruck, they got an SUV. That’s my perception of the outside, but what was it like inside?

RJS: Yeah, I mean, it wasn’t too different from that. We recognized that in order for us to earn the right to exist, we needed to do something that was unique and could stand on its own, and so some of the early things we thought about, we’d originally thought about doing a sports car, we realized that we were just going to be too close to what Tesla had done, and what Tesla had done well, by the way.

So we went through deep soul searching to say, “What are the things we’re passionate about?”, “What are the things we want to enable?”, “What are the things that are going to matter?”, once everything’s electric, imagine every car on the road is electric, you can’t say we’re differentiated because we’re electric. “Why are you differentiated?”, “What is the reason for someone to choose to buy our products?”, and so we went through a lot of those thought processes and came out of it with this idea of preserving and inspiring the adventures that you want to have in your lifetime, the kinds of things you want to take photographs of.

The reason why you want cars is you can go anywhere—

RJS: Yeah, you can do all this, you can go to your grandparents’ house, you can go to the beach, you can go climbing. So that led to this really clear vision, which then led to product requirements. “Okay, if we want a car that’s going to enable and inspire adventure, what does it want to look like?”, “What are the features it needs to have?”, so storage becomes a really big consideration, being able to drive on any type of terrain becomes a big consideration, and then you say, “Okay, what’s the vehicle form factors that are going to do that?” — a re-imagination of a pickup truck and a re-imagination of a large SUV, that’s a great flagship.

So it wasn’t always as direct as that single sentence, sometimes it took us a month to get there or more, but a lot of iteration, a lot of the product concepts, some of the early R1T stuff that we put together looked really futuristic and not inviting, which is the word we use all the time, like inviting you to use it. Not wanting to get dirty, or it didn’t want to get used or you don’t want to put a surfboard on the top so we became really very intentional around, “What is a Rivian?”, “What is not a Rivian?”, so we do all these exercises from a design aesthetic point of view, which of course now we know our aesthetic, but in 2015, we had no idea what a Rivian aesthetic was, so we had to define that. So we do these is/is not exercises, it’s all things that it’s amazing sitting here today to see it having played out where people actually connect and resonate with the brand that we were hoping to.

It’s an incredibly strong brand, you can identify it right away.

Production Challenges

You mentioned the COVID production challenges, there’s also a bit where actually scaling up production is just really hard, even if everything is perfect. How do you distribute the blame between COVID and between the fact that actually, “This is much harder than I thought it would be”, in terms of the challenges in getting out the door?

RJS: Yeah. I think we made a tactical or strategic error, which is we decided to launch three vehicles at the same time.

Yep, that’s one of my questions coming up.

RJS: Launching any vehicle is really hard. So just to put this into perspective, you have around 30,000 discrete components, which you purchase as a company as maybe 3,000 items and the reason there’s more discrete components is you buy something like a headlight as a single assembly, but it has many components in it. But all those tier two, tier three, tier four supply components, any one of those can stop production if they’re missing and so you still have to think about it considering the full complexity of all the parts, every single mechanical part that’s in the vehicle and so turning on a supply chain for the first time is hard for any product.

For a car, it’s really hard. And for a car when the supply chain doesn’t want to work with you, meaning business is thriving, it’s very different than let’s say 2010 or ’11 or ’12 when the suppliers were all beat up from the recession and were willing to take any business. In 2020, they were busy, they didn’t want to take on this new customer Rivian with an unproven brand and unproven product. So it was very, very hard to get them to work with us and so just getting all those suppliers to ramp at the same rate on one car would’ve been tough. And the reason I say same rate, if some ramp faster than others and you have inventory issues.

Right, then you have these working capital problems.

RJS: You have to ramp at the same time so you can make a complete car and sell it, sounds so simple.

(laughing) No, don’t worry. It does not sound simple at all.

RJS: So we were doing that across three vehicles at the same time, that was already a big — the R1T, R1S launched at the same time plus a commercial van. And then on top of that, we had COVID, which made everything more challenging. Yeah, so it was maybe the most perfect of perfect storms for difficulty and so I wouldn’t use COVID as an excuse or I’m not putting blame there, I’m just saying it’s just a reality, it was just a reality.

Oh, for sure. No question, I don’t think anyone is denying that.

RJS: But if I were to do it over again, I would’ve launched probably the SUV first, spaced out maybe 12 months, then launched the truck, spaced out probably 12 months, launched the van, and had smoother launches that consume less capital that would allow us to get to profitability faster. But hey, you learn. And so here we are in R2-

In 2025.

RJS: Yeah. So R2 is like, we’re launching one build combination, we’re launching a launch edition, we’re not launching R3 at the same time. You’ll laugh at this, Ben, there was a lot of debate like, “Oh man, R3 is so cool”, we had thousands of customers like, “Oh, we can’t wait to get an R3 as well, can you guys launch that quicker?”, and we’re like, “Should we try to launch R2 and R3?”, “No, no, don’t do it! Don’t do it! Hold it back!”, and so we held back.

It’s like you needed to hire someone back in 2021 that says, “If we ever consider doing this again, stomp on the table and say ‘No’.”

RJS: As a product person, you have all these ideas, you want to see them out in the world as fast as possible so simplicity and focus has been a major emphasis for us and so the entire business is laser locked on launching R2 and it’s a beautiful thing. We don’t have other programs that we have to manage, it’s like, “Let’s get that, that has to ramp quickly, that’s what’s going to drive us to profitability”. It’s key for cash flow, we have this enormous R&D spend we’ve created intentionally to build out all these vertically integrated technologies, whether it’s our chips, our software, our compute platforms, our high voltage architectures, that the scale that R&D necessitates the scale of more than just R1, more than just a flagship product that needs a mass market product, and that’s what R2 brings us.

Got it. So you mentioned the van. Ideally from a production standpoint, you do SUV, then you do truck, then you do van. The van though came with a lot of money from Amazon, is that a critical component in why you launched it maybe sooner than you should have?

RJS: No, at the time I didn’t think it was sooner. I mean, at the time I thought it was the right thing to launch them all at the same time. Amazon’s still our largest shareholder and they’ve been a great partner and they were an investor in us when we were private, as you said.

But what’s so exciting about that program is it took a space that has the logistics based on last-mile e-commerce space that has such a clear value prop for electrification, meaning the vehicles start and end the day in the same spot, which is a great thing from a charging point of view. You know what they’re going to do, that you can deterministically control what they’re going to do in terms of mileage. You know your 99th percentile route in terms of number of miles, and you know your one percentile, so you can really optimize it for total cost of ownership, and so that’s what we did. So we went about and said, “Let’s make the ultimate delivery van, let’s make it the most cost-effective way to deliver”.

So you talk about the complexity of doing three vehicles. Is that just in terms of getting started or is there a production capability, like you only do so many things, or is that part fine? It’s just the part of getting started?

RJS: Part of the challenge is when you’re launching a manufacturing and supply chain infrastructure for the first time, in our case, we didn’t fully appreciate all the things you need to be really good at to do it and so we tried to very quickly learn how to be able to launch multiple programs at the same time, which eventually Rivian should be able to launch multiple vehicles in the same year at the same time, but we just didn’t have the maturity of process, maturity of our organization or the depth of teams to be able to support that.

The issue is not things you can plan for, it’s all the little things you don’t plan for, and there’s all these little things, each of which requires problem solving. So I used to describe it in 2021 and 2022, is it’s not like there’s some giant unlock. Like, “If we just solve this, we will make more vehicles and we’ll get our cost structure in line”. There was just a stack of thousands, truly thousands of little things that needed to be adjusted or changed or negotiated and I think the thing that compounded all this that was really hard, is a lot of those issues were at our suppliers, and then those suppliers had a lot of leverage over us, because they know that in that time where they couldn’t get enough-

If this doesn’t get done, you’re done.

RJS: We broke up with a lot of these suppliers, but some of them would just say, “We want you to pay us twice what we previously negotiated if you want parts”, and we’d say, “No”. and they’d say, “Okay, fine, we just won’t send you the parts”, and we’d say, “Okay, how about one and a half?”. We just had no leverage.

So that’s changed so dramatically and we see it with R2. R2’s the first, I’d say, clean sheet from a supply chain point. Even with the updates we made to R1, we were able to get rid of a lot of that and they call it inflation-related, COVID-related cost growth that was born out of a lack of leverage that we had, R2 is the first time we were able to really reset the negotiations. You think of it from the perspective of if you’re Volkswagen, the leverage is the other way around, which is Volkswagen has so much scale and so many diverse sets of suppliers that they could say, “Hey, if you don’t bring your costs down, we’re just going to switch to another supplier”, we didn’t have that. We couldn’t say, “Well, look, we’re going to pull this other program from you” — it was no leverage, so we sort of were complete takers in that.

Volkswagen Partnership

Yeah, that makes sense. Before we got into the AI stuff, I did want to ask about the VW partnership. This includes access to your electric vehicle software, electrical architectures, you get supply chain expertise from them. How do you characterize this deal as a whole? I should also mention a sort of massive investment on their side as well, give me the framework of that deal and why it’s important.

RJS: Yeah, it’s a $5.8 billion deal, some of which is technology licensing, some of which are investments.

Right, I was going to ask about that. Some of it is just actually putting money in the company, and some of it is they’re going to license your software and things like that going forward.

RJS: Yeah, and a lot of those are upfront licensing fees, most of which have already been paid and before I get to the business of it, it’s important to talk about the mission of it. We’ve spent a lot of time developing what we call a zonal architecture, but essentially think of it as a number of computers consolidating into one that perform a wide array of functions across a physical zone of the vehicle and it allows us to do things like over-the-air updates very seamlessly because rather than having a bunch of smaller function or domain-based electronic control units, little mini computers run the software for different functions, we run all this software for those functions on one computer on our OS, which makes it much easier to update. And so the strategy there was, “Boy, we’ve spent a mountain of investment building this tech stack, it’d be really nice to see it applied in another way.

Yup. You need to get leverage on that investment and you just don’t have the volume by yourself.

RJS: And it aligns to our mission in terms of enabling more electric vehicles to get highly compelling electric vehicles on the road and then it gives us a lot of scale, scale for sourcing the components that are shared and then it gives us the benefits of other, what we think of as joint sourcing agreements, so sourcing partnerships that can exist with Volkswagen.

It’s been a great relationship, those types of relationships are very, very hard to build because it does require buy-in from the top so one of the things that allowed us to work so well with Amazon, I mean, you think about Amazon and it’s one of the largest companies in the world, certainly the largest e-commerce company in the world, and imagine they go out and say, “We’re going to build our future logistics network around a van that’s being not dual sourced, but single sourced to one company” — this is in 2019 — “has never built a car before at scale, and they’re like a startup”. But that was born out of a great relationship that I had with [Former Amazon CEO] Jeff [Bezos] and Jeff’s trust in supporting us and that enabled them to really lean in with us and lean in in defining the product, defining what it was, that was a really big leap.

So we’ve built, I’d say, organizationally, really great capability of taking the strengths of being a fast-moving startup and working with very large companies as partners and in the case of Volkswagen, my relationship with Oliver Blume, the CEO of the group — so Volkswagen Group is, we think of VW as a brand, but they’re a group — they’ve got Porsche, Audi, Lamborghini, SEAT, Škoda, these are brands that aren’t sold in the United States, but it’s the second-largest car company in the world, largest industrial company in Europe, a huge company. But having Oliver and I aligned just allowed us to really move through the deal mechanics and the deal structuring quite quickly.

So this bit, as you sort of zoom out, the deal makes a lot of sense to me. Actually, I think it makes a lot of sense for both sides.

RJS: Yeah, it’s a win-win.

They get expertise that they’re not going to develop internally. I’ve had plenty of German cars, the software is okay for what it is, I don’t think it’s going to sort of go where you’re going to go. You also have on your side, you can do these huge investments like you talked about last week, and we’re about to transition into that, with the promise of scale that is much more than you can certainly deliver today. Is there a bit of you though is like, “If we had ramped up correctly, if we had not done multiple vehicles, we could actually be at scale, we could keep this all to ourselves”, or is this ultimately the best outcome that you’re sharing with them in the long run?

RJS: I think in hindsight, I wish we’d ramped up more quickly, there’s things you’d change, but they’re also all things you’d learn from. We don’t spend any time lamenting them or anything like that. But to be clear, both in the case of our in house software and zonal controllers, which is what we’ve done with in Infotainment, which is what we’ve done with Volkswagen, as well as our autonomy platform and AI platforms, which is separate from the Volkswagen venture.

Is that part of the deal?

RJS: No, that’s not part of the deal.

Got it.

RJS: That’s 100% Rivian.

Okay.

RJS: But in both cases, we developed them thinking that we would eventually leverage this, not just with our own products, but with other companies as well.

Got it. Okay.

RJS: And so Volkswagen was, in many ways, the ideal first customer. And the reason I say it’s the ideal first customer, 1) it’s huge, as we’ve already described, but 2) it has the complexity of managing across many different brands, and so being able to support a company like Volkswagen Group, which spans very premium brands, like Porsche, down to one of the products that’s been announced that we’re doing together, the Volkswagen ID1, which is a $22,000 EV, it’s the existence proof that we, Rivian, can support working across large complex organizations, across large ranges of price and product features, and across very different vehicle form factors. And if you’re another car company, you couldn’t look at Rivian and say, maybe before you could have, but now you couldn’t, say, “Well, I don’t think you could do this at this price point” — well, actually we cover every price point across the spectrum.

So there’s an opportunity for other car companies to do the same thing.

RJS: Absolutely, yeah. And now on the autonomy front, I think the opportunity there is actually bigger because this is a very, very hard problem to solve, it requires vertical integration in ways that are not typically — it’s just things that OEMs typically don’t do.

Autonomy

Tell me your vertical integration story, because it is really interesting. You’re on last week, you talk about everything from your chip to your sensors to your software, you talked about building your own compiler. We are talking total front-to-back, end-to-end vertical integration. Why is that important?

RJS: Yeah. It’s important to just talk about how autonomy is now being developed, and I do think for anyone listening to this, it’s very, very important to understand this because there’s perhaps some histories to how it was done before.

The idea of a vehicle driving itself isn’t a new idea, that’s been something, it’s been in sci-fi movies for decades, but in terms of actual technology development, it started in, call it early 2010s, in that time range, so roughly 15, 20 years ago.

The early platforms and what was done in terms of the approach up until the very early 2020s was something that was designed around a rules-based approach and so what you would have is you’d have a set of sensors, perception that would identify objects in the world, so all the things in the scene, so that’s cars, people, bikes, kids, balls bouncing on the street, everything that you can see, it would identify all those objects, it would classify the objects as to what they are, it would then associate vectors to those objects, acceleration of velocity, and it would hand all those objects and their classifications and their vector associations to a rules-based planner. The rules-based planner was a team of software developers attempt to codify what are the rules of the road. So, I’m going to oversimplify here, but think of it as a whole series of if/then statements.

Totally deterministic, by the biggest spaghetti code mess you’ve ever seen because there’s so many possibile exceptions and issues.

RJS: It’s a giant, giant code base that’s trying to describe how the world works. And so, it wasn’t actually AI as we think about AI today, there was machine vision.

Machine learning, neural nets, yeah.

RJS: Yeah, there was machine vision for the object detection classification, but in terms of the planning and the actuation of vehicle was very much a rules-based environment. Then along came the idea of neural nets, and the idea of transformers to do encoding, and that happened, of course, in the LLM world, but that’s also happening in the physical world.

Everything can be a token. We think about it, everyone thinks one of the context of letters and words, but everything can be a token.

RJS: Yeah, everything can be tokenized and the whole world changed in self-driving, so everything that was done prior to, call it 2020, 2021 is largely throwaway, meaning the way the systems are now developed is you build, you need to have complete vertical control, it needs to be one developer that controls all the perception, because you don’t want a pre-processed set of outputs from a camera, you want the raw signals from a camera. If you have other modalities like a radar or LiDAR, you want the raw signals from those, you want to feed it in through a transformer-based encoding process early, so fuse all that information early, and build a complex, it’s hard to imagine in our human brains, but it’s a complex multidimensional neural net that describes how the vehicle drives.

Then you want to train that with lots and lots of data, and you’re training it offline. The word that gets used all the time is end-to-end, so it’s trained end-to-end from the vehicle through the human drivers back to the model and so, to do that well, you need a few ingredients, you need this vertically-controlled perception platform, you need a really robust onboard data infrastructure that can both trigger interesting data events, hold them, do something to them to make them a little easier to move off the vehicle, ideally through Wi-Fi, and a worst case through LTE, but mostly through Wi-Fi, all that data gets moved off the vehicles, and this is happening at millions and millions of miles accumulating just in the course of a day.

And so all that data is moving off the vehicle, and then you’re training it on thousands and thousands of GPUs. You’re going around and around and around, and it gets better and better and better. That’s an approach that is so different, as I said from what was done before, but to do that, you need all those ingredients.

Well, you need cars on the road.

RJS: You need cars on the road. So, we looked at it, we launched in 2021 with our Gen 1 architecture, we almost immediately after that realized we needed a complete rethink of our self-driving approach.

Right, that’s exactly what I was going to ask. Was this an issue where in some respects you launched later than you wanted to because all the supply chain issues, but then you actually launched earlier than you wanted to because you didn’t have the right sort of stuff on your cars?

RJS: Well, we launched — and we didn’t realize, and this is the thing, and even some of our Gen 1 customers are not happy with this, but when we developed the Gen 1 system, this was on 2018, 2019, we didn’t know this big technical massive shift was going to happen.

So, our Gen 1 architecture uses a Mobileye front-facing camera, and it uses — it’s a collection of things, it’s very classical rules-based approach, if you’re going to develop something around AI, it’s a completely different architecture, not a single shared line of code, not a single shared piece of hardware.

So we started working in the beginning of 2021, right after launching on a whole new clean sheet, everything new, we didn’t try to morph anything over, it’s a complete melt and re-pour. In that new architecture, we designed cameras, we designed a new radar, we designed a new compute platform, we built, we call this our Gen 2 architecture. We built it around an Nvidia processor, we designed a data flywheel, we designed an offline training program. The vehicle launched in the middle of 2024, the features then were trained on a very small number of miles, which was our own internal fleet and now over the course of last year, we’ve built up enough data that’s allowed us this flywheel starting to spin.

Yep. And that data is only coming from the Gen 2 vehicles, right? Not from the Gen 1 ones?

RJS: Only Gen 2. Gen 1, it’s asymptoted, both in terms of capability and it has no value to us in terms of data, so only Gen 2. And so, in parallel to kicking off this Gen 2 platform, which we said, we need to get this in the field as fast as possible because we need to start the data flywheel, we also need to get better hardware so that when we have the model built, we can run it with a higher ceiling. That kicked off updates to the cameras that are going to our Gen 3 architecture, very importantly, an in-house silicon program.

Why is that very important?

RJS: Compute inference on the vehicle, we wanted to have — what would we have in our Gen 2 is around 200 TOPS [Trillions of Operations per Second], we wanted that to be closer to 200 TOPS per chip, so 400 TOPS total, sparse TOPS. Well, what’s going to be in Gen 3 will be 1,600 sparse TOPS, but importantly, we designed it specifically around a vision-based robotic platform. And so, the utilization of those TOPS is very high, much higher than what we see in other platforms that are more generalized, and then the power efficiency is very high, and then the cost is much lower. So we have a very, very high capability, low cost platform for which we can afford to put enormous compute in.

All that is true, but the actual development of that is very expensive. Is this going to pay off with those lower unit costs, and that increased capability with just your vehicles, or like the VW deal, is this something that you’re going to be looking to sell broadly?

RJS: Well, this is an interesting one. Even on its own within Rivian, just R1 and R2, it’ll pay off because the cost savings are so significant on the chips. But more than that, we believe we’re very, very — we’re spending billions of dollars in developing our self-driving platform, our level of conviction as this being one of the most important, I shouldn’t even say one of the most, the most important shift in transportation and transportation technology means that we want it to control the whole platform. Then once we control the whole platform, it makes it a very interesting system that can be provided to other manufacturers. And so, I think in time, the number of companies that will have all the ingredients to do what I’ve just described, they’d be very limited, I think there’ll be less than five in the West.

Did you get any of this thinking from Jeff Bezos? Because there is a bit here where our cars are where we get out and develop this and prove it out, but the real payoff is to do the platform at scale across other entities. It sounds a little Amazon-like.

RJS: Amazon’s our largest shareholder, and Jeff’s somebody I look to for a lot of inspiration on these kinds of things. So, certainly I think there’s some of that. We think of our vehicles as our own dog food, but we’re going to make a platform that’s so darn good that we think others will-

You’ll sell a lot of vehicles.

RJS: And if others aren’t buying our platform, we’ll monetize it through selling more vehicles, and we’ll grab market share. I think on both sides of that, we can win. I do think that it’s going to move far faster than anyone realizes. I think, the way I describe it is if you look at the last three or four years of development in autonomy, and you try to draw a line to represent the slope of improvement, and you look at the next three or four years, the two lines are completely unrelated.

Totally agree.

RJS: But the acceleration is going to be so fast. And what I’m surprised is people aren’t — I say this, I don’t think people fully realize it, but the LLM space should teach us that.

Yeah. GPT-1, GPT-2, GPT-3, GPT-4.

RJS: But look at the 1.0 architectures.

Oh, which are rules-based. Yeah, to your real point, it’s exactly what it is.

RJS: Rules-based. And look at the progress that was made on Alexa, let’s say, relative to the progress that’s happened on GPT-3, 4 now, and beyond, it’s just like they’re not even closely related. And so the same thing is happening in the physical world with cars, and if you don’t have a data flywheel approach, you’re just not in the game and there’s no way you can compete. And so, very few people have that, far fewer I think is right.

A big differentiator between what you’re doing and what Tesla is doing, and we have to sort of come back to it, they shifted to the pure neural network approach, but they’re doing vision only. Do you just think that’s a fundamentally flawed decision?

RJS: We have a different point of view.

Right. Because you have radar and LiDAR too, is the difference there.

RJS: Yeah. There’s a lot of alignment, and we both agree, and we’re both approaching it as building a neural net. So, I want to call that out that we have a very aligned view.

Right. Your core philosophy is absolutely the same. And I think there’s an extent where Waymo is getting there as well.

RJS: The same philosophy. And then it’s like, “How can we teach the brain as fast as possible?” is our question. They have the biggest fleet of data acquisition in the world, they have fewer cameras, that have far less dynamic range. When I say dynamic range, I mean performance on very low light conditions, and very bright light conditions.

Right, yep.

RJS: We have much better dynamic range that of course adds bill of material cost, but we did that intentionally. And then, we have the benefit of our whole fleet, all Gen 3 R2s, think of those as ground truth vehicles. They’ll have LiDAR and radar on them.

Tesla just has a few ground truth vehicles that do have radar and LiDAR, but they’re trying to service the whole fleet.

RJS: Yeah, I’m looking out the window here at El Camino and you just have to stand at the corner and see Teslas driving around and around everywhere.

One will go by eventually, yeah. So that’s the question, is the benefit of putting radar and LiDAR on all your cars, is that just something you need to do now so you can just gather that much more data that much more quickly? Or is that going to be a necessary component for at scale, everyone has an autonomous vehicle and they need to have radar and LiDAR?

RJS: Yeah, I think, the way I look at it is, in the absolute fullness of time, I think the sensor set will continue to evolve. But in the process of building the models and until cameras can become meaningfully better, there’s very low cost, very fast ways to supplement the cameras that solve their weaknesses. So seeing through fog we can solve with a radar, seeing through dense snow or rain we can solve with a radar, seeing extremely far distances well beyond that of a camera or human eye, we can solve that with a LiDAR, our LiDAR is 900 feet. And then the benefit of having that data set from the radar and the LiDAR is you can more quickly train the cameras. The cameras, when I say train, it doesn’t mean we’re in there writing code to do this.

I think my audience broadly gets how this works, yeah.

RJS: The model understands this and so you feed this in and the neural net understands because you have the benefits of these non-overlapping modalities that have different strengths and weaknesses to identify, “Is that blurry thing out there actually a car?”, “Is it a person?”, “Is it a reflection off of a building?”, and when you have the benefit of radar and the benefit of LiDAR, that blurry thing way off in the distance that the camera sees starts to become — you can ground truth that much faster.

And then you teach your camera to figure out what it is.

RJS: Then your cameras become better, and so that’s our thesis. And of course, that’s important that we have a thesis that’s different than Tesla, if we had an identical thesis to Tesla on perception-

They just have way more cars out there.

RJS: Yeah, the only way to catch up is with building a fleet of millions of vehicles, we want to catch up faster than that.

So is it also sort of this advantage that — to what extent do you feel the auto industry, you start out and you’re sort of the outsider, you can’t get suppliers to help you, they’re ripping you off, all the sorts of problems you talked about. Now you’re like, I can imagine Volkswagen at a minimum is looking at you, “Please figure this out, we have a relationship, we can sort of jump on if need be” — do you get that sense more broadly from the industry? Because I don’t think anyone expects Tesla to share their technology, Google is sort of its own thing, do you have the potential to be the industry champion in some ways?

RJS: We hope. I mean, I think every manufacturer has three c