- 2 days ago
Category
🤖
TechTranscript
00:00RJ, Rivian finally has an AI story. I actually would like to start by asking why now you're
00:06ready to talk a bit more about where Rivian feels its AI competencies are.
00:12Yeah, so Rivian launched our first vehicles in end of 2021. And almost immediately following
00:17that, we began the process of designing the clean sheet approach to how we were going to
00:21integrate AI across the business. That's on the autonomy platform, and that's the backbone level
00:26of the vehicle, and of course, with the enterprise. And in doing that, we redefined the architecture.
00:32And the first embodiment of that was actually in our Gen 2 fleet. And so the Gen 2 fleet was a
00:3610x improvement in the compute, dramatic improvement in the camera platform. But while that was
00:42something we showed publicly and really launched, think of it as like the data flywheel for building
00:47this large training environment for building out our foundation model, we also were developing
00:54hardware in parallel that we didn't talk about. And so that's what's going into our Gen 3
00:58architecture. So that's an in-house processor that really significantly improves the compute
01:04capabilities in the vehicle, further enhancements on the camera platforms we developed in-house,
01:09and then the inclusion of a LiDAR in the vehicle as well.
01:12So those three pieces are the core, but the specifics are important. Wrap 1, in-house custom
01:20silicon, the LiDAR sensor indented design, cool. And then the work on software, that trio is specific
01:28to R2 and then R3. But the point is that it unlocks, to your mind, a not-too-distant future
01:38where those vehicles will offer the owners a genuine set of autonomous capability.
01:44Absolutely. And I think an important point, I'm glad you made the distinction. There's a software
01:48element to this, that with the Gen 2 vehicles, our R1 vehicles, where this is designed around an AI-centric
01:54approach. And so the way that the vehicles are operating, there's a large driving model
01:57that we continually improve through the deployed fleet, which is generating millions and millions
02:04of miles every day. Existing R1s, but the second iteration of it, what you call Gen 2.
02:08What we call Gen 2. So those launched a little more than a year ago. And those Gen 2 vehicles are part
02:13of the training flywheel for the model, but they'll also see enhancements. And so the first of which
02:19we're going to be putting out through an OTA and over there update later this month, where we'll expand
02:25hands-free from today, just under 150,000 miles in North America to over three and a half million
02:31miles in North America. And that's where the vehicle will be able to operate with your hands off the
02:35wheel. And then shortly thereafter, in 2026, we'll add point-to-point navigation, meaning you can get into
02:41the car, plug in the address, and it'll drive entirely on its own to that address.
02:45But require supervision, and there will be attention monitoring.
02:48And so there'll be somebody in the car still, hands off the wheel, but still eyes on the road.
02:52And that's actually the demo that we're showing today. And so that's just a preview of what's to
02:57come in 2026. What the Gen 3 architecture does with compute levels that are dramatically expanded,
03:03so this is, just to put some numbers to this, at the platform level, 1,600 sparse tops.
03:08We can process 5 billion pixels per second. We include, as you said, a beautifully integrated
03:14LiDAR. That raises the ceiling to allow us to take the platform to personal level four.
03:19So the first step before we get to personal level four is beyond hands-off point-to-point.
03:25Maintain those hands-off point-to-point, but take your eyes off the road.
03:29What's your timeline for that, RJ?
03:31That's a great question. So the way we'll launch eyes off is it won't be eyes off everywhere.
03:35First, it'll be eyes off on highway, and that'll be coming after point-to-point. So that'll be in
03:40the 27 timeframe. And we'll expand eyes off highway to eyes off everywhere. And then the next big step
03:47is personal level four. And what I mean by personal level four is the vehicle can operate empty. It can
03:53operate without anyone in the driver's seat. It can pick your kids up from school. It can drop you at
03:57the airport. It can go between houses. So it's a complete shift in how we think about the vehicle
04:04experience. But that's enabled by this very high level of compute we have in the vehicle,
04:09this very large driving model that we're building, and of course, the enhanced perception platform.
04:14So there are different academic schools of thought, right? There is a future where none of us own our
04:20own vehicles. We have purpose-built robo-taxis operated by technology companies through proprietary
04:27ride-hailing fleets, or we do own our own cars. And we have expectations that those vehicles have
04:33capabilities. Which of those camps do you subscribe to?
04:37I think the technology is nearly identical. To do level four capability, be able to operate
04:42empty, without someone in the driver's seat. It's really just a question of which business model
04:51you pursue and when. And for us, the initial focus will be personal level four, but it doesn't
04:56preclude us in any way from participating in rideshare. But it is worth noting, the vast majority of miles on
05:02the road today are in personally owned vehicles. So close to 99% of the miles that we collectively
05:08drive are in vehicles that we or our household owns. But I think the thing that's so exciting about
05:15this AI-centric approach is the model that we have running in our Gen2 R1 vehicles today
05:20will just simply get better. It's accretive with the new sensors and new compute. Think of it like
05:26you know, if you learn to drive and you had bad vision, and suddenly I handed you a set of glasses,
05:32your vision would become a lot better. Right.
05:35Your driving skills would become better. Your ability to learn how to drive a vehicle would improve.
05:39I would see the world with a richer picture.
05:41But you wouldn't lose all the knowledge you had from driving without glasses. Right.
05:45And so what we do with this next generation set of perception and compute is we not only improve
05:51the camera set, but we add these additional modalities. We add a LiDAR on top of it. We add
05:56improved radar. And then we increase the compute of the brain.
06:01You're going to get this question. You're going to have to answer it.
06:05You're committing to R2 being, you know, a certain price point. You're now adding in LiDAR.
06:10Your strategy with the refresh on R1 was to have really leading cutting edge sensors on the camera
06:19side. Those are expensive leading components. Right. So how is that going to be possible, RJ?
06:24Yeah. And we haven't even started discussing the economics of custom silicon yet.
06:27Yeah. Well, there's a lot there. A few questions. So I think first on LiDAR. LiDAR has changed a lot.
06:34And I think there are some there's some histories in our thinking around the cost structure of LiDAR.
06:40So LiDAR costs have come down depending on when you draw the line or when you're measuring from 50
06:46to 100 X. These went from, you know, tens of thousands of dollars down to, you know, we're talking
06:52like low hundreds of dollars at a sensor level. And so it's no longer a big expense. It's a very small
06:59percentage of the vehicle bill of materials. And then on the compute side, on the on the Rivian
07:04processor side, this represents a significant cost savings to us. There is a lot of margin,
07:10of course, in the in the semiconductor space. And, you know, working directly with TSMC,
07:15we have a great relationship with them. And so by removing some of the cost layering or cost
07:22stacking that occurs, we're able to deliver an enhanced level of performance much better than
07:27what we have on our Gen 2 architecture. Part of that bet is scale. At a lower cost. Right.
07:31Because you're the assumptions of models you have is that is that R2 is going to be a much and then
07:36R3. Yeah. Higher volume product than the R1 generation. Yes. Of course, there's a scale assumption
07:41in that. But even just removing some of that margin stacking that occurs, it's a it's a fundamentally
07:48lower cost approach with, you know, it's an interesting one. Usually you can't lower costs and
07:53improve performance. But here we improve performance dramatically and simultaneously lower
07:57cost very significantly by hundreds of dollars per vehicle in terms of savings. It is a bold and
08:01risky bet. You know, you talked about you talked about reinforcement learning and GRPO it through those
08:10reinforcement learning, the ability for an existing piece of software to get better with time in those
08:17second generation R1s you use Nvidia orange chips. Yeah. Nvidia knows what it's doing. Yeah. And you've still
08:22made the decision as I understand it based on cost performance and speed to market. Yeah. To do
08:28in-house custom silicon. Yeah, that's right. And the bold bet to go away from Nvidia. It's also not a
08:34it's not a bet one takes lightly like this is a it's a it's a it's a it's a huge commitment. It's taken
08:41us years to both design the chip and now get it ready to launch, go through many iterations of the chip
08:47itself. And then build the whole middle well middle layer. Yes. For the development and the development
08:55environment are on the chip. And of course, we can leverage a lot of the platforms that are out there
08:59things like PyTorch to speed up the process. But but yeah, it was it was a large, you know, large
09:08commitment we made to the architecture. Going back quite a number of years, I remember conversations we had
09:13in earlier days of that first generation are one, even before one came off the production line,
09:19really, about how difficult it was to convince suppliers to give you the volume of components
09:24you needed. You were basically saying, we can do this. And they weren't forthcoming with with the gear.
09:29Yeah. With TSMC, it sounds like actually, this is a completely different environment and situation.
09:34Yeah, it's I don't think we could have done that. I don't think we would have able to get the
09:40engagement from TSMC. And 2017 or 2018, we had to launch our one, see it be a market success,
09:48see people love the brand. And then on the back of that, develop a relationship for our
09:53Hivewine product with, of course, with R2 leading that. Now, there will be skeptics,
09:58Yeah, existing Rivian owners, investors that will say, you're still struggling a little bit with the
10:05basics of scaling production of vehicles. You know, I remember the conversations we had where
10:11normal Illinois on paper was capable of doing 250,000 vehicles per annum, you might do 50,000
10:18this year. Now, of course, the plan is for a completely different product for the mass market.
10:23Yeah. But to those that do say, is now really the right time? And is this the right thing to do,
10:29focus on autonomy, and not just the basics? What would you say to those people, RJ?
10:34Well, I don't think it's a binary. It's sort of a false binary to have to pick one or the other. So
10:39we, of course, need to get the basics right. So we need to launch R2 that dramatically expands the
10:43business. One of the things I've said a lot is that Rivian's R&D infrastructure and fixed cost
10:49structure isn't designed to be a 50,000 unit company. So we've vertically integrated our entire
10:55software platform. We've built all of our electronics in-house, as we're now showing.
10:59We have a custom silicon platform. We have in-house perception. So these are really big
11:06capital consuming R&D efforts that don't make sense if we're only going to produce 50,000
11:11vehicles a year. Now, they make a lot of sense if we're producing many hundreds of thousands and
11:16ultimately millions of vehicles a year, which is how we've designed and architected the business.
11:20And it's proven true that by taking this first principles-based approach,
11:28vertically integrating in the right areas, it actually does create a better product. And we
11:31saw that in our software platform. We did a $5.8 billion software licensing deal with Volkswagen Group.
11:38I think the benefits that were enabled by vertically integrating our software and our electronics,
11:46is it really gave us this foundation to then think about AI coming in on top of it.
11:52But what we're now doing in self-driving and the AI backbone of the vehicle, I think,
11:56takes it even a step further. It's such a necessary big enabler for a huge level of differentiation at
12:04the product level and at the technology level.
12:06Paradoxically, is Volkswagen going to benefit from this competence in software?
12:11You know, these are Rivian specific capabilities.
12:14Yeah, yeah.
12:15But, you know, Volkswagen, sorry, struggles actually with software historically.
12:20How have you thought about that?
12:22So everything that we're doing in self-driving and AI is actually separate from our joint ventures.
12:26So that's purely 100% within Rivian.
12:29You see a day where you license that technology to others is what I'm asking.
12:34It's funny you ask. In the same way that if you'd asked me in 2020 or 2021,
12:40do we think that our software architecture and zonal approach to ECUs could be licensed to
12:45other OEMs, I would have said yes.
12:50But it wouldn't have been clear who would step up to the plate and how that would play out.
12:54I think we have an even stronger level of conviction that in the self-driving space,
12:58that what we're building is so well architected at a platform level.
13:02The data flywheel and the benefits of data flywheel that we're seeing are so strong
13:06that it's not hard for us to imagine in the next several years that this becomes a platform we also license.
13:13I'd like to end with the product and the buyer of that product, the consumer.
13:18You know, we're today in one of your Palo Alto facilities.
13:22I came all the way here using Tesla full self-driving.
13:24All right, great.
13:25You know, I do hundreds of miles on it a week and I pay a subscription $99 a month, right?
13:30Yeah.
13:30You are pricing Rivian Autonomy Plus as either a subscription at $49.99 a month,
13:38or just up front, $2,500, but with the promise that it gets better with time.
13:44Yeah, yeah.
13:45You know, why is that competitive against an FSD?
13:50And why is that price point kind of right for those early adopters and then the mass market beyond it?
13:57Yeah, I mean, pricing on some of these things is hard to set.
14:00Of course, FSD is a much higher price.
14:04You know, we do think that in time, these are things that, you know,
14:10we think there's maybe the potential to grow pricing, but it's something we don't know.
14:14And so we've spent a lot of time thinking about what's the initial price to charge for this.
14:18If you sign up for $2,500, you lock yourself into that price.
14:23It continues to get better.
14:24Is it a lifetime guarantee?
14:25Yeah, so it's something that's with your vehicle.
14:27If you buy that, it stays with your vehicle.
14:29And if we decide to grow the pricing over time, you know, it's to your advantage to lock it in sooner.
14:34But it's such a big area of investment for us.
14:38So it's not as if this is the price isn't warranted, meaning we spend more on our self-driving
14:44and AI technology than any other category of R&D by far within the business.
14:48So it's our central focus in terms of spending.
14:52It's our central focus in terms of R&D.
14:55And the challenge we've had with that is it's so much energy goes into it,
14:58but you have to build all this plumbing.
15:00You have to validate it.
15:01You have to go through the iteration cycles.
15:04And you don't see the fruit of that until a lot of work has been done.
15:08And so that's where we are today.
15:09And that was the reason we decided to have this AI days to pull the curtain back on all the things
15:14we're doing.
15:15You know, as I said jokingly to you as we sat down, it's shocking to me some of these things didn't leak.
15:19You know, there's a whole set of buildings behind us that are full of people.
15:22And we're working on this for years.
15:24I'm glad it didn't leak, but now we're able to fully expose it
15:29and fully pull back the curtain on what we're doing.
Be the first to comment