Rivian Automotive, Inc. (RIVN) Earnings Call Transcript & Summary

Hello, everyone. I'm John Saager, Director Global Autos here at Evercore ISI. With me today is James Philbin, the SVP of Autonomy and AI for Rivian; and Chip Newcom, the VP of IR. James and Chip, I'll leave it to you guys to kick it off. Maybe give us an overview of Rivian and some of the key talking points you've been dealing with the investors.

Yes. Well, first, thank you so much for hosting us today, and great to be here to be able to talk to investors. So -- so in terms of the key areas, I can talk a little bit and then James will pass it over to you on autonomy. But really what our investors are focused in on right now. First is -- sorry, I'm getting some feedback all about the ramp of the R2 right now. So that's our midsized SUV that we launched back in June. And so as you can imagine, this Q3 being the first quarter that we'll have really had a full quarter of production, lots of questions that I have no doubt will talk about how is the ramp going from a production and a supply chain perspective, the targets that we laid out on the Q2 earnings call, and then, of course, a lot of focus on autonomy and our path there.

Yes. So yes, the team has been laser-focused on getting point-to-point ready this year. So we're very much on track there for initial rollout towards the end of this year. And we've really been on this sort of autonomy journey since the launch of Gen 2. And Gen 2 was built around beside our vertical integration in a very strong data loop so that we can collect customer data, customer sensor data together with the driving action. And that's really what's allowed us to drive the autonomy developments. We launched Universal Hands-Free last year, end of last year. That's the nation's, like, largest-scale hands-free system that you can get. And with point to point, we really think this will be the next step, important step in our autonomy journey as we iterate towards those kind of higher levels of autonomy in the stack.

Maybe you could talk about the development of your AV stack over time and to date and why you think multimodality is the correct approach?

Yes. So we believe very strongly in multimodality. So we think that the -- even if you accept the premise that cameras could be comparable to a human level perception. We think there's many cases where actually that's not enough where humans also make mistakes. And we think that the sort of societal bar will almost certainly be better than a human level for widespread autonomy. And so yes, for those reasons, we think that, yes, multi-modality really helps in the field. It also helps us accelerate our progress. So what we see is that when we have radars with cameras, it actually makes the system that much better. When we add LiDAR, it makes it better again. And so this is really a way to accelerate our progress as we're building the scale with R2 as well as providing a system -- a set of sensors that really provides a safe and performant and robust capability to our customers.

And how has that developed over time?

Yes. So we started, let's say, 3 years ago, where we brought the full autonomy stack in-house. So we -- everything from the lowest level camera drivers and ISPs all the way to the perception models running on the vehicle. So Gen 1 had a Mobileye system. With Gen 2 when we launched that, we pulled that system out and replaced it. And then as we've been gathering that data developing our perception models, we've been steadily improving first of all, with a Highway Assist feature then Universal Hands-Free and now the big focus is on point-to-point.

What are some of the technical hurdles that are still remaining for you guys as you look to roll out point-to-point by the end of the year?

Yes. So autonomy is really about kind of the edge cases. And so a big focus for us has been not how can we get more boring highway miles, right, which are not very informative for the model, the sort of gradient you get back when you train with that data is very, very minimal. But how we can find more interesting cases, which are really sort of driving kind of different behaviors from the models, so the gradient is large. So as part of that, we developed this grid system where we can actually using VLMs, modern AI techniques, we can scale up the collection of edge cases and so we currently have like a 3,000 query set of very long tail edge cases that are constantly querying against the customer fleet. And then we're using that to train the model. So I think that's a big part of the development process is actually getting those edge cases getting new things that the model hasn't seen before so that it can refine its driving behavior.

So you had mentioned point to point by the end of the year. Where does the autonomy program go from there? What are the next steps and ODDs that you're looking to unlock?

Yes. So we'll begin the rollout of point-to-point at the end of the year to probably our early adopter customers, and then we'll continue that rollout into next year with the aim of having it rolled out to the full Autonomy Plus customer fleet later next year. As obviously, we're rolling that out. We'll continue developing and improving the system. So we'll get improved disengagement rates, more scenarios with better handling. There's also some features. We probably will have out of ODD for the first version that we'll aim to add in, for example, parking lots. And then a big focus in on the Uber robotaxi partnership, where we're then iterating on that system getting it to a fully driverless level of performance. And that's both on the model side but also on the validation side, so increasingly leaning into simulation, large-scale simulation to measure and improve the system.

And the Uber robotaxi deployment, will that be domain defined in certain specific roads and areas? Or will that be something that you can use sort of everywhere?

So yes, the system is really like a single stack. So we will be training the model to handle everywhere. Now it may have fine-tuning data that is specific to particular cities or ODDs. And we've actually started building that through expert fleet deployments in our first 3 cities, San Francisco, Chicago and Miami. And we actually have an expert fleet driving around here right now. So all of that data will be fed into the model to really help kind of get that expert that Rivian driver policy dialed in, especially for point-to-point urban driving dense city driving.

And will the first point-to-point release be Gen 2 hardware only? Or does it require Gen 3?

No, it will be the whole on R1 Gen 2 and our Gen 2 Phase II and Phase III fleets.

Okay. But where drivers, Hands-Free, eyes-off where drivers don't really need to be actively involved is sort of a major milestone in the industry. What's the largest gating factor for you guys to get there? Is it validated miles, regulatory liability sign off? Or is there some sort of technical hurdle that we're having?

We have Yes, there's probably 3 pieces to it. I think one is just raw model performance as we're improving the model performance of say in the long tail, that will help. There's then the validation piece like how you can actually measure that model performance. And you really don't want to do that by having to drive every release, you want to do it offline in the cloud. And then the third piece, yes, is regulatory. Currently, it's a complete patch work in the U.S. So a lot of states view eyes off essentially in the same bucket as L4. And so that means in some states, you can ship that system, others, there is a pathway, but you have to jump through a certain number of -- go through a certain regulatory process. And in other states like New York, currently, there is no pathway. So it's a patchwork.

So you've got sort of navigate that.

Yes.

Playing hopscotch.

Yes, definitely.

So paid autonomy is the $2,500 upfront or $50 a month launched in April and you guys have talked about that being encouraging take rates to date. Can you frame the attach rate on new deliveries versus the installed base? Is there a gap or a difference there? And then the sort of upfront versus the subscription mix?

So we haven't quantified that as of yet. For everyone who's buying an R2 right now as part of the Launch Edition package, they are getting Autonomy Plus as part of their purchase. So we'll see as we shift beyond the launch edition package, what that take rate looks like for the R2. But as you pointed out, for the R1, it's been better than we expected.

Yes. Can you just frame out how much Autonomy Plus contributes to your software revenues?

It's still relatively small, just given the size of the car parc and the monthly price or the sort of full price if someone is purchasing it upfront.

Okay. And then RAP1 development is sort of on track and in the final testing phases of the production. So like, what's the BOM delta of the dual RAP1 plus all the sensor suite, the 11 cameras and radars? Is that cost carried in the R2 margin? Or is that offset by autonomy costs sort of pricing?

So from a revenue recognition perspective, it's going to be recognized within autos and auto COGS. So there will be a cost around putting LiDAR into the vehicle. Now the cost of LiDARs has come down significantly. So you're talking on the order of a few hundred dollars versus if you went back several years ago where it was thousands or even tens of thousands of dollars within the vehicle. So that will be beared within the cost of the vehicle. And similarly, as we launch our Gen 3 computer in the vehicle, we actually see a benefit in the sense that we're no longer paying away that margin to our silicon provider. It will be our own in-house silicon.

From an autonomy perspective, how do you think about your position relative to the competition, most specifically Tesla, but also Waymo and all the other players that have spoken at the conference today. Where do you think you fit? And how do you think you either catch up or surpass some of those other players?

Yes. I think we're -- I mean, a, we're vertically integrated. So I think we're both building the vehicle and the autonomous stack. And I think we have the benefit that we have a very large shipped car parc already collecting data. So that means that we get huge diversity already from our customer fleet vehicles. I think in comparison, a Level 4 company like Waymo, they actually have to go city by city to really get that data and do their validation we get the sort of benefit of being able to train the model and from the get go with all of that data. I think the second is that we are really leaned into like the end-to-end approach. We think that is the right way to build the autonomy systems longer term. amazing when you see the capabilities that kind of emerge from that system. I'll give you an example. So narrow negotiations, you're on a narrow roads, 2 cars coming and someone needs to pull out. So I actually worked on these systems at Level 4 companies in the past, very, very complicated set of heuristics or the things that you have to put in place to get that to work. And this model kind of -- it just falls out of the model. So there's nothing extra needed -- so I think when you see things like that, you realize, okay, this really is the right way to build this system the right scalable approach. And then -- but in comparison to say, Tesla, we really believe in multimodalities. And don't think there's really a tension between training end-to-end or and using multimodal sensing. The model sees every sensor essentially the same way. It's a bag of numbers and it encodes them using its own encoders. So it's able to figure out those cross correlations and how to best benefit the output using its inputs.

Right. You mentioned the large and growing fleet bringing in data today, how do you filter through what data is useful and what data is not? I mean, you mentioned the 3,000 parameters that you're looking at. Is that the filter process for you to collect all of the data -- like I drive Rivian. I come home. When does that data get downloaded? And how does that work? And how does that play into your advantage?

Yes. So it has a whole trigger system. So initially, when we launched the triggers were relatively simplistic. So it would simply be on things like gear changes or indicators. We now have a much more sophisticated trigger system that actually has machine learning in the loop as well. So we've found triggers that are interest -- we found events that are interesting, actually train a model to trigger on those and then ship that model to customer vehicles. And in terms of how the uploads happen, we have a tiering method. So some triggers are considered critical. They'll actually go up by LTE. Other triggers are less critical. They'll go up via WiFi and then LTE only if we have to. And we can also push triggers dynamically. So we can actually -- you can sit at your desk, say, write a new trigger, push it and then you immediately get -- start getting that data back. So it doesn't have to wait for the OTA cycle to get the new triggers to the fleet.

And then how do you think about that sort of real-world data relative to the simulation data and the simulation process what is the efficacy of one versus the other? And I would imagine it's a spectrum that everybody is operating across where you're either more SIM heavy or you're more real world heavy. You guys are falling more and more on the real-world heavy side? Or -- how should I think about that?

Yes. I mean for me, there's no real replacement for real-world data. I think simulation can augment, but the real world always has the most interesting edge cases you can't really come up with. Sitting you can't really imagine them and they're often -- it often isn't the one thing. It's actually a -- it's a combination of multiple long-tail events that all sort of congregate into a single scenario. And so those things are very difficult to even think up and then actually quite difficult to simulate as well, -- so for us, it's really -- the real-world data is the kind of foundation. Now we do have abilities to resimulate those real scenarios. And so that's something we believe in. But for us, we have a huge data hose essentially of data, and that's what we think is the right way to do these validations.

kind of the primary driver there. You've said RAP1 can easily be extended to solve other physical AI problems such as robotics. We've recently written a report talking about how we feel like the automotive industry, specifically in the U.S. is behind in terms of using their capabilities in the automotive space to enter into other spaces where they have similar applications. So can you talk about that opportunity for Rivian more broadly to focus on other end applications? Is that something that you take seriously now? Or is something that you see happening down the line because we've seen -- I think there's more than 15 Chinese OEMs, there's Hyundai, there's a couple of the European OEMs have all made an entry in some form or fashion into sort of the embodied AI space to leverage the capabilities that they already have from automotive.

Yes. I mean I think in Rivian's case, part of what we did is set up Mind Robotics, where as RJ was looking at the current robotic solutions, very specifically around manufacturing and the use case there for what we're thinking about that was part of the genesis then of thinking, how can we have extensibility into developing robotic solutions for manufacturing. Now it's not like a general purpose thing that we're trying to create. It's really focused on that specific area. But to a certain extent, we'd have to defer to the Mind team on where they're looking to head at this point. But part of the partnership with them there is going to be around the fact that we have an in-place manufacturing line and then being able to use that data as they think about developing their own solutions.

Now you can sort of deploy it in your own factories and then they get the data flywheel going very similar to the autonomous approach.

Exactly.

So what do you expect initial production capacity look like? And how do you prioritize RAP1 for Rivian before entertaining outside supply contracts?

Well, right now, the priority for RAP1 is going to be going into the R2. So for employees, the plan is to start rolling out RAP1. So our third-generation computer plus the LiDAR to go into the vehicle later this year for employees and then start rolling that out to customers next year.

Okay. Uber had invested. I think it was up to $1.25 billion through 2031 with $300 million at signing and then sort of a milestone based, you have $250 million targeted by the end of this year. what specifically triggers the $250 million? And then there is the 10,000 autonomous R2 purchase commitment? Is that a firm order commitment? Or is it contingent on some specific autonomy metrics?

Well, so we haven't given specific details in terms of what's going to unlock each of the funding milestones. The way you can think about it is going to be around achieving certain technical milestones out of James' team on our march towards getting to an L4 capable robotaxi and then the expansion of that service where we said the goal is by 2031 to have rolled out the service into 25 markets including at least 1 in Europe. So that's what would unlock each of those triggers. But our expectation remains that we'll be able to unlock the second trigger for the $250 million investment here in the fourth quarter. In terms of the actual purchase of the initial 10,000 vehicles from Uber, clearly, we need to have gotten to an L4 capable service before they would want to purchase those vehicles. So we would need to make progress there.

Right. So -- and you haven't given any sense of timing on that publicly?

So we've said that the goal is to have the service available by late 2028.