Rivian's shift to neural-net autonomy was a clean-sheet rebuild, not an upgrade.
At the end of 2021 Rivian scrapped its rules-based Gen One perception-and-planner stack entirely and rebuilt from scratch for Gen Two, with no shared code or hardware between the two systems. Scaringe frames the whole industry's shift to transformer-based, AI-native autonomy the same way: the vast majority of pre-2022 rules-based investment across the industry is throwaway, not a foundation to build on.
autonomous-driving
Scaringe believes fewer than five companies outside China have the ingredients to win at AI-native autonomy.
He names the requirements as full control of the perception platform (no intermediary processing sensor data), a large enough vehicle fleet to generate training-grade data, in-house or dedicated GPU training capacity, and a robust vehicle data architecture. He puts the real number at one to three or four companies, explicitly including Rivian, Tesla, and Waymo, and says companies stuck on rules-based 1.0 approaches have close to a 0% chance of becoming competitive.
autonomous-driving
Rivian built its own onboard inference chip for cost, not capability, reasons.
Cameras, radar, and LiDAR are now cheap; Scaringe says the onboard inference compute is the most expensive part of the autonomy stack, more expensive than the entire perception hardware combined. Rivian brought chip design in-house to strip cost out of the system so high levels of autonomy compute can ship standard on every car, not just premium trims.
autonomous-driving
The old level-2-versus-level-4 autonomy distinction is dissolving into a single spectrum of corner-case handling.
A few years ago level-2 (camera-plus-a-few-radars) and level-4 (camera-plus-heavy-LiDAR) systems were architecturally distinct products aimed at different markets. Scaringe says that gap has closed: perception and compute stacks now look similar across levels, and the real differentiator is how well the model covers rare, long-tail corner cases, since a level-2 and level-4 system feel identical 99.9999% of the time.
autonomous-driving
Rivian expects autonomy to become a mandatory feature, not an option, within a few years.
Scaringe predicts that by around 2030 it will be as inconceivable to buy a car without high levels of self-driving as it is today to buy one without airbags or air conditioning, comparing the shift to those earlier features moving from optional to standard.
autonomous-driving
Legacy automakers' domain-based electronics architecture traces back to 1960s fuel-injection outsourcing and now makes coordinated feature updates prohibitively expensive.
Car companies decided fuel-injection computers weren't a core competency and pushed that first ECU to suppliers like Bosch. Over 70 years every new function got its own ECU, so a modern car has 100-150 of them, each running code from a different supplier (often a tier-two sub-supplier), making even a simple feature like walk-up unlock-plus-climate-plus-seat-adjustment require coordinating ten-plus separate teams.
software-defined-vehicles
Rivian and Tesla are, by Scaringe's account, the only two automakers running a zonal architecture instead of a domain-based one.
A zonal architecture consolidates control into one to three computers running a single operating system, which is what lets Rivian change a whole feature sequence in minutes to an hour and ship roughly one OTA update a month with new features and refinements, versus the multi-team coordination cost of a domain-based system.
software-defined-vehicles
Rivian licensed its zonal architecture and ECU topology to Volkswagen Group for $5.8 billion, betting most legacy OEMs cannot build software-defined architecture in-house.
Scaringe frames car companies as facing three choices: accept they will shrink, build a software-defined architecture and autonomy themselves (hard, because it isn't their organizational DNA), or source it from a third party - where most existing third parties are legacy rules-based autonomy vendors poorly suited to a constantly-learning, constantly-updating data flywheel model.
vertical-integration
Scaringe attributes the US's ~8% EV adoption rate to a lack of product choice, not lack of demand.
He notes there are 300-plus vehicle models available under $70,000 in gas cars, but in EVs there are effectively only one to three genuinely distinct good choices, dominated by Tesla's Model 3 and Model Y (roughly half of US EV market share). Because of Model Y's success, most other EV entrants copy its silhouette rather than offering a different form factor, which he calls 'a shocking lack of choice.'
ev-market-strategy
R2, starting near $45,000, is Rivian's first mass-market bet, aimed at the $45,000-$55,000 range where the average new car sells.
R1 has a roughly $90,000 average selling price and is the best-selling premium electric SUV in the US and California, but its price caps achievable volume. R2 is positioned to pull buyers out of internal-combustion vehicles by competing at the actual average new-car price point rather than the premium segment.
ev-market-strategy
Rivian's product philosophy is to build the best vehicle first and let EV conversion follow, rather than build 'a good EV.'
Scaringe says the majority of R1 buyers were first-time EV owners, which he treats as evidence the strategy works: a genuinely compelling vehicle (performance, range, storage, dynamics) draws in buyers who weren't shopping for an EV specifically, expanding the EV market instead of just moving customers between existing EV brands.
ev-market-strategy
Unlike language models, autonomous driving has no shared 'internet of data,' so fleets and sensor choices - not data availability - are the moat.
Scaringe contrasts LLM training, where models converge because they train on roughly the same internet corpus, with driving, where there is no equivalent public dataset. Rivian's differentiation is a heavier sensor investment than Tesla (adding LiDAR with R2, high-dynamic-range and low-light cameras, radar) used to generate ground truth and train the camera-based model at scale across its growing fleet.
autonomous-driving
Companies
Rivian - Scaringe's EV maker; subject of the whole conversation on autonomy strategy, R2, and architecture
Tesla - Repeatedly cited as one of the few companies (with Rivian and Waymo) that has the ingredients to compete on AI-native autonomy; Model 3/Model Y dominate US EV market share
Waymo - Named as one of the small handful of companies with the data, capital, and compute to build competitive neural-net autonomy
Volkswagen Group - Rivian sold VW a $5.8 billion software license for its zonal network architecture and ECU topology across VW's brands
Nvidia - Cited as a vendor Rivian could have bought onboard inference chips from instead of building its own
Bosch - Named as an example of the fuel-injection-era supplier model that produced today's fragmented ECU architecture
Techniques and frameworks
Zonal / software-defined vehicle architecture - A small number of computers (one to three) running one OS across the car, replacing the 100-150 separate ECUs in a traditional domain-based architecture; enables fast OTA updates
Neural-net-based (end-to-end) autonomy - Rivian's post-2022 approach: a single trained model replaces the rules-based perception-plus-planner stack of its 1.0 system
In-house data flywheel / ground-truth fleet - Every vehicle captures triggered events and sends them offboard over Wi-Fi to train the model; LiDAR on the fleet is used to generate ground truth for training the camera-only inference path
Vertical integration - Rivian builds its own perception platform, in-house inference chip, motors, inverters, and software rather than buying from suppliers
Summary
Sarah Guo talks with Rivian founder and CEO RJ Scaringe about the company's bet that autonomy, not just electrification, is the core of its business. Scaringe walks through Rivian's abrupt pivot at the end of 2021 from a rules-based, camera-plus-third-party perception stack to a fully in-house, neural-net-based system built clean-sheet for its Gen Two vehicles. He argues the whole industry is mid-way through the same discontinuity the LLM world went through with transformers: architectures that were state of the art in 2020 are now close to worthless, and companies still relying on 1.0, rules-based autonomy have close to no path to competitiveness. He estimates only one to four companies outside China - Rivian, Tesla, and Waymo among them - have the full ingredient list (proprietary perception, fleet-scale data, training compute, in-house inference) to compete.
A recurring thread is why Rivian vertically integrates almost everything, including a proprietary onboard inference chip. Scaringe is explicit that the chip decision was about cost, not capability - the onboard compute needed to run an autonomy model has become the single most expensive part of the system, more than the entire camera/radar/LiDAR sensor stack, and Rivian wants that capability standard on every car rather than gated to premium trims. He also collapses the old level-2/level-4 autonomy distinction: those categories used to imply very different hardware, but now the hardware has converged and the real differentiator is how well a system's neural net handles the long tail of rare corner cases that show up only after 99.9999% of ordinary driving.
The conversation's second half is about software-defined vehicle architecture, which Scaringe treats as the precondition for autonomy rather than a separate topic. He traces the mess of modern car electronics - 100 to 150 separate ECUs, each written by a different supplier or sub-supplier - back to a decision in the 1960s to outsource the first fuel-injection computer, arguing that decision snowballed into an architecture so fragmented that even small feature changes require coordinating ten-plus teams. Rivian and Tesla, in his telling, are the only two automakers that instead run a zonal architecture with one to three computers on a single OS, which is what lets Rivian ship roughly monthly OTA updates. Rivian has since turned this into a business line, licensing its zonal architecture and ECU topology to Volkswagen Group for $5.8 billion - a bet that most legacy OEMs cannot build this in-house and have no good third party to buy it from instead.
On the EV market itself, Scaringe rejects the idea that Americans don't want EVs, attributing the ~8% US adoption rate to a lack of genuine product choice rather than demand: outside Tesla's Model 3 and Model Y, which alone command roughly half of US EV sales, most other EV entrants are visually and functionally near-clones of the Model Y. He frames R2, launching around $45,000, as Rivian's answer - its first true mass-market product aimed at the price band where the average new car actually sells, in contrast to R1's roughly $90,000 average selling price. The stated philosophy is to build "the best possible vehicle" rather than "a good EV," pointing to R1 data showing most buyers were first-time EV owners as evidence the approach expands the EV market rather than just poaching Tesla customers.
The episode closes on a more reflective note about cars as identity and self-expression, with Scaringe suggesting that as autonomy and robotaxis reduce the purely functional role of car ownership, brands will increasingly compete on how a vehicle makes its owner feel and what it inspires them to do, rather than on raw utility alone.
Notable Quotes
"By 2030, it'll be inconceivable to buy a car and not expect it to drive itself." - RJ Scaringe
"The world doesn't need another Model Y. The world needs another choice." - RJ Scaringe
"Radars are extremely cheap, LiDARs are now very, very cheap. But the really expensive part of the system is actually the onboard inference." - RJ Scaringe
"The companies that do this well will exist. The companies that don't do this well - like I feel really strongly this - they will not exist. They will shrink to shrink to nothing, asymptotically approach zero." - RJ Scaringe
"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." - RJ Scaringe