Summary

Rivian’s autonomy bet rests on a 2021–22 decision to discard its rules-based stack rather than extend it. Gen 2 launched from a hardware perspective in mid-2024 with “not a single line of shared code” or common perception hardware; Rivian then had to grow its car park into the data flywheel for neural-network training. Scaringe argues the progress curve through 2029–30 will look radically steeper because the architecture is now “truly AI,” not human-codified rules.
Scaringe believes only a handful of companies outside China possess the capital, GPUs, sensor control, and large car park required to remain competitive. He puts the field at “more than one, less than five,” potentially three or four, and agrees Rivian, Tesla, and Waymo belong in it. His stark call: legacy 1.0 autonomy systems have a “truly 0% chance” of progressing to competitiveness with neural approaches, while companies that fail at autonomy will shrink to nothing, “asymptotically approaching zero.”
Rivian built its own inference chip primarily to make advanced autonomy economical across every vehicle, not merely to prove technical capability. Guo characterized radars and LiDARs as cheap and onboard inference as roughly an order of magnitude more expensive than the perception stack; Scaringe called the “brain” the most expensive part and said bringing it in-house is how Rivian intends to put very high autonomy capability in every car.
The distinction between Levels 2, 3, and 4 is collapsing into a contest over rare corner cases and the final safety “nines.” Systems can feel identical through 99.9999% of driving, yet the fifth, sixth, or seventh nine can separate a routine trip from a severe collision. Scaringe expects that by 2030—“maybe sooner”—buying a car that cannot drive itself will feel like buying one without airbags or air conditioning.
Software-defined architecture is the foundation beneath autonomy, and Rivian is already licensing it through its $5.8 billion Volkswagen Group deal. Traditional cars contain 100–150 supplier-written ECU “islands”; Rivian’s zonal approach consolidates control into a few computers and supports roughly monthly over-the-air improvements. Incumbents can build, source, or accept that they will shrink, but Scaringe doubts an arm’s-length supplier relationship can sustain a continuously learning fleet.
R2 is Rivian’s attempt to move from a roughly $90,000 flagship into the heart of a U.S. new-car market averaging about $50,000. Starting at $45,000, it targets buyers in the $45,000–$55,000 range who are underserved by an EV market where Model 3 and Model Y represent roughly half of sales while overall EV adoption remains about 8%. Scaringe’s demand thesis is categorical: “The world doesn’t need another Model Y. The world needs another choice.”
Rivian expects autonomy products to differ through proprietary data and product behavior, even as safety remains the baseline. R2’s higher-capability camera, radar, and LiDAR stack is designed to make every vehicle a training platform—lighter than Waymo’s sensor approach but heavier than Tesla’s—while settings such as “mild, medium, and spicy” can express user preferences. Asked whether robotaxis could make vehicles more utilitarian, Scaringe said cars should continue to provide freedom and identity: vehicles should both enable and “inspire” experiences worth remembering.

Deep dive

1. Rivian treated its first autonomy stack as disposable

Scaringe says Rivian was conceived as a transportation and mobility company before its first product was defined, so autonomy was always strategic: personal transportation would eventually mean vehicles capable of driving themselves.
R1 launched at the end of 2021 with a third-party front camera feeding a rules-based planner. Rivian knew “the moment we launched” that the approach was wrong and decided in late 2021 or early 2022 on a clean-sheet reset. Its initial approach was Mobileye-centric.
Gen 2 arrived from a hardware perspective in mid-2024 with “not a single line of shared code” and no common perception or compute hardware. Rivian then needed enough cars on the road to create the data flywheel behind the capabilities shown in late 2025.
Guo’s challenge—grounded in autonomy investments she saw eight to ten years earlier—was that making the architectural and partner shift is genuinely hard. Scaringe agreed the transformer-driven shift was not gradual: for companies built around classical systems, “the vast majority of it is going to be pure throwaway.”

2. The key asset is a closed fleet-to-training loop

Scaringe’s vertical-integration case begins with complete control of cameras, radar, and LiDAR, including raw signals. Vehicles must identify noteworthy events, save them locally, upload the large payload over Wi-Fi when possible because LTE is expensive, and feed GPU training without an intermediary processing the evidence.
Independent autonomy vendors typically lack a sufficiently large car park; companies developing only a sensor set typically lack the vehicle architecture and fleet. Scaringe estimates “more than one, less than five” viable companies outside China—perhaps three or four—and agrees that Rivian, Tesla, and Waymo belong in the group while allowing “one or two others.”
His sharpest competitive call is that 1.0 systems stuck in that framework have a “truly 0% chance” of progressing to be competitive with a neural-network-based approach. Guo characterized the sensors as cheap and inference as roughly an order of magnitude more expensive than the perception stack; Scaringe called the “brain” the most expensive part and said Rivian built its own inference chip largely for cost, so every Rivian can support very high autonomy.

3. Autonomy levels are converging around the last safety “nines”

Guo summarized the old split: Level 2 meant camera-heavy consumer hardware, while Level 4 carried tens of thousands of dollars in perception. Scaringe said Level 4 was overbuilt for every consumer vehicle. Those worlds are now merging as the perception and compute distinction fades.
The remaining distinction is coverage of rare events. Level 2, 3, and 4 can feel identical through “three or four nines”; the fifth, sixth, or seventh nine contains obscure cases that consumers rarely encounter but that can produce “really terrible” collisions.
Development fleets have consequently expanded from hundreds of dedicated vehicles to thousands, with every car on the road contributing to the data fleet by identifying corner cases and models tested against both captured and simulated scenarios. By 2030—potentially sooner—Scaringe expects self-driving capability to be as expected as airbags.
Guo asked whether driving models will converge like LLMs. Scaringe’s answer: “There is no internet of driving data.” Rivian is going heavier on perception than Tesla but lighter than Waymo, adding LiDAR to R2 so it can help train the models and the whole fleet can become a training and data-acquisition platform; behavior can then vary through preferences such as “mild, medium, and spicy.”

4. Software-defined architecture leaves incumbents three choices

Traditional vehicles contain 100–150 ECUs, each running a supplier—or supplier-to-supplier—software island. Scaringe says this domain-based architecture exists in virtually every car on the road except Tesla and Rivian. It makes debugging and updates difficult because even a walk-up sequence spanning locks, HVAC, seats, lights, exterior sound, and audio may require coordination among ten parties.
Rivian’s zonal model uses one, two, or three computers running one operating system. The same sequence can be changed in minutes or an hour and shipped over the air; Rivian issues roughly one update monthly, typically adding features and refinements that make the car “notably better.”
Scaringe traces the incumbent architecture to outsourced fuel-injection computers that grew over 60–70 years into “a field of weeds.” Rivian’s $5.8 billion software-licensing deal licenses its network architecture and ECU topology to Volkswagen Group; other automakers must build, source, or accept that they will shrink, yet he doubts an arm’s-length vendor can operate a continuously learning autonomy loop.

5. R2 ties mass-market scale to more choice, not another imitation

R1’s average selling price is about $90,000, limiting volume despite Scaringe’s claim that R1S outsells the Tesla Model X roughly two-to-one. He calls R1S the best-selling premium electric SUV in the country among electric SUVs over $70,000, and the best-selling premium SUV, electric or non-electric, in California. R2 starts at $45,000, directly addressing the $45,000–$55,000 band around America’s roughly $50,000 average new-car price.
Guo’s blunt question was whether Americans actually want EVs. Scaringe pointed to only about 8% adoption: buyers below $70,000 can choose among well over 300 combustion-model lines, yet he sees only “more than one, less than three great choices” in EVs, with Model 3 and Model Y capturing roughly half the category.
His diagnosis is an “extreme lack of choice,” compounded by automakers copying the Model Y’s profile. Rivian respects that vehicle, but “the world doesn’t need another Model Y”; compelling alternatives must pull people from combustion vehicles, as the vast majority of R1 customers entered Rivian as first-time EV owners.
Asked whether autonomy and robotaxis could make transportation more utilitarian, Scaringe said cars should retain some of their emotional value, though that relationship will evolve, because they enable freedom and express identity. Rivian’s design test is whether a vehicle both enables and inspires memorable experiences: even the flashlight in the door is “an invitation to explore,” not merely another feature.

The AI Frontier: from Gemini 3 Deep Think distilling to Flash — Jeff Dean Anthropic's Super Bowl Ad: Who Won & Lost? | Sierra Hits $150M ARR: Is Customer Support Too Crowded?