Rivian’s December 2025 Autonomy & AI Day brought custom compute, an expanded sensor stack and its Large Driving Model into one public architecture story. The company reported 1,600 sparse INT8 TOPS for its Gen 3 Autonomy Computer, not for a single RAP1 processor.
The useful question is what the architecture asks of the data underneath it. Additional sensing can provide another view of a scene, but it also creates more synchronization, calibration and validation work. A faster processor does not remove those requirements. Nor does a sensor list establish which driving system is safer.
The silicon is designed around data movement, not peak compute
Rivian’s Q4 2025 shareholder letter specifies an 800-TOPS neural-network engine in each RAP1, with two processors collectively providing 1,600 sparse INT8 TOPS. It describes a 5nm design, tightly coupled memory and the RivLink interconnect. These are company-reported hardware specifications, not an independent workload benchmark.
Arm’s announcement identifies Armv9 and Cortex-A720AE in the platform, alongside processors dedicated to real-time safety functions. That supports a discussion of the compute architecture. It does not establish a vehicle-level safety outcome.
Our engineering reading is that the interesting measurement is the complete workload: sensor input, preprocessing, inference, memory transfers and control deadlines. Peak arithmetic throughput leaves out sustained performance, power, thermal limits and scheduling. Comparisons need the same precision, sparsity assumptions and representative workload. We would not rank systems from headline TOPS alone.
LiDAR is a training-data decision as much as a safety decision
The event announcement describes adding LiDAR to future R2 models alongside cameras and radar, including a training-data role. The later shareholder letter specifies 11 cameras, five radars and one LiDAR for the announced third-generation suite. Those statements describe the intended configuration; they are not evidence that every R2 already carries it.
LiDAR supplies range measurements, but a range return is not automatically a correct object label or a complete description of a scene. A training pipeline still needs to account for calibration, timing, coverage and uncertain observations. An image annotation projected into a point cloud can be wrong even when both individual sensors are functioning.
The useful deliverable is therefore more than a collection of files. It includes the sensor configuration, timestamps, coordinate frames, calibration version and review decisions needed to interpret the observations together. Where a signal is estimated or missing, that distinction should remain visible.
The Large Driving Model joins the end-to-end convergence
Rivian described an end-to-end model and a fleet-data feedback process at the event. That gives a high-level view of the proposed learning architecture. It does not disclose enough about the complete training corpus, selection process or evaluation setup to reproduce its results.
For data teams, end-to-end learning changes the questions rather than eliminating them. Which observations and actions are paired? What was the driver or controller trying to do? Where did control change hands? Which sequences belong in training, and which must remain held out?
Those questions require a record of each sequence and the transformation applied to it. They do not imply that every model needs the same label taxonomy, or that a human must annotate every frame. The right supervision depends on the learning objective and should be tested against it.
Fleet size is not an evaluation result
A fleet-mile total measures exposure, not necessarily retained training data or useful coverage. To compare learning systems, we would want to know what was recorded, what survived selection, how often relevant conditions occurred and how the evaluation set was separated from training.
The original article inferred that a change in model architecture discounted an established fleet’s data advantage. The public sources reviewed here do not establish that conclusion. Older recordings may remain useful, and a new sensor configuration may require additional data. Neither fact supplies a conversion rate between two fleets’ miles.
A stronger evaluation would compare defined driving conditions, interventions and failure categories under a documented protocol. Aggregate mileage, sensor count and benchmark scores answer different questions. None should silently substitute for the others.
Custom silicon economics need an explicit model
The original article gave development-cost, staffing and production-volume thresholds without a cited cost model. Those figures are removed. The reviewed announcements do not establish a universal break-even volume or prove that a particular partnership makes the investment economical.
A useful model would separate development and validation expense, recurring hardware costs, software maintenance, supplier commitments and expected production. It would also value the engineering flexibility gained or lost by owning the design. Those inputs need evidence and sensitivity ranges; a single annual vehicle threshold hides too much.
This is a build-versus-buy question with several possible answers. Public architecture descriptions alone are insufficient to judge the financial return.
The bottleneck moves to multi-modal data infrastructure
Our narrower thesis is that a richer sensor stack increases the importance of preserving context. It does not establish data as the only bottleneck, or guarantee that adding sensors improves a trained policy.
Four checks help turn that thesis into an engineering plan:
- Temporal consistency: preserve timestamps, sequence boundaries and object identity across the interval being evaluated.
- Sensor alignment: state coordinate frames, calibration dependencies and how uncertain or missing observations are handled.
- Scenario coverage: define the conditions the program must capture, including failures and interventions relevant to its objective.
- Evaluation separation: document which locations, sequences or other related groups are held out and how leakage is checked.
These are evaluation questions, not claims about a particular manufacturer’s internal process. The autonomy workflow and labeling reference provide the relevant Avala entry points for discussing a scoped program.
The bet we would take
The announcement is interesting because it connects sensing, compute and learning. The case is weaker when it becomes a claim that one modality guarantees superior safety, that an architecture change resets the competitive field, or that a projected launch is already a delivered capability.
Our bet is on teams that make the loop measurable: identify a failure, define the missing evidence, acquire or review the relevant data, and test the next model against a held-out evaluation. Hardware specifications inform that work. They do not replace the result.
The December 2025 launch targets were forward-looking statements. This article does not verify their subsequent delivery or establish current vehicle availability. Readers evaluating a vehicle or feature should consult the manufacturer’s current specifications and operating requirements.
Avala Robotics explores the technical foundations of Physical AI, with a focus on the data infrastructure that connects sensor recordings to model evaluation.


