NVIDIA projected that the global robotaxi market will reach $400 billion by 2035, encompassing more than 6 million commercial vehicles operating worldwide. Every major commercial robotaxi program operating today relies on the company’s modular computing stack across AI model training, simulation, and in-vehicle processing, NVIDIA announced on Thursday.
NVIDIA structured its platform around three computing tiers to support fleet scaling. The setup divides workloads between off-board model training, cloud-based simulation and validation, and real-time computing inside the vehicle.
The three-computer architecture
Developers train driving models using NVIDIA DGX systems. The workflow incorporates the Alpamayo portfolio of open reasoning vision-language-action models, physical AI datasets, and reinforcement learning blueprints. These reasoning models break driving problems into sequential steps to plot safe vehicle trajectories during long-tail edge cases.
Engineers manage simulation and closed-loop validation on NVIDIA RTX PRO servers. NVIDIA Omniverse NuRec models reconstruct recorded real-world driving data, while NVIDIA Cosmos foundation models generate variations in traffic, weather, and lighting. The accompanying AlpaSim framework evaluates reasoning models in virtual environments before road testing begins.
Vehicles handle onboard processing through the DRIVE Hyperion 10 reference architecture. The hardware integrates dual DRIVE AGX Thor systems-on-a-chip, built on the Blackwell platform, alongside 14 cameras, nine radars, three lidars, and 12 ultrasonic sensors. NVIDIA Halos provides the Halos operating system and safety frameworks to maintain fail-operational performance if a sensor or chip fails.
Global fleet expansion
Uber plans to expand its DRIVE Hyperion fleet to 28 cities by 2028. Uber and NVIDIA are also constructing a data factory on NVIDIA Cosmos to curate driving records of rare road events, working alongside partners including Wayve, Momenta, Zoox, Waabi, and Avride.
Automakers are integrating the architecture into passenger production lines. Mercedes-Benz is developing an S-Class robotaxi fleet with Uber using DRIVE Hyperion and Alpamayo tools, while Lucid, Nuro, and Uber are building a driverless service powered by DRIVE AGX Thor. Stellantis is collaborating with Wayve and Uber on driverless mobility services, and Tesla trains its autonomous driving neural networks on NVIDIA supercomputers.
Equipment manufacturers are deploying the platform across regional transit networks. TIER IV and Isuzu are launching Level 4 autonomous buses on DRIVE Hyperion, Lenovo is supplying its Thor-based AD1 domain controller to SWM, and Geely brand Zeekr has adopted DRIVE AGX Thor for centralized vehicle computing.
