NVIDIA Says Its AI Stack Powers Every Major Commercial Robotaxi Program
NVIDIA says every major commercial robotaxi program uses its AI stack, as Uber targets 28 cities by 2028 and the market eyes $400 billion by 2035 globally.
Summary
Published September 10, 2026, the outlook calls robotaxis physical AI’s first commercial breakthrough and projects a $400 billion market exceeding 6 million commercial vehicles by 2035. NVIDIA says every major commercially scaled program uses its three-computer platform: DGX training; Omniverse and Cosmos on RTX PRO Servers for simulation and validation; DRIVE Hyperion with DRIVE AGX onboard. Alpamayo provides open vision language action models, AlpaSim, datasets, reinforcement learning blueprints, post-training and distillation tools. Meta-action and chain-of-thought data cut minimum average displacement error 43%, from 2.08 to 1.18. NuRec rebuilds sensor scenarios; Cosmos multiplies thousands of corner cases into millions of variants; Halos OS supports inspection, certification, simulation and continuous cloud-to-car testing.
Level 4-ready Hyperion 10 pairs two Blackwell-based DRIVE AGX Thor chips with 14 HD cameras, nine radars, three lidars and 12 ultrasonics, providing redundant 360-degree sensing and fail-operational driving. Uber targets 28 cities by 2028 and is building a Cosmos data factory with NVIDIA. Partners include Autobrains, Avride, Lucid, May Mobility, Mercedes-Benz, Momenta, Nissan, Nuro, Pony.ai, Stellantis, Waabi, Wayve, WeRide and Zoox. May Mobility serves Uber and Lyft, including Atlanta; Bolt targets Europe; Lyft plans Hyperion fleets; Grab and WeRide plan GXR deployments in Southeast Asia; Waymo partners on autonomous computing.
Wayve, Nissan and Uber prototype globally; Autobrains works with Uber in Munich and VinFast in Southeast Asia; Zoox uses DRIVE in vehicles and cloud; Momenta uses AGX and DriveOS; Pony.ai, DeepRoute.ai and Zeekr use Thor; Tensor’s level 4 Robocar uses eight Thor chips; Waabi deploys with Uber; TIER IV and Isuzu build level 4 buses; Lenovo supplies SWM’s AD1 controller. Tesla trains on NVIDIA supercomputers. Mercedes-Benz, Uber and NVIDIA target an S-Class L4 service; Stellantis, Wayve and Uber plan L4 mobility; Lucid, Nuro and Uber plan a global service; Hyundai Motor, Kia and potentially Motional expand Hyperion work; Geely partners plan commercialization.
Positives
- $400 billion in projected market value and more than 6 million commercial vehicles by 2035 signal substantial robotaxi growth.
- 43% lower minimum average displacement error, from 2.08 to 1.18, followed the addition of meta-action and chain-of-thought reasoning data.
- 28 cities by 2028 is Uber’s target for scaling its NVIDIA DRIVE Hyperion fleet.
- Two DRIVE AGX Thor chips and 38 sensors give Hyperion 10 redundant, 360-degree perception for fail-operational driving.
- Thousands of captured corner cases can become millions of synthetic variations through NuRec and Cosmos.
Risks & concerns
- Thousands of fleet vehicles require enormous computing capacity across model preparation, training, simulation, validation and real-time driving.
- Physical road miles alone cannot capture enough rare, long-tail scenarios for comprehensive robotaxi validation.
- Complex driving situations remain difficult enough to require stepwise reasoning models that select among possible trajectories.
- Many announced collaborations lack deployment dates, fleet sizes or launch markets, leaving commercialization details limited.
- Tensor’s level 4 Robocar uses eight DRIVE AGX Thor chips, highlighting the substantial onboard computing required.