Autonomous Driving - AI Infrastructure

Autonomous vehicle (AV) developers train their driving models on dedicated AI compute clusters. This page compares the clusters run by Tesla, Waymo, XPeng, Huawei, NIO and Rivian, the chips inside them, and the Nvidia training and simulation platform that other developers build on.

Datacenter build-out, power, cooling and facility design are covered on DatacentersX. See: Autonomous Vehicle AI Datacenters - DatacentersX.

How Compute Scale Affects Model Development

Rule-based ADAS (advanced driver assistance systems) improved mainly with engineering hours. End-to-end neural networks improve with training data volume, model architecture and compute scale, and compute is the most capital-intensive of the three and the hardest to replicate quickly.

• Iteration speed: a training job that takes four weeks on a small cluster may take three days on a large one, allowing more experiments per quarter.
• Model size: Tesla FSD v14.3's neural network is reportedly 10x larger than prior versions, which needs a proportionally larger training cluster.
• Simulation: Waymo and Tesla run millions of simulated miles per day on GPU or TPU capacity separate from model training.
• Auto-labeling: pipelines that turn raw fleet video into training data need sustained high-throughput compute; a bottleneck here slows the whole over-the-air (OTA) update loop.
• Reinforcement learning (RL): introduced in Tesla FSD v14.x and XPeng VLA 2.0, RL needs continuous rollouts and reward computation and uses much more compute than supervised learning.

Tesla: Cortex and Dojo

• Clusters: Cortex (primary training cluster) and Dojo (custom training supercomputer) at Giga Texas, Austin, with additional capacity in Palo Alto
• Cortex silicon: Nvidia H100 (80 GB HBM3), about 50,000 H100-equivalent GPUs as of 2024 and still expanding
• Dojo silicon: Tesla D1 training chip (362 TFLOPS BF16); ExaPOD architecture with 120 D1 chips per training tile and 9 TB/s die-to-die interconnect per tile
• AI5: Tesla's next inference chip, announced 2025-2026 for both vehicle inference and datacenter training; Musk describes it as co-designed with the FSD software stack
• Terafab: Tesla, SpaceX and xAI joint venture announced March 2026 for chip fabrication at Giga Texas, aimed at reducing dependence on TSMC and Nvidia supply
• Workloads: FSD end-to-end training, auto-labeling, simulation and Optimus robot model training
• Training data: about 6 million fleet vehicles worldwide; shadow mode collects data beyond enrolled FSD users
• Release cadence: FSD major releases every two to six weeks, with more frequent point releases

Waymo: Google TPU Infrastructure

• Silicon: Google TPU v4 and v5 pods through Google Cloud, available to Waymo as an Alphabet company. TPU v5p delivers about 459 TFLOPS per chip with high-bandwidth interconnect within a pod.
• Simulation City: Waymo's main workload; it reconstructs real-world scenarios and runs synthetic variants, which produce most of Waymo's effective training miles
• Fleet: about 1,500 vehicles across all cities as of 2025; simulation makes up for the small physical fleet
• Data strategy: each vehicle carries LiDAR, cameras and radar and produces high-fidelity 3D data, which Waymo argues is worth more per mile than camera-only data
• Constraint: depends on Google Cloud and shares TPU access with other Alphabet workloads; less vertically integrated than Tesla

XPeng: Turing Cluster

• Silicon: about 3,000 Nvidia H800 GPUs (the China export-compliant version of the H100) as of 2024, still expanding
• In-house chip: the Turing AI chip, used for vehicle inference in the XNGP system and planned for the training cluster to reduce Nvidia dependence; the cluster is named after it
• Iteration speed: XNGP model updates every one to two days
• Training data: more than 1 billion km of video; 429,000 vehicles delivered in 2025 (125% growth year over year)
• Volkswagen: VLA 2.0 open-sourced to VW, and the Turing chip supplied to VW models from 2026, extending XPeng's data input beyond its own fleet
• Export control exposure: tighter US restrictions on the H800 could limit cluster expansion

Huawei: Ascend Cluster

• Silicon: Huawei Ascend 910B and 910C, developed as a domestic alternative to Nvidia under US export controls; the 910C is reported at about 256 TFLOPS (FP16)
• Scale: undisclosed; HIMA alliance training runs on dedicated Ascend infrastructure, and Huawei Cloud sells AI compute more broadly
• Vertical integration: Huawei builds the training chip (Ascend), inference chip (HiSilicon/Kirin), training cluster, ADS driving software and cockpit OS, which insulates it from export control disruption more than other Chinese players
• Fleet data: 7.28 billion km from HIMA alliance vehicles as of January 2026; 28 vehicle models across five OEM partners, with 80+ models targeted by end of 2026
• Ascend vs. Nvidia: the 910C is broadly competitive with the Nvidia A100 for training but trails the H100
• Constraint: the CANN (Compute Architecture for Neural Networks) software stack is less mature than Nvidia CUDA and has a smaller model ecosystem, though a captive workload like ADS training needs less ecosystem breadth

NIO: Adam Supercomputer

• Cluster: Adam, announced in 2021 with about 1,000 Nvidia A100 GPUs and expanded since
• Sensor platform: NIO Aquila, with 33 sensors including LiDAR, cameras, radar and ultrasonic, producing high-fidelity training data per vehicle
• Business model: NIO Autonomous Driving (NAD) is sold as a subscription combining Aquila hardware and Adam compute
• Constraint: smaller fleet than XPeng or Tesla, a modest cluster and dependence on Nvidia supply

Rivian: RAP1 and Training

• In-house chip: RAP1 (Rivian Autonomy Processor 1), dual-chip in the Gen 3 platform with 1,600 TOPS total; designed for vehicle inference and as a basis for training acceleration
• Training cluster: not publicly disclosed; the Volkswagen joint venture (announced 2024, $5.8B) adds software infrastructure investment
• Data: LiDAR point clouds from the R1 and R2 consumer fleet supply 3D ground truth; the Uber partnership (50,000 robotaxi vehicles) will add fleet data after 2028
• Large Driving Model: trained with Group-Relative Policy Optimization (GRPO), the RL technique used in DeepSeek R1, adapted for driving policy
• Constraint: a small consumer fleet relative to Tesla and XPeng; the Georgia robotaxi factory is still under construction

Training Cluster Comparison

Cluster Notes

• Tesla: In-house D1, with AI5 pending. Fleet data from ~6M vehicles globally. Low export control risk as a US-based OEM.
• Waymo: No in-house chip; uses Google TPU, with scale undisclosed inside Google Cloud. Fleet data from ~1,500 vehicles plus simulation. Low export control risk as a US-based operator.
• XPeng: In-house Turing AI Chip. Fleet data from ~429K vehicles (2025). Medium export control risk because H800 supply is at risk.
• Huawei: Fully domestic silicon. Fleet data of 7.28B km from the HIMA fleet. No export control risk because the stack is fully domestic.
• NIO: No in-house training chip. Fleet data from the NIO fleet in China. Medium export control risk because the A100 is restricted.
• Li Auto: No in-house training chip. Fleet data from the Li Auto fleet in China. Medium export control risk because H800 supply is at risk.
• Rivian: In-house RAP1 is an inference chip; training silicon is undisclosed. Fleet data from the R1/R2 consumer fleet. Low export control risk as a US-based OEM.

Nvidia AV Training and Simulation Platform

Nvidia does not build vehicles or its own AV software stack. It supplies training compute, simulation and pre-trained models to OEMs using HD maps, vision-only networks and sensor fusion alike. Nvidia describes a three-computer model: DGX for training, Omniverse and Cosmos for simulation and synthetic data, and DRIVE AGX for in-vehicle inference. Its automotive business is projected at about $5 billion revenue in fiscal 2026.

DGX (Training)

DGX systems are sold on-premise or as DGX Cloud, a managed H100 service on Microsoft Azure, Google Cloud, Oracle Cloud and AWS priced at about $36,999 per instance per month. Supporting software includes the Physical AI Data Factory Blueprint for dataset curation, CUDA-X AI, NGC containers, TensorRT, Triton, and Run:ai (acquired by Nvidia in 2024 for about $700M) for GPU workload scheduling.

Users include Volvo Cars / Zenseact, NIO, Waabi and Uber (DGX Cloud, early 2025).

Cosmos and Omniverse (Simulation)

Cosmos is a world foundation model that turns real sensor data into simulation and generates synthetic sensor data. A developer with a small fleet can use it to produce rare scenarios such as construction zones, unusual pedestrian behavior, bad weather and sensor failures.

• Cosmos Curator: a DGX Cloud service that segments raw fleet video into clips, generates embeddings and auto-labels them (Stage 3 of the OTA loop)
• Omniverse NuRec: rebuilds real drives as interactive 3D scenes for closed-loop testing, where the simulated world reacts to the vehicle
• OVX systems: dedicated hardware for Omniverse and Cosmos rendering and physics simulation, separate from DGX
• Users: Waabi; Toyota, Aurora and Continental were announced as Cosmos simulation partners in September 2025

Alpamayo (Open Models)

An open set of pre-trained Vision-Language-Action (VLA) models, simulation frameworks and datasets that OEMs fine-tune on their own fleet data. Fine-tuning cuts the compute and data smaller developers need and ties adopters more closely to Nvidia's ecosystem.

DRIVE AGX (In-Vehicle Compute)

• Orin: current production SoC, 254 TOPS, safety-certified; in the Volvo EX90, JLR vehicles from 2026 and others
• Thor: 2,000 TOPS for L4 workloads, in production ramp 2025-2026; Volvo is migrating to it
• DriveOS: safety-certified OS (ISO 26262) running CUDA, TensorRT, NvMedia and NvStreams
• DRIVE Hyperion: reference architecture pairing DRIVE AGX with a validated camera, radar and ultrasonic sensor suite, from L2++ to L4
• Rivian: DRIVE AGX partnership confirmed

Halos (Safety)

A safety framework spanning DGX training, Cosmos simulation and DRIVE AGX deployment, built to support ISO 26262 and SOTIF (ISO 21448) compliance. Nvidia positions it as its answer to proving to regulators that an end-to-end neural driving system is safe.

Nvidia AV Platform Summary

Related Coverage

• Fleet Autonomy Sensors & Perception
• Sensor Fusion Approaches
• Camera Systems for Autonomy
• Radar Systems for Autonomy
• LiDAR Systems for Autonomy
• Autonomy Compute Platforms
• Autonomy Stack
• Autonomous Driving Technology