Inference and Compute Manufacturers

driving SoCs · vehicle computers · simulation · toolchains

Every autonomous vehicle and humanoid runs on an inference chip from a small set of vendors. NVIDIA leads, and Chinese OEMs are designing their own chips to replace it.

Bar chart on a log scale of peak AI performance per driving chip, in INT8 TOPS, from vendor figures: Mobileye EyeQ6 High 34, Mobileye EyeQ Ultra 176, NVIDIA DRIVE Orin 254, Horizon Robotics Journey 6P 560, XPeng Turing about 750 (claimed), NVIDIA DRIVE Thor 1,000 and NIO Shenji NX9031 more than 1,000. Chinese chips in red, US and Israeli chips in blue. Vendors use different precision and sparsity assumptions. DRIVING SOC PEAK AI PERFORMANCE, INT8 TOPS PER CHIP (LOG SCALE) Mobileye EyeQ6 High34 Mobileye EyeQ Ultra176 NVIDIA DRIVE Orin254 Horizon Journey 6P560 XPeng Turing (claimed)~750 NVIDIA DRIVE Thor1,000 NIO Shenji NX90311,000+ 10 100 1,000 Vendor figures. Red: China. Blue: US, Israel.
Thor peak
1,000 TOPS
OEMs with in-house chips
Tesla, XPeng, NIO, Rivian
Leading Chinese vendor
Horizon Robotics
Hub
OEMs

Finding

Driving compute is consolidating on NVIDIA, Qualcomm and Horizon Robotics, while Tesla and Chinese EV makers move to in-house chips

NVIDIA's DRIVE Orin and Thor run the higher-end driver assistance of most premium Western and Chinese brands; Mobileye's EyeQ chips run basic ADAS in most of the world's cars; Horizon Robotics' Journey 6 is the volume choice in China. Tesla, XPeng, NIO and Rivian now design their own driving chips.

In-house silicon follows from vision-heavy end-to-end driving models, which need large on-chip memory bandwidth and a toolchain tuned to one model. Tesla's AI5 is built by TSMC and Samsung; XPeng's Turing runs its G7 and IRON humanoid and is supplied to Volkswagen China; NIO's 5 nm Shenji NX9031 matches four Orin-X chips by NIO's figures. US export controls add a second motive for Chinese makers: domestic chips avoid supply risk on NVIDIA parts.

The same silicon is moving into robots. NVIDIA's Jetson Thor shares the Blackwell architecture with DRIVE Thor and targets humanoids; Tesla's Optimus runs on its vehicle chips; XPeng's IRON uses Turing. Training happens elsewhere, on data center clusters such as Tesla's Cortex in Austin and xAI's Colossus in Memphis; see Autonomy Training Clusters.

SoCs

Driving and robot SoCs

Linked names have a full profile.

Automotive and robot SoCs
CompanyHQSiliconFocusCustomers and notes
NVIDIAUSDRIVE Orin, DRIVE ThorCentral vehicle computer, ADAS to L4Mercedes-Benz, Volvo, Zeekr, Li Auto, BYD, Lucid, robotaxi builders
QualcommUSSnapdragon Ride, Ride Flex, Ride EliteADAS plus cockpit on one chipBMW (Neue Klasse ADAS), GM, Chinese brands
MobileyeIsraelEyeQ6 Lite and High, EyeQ UltraADAS to eyes-off drivingVolkswagen Group, most global OEMs for basic ADAS
Horizon RoboticsChinaJourney 6 family (6B, 6E, 6M, 6P)ADAS to urban NOALi Auto, BYD, Volkswagen (Carizon JV), Chery, SAIC
HuaweiChinaAscend-based MDC computersQiankun ADSHIMA brands, Changan, GAC, Dongfeng
Black Sesame TechnologiesChinaHuashan A1000, Wudang C1200ADAS, cross-domainDongfeng, Geely Lynk & Co
TeslaUSAI4 (HW4), AI5In-vehicle FSD inferenceTesla vehicles, Cybercab, Optimus; AI5 built by TSMC and Samsung at Taylor, Texas
XPengChinaTuring AI chipIn-vehicle inference, robotsXPeng G7 and P7, IRON humanoid; supplied to Volkswagen China
NIOChinaShenji NX9031 (5 nm)In-vehicle inferenceNIO ET9, ES8 and newer models
RivianUSRAP1 autonomy processorIn-vehicle inferenceRivian Gen 3 autonomy platform
RenesasJapanR-Car Gen 5 (X5H)Central compute, gateway, ADASJapanese and European OEMs
NXPNetherlandsS32N, S32G, S32KVehicle networking, zonal and safety controllersMost global OEMs
AmbarellaUSCV3-AD familyPerception SoCsTier 1 ADAS suppliers
AMDUSVersal AI Edge Gen 2, Ryzen EmbeddedCockpit, ADAS sensor fusionTesla cockpit (Ryzen), Subaru EyeSight

Computers

Vehicle computers and modules

Tier 1 boxes built around the SoCs above.

Vehicle computers and compute modules
CompanyHQProductsCustomers
Desay SVChinaIPU04 (Orin), Thor-based controllersLi Auto, XPeng, Geely, Toyota China
BoschGermanyVehicle computers, ADAS domain controllersGlobal OEMs
ZFGermanyProAI supercomputerGlobal OEMs, commercial vehicles
AumovioGermanyHigh-performance computers, cockpit controllersVolkswagen, BMW, others
AptivSwitzerlandGen 6 ADAS platform, central computeGlobal OEMs
MagnaCanadaADAS domain controllersGlobal OEMs
Hyundai MobisSouth KoreaADAS and cockpit controllersHyundai, Kia, Genesis
ECARXChinaAntora and Skyland computersGeely brands, Volvo, Polestar, smart
NVIDIA JetsonUSJetson Thor, Jetson Orin modulesHumanoids, AMRs, drones

Simulation

Training and simulation platforms

Training and simulation platforms
CompanyHQPlatformUse
NVIDIAUSOmniverse, Cosmos world models, DRIVE Sim, Isaac SimSynthetic data and closed-loop simulation for vehicles and robots
Applied IntuitionUSVehicle OS, simulation and validation suiteOEMs, trucks, defense
ForetellixIsraelForetifyCoverage-driven safety validation
dSPACEGermanyAURELION, HIL systemsSensor simulation, hardware-in-the-loop
MathWorksUSMATLAB, Simulink, RoadRunnerModel-based design and scenario authoring
51SimChina51Sim-OneChinese OEM simulation and data
WaabiCanadaWaabi WorldNeural simulation for driverless trucks
TeslaUSCortex training cluster at Giga TexasFSD and Optimus training; Dojo program ended in 2025

Toolchains

Toolchains, operating systems and middleware

Toolchains and middleware
VendorHQStackUse
NVIDIAUSCUDA, TensorRT, DriveOS, DriveWorksInference and AV software on NVIDIA silicon
QualcommUSSnapdragon Ride SDKADAS and AV software on Ride
BlackBerry QNXCanadaQNX OS for Safety, hypervisorSafety-certified RTOS in most cars
Horizon RoboticsChinaOpenExplorer, TogetheROSModel deployment on Journey chips
HuaweiChinaCANN, MindSpore, HarmonyOS cockpitAscend toolchain and cockpit OS
Apex.AIUSApex.OS, Apex.GraceSafety-certified ROS 2 middleware
VectorGermanyAUTOSAR stack, toolsClassic and adaptive AUTOSAR
Autoware FoundationGlobalAutoware (ROS 2 based)Open-source AV stack for shuttles and research
Open RoboticsGlobalROS 2Robot middleware standard

Sources

Sources and scope

Scope
Inference and Compute Manufacturers
Sources
Vendor product briefs and launch disclosures (2024-2026); AustinIO
Updated
October 2026