NVIDIA DRIVE Hyperion
Last reviewed
Jun 2, 2026
Sources
9 citations
Review status
Source-backed
Revision
v1 · 2,130 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 2, 2026
Sources
9 citations
Review status
Source-backed
Revision
v1 · 2,130 words
Add missing citations, update stale details, or suggest a clearer explanation.
NVIDIA DRIVE Hyperion is a production-ready reference architecture for autonomous vehicles developed by NVIDIA. It packages a validated in-vehicle compute system, a qualified multimodal sensor suite of cameras, radar, lidar and ultrasonic sensors, and a safety-certified software stack into a single design that automakers and mobility companies can adopt rather than building an autonomous vehicle platform from scratch. The platform has gone through several generations since its debut around 2019, scaling from advanced driver assistance up to Level 4 autonomy, and at NVIDIA GTC Taipei during COMPUTEX 2026 on June 1, 2026, the company positioned DRIVE Hyperion as "the global platform for a robotaxi-ready world."[1][2][3]
| Attribute | Detail |
|---|---|
| Developer | NVIDIA |
| Type | Autonomous-vehicle reference architecture (compute + sensors + software) |
| Core compute (current) | Two NVIDIA DRIVE AGX Thor systems-on-chip, built on the Blackwell architecture |
| Operating system | NVIDIA Halos OS, built on the safety-certified NVIDIA DriveOS |
| Autonomy range | Level 2 ADAS through Level 4 (current generation) |
| Functional safety | ISO 26262 ASIL-D capable; assessed by TÜV SÜD and TÜV Rheinland |
| Current generation | DRIVE Hyperion 10 (unveiled at CES 2026) |
| 2026 positioning | "The global platform for a robotaxi-ready world" (GTC Taipei, June 1, 2026) |
| Notable adopters | Mercedes-Benz, Lucid, Stellantis, BYD, Geely, Nissan, Isuzu, Foxconn, Uber, VinFast |
DRIVE Hyperion is not a single chip or product but a complete development platform and reference design. NVIDIA describes it as combining high-performance DRIVE AGX in-vehicle compute, a software foundation built on the safety-certified DriveOS operating system, a compatible multimodal sensor suite, and NVIDIA DRIVE autonomous-vehicle (AV) software purpose-built for highly automated and autonomous driving. The intent is to give carmakers a known-good starting point: a sensor placement and compute architecture that NVIDIA has already validated for performance, redundancy and functional safety, so that engineering teams can focus on their own driving software and vehicle integration rather than re-solving the underlying platform.[1][4]
A useful way to think about Hyperion is as the in-vehicle half of NVIDIA's three-computer approach to autonomy. One computer (in the data center) trains the neural networks, a second simulates and validates them, and the third, the car itself, runs them in real time. Hyperion is the specification for that third computer plus the sensors that feed it. The platform is designed to be programmable and upgradeable, so a vehicle shipped with driver-assistance features can later receive more capable autonomous-driving software over the air as it is validated.[4][5]
The defining characteristic of Hyperion is its qualified sensor set, which provides overlapping, diverse coverage around the vehicle for redundancy. The current DRIVE Hyperion 10 generation specifies a suite of 14 cameras, 9 radars, 1 lidar, 12 ultrasonic sensors, 4 interior cameras and an exterior microphone array. Earlier Hyperion 8 used a smaller configuration of roughly 12 cameras, 9 radars, 1 lidar, 12 ultrasonic sensors and 3 interior cameras. By pre-validating the sensor types, mounting positions and data interfaces, NVIDIA lets partners reuse the same perception software across different vehicle models.[4]
At the core of the current platform sits NVIDIA DRIVE AGX Thor, the automotive system-on-chip built on the Blackwell GPU architecture and closely related to NVIDIA's Jetson Thor robotics module. NVIDIA rates each DRIVE AGX Thor SoC at up to 1,000 INT8 TOPS and 2,000 FP4 teraflops, and DRIVE Hyperion 10 pairs two of them. That headroom is meant to run the large transformer-based and generative AI models, including vision-language models, that increasingly drive modern AV perception and planning, while leaving spare capacity for redundancy and future feature growth. Thor succeeded the earlier DRIVE Orin SoC, which powered the Hyperion 8 generation.[4][6]
Running on top of the hardware is NVIDIA DriveOS, the company's safety-certified automotive operating system, which NVIDIA states conforms to ISO 26262 ASIL-D, the highest automotive functional-safety integrity level. In 2025 and 2026, NVIDIA reframed Hyperion's software around NVIDIA Halos, which it calls a full-stack safety system for physical AI spanning "cloud to car." The Halos OS, built on DriveOS, is now described as the software foundation of the platform. On top of that runs NVIDIA DRIVE AV, the autonomous-driving application software. NVIDIA's open driving models, including NVIDIA Alpamayo reasoning models, are designed to run on this stack, and the platform connects to NVIDIA's data-center tools such as the NVIDIA Cosmos world foundation models and NVIDIA Omniverse for generating and curating training and simulation data.[1][6][7]
NVIDIA has shipped DRIVE Hyperion as a series of numbered generations, each tied to the compute SoC available at the time. The platform first appeared around 2019, but the versions that drove broad industry adoption were Hyperion 8 and the current Hyperion 10.
| Generation | Introduced | Core compute | Sensors (typical) | Autonomy target |
|---|---|---|---|---|
| DRIVE Hyperion 8 | Announced 2021 (production from ~2023) | Two DRIVE Orin SoCs (~508 INT8 TOPS combined) | 12 cameras, 9 radars, 1 lidar, 12 ultrasonics, 3 interior cameras | Level 2 ADAS through Level 3 |
| DRIVE Hyperion 9 | GTC, March 2022 | DRIVE Thor central computer (~2,000 TFLOPS, ~8x Orin) | High-resolution cameras, radars, lidars, ultrasonics | Level 3 urban, Level 4 highway |
| DRIVE Hyperion 10 | CES, January 2026 | Two DRIVE AGX Thor SoCs (Blackwell) | 14 cameras, 9 radars, 1 lidar, 12 ultrasonics, 4 interior cameras, microphone array | Level 2 ADAS through Level 4 |
DRIVE Hyperion 8 was the generation that turned Hyperion into a widely adopted commercial platform. It was built around two DRIVE Orin processors and a full production sensor set, scaling from Level 2 driver assistance up to Level 3 conditional automation, and NVIDIA marketed it for everything from private passenger cars to robotaxis and autonomous trucks. Through 2021 and 2022, a long list of automakers committed to Hyperion 8 or the broader DRIVE platform, including Mercedes-Benz, Jaguar Land Rover, Volvo Cars, Polestar, and a cluster of Chinese EV makers such as NIO, Xpeng, Li Auto and IM Motors. At GTC in March 2022, NVIDIA added BYD and Lucid Group as customers, with BYD slated to build vehicles on Hyperion starting in 2023.[8][9]
At GTC in March 2022, Jensen Huang also previewed DRIVE Hyperion 9 as the next-generation platform, built on the then-new DRIVE Thor central computer. NVIDIA said Thor would deliver roughly 2,000 teraflops, about eight times the performance of the Orin-based design at the same power envelope, with an expanded, higher-resolution sensor suite for greater redundancy. Hyperion 9 was described as enabling Level 3 urban and Level 4 highway driving and was slated for vehicles arriving around 2027.[5]
At CES 2025, NVIDIA announced that DRIVE Hyperion had cleared a set of industry safety and cybersecurity assessments, which it presented as evidence the platform was ready for production AV programs. TÜV SÜD granted an ISO 21434 cybersecurity-process certification covering NVIDIA's automotive SoC, platform and software engineering, and TÜV Rheinland completed an independent safety assessment of the platform's complex electronic systems against UNECE requirements. NVIDIA also stated that DriveOS 6.0 conforms to ISO 26262 ASIL-D. At the same show NVIDIA confirmed that the DRIVE AGX Thor SoC, built on Blackwell, would be available in the first half of 2025.[7]
DRIVE Hyperion 10 became the current generation at CES in January 2026. Built on two DRIVE AGX Thor SoCs, it extended the platform's reach to a full Level 2-through-Level 4 range and integrated the NVIDIA Halos safety system. NVIDIA paired the hardware announcement with a wave of adopters spanning passenger cars, trucking and robotaxi developers: Mercedes-Benz, Lucid and Stellantis among automakers; Aurora, Volvo Autonomous Solutions and Waabi in autonomous trucking; and ecosystem members including Avride, May Mobility, Momenta, Nuro, Pony.ai, Wayve and WeRide. NVIDIA also debuted its DRIVE AV software in the production Mercedes-Benz CLA, described as the first production vehicle to feature NVIDIA's Alpamayo-powered autonomous-driving capabilities.[6][4]
This generation built on a robotaxi partnership with Uber that NVIDIA had announced in late October 2025 at GTC Washington, under which Uber set a target of deploying on the order of 100,000 autonomous vehicles over time, beginning in 2027, using Hyperion 10-class vehicles and a joint AI data factory built on NVIDIA Cosmos.[6]
On June 1, 2026, at NVIDIA GTC Taipei held alongside COMPUTEX in Taiwan, NVIDIA announced a major expansion of the DRIVE Hyperion ecosystem and positioned the platform as "the global platform for a robotaxi-ready world." The company framed Hyperion, paired with the Halos OS, as a common foundation that automakers and mobility operators across regions could standardize on, and said that DRIVE Hyperion-equipped vehicles would connect into mobility services that already represent approximately 97 percent of the world's mobility services. NVIDIA founder and CEO Jensen Huang summarized the moment this way: "Autonomous mobility is entering its industrial scaling moment. Vehicles are becoming robots, and robotaxi fleets will require AI infrastructure that can perceive, reason and operate safely in the real world."[1][2]
The announcement brought together automakers, contract manufacturers, AV-software companies and ride-hailing providers around the Hyperion standard:
| Partner | Role announced at GTC Taipei 2026 |
|---|---|
| Foxconn | Combining its contract design and manufacturing with DRIVE Hyperion to integrate and scale Level 4 electric robotaxis, starting in Taiwan with Kaohsiung as an early deployment city and expanding across Asia; service targeted for 2028 |
| Uber | Integrating multiple DRIVE Hyperion-powered AV fleets into its global ride-hailing network |
| Autobrains | Providing autonomous-driving software; launching a Munich robotaxi program with Uber on Hyperion in late 2026, and supplying software for VinFast |
| VinFast | Working with Autobrains to bring Level 4 vehicles built on DRIVE Hyperion to the Southeast Asia market |
| HUMAIN | Bringing DRIVE Hyperion-powered robotaxis to Saudi Arabia, extending the platform into the Middle East |
These were not the only automakers committed to Hyperion for Level 4 work. In March 2026, NVIDIA had separately announced that BYD, Geely, Isuzu and Nissan were adopting DRIVE Hyperion for Level 4 vehicles, with Isuzu collaborating with TIER IV on an autonomous bus built on DRIVE AGX Thor and Nissan developing Level 4 programs using Wayve software. At that event NVIDIA also reported that its open Alpamayo reasoning model had been downloaded more than 100,000 times.[3]
DRIVE Hyperion reflects NVIDIA's strategy of supplying a horizontal, reusable platform to the automotive industry rather than building its own cars. By standardizing the compute, sensors and safety-certified software, NVIDIA aims to let many manufacturers reach Level 4 autonomy on a shared technical base, much as a common silicon platform underpins many different consumer devices. The 2026 robotaxi positioning extends that idea from individual vehicles to entire fleets and mobility networks.[1][6]
Several claims around the platform should be read as NVIDIA's own characterizations rather than independently verified facts. The figure that Hyperion-equipped vehicles connect into mobility services representing about 97 percent of the world's mobility, the "global platform" and "robotaxi-ready" framing, and partner deployment timelines such as a 2026 Munich launch or 2028 service in Taiwan are forward-looking statements from NVIDIA and its partners. Performance numbers such as 2,000 FP4 teraflops per Thor SoC are NVIDIA specifications. Achieving large-scale Level 4 robotaxi operation also depends on regulatory approval, real-world validation and commercial execution that extend well beyond the reference platform itself.[1][2][3]