GTC 2026 Automotive Special Address

Category: Automotive Special Address · Year: 2026 · ▶ Watch

Switch language → 中文

Segments (11)

  • 00:00 · Introduction and Vision
    • Xinzhou Wu introduces the rapid progress in AI and the challenge of achieving Level 4 autonomy.
  • 03:18 · Alpamayo Reasoning Demo
    • A video demonstration of the Alpamayo reasoning model powering a Mercedes prototype in San Francisco.
  • 07:04 · The Era of Physical AI
    • Discussion on the evolution of AI towards physical AI and the massive market opportunity for autonomous vehicles.
  • 09:55 · NVIDIA DRIVE Full Stack
    • Overview of the NVIDIA DRIVE platform, including infrastructure, applications, models, OS, and hardware.
  • 12:28 · DRIVE Hyperion and Thor
    • Details on the DRIVE Hyperion reference architecture and the DRIVE AGX Thor in-vehicle supercomputer.
  • 16:08 · Halos OS
    • Introduction to Halos OS, a unified software safety foundation for Level 4 autonomy.
  • 18:55 · Announcing Alpamayo 1.5
    • Unveiling the new version of the open reasoning VLA model for autonomous vehicles.
  • 21:30 · Training and Simulation
    • How NVIDIA uses massive datasets, Cosmos, and Omniverse NuRec to train and simulate AV models.
  • 25:20 · Omniverse NuRec and Cosmos Transfer
    • Showcasing tools for neural reconstruction and environment transfer to generate synthetic training data.
  • 29:14 · DRIVE AV Stack and Roadmap
    • Explanation of the hybrid end-to-end and classical stack, and the roadmap to L4 full autonomy.
  • 39:04 · Ecosystem and Partnerships
    • Announcements of new OEM adoptions and a worldwide L4 Robotaxi rollout partnership with Uber.

Product Announcements (2)

  • [18:55] NVIDIA Alpamayo 1.5
    • The world’s first open reasoning VLA model for AVs, with new developer-driven upgrades.
    • specs: 10B parameters, routing capability (waypoints/nav guidance), text prompts, multiple camera configurations.
    • availability: Announced as being released.
  • [25:20] NVIDIA Omniverse NuRec
    • A tool for generating 3D simulations from real-world data at scale.
    • specs: Neural reconstruction, object insertion (Fixer/Harvester).
    • availability: General Access in the next few weeks.

Specific Numbers (9)

Timestamp Metric Value Context
09:11 Annual Global Vehicle Miles 13 Trillion Total miles driven globally across all segments.
09:29 AV Driverless Miles Percentage 0.006% Percentage of total global miles currently driven autonomously.
13:40 Compute Performance Up to 2000 FP4 TFLOPS Performance of the DRIVE AGX Thor supercomputer.
19:14 Model Size 10 Billion parameters Size of the Alpamayo 1.0 and 1.5 models.
19:39 Downloads >160,000 Number of times Alpamayo 1.0 was downloaded on Hugging Face.
21:41 Training Data 80,000 Hours Amount of AV data used to train the Alpamayo model.
22:08 Training Data 20 Million Hours Amount of real-world videos used to train the Cosmos World Foundation model.
24:33 Open AV Dataset 7,000 Hours Size of the open-sourced physical AI dataset.
26:28 Simulation Scale 2 Million+ Number of simulation tests run daily based on NuRec.

Benchmark Claims (3)

  • [13:40] LLAMA-7B context: 9x faster
    • vs: Orin (est.)
    • gain: Significant speedup for large language model inference in-vehicle.
  • [13:40] SPECrate 2017_int_base: 2.3x performance
    • vs: Orin (est.)
    • gain: Improved high single-thread CPU performance for real-time vehicle control.
  • [18:55] LingoQA: #1
    • vs: Other AV Reasoning models
    • gain: Top ranking for AV reasoning capabilities.

Customer Stories (3)

  • [03:18] Mercedes-Benz
    • Integrated the Alpamayo reasoning model into a prototype vehicle for testing in San Francisco.
    • outcome: Demonstrated real-time reasoning, narration, and safe handling of complex urban driving scenarios.
  • [13:21] Uber
    • Partnering with NVIDIA to build a fleet using the Hyperion 10 sensor set for data collection.
    • outcome: Jumpstarting the industry and planning a worldwide L4 Robotaxi rollout in 2028.
  • [39:04] BYD, Geely, Nissan, Hyundai
    • Adopted the NVIDIA DRIVE Hyperion platform for their Level 4 autonomous vehicles.
    • outcome: Expanding the ecosystem of OEMs utilizing NVIDIA’s unified sensor and compute architecture.

Key Technologies (6)

  • DRIVE Hyperion: A unified vehicle reference architecture providing a common sensor suite and compute platform for L4 autonomy.
  • DRIVE AGX Thor: An in-vehicle AI supercomputer based on the Blackwell architecture, powering next-gen compute.
  • Halos OS: A unified software safety foundation and operating system that interacts between models and hardware.
  • Alpamayo: An open reasoning Vision-Language-Action (VLA) model for autonomous vehicles.
  • Omniverse NuRec: Generates 3D simulations from real-world data using neural reconstruction.
  • Cosmos Transfer: A foundation model platform that can render sequences into different environments, weather, and lighting conditions.

Demos Shown (4)

  • [03:18] A Mercedes prototype driving autonomously in San Francisco while the Alpamayo model narrates its reasoning and responds to voice prompts.
    • True
  • [25:44] Omniverse NuRec reconstructing pixels and scenes from different camera poses based on real-world data.
    • True
  • [27:48] Fixer and Harvester tools inserting objects like scooters and cones into reconstructed scenes to modify trajectories.
    • True
  • [28:50] Cosmos Transfer altering the weather, lighting, and time of day of a driving sequence.
    • True

Predictions / Commitments (4)

  • [08:26, Coming decade] Building and safely deploying physical AI is going to be the defining challenge of the coming decade.
  • [09:36, Near future] In the near future, every mile, everything that moves will become autonomous.
  • [29:14, 2026] L2++ Address to Address capability rollout.
  • [39:53, 2028] Worldwide L4 Robotaxi rollout in partnership with Uber.

Companies Mentioned (8)

Mercedes-Benz · Uber · Lucid · JLR · BYD · Geely · Nissan · Hyundai

Notable Quotes (3)

We believe that building and safely deploying physical AI is going to be the defining challenge of the coming decade. — Xinzhou Wu @ 08:26

We do believe that in the near future, every mile, everything that moves will become autonomous. — Xinzhou Wu @ 09:36

We are eating our own dog food. And I am very confident about the quality of this tool. — Xinzhou Wu @ 26:38

Key Topics

Autonomous Vehicles · Level 4 Autonomy · Physical AI · Vision-Language-Action (VLA) Models · Simulation · Neural Reconstruction · AI Supercomputers · Robotaxis · Generative AI · Automotive Ecosystem · Software-Defined Vehicles · Synthetic Data

Takeaways

  • NVIDIA is pivoting to provide a comprehensive, full-stack solution for Level 4 autonomy, encompassing hardware, OS, models, and infrastructure.
  • The newly announced Alpamayo 1.5 is an advanced open reasoning VLA model that brings human-like decision-making and narration to AVs.
  • Scaling AV training requires massive amounts of data, which NVIDIA is addressing through simulation tools like Omniverse NuRec and Cosmos Transfer.
  • Safe deployment of L4 autonomy requires a hybrid software stack that combines the capabilities of end-to-end AI models with strict classical safety guardrails.
  • NVIDIA’s DRIVE Hyperion platform is seeing strong adoption, with major OEMs like BYD, Geely, and Nissan, as well as Uber for a 2028 Robotaxi rollout, committing to the architecture.