Real Time
NVIDIA GTC 2026: Jensen Huang Unveils Alpamayo 2 Super, Vera CPU, and DSX AI Factory Blueprint
🚗 NVIDIA Unveils Alpamayo 2 Super: The Autonomous Driving "Black Box" is Finally Opened"Alpamayo marks the era where vehicles transitio...
3 weeks ago
•
June 2, 2026
•
52 views
During his nearly two-hour keynote, Huang not only introduced products like the Agent-designed Vera CPU, the AI Factory prACTical guide DSX Platform, and the RTX Spark superchip, but also launched a highly anticipated open reasoning model for the automotive industry—NVIDIA Alpamayo 2 Super. This 32-billion parameter open Vision-Language-Action (VLA) reasoning model is capable of reasoning, planning, and acting within a complete driving stack, providing support for safer and scalable L4 autonomous driving development. This represents a fundamental paradigm shift from "imitation driving" to "safe reasoning".
The product lineup unveiled at the conference—spanning from underlying chips to upper-level infrastructure, data centers to Personal PCs, and cloud intelligence to physical robots—signals NVIDIA's fundamental transformation from a "GPU supplier" to an "AI infrastructure operator".
🧠 Alpamayo 2: Autonomous Driving Enters a New Stage of Reasoning
The core breakthrough of Alpamayo 2 lies in its "explAInability". Traditional end-to-end autonomous driving models are like silent veteran drivers; their decision-making processes are opaque. In contrast, Alpamayo 2 acts like a "chatty driver," allowing the vehicle to explain every decision in real-time using natural language. For instance, it can state, "Slight left adjustment due to a stationary vehicle blocking the lane," "Yielding to a vehicle cutting in from the left," or "Stopping to yield to crossing traffic." This ability to externalize its "chain of thought" is crucial for building public trust. Key Technical Upgrades:
360-Degree Panoramic Perception: Expands from front-facing cameras to include front, side, and rear views, providing complete scene Information for lane changes, merging, and interSECtion navigation.
Meta-Action OuTPUt: Introduces high-level macro decisions such as yielding, changing lanes, and stopping, providing advanced driving instructions for downstream planning.
inference-based Auto-Labeling: Features 2D localization to compress data labeling cycles from months to days, reshaping the cost structure of assisted driving data pipelines.
A notable feature is its "Teacher-Student" distillation architecture. The 32-billion parameter Alpamayo 2 Super acts as the "teacher" model and can be distilled into a compact model to run on the NVIDIA DRIVE AGX Thor in-vehicle computing platform. This allows automakers to gain "plug-and-play" reasoning capabilities through NVIDIA's open ecosystem without building large models from scratch.
"Alpamayo marks the transition of cars from 'simply driving' to 'safe reasoning.' Only NVIDIA can provide the open models, simulation environments, real-world data, and agent skills to support the global robotaxi ecosystem in developing L4 capabilities that underStand edge cases, explain decisions, win public trust, and safely scale to millions of vehicles," Huang stated.
🎮 The "Virtual Driving School": AlpaGym and OmniDreams
If Alpamayo 2 is the "thinking driver," then the simultaneously launched AlpaGym and OmniDreams serve as its "virtual driving school."
AlpaGym: An open-source, high-throughput closed-loop reinforcement learning Framework. Unlike traditional open-loop training that generates single actions based on recorded data, AlpaGym allows the model to undergo continuous decision-observation cycles within the NVIDIA AlpASIm environment. Every brake, turn, and navigation choice has a real impact on the environment, exposing compound errors and edge failures overlooked by static datasets. This is equivalent to letting the AI driver experience millions of extreme road conditions in a "parallel universe" at zero cost of failure.
OmniDreams: A new generative world model capable of creating realistic closed-loop assisted driving scenes. It supports developers in大规模 simulating rare, long-tail driving scenarios. Combined with NVIDIA Omniverse NuRec neural reconstruction technology, developers can rebuild real fleet data into Realistic 3D scenes adaptable to different vehicle sensor configurations.
NVIDIA has also open-sourced the Causal Chain Auto-Labeling Pipeline on GitHub. This tool automatically generates decision-based causal chain labels from raw driving CLIps without the need for manual labeling. This "Simulation-Training-deployment" closed loop signifies a shift from "road-test driven" to "simulation driven" development. Automakers no longer need to deploy thousands of test vehicles to accumulate data on real roads; instead, they can complete over 90% of edge case validations in a digital twin environment, significantly reducing R&D costs and shortening time-to-market.
💻 The "Super Brain": Vera Rubin and Vera CPU
Huang also provided an update on Vera Rubin, the world's first multi-rack Pod-level supercomputing system designed specifically for agent AI, which has now entered full-scale production.
Vera Rubin reduces the latency of "inference" and "tool calling" to a nanosecond-sensitive level—the real-time foundation required for agent decision-making. Its most disruptive component is the Vera CPU, designed specifically for agents.
"AI Agents will become the largest consumers of computing resources. Vera is the first CPU built for this future—it offers superior performance, energy efficiency, and programmability, born to run agent AI at a hyperscale," Huang explained. Vera CPU Specifications:
Equipped with 88 Olympus cores using spatial multithreading technology.
Features an LPDDR5X mEMOry subsystem with bandwidth up to 1.2TB/s.
Designed with a philosophy distinct from traditional CPUs: while traditional CPUs are designed for humans (insensitive to second-level delays), Vera is built for agents that require nanosecond responses.
Performance benchmarks show the Vera CPU is 3x faster in SQL query speeds than top-tier x86 processors, 6x faster in real-time stream processing for the NYSE, and achieves 1.8x the agent sandbox performance of x86. Through second-generation NVLink-C2C interconnect technology, it achieves a coherent bandwidth of up to 1.8TB/s between CPU and GPU, extending NVIDIA confidential computing to the full rack scale. The Vera BlueField-4 STX processor further integrates the Vera CPU with high-performance networking, stoRAGe acceleration, and chip-level security, building a "secure by design" AI-native data platform.
🏭 DSX: The Blueprint for AI Factories
How can these hardware components translate into tangible returns for customers? NVIDIA's answer is DSX—a reference design blueprint for building AI factories.
DSX (data center Scale eXtended) is a complete practical guide for building AI factories from the ground up. It integrates open-source modular software libraries, APIs, reference designs, NVIDIA accelerated computing platforms, and partner technologies to create a universal co-design platform for the design, deployment, and Operation of AI factories. DSX MaxLPS (Lowest Power per token System): Addresses a major pain point in the autonomous driving industry: maximizing token output per megawatt within a set power budget. By combining 45°C liquid cooling technology with rack-level optimizations, DSX MaxLPS allows operators to run GPUs at their peak efficiency with minimal impact on workload performance, enabling the deployment of up to 40% more GPUs. For automaker Data Centers facing tight power resources and sensitive compute costs, this means supporting more simulation training miles with the Same electricity bill.
DSX Sim: Provides a high-fidelity simulation layer for the entire AI factory lifecycle. It helps NVIDIA, partners, and clients model, validate, and optimize infrastructure decisions from planning and design to deployment and operation. "With the DSX platform, you can simulate the entire factory before spending a single dollar and validate performance before a single cabinet is installed," emphasized Huang.
DSX Flex: Connects AI factories with grid services, allowing them to dynamically adjust workloads based on grid signals like load shedding, demand response, and price fluctuations. This echoes the "V2G (Vehicle-to-Grid)" vision: future autonomous driving data centers will not only be power consumers but also "flexible loads" for the grid, training models at full capacity during off-peak hours and feeding power back during peaks.
The final segment of the keynote pointed to the critical leap of agents moving from the digital to the physical world. Huang noted that the core challenge of physical AI is data; internet text is mostly in a "third-person perspective," whereas robots require "first-person perspective" physical world data. NVIDIA's solution is Cosmos 3—an open physical world Foundation Model. It serves as a vision-language model to understand physical scenes, generates physically accurate synthetic videos to complete the strategy training loop as a simulator, and acts as the foundation for the Omniverse Digital Twin platform. It supports all types of robot and physical system development, is fully open, and allows for user customization. In the autonomous driving sector, NVIDIA introduced Alpamayo 2—the world's first reasoning-capable open autonomous driving model. Based on the Hyperion platform (adopted by 80% of global automakers and interfacing with 97% of mobility services), it supports end-to-end reasoning and planning. Vehicles can explain every step of their decision logic in natural language, transforming "black box models" into "explainable AI".
Also worth mentioning is the Isaac GR00T humanoid robot reference platform. standing 6 feet tall and weighing 150 pounds, it features 25 degrees of freedom in its hands and 31 degrees of freedom across its body. It integrates a full suite of software stacks for data generation, simulation, training, and operation, aimed at univeRSIties and research institutions. What previously took months of setup can now be initiated in just hours, aiming to lower the bARRier to entry for humanoid robot research and drive the development of the entire field. Bottom Layer: computing power supply via Vera Rubin/Vera CPU.
Middle Layer: AI factory blueprints and enterprise agent toolkits via DSX.
Top Layer: Personal agent entry points via RTX Spark, and a Physical AI ecosystem consisting of Cosmos, Alpamayo, and GR00T.
Thus, NVIDIA's transformation path is clear: from selling GPUs, to selling systems, to helping clients build "profitable AI Infrastructure".
Be the first to rate this article.
Comments & Questions (0)
No comments yet
Be the first to comment!