NVIDIA unveils Rubin platform with 50 petaflops inference capacity and suite of open models across healthcare, robotics, and autonomous driving

NVIDIA CEO Jensen Huang announced Rubin, the company's first extreme-codesigned six-chip AI platform, now in full production. The platform integrates Rubin GPUs delivering 50 petaflops of NVFP4 inference capacity, Vera CPUs for data movement and agentic processing, NVLink 6 scale-up networking, Spectrum-X photonics, ConnectX-9 SuperNICs, and BlueField-4 data processing units. The company claims this tight integration across components reduces inference token generation costs to approximately one-tenth that of its predecessor, the Blackwell architecture.

Alongside Rubin, NVIDIA unveiled Alpamayo, an open portfolio of reasoning vision-language-action models designed for autonomous vehicle development. Alpamayo R1, described as the first open reasoning VLA model for autonomous driving, is paired with Alpamayo, a fully open simulation blueprint for autonomous vehicle testing. NVIDIA stated that Alpamayo models perform physical reasoning and trajectory prediction to enable level-4 capable autonomy.

NVIDIA will supply the Rubin-based DRIVE platform to Mercedes-Benz for the all-new CLA. The company announced that the CLA will be the first passenger car featuring Alpamayo-based AI-defined driving, with availability in the U.S. market targeted for this year. The vehicle recently achieved a five-star EuroNCAP safety rating.

The company released open models across six technical domains: Clara for healthcare, Earth-2 for climate science, Nemotron for reasoning and multimodal AI, Cosmos for robotics and world simulation, GR00T for embodied intelligence, and Alpamayo for autonomous driving. NVIDIA stated these models are trained on its supercomputers and released openly to enable adoption across industries.

NVIDIA demonstrated DGX Spark, a desktop supercomputer claiming up to 2.6x performance improvement for large model inference compared to prior versions. The company showed a prototype featuring a personalized AI agent running locally on DGX Spark hardware and embodied through a Reachy Mini robot using Hugging Face models, highlighting local execution capabilities for agentic systems.