NVIDIA expands Cosmos platform with Data Factory blueprints and Cosmos 3 Nano simulation models
At NVIDIA GTC Berlin on October 21, 2026, NVIDIA announced a major expansion of its Cosmos physical AI platform, introducing new synthetic data curation blueprints, simulation checkpoints, and optimized inference microservices. The release expands Cosmos across data generation, policy training, and deployment of world models that let robots, autonomous vehicles, and industrial vision systems comprehend and simulate physical environments.
A core operational addition is the Cosmos Data Factory, which introduces five cloud-native blueprints distributed via GitHub, Microsoft Azure, Nebius, and CoreWeave. The blueprints provide standardized workflows for Autonomous Vehicle Dataset Generation, Image Attribute Augmentation, Event Video Generation, Defect Image Generation, and Video Data Augmentation. Concurrently, the GPU-accelerated Cosmos Curator tool gained support for multimodal, multisensory data streams.
For policy evaluation, NVIDIA released Cosmos 3 Nano model variants for closed-loop simulation, allowing developers to test autonomous navigation and bimanual robotics policies in repeatable, multi-camera scenarios before hardware deployment. NVIDIA also updated Cosmos 3 NIM microservices, claiming up to 2x faster inference in validated configurations.
Alongside Cosmos, the new NVIDIA Metropolis Blueprint for Video Search and Summarization (VSS) 3.3 enables teams to monitor 40% more camera streams per GPU. Early adopters include Porsche, which is using Cosmos Data Factory via Voxel51 to build a shared autonomous vehicle data flywheel across Volkswagen Group brands, and ANYbotics, which generated over 1,500 synthetic clips with Encord to train industrial inspection robots.
For robotics engineering directors, autonomous vehicle systems leads, and industrial computer vision teams, these additions lower the operational friction of training physical world models. Pre-packaged blueprints and NIM microservices let engineering teams bypass building bespoke synthetic data pipelines and custom simulation harnesses.
Buyers must distinguish advertised performance targets from production reality. The 2x inference speedup and 40% stream capacity gains represent vendor-validated reference configurations under specific hardware setups rather than universal field performance. NVIDIA has not published standalone commercial pricing for Cosmos blueprints or NIM enterprise licensing, which typically require paid NVIDIA AI Enterprise subscriptions, and synthetic data augmentation requires ongoing real-world validation to prevent simulation-to-reality transfer errors.