NVIDIA expands Cosmos with 5 data blueprints for physical AI
Porsche expects to decrease test drives by a maximum of 90%, but this is an estimate and NVIDIA gives no test by a different group.
NVIDIA expanded its Cosmos platform with new training tools, simulation checkpoints and data blueprints. Cosmos helps developers make world models, which are AI systems that simulate the physical world for robots and vehicles. The 5 blueprints make or change training data. Developers can get them from cloud companies. NVIDIA says this makes the time from first tests to use short.
New parts
The 5 blueprints help to make or change training data:
- Vehicle data: select and mark road data, and add synthetic hard cases
- Image changes: change how objects look in images
- Incident video: change images that do not move into synthetic videos of incidents that are not usual
- Defect images: make images of defects in circuit boards, metal and glass
- Video changes: change weather and light in video, keep the scene
The numbers
Porsche expects to decrease test drives by a maximum of 90% and development time by 85%. NVIDIA says its new microservices give a maximum of 2x the speed of inference with the same quality in examined setups. MicroAGI made a 32 billion parameter model into an 8 billion parameter model with the same accuracy.
Not known
All numbers come from NVIDIA or its partners. The source gives no test by a different group and no cost.
Insight: NVIDIA opens the software around its GPUs
In October 2026, NVIDIA also made open source its full platform that gives GPU resources to AI work. The new tools are open too and deploy on its GPUs. NVIDIA seems to open its software to make users select its GPUs for AI work. Open software is not always neutral.
Sources
Posted
