TLDR: Industry leaders like CrowdStrike, Uber, and Zoom are leveraging NVIDIA’s Nemotron and Cosmos reasoning models to develop more intelligent AI agents for enterprise and physical AI applications. This initiative, announced at SIGGRAPH, aims to unlock significant revenue gains and cost savings, with AI agents projected to contribute up to $450 billion by 2028.
At SIGGRAPH, NVIDIA announced a significant expansion of its Nemotron and Cosmos reasoning model families, which are being adopted by industry pioneers such as CrowdStrike, Uber, and Zoom to build smarter AI agents for enterprise and physical AI applications. This development comes as AI agents are poised to deliver substantial economic benefits, with Capgemini projecting up to $450 billion in revenue gains and cost savings by 2028. Developers are increasingly turning to higher-performing reasoning models to enhance AI agent platforms and physical AI systems.
NVIDIA Cosmos is highlighted as a platform of state-of-the-art generative world foundation models (WFMs), guardrails, and an accelerated data processing and curation pipeline. It is designed to accelerate physical AI development for autonomous vehicles (AVs), robots, and video analytics AI agents. These open reasoning models are crucial for enabling faster and more extended thinking, leading to smarter outcomes for AI agents across various sectors including customer service, manufacturing, logistics, and robotics.
The Cosmos platform includes several specialized models: Cosmos Predict, a world state prediction model capable of generating up to 30 seconds of continuous video from multimodal inputs; Cosmos Transfer, a multicontrol model for scaling simulations or spatial videos across diverse environments and lighting conditions; and Cosmos Reason, a fully customizable reasoning vision language model (VLM) that excels in understanding the physical world through structured reasoning on videos and images. Cosmos Reason is particularly built to power video analytics AI agents with spatiotemporal understanding for real-time question-answering, rapid alerts, and rich contextual insights.
To support the development of these advanced AI agents, NVIDIA also offers NeMo Curator, a framework that enables developers to efficiently filter, annotate, and deduplicate large volumes of sensor data, creating tailored datasets for physical AI model development. The applications of Cosmos extend to robot learning, where it helps generate controllable, high-fidelity synthetic data for training robot perception and policy models; autonomous vehicle training, by amplifying existing data diversity with new weather, lighting, and geolocations; and video analytics AI agents for enhanced automation and safety in industrial and urban environments.
Also Read:
- NVIDIA Advances Robotics with New AI Models for Human-Like Reasoning
- Tabnine Enhances AI Agent Capabilities with Integration of NVIDIA Nemotron Reasoning Models
NVIDIA emphasizes its commitment to the physical AI community by making Cosmos models, guardrails, and tokenizers openly available on Hugging Face and GitHub under an NVIDIA Open Model License. Furthermore, the NVIDIA Blackwell architecture, specifically NVIDIA RTX PRO 6000 Blackwell Series Servers and NVIDIA Blackwell GB200, are designed to accelerate physical AI development, offering peak performance for Cosmos world foundation models across training, synthetic data generation, simulation, and inference workloads.
