Physical artificial intelligence with NVIDIA Cosmos

NVIDIA Cosmos is a platform for physical Artificial Intelligence that provides open world foundation models, guardrails, and an accelerated data pipeline for developers building robots, autonomous vehicles, and video analytics agents. Models, tooling, and datasets are available on GitHub and Hugging Face under the NVIDIA Open Model License.

NVIDIA Cosmos is presented as a platform purpose-built for physical Artificial Intelligence. It combines pretrained multimodal world foundation models, guardrails, and an accelerated data processing and curation pipeline to help developers simulate, reason, and generate synthetic data for downstream workflows. The site emphasizes Cosmos as a developer-facing stack for autonomous vehicles, robots, and video analytics agents, and points users to GitHub and Hugging Face where Cosmos models, tokenizers, and resources are openly available under the NVIDIA Open Model License.

The Cosmos model family contains three core world foundation models. Cosmos Predict is a world state prediction model that can generate up to 30 seconds of continuous video from multimodal inputs to support forecasting and scenario planning. Cosmos Transfer is a multi-control style transfer model that scales single simulations across environments and lighting conditions, ingesting physics-based simulator outputs from tools such as CARLA or NVIDIA Isaac Sim to accelerate controllable synthetic data generation. Cosmos Reason is a reasoning vision language model designed to answer text queries about image and video inputs, applying prior knowledge, physics understanding, and common-sense reasoning to annotate, critique, or create prompts for the other models.

Data and tooling are highlighted as integral to the platform. NVIDIA Cosmos Curator is a framework for filtering, annotating, and deduplicating large sensor datasets, and Cosmos Dataset Search lets developers query and retrieve targeted scenarios for post-training. The site links to the Cosmos Cookbook and PyTorch post-training scripts to customize models, and references complementary NVIDIA technologies including NeMo Curator, the Cosmos tokenizer, NVIDIA NeMo, NIM microservices, and NVIDIA DGX Cloud. For infrastructure, the page recommends NVIDIA RTX PRO 6000 Blackwell Series servers and the NVIDIA Blackwell GB200 for peak performance on post-training and inference workloads. The ecosystem section lists partners and adopters across robotics, autonomous vehicle, and vision AI industries and points developers to demos, sessions, and news to get started with Cosmos.

58

Impact Score

El Capitan retains top spot on TOP500 list

The 66th TOP500 list, unveiled at the SC25 conference in St Louis, confirms El Capitan at Lawrence Livermore National Laboratory as the fastest supercomputer following benchmark remeasurements. The list also highlights continued U.S. leadership, historic European milestones, and increasing global diversity in architectures and energy efficiency.

Microsoft rolls out Cobalt 200 cpu with 132 Arm cores

Microsoft introduced the Cobalt 200, a custom server processor for Azure that uses Artificial Intelligence-driven simulations during design. The chip promises roughly 50 percent better performance than the prior Cobalt 100 and wider customer availability in 2026.

NVIDIA demand may double lpddr5x and server dram prices in 2026

Counterpoint Research warns memory prices will keep climbing into early 2026 as demand from NVIDIA and other Artificial Intelligence deployments tightens LPDDR supply. The firm flags steep short-term increases and wider risks from tariffs, geopolitics, and labor costs.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.