All companies

Local open source · Adjacent progress · Company profile

NVIDIA

NVIDIA Cosmos is a world foundation model platform for physical AI, synthetic data, and robot-training workflows, not a consumer 3D world creator.

What this company is for

Cosmos explains the physical-AI infrastructure lane.

Consumer access. Local open source

Visitor can see. Platform story: synthetic data, world foundation models, robot-training workflows.

Where the claim stops. For robotics and physical-AI workflows, not a 3D creator.

This company's models

The model NVIDIA ships

World foundation model platform

Cosmos

Physical-AI world foundation model platform; Cosmos 3 arrived March 2026.

Public GitHub repos: Predict, Transfer, Reason, Curate, Evaluator, RL, Cookbook.

Not the same category as consumer 3D world generation tools.

Open the Cosmos dossier

What to open first

Cosmos is infrastructure for physical AI, not a Marble alternative.

NVIDIA launched Cosmos as a world foundation model platform for physical AI: synthetic data, simulation, robotics, vehicles, and embodied reasoning. Cosmos 3 arrived as a March 2026 preview direction across generation, reasoning, and action simulation.

The public GitHub organization is the inspectable surface: Predict, Transfer, Reason, Curate, Evaluator, RL, Cookbook, and related repos. That is a platform catalog, not a single consumer demo.

Why it is adjacent

The word world here means a trainable environment, not a shareable scene.

A Cosmos workflow can matter enormously to a robotics team and still be the wrong click for a game artist. This site keeps NVIDIA in the adjacent-progress lane so that physical-AI progress does not get sold as an explorable 3D product.

Hardware, SDK, and enterprise context still matter. Preview and available components should be named from current NVIDIA posts and the GitHub stack, not flattened into one ship date.

How to verify

Use NVIDIA Newsroom plus the Cosmos GitHub org.

Launch posts, Cosmos 3 coverage, the Physical AI Data Factory Blueprint, and the GitHub organization are the primary sources. NVIDIA's world-models glossary is useful framing, not a product spec.

If the reader wanted Gaussian splat export or a 360 skybox, send them back to Marble or Skybox AI.

Keep reading

Sources