🚀 One of the big themes at GTC 2026: shifting from “we only have a little data” → “ok wow… this model is smashing leaderboards.” Why? Because with modern world models, the loop gets interesting:
Better models → generate better synthetic data → which trains even better models.
Rinse. Repeat. Improve. 📈
At GTC, we announced the Physical AI Data Factory Blueprint, an open reference architecture that unifies data curation, augmentation, evaluation, and orchestration. So teams can turn small, fragmented datasets into large, diverse, model-ready training sets.
It also includes a new framework - Cosmos Evaluator, for reviewing and scoring outputs from generative video systems.
- Read the press release 📖 NVIDIA Announces Open Physical AI Data Factory Blueprint to Accelerate Robotics, Vision AI Agents and Autonomous Vehicle Development | NVIDIA Newsroom
We also announced the latest updates to the Cosmos Platform, including:
Cosmos Reason 2 new checkpoint - improved benchmarks + reduced hallucinations. Use Cosmos Reason 2 with the latest NVIDIA Metropolis VSS Blueprint v3 to build advanced video analytics AI agents.
Cosmos Predict 2.5, Cosmos Transfer 2.5 → shredded some layers for faster inference & lower GPU cost.
- Download on GitHub 👉 NVIDIA Cosmos · GitHub
- Start with the Cosmos Cookbook 🧑🍳 Cosmos Cookbook
And did you hear…. Cosmos 3 is coming 🎤
🩺 Bonus: Cosmos-H-Surgical
We are just getting started, and the next stop is surgical robotics built on Cosmos.
- Check it out 👉 NVIDIA-Medtech · GitHub