NVIDIA's GR00T N1: the open model that wants to be every humanoid's brain.

NVIDIA's move at GTC 2025 was to commoditize the layer everyone else was trying to own. Where Figure and Tesla are building proprietary, vertically integrated intelligence, NVIDIA released GR00T N1 as an open, customizable foundation model any humanoid maker can download, fine-tune, and deploy. It is a dual-system vision-language-action model — a slower reasoning module paired with a fast diffusion-transformer motor policy — trained on a heterogeneous mix of real robot trajectories, human video, and synthetic data. The pitch to a hardware company is blunt: do not spend two years and a fortune building a brain from scratch when you can start from ours.

The early-access roster is the tell. NVIDIA named Agility Robotics, Boston Dynamics, Mentee Robotics, and NEURA Robotics among first partners, which spans warehouse workhorses to research-grade platforms. For a category where every startup claims a proprietary AI edge, an open baseline that credible names are willing to build on resets expectations. It means a buyer evaluating a mid-tier humanoid vendor can reasonably ask whether that vendor's in-house model actually beats a free, well-funded, continuously improving alternative — and many will not.

Just as important was the surrounding scaffolding: the Isaac GR00T Blueprint for generating synthetic training data, and Newton, an open-source physics engine built with Google DeepMind and Disney Research. The scarce input in humanoid AI is training data, because you cannot cheaply collect millions of hours of real-world manipulation. NVIDIA's answer is to let developers manufacture that data in simulation. Whoever controls the simulation-to-real pipeline shapes how fast the whole field learns, and NVIDIA is positioning to sit under everyone — selling the chips, the sim, and now the model.

For buyers, GR00T N1 lowers the odds that any single humanoid vendor achieves a durable software monopoly, which is good for pricing and optionality. The caveat is that an open baseline raises the floor without guaranteeing the ceiling; the winning robots will still be the ones whose makers fine-tune, deploy, and support the model in real workcells. But after March 2025, 'we have our own AI' stopped being a differentiator by default. The differentiator became what you do with the AI everyone can now get.