Nvidia bets big on physical AI at GTC Taipei with a new world model, driving brain, and open humanoid robot

The open humanoid reference design is the piece most coverage is glossing over. By publishing a reference architecture rather than a finished product, Nvidia is playing a platform-layer move borrowed from its GPU playbook: let others build the robots, but ensure every robot runs on Nvidia silicon and software.

This lands on the same day as two directly relevant stories in our archive. The Hugging Face writeup on Cosmos 3 confirms the open-weights angle is real and deliberate, not just marketing framing, which strengthens the case that Nvidia is genuinely trying to seed adoption across research labs rather than just announce. Meanwhile, the OpenAI robotics story from the same day makes the competitive framing concrete: two well-capitalized organizations are now publicly racing to own the embodied AI infrastructure layer, with meaningfully different approaches. Nvidia is betting on open models and hardware reference designs to pull developers in; OpenAI is betting on simulation-to-real transfer from foundation models. These are testable, divergent bets, and the outcome will likely hinge on which approach produces reliable manipulation and locomotion at scale first.

Watch whether third-party humanoid manufacturers (specifically those currently using non-Nvidia compute) publicly adopt the reference design within the next two quarters. Adoption by even one mid-tier OEM would confirm the platform strategy is working; silence would suggest the open framing is more positioning than traction.