Jim Fan in Three Minutes

The Room · Jim Fan Talks Physical AI

The evening was hosted by Nebius, the AI-cloud company training a growing share of the world's robot foundation models, as one stop on its RoboTour series – and co-hosted by Lukas Ziegler, who curated the room together with the RAISE Summit and MACHINA Summit teams. Partners in the mix: NVIDIA, Encord, Enchanted Tools and Genesis AI. The guest list hit capacity before the doors opened (event page).

Then Jim Fan, Director of AI & Distinguished Scientist at NVIDIA and one of the driving forces behind the GR00T models, opened his three minutes on physical AI bluntly: "The time for robotics is now – and we're still early in the game." Three numbers, three strategies, three minutes. Watch it here – then read on for the notes.

Three Numbers

The case for humanoids starts with the macro picture and ends on his own download counter.

• –1.5% · the zero-trillion-dollar business
  The world is aging and there are not enough humans for the jobs required. Fan puts the loss from vacant jobs at 1.5% of GDP – more than $1 trillion. "We are in a zero-trillion-dollar business: zero today, one trillion ahead."
• $40K · a humanoid at Walmart prices
  A Unitree G1 is now listed on Walmart for $40K – roughly the price of a car, against the ~$2 million Honda's ASIMO cost twenty years ago. That is "about the highest price humanoids will ever have". From here it only falls, toward the cost of the raw materials.
• 10 million · an ecosystem forms
  NVIDIA's open-source GR00T models and physical-AI datasets have passed 10 million downloads. The whole community is building physical AI together.

Three Strategies

If the numbers say why now, the strategies say how Jim Fan expects physical AI to get built: a new model paradigm, two new data engines, and a learning loop borrowed from the LLM world.

• Models · world action models
  The next modeling paradigm is the world action model: physical AI that produces pixels, learns physics from vast amounts of video, and predicts the next frame.
• Data · human data + simulation
  Two engines feed it. Egocentric human data – sensors on humans, taking the robot out of the loop – which Fan calls "the new FSD": full-self-driving-style data collection at fleet scale. And simulation data, ever more realistic and ever faster, with Cosmos using neural world models as the new simulator.
• Learning · agentic robotics
  "The LLM folks have had all the fun – so we're going to take some of their toys." Apply the latest agentic techniques to robots so they learn skills automatically, building ever-expanding libraries of novel skills – and eventually, fully autonomous research in the physical world. General robotics is the next learning paradigm.

The Full Version · From the Machina Stage

Later that week, at Machina 2026 at Station F, Fan gave the long cut – opening with what he calls the Great Parallel: 2016, the first DGX and Jensen with Elon; 2020, GPT-3; 2022, InstructGPT; 2024, o1 surpassing imitation learning; 2026, AutoResearch. The LLM endgame is here – "they have a party. Robotics should have some fun too."

His new bet is the World Action Model. Instead of a traditional VLA taking a frame and predicting an action, the robot dreams – it generates video a few seconds into the future. Fan calls it the GPT-2 moment for robotics, with Cosmos simulating buoyancy, reflection and collision.

On data: the golden era of teleoperation (2022–2025) is bounded by 24 hours per day – in practice about 3. NVIDIA is going all in on egocentric data with EgoScale: 99.99% human videos, 20,000 hours, no robot data in pre-training. Add 50 hours of glove data and 4 hours of teleop, and EgoScale can sort cards or manipulate syringes. Physical RL comes next: scan 3D worlds with an iPhone and simulate them in Isaac Lab. The endgame is fully autonomous robot research – already demoed in NVIDIA's lab.