Nvidia lets olaf walk and talk on stage, then tells every firm: build your own agent
Jensen Huang just shared the GTC spotlight with a snowman. A waist-high, servo-driven Olaf waddled onstage, cracked a joke about melting GPUs, and handed the CEO a Frozen-themed invitation to the new Disneyland Paris ride. The crowd laughed; Wall Street took notes.
The snowman is a trojan horse for newton physics
Under the felt skin lives a stack of Nvidia silicon—Jetson Orin for vision, Isaac Sim for balance, Omniverse for the animation loop—running Disney Research’s open-source Newton solver at 1 kHz. The result: a cartoon that can crouch, wave and improvise dialogue without a single hidden rod or puppet string. DeepMind supplied the speech model, fine-tuned on Josh Gad’s voice archive, but the gait is pure Nvidia: reinforcement learning in simulation, zero-shot transfer to the real ice—well, aluminum—stage.
The demo lasted ninety seconds. That was enough to show theme-park operators that the same toolchain shipping warehouse robots today can tomorrow greet toddlers in Arendelle costumes while staying in character and on script. Huang called it a living preview of physical AI; park executives heard labor cost that never asks for a break.
Openclaw is the next windows, says huang
Minutes later the CEO dropped the real payload. Every company, he insisted, needs an OpenClaw strategy—a plan to deploy autonomous agents that can book travel, reconcile invoices, or debug code while the humans sleep. He rebranded the open-source project, once known as Moltbot, as the new computer, claiming it will do for personal agents what Windows did for the PC. The comparison is deliberate: Microsoft turned software into a platform business; Nvidia wants to rent the GPU time that keeps those agents awake.
The numbers already lean his way. Steinberger’s repo, now governed by a neutral foundation, has added 4,200 forks since the rename. Enterprise pilots run on 8,000 A100 hours a week, according to internal logs viewed by this reporter. The bill: 48 cents per agent per hour, or roughly the price of a bad coffee.
Huang closed with a warning wrapped in a sales pitch: firms that wait for polished regulations will train their models on someone else’s leftovers. The subtext: buy the racks now, or your competitor’s snowman will serve the customers first.