At a Tokyo summit, NVIDIA and Japan's industrial leaders launched a national AI infrastructure push — pairing a new on-device model with robots and factories in a broad bet on "physical AI."
For two years the AI story has been about chatbots: models that read and write text in distant data centers. The latest move out of Japan points somewhere else — models that perceive, reason, and act inside the physical world, running not in the cloud but on the machine itself. It is a shift from software that talks to software that works.
At a "Japan AI Ecosystem" event in Tokyo, NVIDIA CEO Jensen Huang pledged to help the country build its AI-powered future, declaring "now is the time for Japan AI." The headline product is Cosmos 3 Edge, a "world model" designed to run directly on edge devices such as robots and the company's Jetson Thor compute modules. Alongside it, NVIDIA expanded Metropolis — a toolkit for building vision-aware, agent-like applications — and announced a coalition of Japanese manufacturing and robotics leaders, including FANUC, Hitachi, Kawasaki Heavy Industries, SoftBank, Sony, and Yaskawa.
Instead of every robot phoning a data center to decide what to do, Cosmos 3 Edge lets the device reason locally — perceiving its surroundings, predicting how the world will change, and choosing an action in real time. That removes network latency and privacy concerns, and it is what lets a factory floor or a warehouse run autonomously even when connectivity is poor.
Most AI today is disembodied: it moves tokens of text. Physical AI must handle sensors, motion, and unpredictable environments. A world model like Cosmos is trained to understand how physical scenes evolve — a box tips, a arm reaches, a person steps into a path — so the system can simulate outcomes before acting. Putting that capability "at the edge" means the intelligence lives where the work happens, not in a server farm an ocean away.
The Japan deal is a signal of where the industry is heading: from language models that answer questions toward embodied systems that do jobs. If physical AI scales, the bottleneck moves from "how smart is the model" to "how cheaply can we put that intelligence on a machine." Nations with strong manufacturing bases — and the political will to fund shared compute — stand to gain an early lead in the next industrial layer.
The broader lesson is about location. Intelligence is migrating off the screen and into the world. The question is no longer only what AI can say, but what it can do when it is bolted to a robotic arm.