Nvidia has recruited most of Japan’s industrial robotics establishment into the Cosmos Coalition, the open world-model programme it uses to seed its physical AI stack, in announcements timed to Jensen Huang’s week in Tokyo.
Twenty-two companies are named: AIRoA, classmethod, Enactic, FANUC, Fujitsu, GROOVE X, Hitachi, Honda R&D, Kawasaki Heavy Industries, Kubota, Mitsui & Co., Mitsubishi Corp., Mujin, NEC, Preferred Networks, SoftBank Corp., Sony Group, Telexistence, TIER IV, TRON K.K., Turing, and Yaskawa Electric.
All of them, in Nvidia’s wording, “intend to join”. No binding commitment has been disclosed, and no money has been mentioned in either direction.
That roll call is the news. FANUC and Yaskawa are the two largest industrial robot makers on earth by installed base, and they have spent decades running proprietary control stacks. Signing them up to a coalition organised around someone else’s open models is not a small thing, even at the level of intent.
It is also the same play Nvidia ran with Hyundai around the Atlas humanoid, and is running against Tesla’s vertically integrated Optimus programme: supply the stack, let the industrial partner own the deployment.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!
The technical centrepiece is Cosmos 3 Edge, a four-billion-parameter model built on Nvidia’s Nemotron family that runs on Jetson edge hardware rather than in a data centre.
Nvidia says developers can adapt it to a specific robot, vehicle, or sensor rig in about a day, and that it will deploy across RTX GPUs, DGX systems, and the newly announced Jetson T2000 and T3000 modules.
Fujitsu is leading the most concrete piece of the programme, a collaborative control platform being explored with FANUC, Yaskawa, and Kawasaki that would bridge digital and physical operations across industrial sectors.
It is built on Cosmos world foundation models, the Isaac robotics platform, Omniverse NuRec libraries, and the Newton physics engine, and is meant to handle digital twins, robot learning, and simulation-to-real validation before anything touches a factory floor.
Elsewhere in the announcement the applications get specific in a way these things usually do not. Kubota is looking at Cosmos for autonomous agriculture, and Enactic is fine-tuning Nvidia’s Isaac GR00T model for elder-care semi-humanoid robots.
Shimizu Corporation is using Metropolis for construction-site safety, GROOVE X is building Jetson-powered companion robots, and Kawasaki is spreading the technology across healthcare, shipbuilding, transportation, aerospace, and energy.
“Japan invented modern manufacturing,” Huang said in the release. “Now, it has the opportunity to reinvent it for the age of intelligent industries.” He called physical AI “a once-in-a-generation opportunity” for the country.
A second release the same day covered the language layer. Institute of Science Tokyo built its Swallow open models on Nemotron datasets, and SB Intuitions, SoftBank’s generative AI subsidiary, trained its Sarashina series using Nemotron libraries.
Sarashina3 mini has since been picked up by Japan’s Digital Agency. SoftBank Corp. has deployed a Large Telecom Model for autonomous network operations, and NTT DATA used Nvidia’s Japanese personas dataset to augment training for its tsuzumi 2 model.
The framing throughout is sovereignty. “Every nation and every company should own and control its intelligence infrastructure,” Huang said. “Open models make that possible.”
It is a pitch that lands differently depending on where you sit, given that the open models in question run best on hardware sold by one company. Nvidia has made the same argument in Europe, where it recently unveiled 35 new AI supercomputers.
Japan’s interest is less abstract than the rhetoric suggests. Nvidia’s release points to the ageing population and workforce transition as the driver, which is the polite version of a labour shortage that Japanese manufacturers have been managing for a decade.
Sakana AI, meanwhile, is wiring Nemotron into its Fugu model-routing platform, which picks the best model for each task rather than betting on one.
What none of it settles is whether world models actually shorten the road from demonstration to deployment. That question has been sitting unanswered through China’s crowded robot boom, and a coalition of intent does not answer it either.
Get the TNW newsletter
Get the most important tech news in your inbox each week.