Japan and NVIDIA announce a national factory for physical AI
Japan just bought a national training stack for robots that move steel, not chatbots that move tokens. On July 16, NVIDIA and Noetra Corp. announced a Vera Rubin AI factory backed by Japan’s Ministry of Economy, Trade and Industry — pitched as the world’s first national infrastructure built for physical AI.
Per NVIDIA’s investor release, the plant will run 13,750 Vera CPUs and 27,500 Rubin GPUs, deliver about 140 megawatts of data-center capacity on the NVIDIA DSX platform, and scale over Spectrum-X Ethernet. The buyer is industrial Japan. The product is multimodal models for factories, logistics, healthcare, and telecom.
What the Vera Rubin factory is for
Noetra will operate the factory. NVIDIA supplies Vera Rubin NVL72 racks, BlueField DPUs, and the DSX reference design meant to cut time-to-production and raise tokens per megawatt. The job is training open multimodal foundation models that power AI agents, digital twins, robotics, and other physical-AI applications.
Pretrained weights from Noetra’s models are supposed to go out broadly to domestic developers and enterprises, alongside NVIDIA software such as Nemotron, Cosmos, Isaac GR00T open models, and NeMo libraries. That is the sovereignty play in plain language: Japanese industry keeps a shared model layer on Japanese-facing infra, even if the silicon brand is American.
Jensen Huang framed it as industrial succession. “Japan invented modern manufacturing. Now, it is building the AI factories that will power the next industrial revolution,” he said in the release. METI minister Ryosei Akazawa cast FRONTia as the core of the country’s physical-AI ecosystem — global vendors plus Japan’s onsite manufacturing know-how.
FRONTia multimodal robotics models
The compute feeds METI’s FRONTia Project — formally “Development of Multimodal Foundation Models with a View to AI Robotics and Physical AI.” The brief is reliability under industrial constraints: real factory data, real robot kinematics, models that do not hallucinate a gripper into a human’s hand.
NVIDIA says the site will eventually support trillion-parameter-scale training. That is a capacity claim, not a shipped model card. The useful pipeline is national-scale pretraining, then open weights and NVIDIA’s robotics stack for sector fine-tuning. If FRONTia works, a logistics firm or hospital system does not start from a U.S. chat model and hope sim-to-real holds.
Cosmos Coalition industrial partners
The factory sits next to a wider industrial stack. In a July 15 NVIDIA newsroom release, NVIDIA said Japanese physical-AI leaders including FANUC, Fujitsu, Kawasaki Heavy Industries, Yaskawa Electric, plus SoftBank Corp., Sony, Hitachi, NEC, Kubota, and others, intend to join the NVIDIA Cosmos Coalition to help build open world models.
Fujitsu is exploring a collaborative control platform with FANUC, Yaskawa, and Kawasaki, integrating Cosmos, Isaac, Omniverse, and the Newton physics engine for digital twins and sim-to-real validation before robots hit the floor. That is the industrial layer above the Noetra factory: shared control and simulation, not just more GPUs. “Intend to join” is not a signed JV with a go-live date.
Why Japan is buying physical-AI infra now
Japan’s AI Robotics Strategy, cited in the IR note, aims for more than 30% of the global AI robotics market by 2040 — an estimated $133 billion opportunity. Demographics do the rest of the arguing. Aging labor pools and precision manufacturing are a match for robots that can see, plan, and act, if the models are trained on Japanese industrial reality.
This is also NVIDIA locking the next demand curve. Language-model training is crowded. Physical AI needs denser simulation, robotics data, and factory integration — exactly where FANUC and Yaskawa already own customer relationships. Japan gets a national compute story. NVIDIA gets another sovereign-scale Rubin deployment. The AI infra race did not end at chat. It moved onto the shop floor.
