Japan Plans a Major Rubin Chip Buildout for Domestic Physical AI

Japan is moving ahead with a large AI infrastructure project built around Nvidia’s next-generation Rubin chips. Nvidia said it is working with Noetra Corp. on an AI factory that will use 27,500 Rubin GPUs and 13,750 Vera CPUs, with 140 megawatts of data-center capacity, to support Japan’s FRONTia project. The company described it as the world’s first national AI infrastructure for physical AI.

The project is not aimed at a general-purpose chatbot. Its stated focus is multimodal foundation models for robotics, digital twins, AI agents, and industrial uses such as manufacturing, logistics, healthcare, and telecommunications.

Nvidia said the system will be built on its Vera Rubin NVL72 racks, DSX AI factory architecture, Spectrum-X Ethernet networking, and BlueField DPUs.

Japan’s Ministry of Economy, Trade and Industry, or METI, said the FRONTia project is formally tied to a domestic multimodal foundation model program for AI robotics and physical AI.

In a June 30 release, METI said the work is intended to use Japan’s onsite industrial expertise and manufacturing data to build highly reliable models. The ministry also approved an initial budget of ¥387.3 billion, with up to ¥1 trillion available over five years.

Reuters reported that Noetra, a government-supported company partly backed by Sony, plans to buy the 27,500 Rubin chips as part of the effort. The news agency also said construction is expected to begin in April 2027, with operations set to start in June 2028.

Reuters added that Nvidia is partnering separately with Japanese robotics companies including Fanuc and Yaskawa Electric to advance robotics and AI development.

Noetra’s own company announcements show that the venture is being built around a broad domestic coalition. Sony said Noetra launched full-scale research and development for a Japan-developed multimodal foundation model, and identified Sony Group, SoftBank, NEC, and Honda as core member companies and investors.

NEC issued a similar statement, saying the model will serve as the foundation for AI-enabled robots and physical AI, with collaboration across a wider network of Japanese partners.

The scale of the hardware matters because it places Japan in a growing race to build sovereign AI systems. Nvidia’s release said the project is meant to give domestic developers and enterprises access to pretrained model weights, along with software tools such as Nemotron, Cosmos, Isaac GR00T, and NeMo.

That arrangement suggests the country wants more than imported compute: it wants a local stack that can support homegrown model development.

The policy backdrop is equally important. METI said Japan’s AI Robotics Strategy, released in March, targets the introduction of about 10 million robots by 2040 and broader social deployment across 18 areas. The ministry’s June release framed physical AI as a practical answer to Japan’s industrial strengths and demographic pressure.

Japan’s approach is also specific about the kind of AI it wants. Rather than treating AI as a text-only service, the country is building for systems that can process language, images, video, and sensor data together.

Nvidia said the expanded factory will eventually support trillion-parameter-scale model training, while METI said the work should help create highly reliable multimodal foundation models for real-world use.

The timing is notable because global AI investment is increasingly tied to national strategy as much as corporate competition. Japan’s project shows how governments and industrial groups are trying to keep strategic control over compute, data, and model development rather than relying only on overseas platforms. In this case, the goal is not just to buy chips; it is to build a domestic base for robotics and industrial AI. That is the real point of the plan.