NVIDIA and Noetra Corp Plan Vera Rubin AI Factory in Japan — a World First for Physical AI

A National Bet on Physical AI

NVIDIA is working with Noetra Corp to launch an NVIDIA Vera Rubin AI factory in Japan, an initiative that marks the world's first national AI infrastructure for physical AI. The facility will support Japan's FRONTia Project, the METI-launched effort to develop highly reliable multimodal foundation models for physical AI — models aimed squarely at robotics and machines that operate in the real world.

For manufacturing professionals, the signal here is the framing: this is national infrastructure, not a single enterprise deployment, and its purpose is physical AI rather than general-purpose compute.

The Hardware Build

The numbers are substantial. The facility will field 13,750 NVIDIA Vera CPUs and 27,500 NVIDIA Rubin GPUs, built on the NVIDIA Vera Rubin DSX AI factory architecture using the NVIDIA DSX platform. Compute will be racked in NVIDIA Vera Rubin NVL72 systems, connected via the NVIDIA Spectrum-X Ethernet networking platform and NVIDIA BlueFieldDPUs.

All told, the AI factory will deliver 140 megawatts of data center capacity. That is a serious power envelope, and it underscores the scale at which frontier model training now operates.

Open Weights for Domestic Developers

The more interesting detail for engineers may be the access model. Noetra's multimodal foundation model weights will be made available to domestic model developers and enterprises, alongside software including NVIDIA Nemotron, NVIDIA Cosmos, NVIDIA Isaac GR00T and NVIDIA NeMo.

That matters. Pretrained weights plus an open software stack lowers the barrier for domestic teams to build on the FRONTia Project's output rather than starting from scratch — a deliberate ecosystem play, not just a hardware announcement.

The Stated Payoff

NVIDIA and Noetra Corp say the factory will provide breakthrough AI performance and lower token costs. The underlying goal remains constant across the announcements: highly reliable multimodal foundation models for physical AI, trained and deployed on sovereign Japanese infrastructure.

Whether the open-weights approach accelerates adoption as intended is the question worth watching. The compute is committed. Now the ecosystem has to use it.