Japan's Noetra Consortium Commits to 27,500 Nvidia Rubin GPUs for Physical AI

Japan is building a 140MW AI factory dedicated to physical AI. The facility, operated by Noetra Corp, will run 13,750 Nvidia Vera CPUs and 27,500 Rubin GPUs in Vera Rubin NVL72 racks, connected via Spectrum-X Ethernet and BlueField DPUs. This is the first concrete national AI infrastructure for robotics, according to Nvidia.

Noetra is a private consortium majority-owned by SoftBank, NEC, Sony Group, and Honda. It is not a government body. The consortium was formed to pool resources for physical AI, which requires enormous compute, data, and foundational technologies that no single company can handle alone, as Noetra CEO Hironobu Tamba stated.

Funding Structure: ¥1 Trillion Ceiling, Stage-Gate Reviews

The Japanese government, through METI and NEDO, commissioned Noetra and AIST to deliver the FRONTia Project (Development of Multimodal Foundation Models with a View to AI Robotics and Physical AI). METI committed up to ¥1 trillion (~$6 billion) across fiscal 2026-2030, but only the first two years are locked. The first tranche is ¥387.3 billion from Japan's GX Economy Transition Bonds. The remaining funds require annual stage-gate reviews, making ¥1 trillion a ceiling, not a promise.

Strategic Target: 30% of Global AI Robotics Market by 2040

Japan's AI Robotics Strategy, published in March, aims for more than 30% of the global AI robotics market by 2040, estimated at $133 billion. Separately, METI set a target of 10 million AI-equipped robots across 18 sectors by the same year. These numbers come from different documents but point to the same goal: embedding intelligence into Japan's existing manufacturing strength.

Technical Specifications: Vera Rubin NVL72, Spectrum-X, BlueField

The AI factory will use Nvidia's Vera Rubin NVL72 racks, which integrate 72 GPUs per rack with high-speed NVLink. The network relies on Spectrum-X Ethernet for scalability and BlueField DPUs for data processing acceleration. No purchase price or deal value was disclosed. Nvidia's release describes working with Noetra to launch a facility Noetra is establishing, not a direct government purchase.

Comparison to Europe: Faster Execution, Single Vendor Risk

Japan's approach contrasts with Europe's stalled gigafactory programme. Japan has a named operator, ministry, bond-funded first tranche, and rack count. However, it handed the entire stack to a single American vendor, which Europe has spent years debating. Sovereignty here means owning the models and deployment while renting the silicon.

Next Steps for Developers

If you work in robotics or AI, watch Noetra's open multimodal foundation models. They will be trained on this infrastructure and released as open models. For Japan-based developers, consider contributing to the FRONTia Project or building on top of these models. For hardware engineers, the Vera Rubin NVL72 configuration sets a new benchmark for AI training clusters.

Why This Matters for Developers

This project provides a rare concrete specification for a national AI infrastructure. Developers can expect open foundation models for physical AI, digital twins, and robots. The scale (27,500 GPUs) means significant advancements in multimodal AI, and the open nature of the models could accelerate robotics development globally.