Japan and NVIDIA Launch World's First National AI Infrastructure: 27,500 Rubin GPUs, 140MW
NVIDIA and Japan’s Noetra Corp announced on July 16 the deployment of what both parties are calling the world’s first national AI infrastructure for physical AI. The facility is backed by Japan’s Ministry of Economy, Trade and Industry (METI) and will run on the NVIDIA Vera Rubin NVL72 rack system — the same hardware CoreWeave was first to deploy at rack scale two months ago.
The Hardware Stack
The Noetra AI factory will be built on NVIDIA’s DSX platform and configured as follows:
- 13,750 NVIDIA Vera CPUs
- 27,500 NVIDIA Rubin GPUs (2 per CPU, in NVL72 racks)
- 140 megawatts of data center capacity
- Architecture: NVIDIA DSX platform, Vera Rubin NVL72 racks
The Vera Rubin NVL72 packs 72 Rubin GPUs per rack with liquid cooling and high-bandwidth NVLink interconnects. At 140MW, the Noetra facility is a serious compute cluster by any measure — equivalent in power to roughly 1.4 large enterprise data halls at traditional sizing, though AI factory power density runs considerably higher.
Scope and Governance
The Noetra consortium is backed by Japan’s industrial and technology sector, with METI providing government support under the FRONTia Project (Framework for Realizing Open and Novel Technologies in Industry and AI). The initiative’s stated targets:
- Manufacturing — AI agents and simulation for production systems
- Logistics — physical AI for warehouse and transport automation
- Healthcare — multimodal clinical AI and diagnostics
- Telecommunications — network-optimised inference and agent infrastructure
Noetra will create open multimodal foundation models designed for AI agents, digital twins, robotics, and physical AI applications. The open model framing matters: it positions Japan’s national infrastructure as a training resource for domestic industry rather than a captive compute pool for a single lab.
Why Rubin, Why Now
The timing is deliberate. NVIDIA’s Rubin GPU (Vera Rubin NVL72 configuration) only entered production in late 2026, with CoreWeave as the first to deploy at rack scale and set MLPerf v6.0 records on the hardware. Japan’s Noetra move locks in next-generation silicon before the Rubin supply tightens — hyperscalers are already committing multi-gigawatt Rubin buildouts through 2028.
For NVIDIA, the Noetra deal is more than a hardware sale. It is the clearest articulation yet of the “national AI factory” model NVIDIA CEO Jensen Huang has been pitching to sovereign governments for two years. Japan becomes the reference customer for that thesis — a METI-backed facility running NVIDIA silicon as public infrastructure rather than commercial cloud.
What Japan Gets
Japan has been outspent on AI compute by the US and China, with domestic model development constrained by access to frontier hardware. The Noetra factory changes the equation: 27,500 Rubin GPUs gives Japanese researchers and manufacturers access to compute at a scale that wasn’t domestically available under prior hardware generations.
The open multimodal foundation model deliverable is the specific bet METI is making — that the bottleneck for Japan’s AI adoption is not software or research talent, but pre-trained models fine-tuned to Japanese-language industrial data. If those models exist and are open, the FRONTia Project succeeds at its policy goal.