Japan Bets on Physical AI in a 140-Megawatt “AI Factory”

What happened

On July 16, Japan’s Noetra announced the full-scale development of domestic multimodal foundation models together with companies including Sony Group, SoftBank, NEC and Honda. With NVIDIA’s support, the company plans to build an AI computing infrastructure equipped with about 27,500 Rubin GPUs: construction is scheduled to begin in April 2027, with operations expected to start in June 2028. NVIDIA said the system will provide 140 megawatts of data center capacity and serve as the computing foundation for Japan’s government FRONTia project for physical AI.

This is not just a chip purchase. Noetra plans to start with Japanese-language understanding, logical reasoning and instruction execution, then gradually move toward multimodal models capable of handling text, images, video and audio, and by 2030 explore “real-world-native AI” that can understand spatial relationships and serve real environments. Model outputs will be offered externally in phases, with the exact timetable depending on R&D and deployment progress.

Why it matters

Over the past two years, the main thread in AI infrastructure has largely been driven by U.S. cloud providers and model companies. Japan’s proposal, by contrast, puts national industrial policy, manufacturing data and compute capital expenditures on the same balance sheet. In its AI robotics strategy announced at the end of June, Japan’s Ministry of Economy, Trade and Industry set a target of about 10 million robot deployments by 2040 and named domestic multimodal foundation models as a key project. Noetra said 44 companies and institutions are already participating as investors or collaborators.

For the market, the key change is that the definition of compute demand is expanding. Training a model is only the starting point; real commercial value comes from factories, logistics, healthcare and robotic systems continuously invoking the model. That means power, data centers, networking equipment, industrial software and sensors could all become part of physical AI infrastructure. SoftBank previously disclosed that it is building a 140-megawatt-scale AI data center in Japan; the Noetra project echoes that power and data-center construction cycle, and also shows that Japan is trying to form a domestic chain from energy and compute to industrial applications.

But this is still a long-term capital allocation plan, not yet a validated revenue growth story. GPU counts, data center capacity and government support can ease supply constraints, but they do not automatically solve for high-quality industrial data, model reliability, deployment costs or customer willingness to pay.

What to watch next

First, whether the 2027 start of construction and 2028 launch can proceed on schedule, especially power interconnection, network connectivity and Rubin system delivery. Second, whether Noetra can turn its “domestic models” into tools that manufacturing customers are willing to buy on a long-term basis, rather than leaving them as public R&D projects. Third, whether the model weights and software are truly open, and how that openness affects Japanese developers and overseas partners. Finally, it will be important to see whether Japan’s robot adoption subsidies, data governance and liability rules can move in step with infrastructure construction.

If these pieces can come together, Japan’s competitive focus will not just be catching up on general-purpose models, but using manufacturing sites, robots and industrial data to build a physical AI supply system that is closer to a “national industrial operating system.” For now, the market can confirm the direction of investment, but not the pace of returns.

Sources

• Noetra: Launch of domestic multimodal foundation model development
• NVIDIA: Japan government, industry leaders and NVIDIA launch the world’s first national AI infrastructure
• Japan Ministry of Economy, Trade and Industry: Press conference on AI robotics strategy
• SoftBank: FY2025 investor presentation materials