Training a physical AI system — one that has to handle a factory floor that doesn't match its training conditions — requires orders of magnitude more compute than running a chatbot, which is precisely why Japan committed at this scale. The facility, built on Nvidia's Vera Rubin DSX architecture with Spectrum-X networking, is designed to support trillion-parameter model training as capacity expands, and Nvidia CEO Jensen Huang personally traveled to Tokyo for the announcement.
The project sits inside METI's FRONTia initiative and Japan's March 2026 AI Robotics Strategy, which targets more than 30% of the global AI robotics market — an estimated $133 billion opportunity — by 2040, alongside a separate national goal of deploying 10 million AI-equipped robots across 18 sectors by 2040. Noetra, led by Tamba (who previously built SoftBank's large-language model), plans to release its first foundation model by March 2027 and will make pretrained weights available to Japanese developers and enterprises, with SoftBank, Sony, and dozens of other manufacturing and technology companies joining as partners.
"Bringing physical AI into the real world requires enormous computing, data and foundational technologies — challenges no single company can solve alone," Tamba said at the launch. Manufacturing, logistics, healthcare, and telecommunications are the initial target sectors.
