SoftBank assembles Japan Inc to build domestic AI model

A SoftBank-led consortium bringing together Japanese industrial groups and major banks is planning a domestic “foundation model” program aimed at reducing reliance on US and Chinese AI systems. The Decoder reports that eight Japanese corporations—including NEC, Honda, Sony, Nippon Steel and Kobe Steel, alongside three major banks—have invested in a new SoftBank unit that says it wants to build a roughly one‑trillion‑parameter model by the end of the decade.

The pitch is not consumer chatbots. The project is framed around “Physical AI”: models designed to control robots and machinery in factories and other industrial environments. That focus explains why steelmakers and manufacturers are at the table. Japan’s competitive advantage is still embedded in production systems and process know‑how, and the moment AI becomes part of the control layer, the training data stops being marketing copy and starts being operational telemetry.

Nikkei, cited by The Decoder, describes growing concern among Japanese firms that sensitive industrial data used for model training could flow to foreign servers when companies rely on OpenAI, Anthropic, or Chinese providers. The proposed fix is jurisdictional as much as technical: the consortium says data processing will take place on Japanese soil, including at a data centre SoftBank is building in a former Sharp LCD factory in Sakai near Osaka. The physical location is part of the product—compute as a domestic utility, not a cloud feature.

The initiative is also designed to match the state’s funding channels. Through NEDO, Japan’s government is expected to direct roughly one trillion yen over five years into national AI development, and SoftBank’s unit is described as a leading candidate for that money. When public funding is large and earmarked, the private incentive is to form a vehicle that looks like national infrastructure: broad corporate membership, domestic data residency, and an industrial use case.

Japan is entering the AI race late, but with a different playbook. Rather than betting on a single startup to out‑iterate Silicon Valley, it is assembling a coalition of firms that already control factories, balance sheets, and procurement. That can accelerate deployment if the model works—and it can also lock in a standard that is expensive to reverse.

SoftBank’s data centre will be built in a repurposed electronics plant, and the model is supposed to be trained on domestic compute. The consortium’s first deliverable is therefore not a product launch but a claim about where the data will sit when the robots start listening.