- SoftBank has formed a new subsidiary backed by eight major Japanese corporations — including NEC, Honda, Sony, Nippon Steel, Kobe Steel, and three large banks — to develop a domestic AI foundation model.
- The consortium is targeting a model of approximately one trillion parameters, with a stated focus on “Physical AI” for autonomous control of industrial robots and machinery.
- All data processing will be conducted on Japanese soil, anchored by a SoftBank data center under construction at a former Sharp LCD factory in Sakai, near Osaka.
- Japan’s NEDO funding agency is expected to direct roughly one trillion yen ($6.7 billion) into national AI development over five years; SoftBank’s unit is a leading candidate for that allocation.
What Happened
SoftBank has established a new business unit with equity participation from eight major Japanese industrial and financial corporations — NEC, Honda, Sony, Nippon Steel, Kobe Steel, and three unnamed major banks — aimed at building a sovereign Japanese AI foundation model, according to reporting by Maximilian Schreiner in The Decoder on April 13, 2026, which cited Japan’s Nikkei. The project targets a model with roughly one trillion parameters, with a delivery window of before the end of the decade.
The initiative is an explicit response to Japan’s deepening reliance on foundation models from American and Chinese providers, including OpenAI, Anthropic, and Alibaba.
Why It Matters
Japan’s industrial base — spanning automotive production, steelmaking, and finance — has been integrating third-party foundation models into operational workflows. As Nikkei reported, concerns are mounting over sensitive data, including the operational status of industrial facilities, being processed on servers outside Japan’s legal jurisdiction.
The effort follows a broader pattern of sovereign AI investment: France backed Mistral AI with state support, the UAE established the Technology Innovation Institute, and the European Union has pushed AI infrastructure investment through the AI Act compliance framework. Japan’s approach is distinct in its emphasis on physical-world industrial automation rather than general-purpose language or multimodal capability.
Technical Details
The consortium’s declared focus is on what it describes as “Physical AI” — systems designed to autonomously control robots and industrial machinery, a domain distinct from the text generation and multimodal reasoning that characterizes most current frontier deployments. The one-trillion-parameter target places the project in the same scale class as large models currently operated by OpenAI and Google, though no architecture, training methodology, or hardware stack has been disclosed.
Data sovereignty is an explicit engineering constraint. SoftBank is converting a former Sharp LCD manufacturing facility in Sakai, near Osaka, into a dedicated data center. All model training and inference is planned to remain on Japanese infrastructure.
The project timeline extends through the end of the decade, consistent with multi-year development cycles seen in other frontier model programs at comparable parameter scales.
Who’s Affected
The immediate stakeholders are Japan’s heavy manufacturing and banking sectors. Companies such as Honda and Nippon Steel would gain access to an AI platform controlled domestically, reducing exposure to foreign data governance regimes and geopolitical supply risk.
Existing foundation model providers — specifically OpenAI, Anthropic, and Alibaba — could face reduced enterprise adoption among major Japanese customers if the consortium produces a competitive model before 2030. Smaller Japanese software vendors that have built products on top of foreign API infrastructure would also face transition decisions.
What’s Next
Japan’s New Energy and Industrial Technology Development Organization (NEDO) is expected to channel approximately one trillion yen (roughly $6.7 billion) into national AI infrastructure over the next five years, according to Nikkei. SoftBank’s new unit is described as a leading candidate for that public funding, though formal allocation decisions have not been announced.
No technical milestones, interim model checkpoints, or benchmark targets have been published. The consortium has not disclosed a lead researcher, technical director, or the computational resources committed to the first training run.
