Japan's Noetra Partners with Sony, SoftBank and 44 Companies to Advance Sovereign AI, Plans to Deploy ~27,500 NVIDIA Rubin GPUs
0xBroomberg
Japanese AI firm Noetra has launched the country's largest sovereign-AI initiative with Sony, SoftBank, NEC, Honda and 44 corporate backers, planning a compute cluster armed with 27,500 Nvidia Rubin GPUs to build a full-stack AI capability — from chips to models — on home soil.
What exactly is Japan trying to build?
Noetra's goal is a complete domestic AI chain — from compute infrastructure to multimodal foundation models — built and operated in Japan.
This means → Japan no longer wants to rely solely on American models and cloud compute; it wants to own the full stack, hardware through model weights.
In plain terms = Japanese companies have mostly called OpenAI and Google APIs; this time they are training their own model from scratch.
Who is paying and who is building?
Four core members: Sony, SoftBank, NEC and Honda — spanning entertainment, telecom, IT and automotive.
The broader coalition numbers 44 investors, including manufacturing heavyweights (Fanuc, Yaskawa, Mitsubishi Electric, Tokyo Electron), Japan's three mega-banks (MUFG, Mizuho, SMBC), and tech firms such as Rakuten and Sakana AI.
This reflects a national-scale bet: sovereign AI is not one company's project but a collective wager by manufacturing, finance and energy.
Where will the compute come from?
Noetra plans to partner with Nvidia to build a cluster housing roughly 27,500 Rubin GPUs — Nvidia's next-generation architecture after the current Blackwell line, not yet in mass production.
Construction is set for April 2027; the cluster should go live by June 2028.
Until then, model development will run on existing infrastructure from Japanese AI-compute providers.
This means → the team faces a roughly three-year window where it must prove the fundamentals on limited compute before the big cluster arrives.
A three-phase model roadmap — is it realistic?
Phase 1 (FY 2026): ship a reasoning foundation model focused on Japanese-language comprehension, logical reasoning and instruction-following — solving the pain point that Japanese-language AI still lags.
Phase 2 (FY 2028): a full multimodal foundation model handling text, image, video and audio.
Phase 3 (FY 2030): "real-world native AI" — models that understand spatial perception and physical properties, designed to operate in the real world (think robotics and autonomous driving).
In plain terms = first, build a smart Japanese chatbot; then teach it to see pictures and video; finally, let it walk into the physical world.
What is the real test for Japan's sovereign AI?
The critical milestone is June 2028 — whether the compute cluster ships on time determines if phases two and three can start on schedule.
A 44-member alliance offers breadth, but the bigger the coalition, the higher the coordination cost. The actual R&D team is staffed by engineers seconded from core members, supported by Japan's AIST research institute and Preferred Networks.
This means → funding and political will are in place, but delivering the hardware and a working model within three years is the sole yardstick for whether this path holds.
Content is for reference only, not financial advice.