NVIDIA said on July 15 that a group of Japanese enterprises, startups, and research institutions are building industry-specialized AI models on its open Nemotron model family, an effort it frames as addressing Japan’s demographic and workforce pressures directly through domestically controlled AI infrastructure. The list of participants is unusually broad for a single national push: the Institute of Science Tokyo, SoftBank Corp. and SB Intuitions, Stockmark, avatarin, ENEOS Holdings, Hitachi, NTT DATA, and Sakana AI.
What each organization is building
The Institute of Science Tokyo used Nemotron to build its “Swallow” models for financial-document translation and asset-management reporting. SB Intuitions built “Sarashina” models, with the compact Sarashina3 mini already selected for use by Japan’s Digital Agency. SoftBank deployed a Large Telecom Model for autonomous network operations, while Stockmark released a Japanese-language document-understanding model. NTT DATA used Nemotron to enhance its own “tsuzumi 2” model, Hitachi built physical AI for coordinating IT and operational technology, ENEOS applied it to materials research and molecular screening, and Sakana AI integrated it into its Fugu model-orchestration platform. NVIDIA CEO Jensen Huang framed the underlying strategy plainly: “Every nation and every company should own and control its intelligence infrastructure.”
Why it matters to enterprise AI buyers
This is a concrete, multi-industry case study of what “sovereign AI” looks like in practice, not as policy rhetoric but as production deployments across finance, telecom, energy, and public-sector software, all built on the same open model family rather than eight separate one-off contracts. For enterprise buyers outside Japan weighing build-vs-buy on model infrastructure, or negotiating data-residency and model-control terms with a foundation-model vendor, this is a working reference for what a national- or enterprise-scale “own the stack” strategy actually requires: an open base model, a chip and systems partner, and industry-specific fine-tuning done in-house rather than purchased as a finished product.
Source: NVIDIA Newsroom
