NVIDIA expands AI projects in Japan across banking and robotics

NVIDIA is working with Japanese banks, manufacturers, researchers, and government agencies on open AI models, financial infrastructure, and robotics.

NVIDIA has outlined a series of artificial intelligence projects in Japan involving banks, manufacturers, research institutions, technology companies, and government agencies. The work covers Japanese-language models, financial AI infrastructure, industrial research, robotics, and edge computing.

Japanese organizations are using NVIDIA's Nemotron open models, datasets, and software libraries to develop systems for local language and industry requirements. The models are released with open weights and supporting resources, allowing organizations to customize and deploy them within their own computing environments.

"Every nation and every company should own and control its intelligence infrastructure," NVIDIA founder and CEO Jensen Huang said. "Open models make that possible."

Japanese groups build specialized AI models

The Institute of Science Tokyo used Nemotron datasets and NVIDIA's NeMo software stack to develop its Swallow family of foundation models. The models are designed to improve Japanese-language and reasoning performance while retaining English, mathematics, and coding capabilities.

Organizations are adapting Swallow for financial-document translation and asset-management report generation. SB Intuitions, SoftBank's generative AI research subsidiary, also used Nvidia software, including NeMo RL and Megatron-LM, to train its Sarashina model series.

Japan's Digital Agency selected Sarashina3 mini for specialized AI applications. SoftBank has also combined Sarashina and Nemotron technology in a Large Telecom Model intended to support autonomous telecommunications network operations.

Stockmark released a Japanese-language document-understanding model based on Nemotron 3 Nano Omni. The company is also developing enterprise knowledge applications for manufacturers and companies in the energy and chemical sectors through Japan's Generative AI Accelerator Challenge project.

ENEOS Holdings is using Nemotron for document retrieval, visual and language processing, and simulation-based molecular screening in energy and materials research.

NTT Data used a Japan-focused Nemotron dataset to expand the training data for its tsuzumi 2 model. The company said the work improved question-answering performance, particularly for prompts requiring additional knowledge.

Hitachi is combining Nemotron and NVIDIA's Cosmos world models with its information technology and operational technology expertise. Its work focuses on coordinating digital systems and industrial operations through multi-agent software.

Sakana AI is integrating Nemotron into Fugu, its model-orchestration platform. Fugu assigns tasks to different open or proprietary models based on factors including accuracy, performance, and cost.

Mizuho plans to deploy an on-premises AI system starting with DGX B200 hardware before expanding it into a larger cluster.

The bank expects to use the infrastructure for information retrieval, document preparation, analysis, and software development support. The on-premises setup is intended to keep regulated data within the bank's environment while supporting governance and auditing requirements.

The Japan Research Institute, the main information technology company within SMBC Group, has deployed an AI computing environment using Nemotron models. The infrastructure is intended to support wider use of AI applications across the financial group.

Rakuten Bank plans to develop transaction-focused foundation models using NVIDIA's Agent Toolkit. The work will draw on Rakuten Group's service network, which includes more than 18 million bank accounts, 33 million credit cards, and 14 million brokerage accounts.

Japanese AI company Ippu Senkin is also working with a financial institution on locally operated systems for secure payment processes. Its Local AI Agent uses Nemotron, Agent Toolkit, and NemoClaw and is intended to operate within the institution's own computing environment.

Manufacturing, robotics, and edge computing

NVIDIA's work in Japan also includes manufacturing and supply chain companies. Huang met more than 30 senior executives from 16 Japanese companies involved in semiconductor equipment, memory, materials, and electronic components during a gathering in Tokyo's Kanda district.

Participants included Advantest, Tokyo Electron, Kyocera, Mitsubishi Electric, Murata Manufacturing, Panasonic, Renesas Electronics, Sumitomo Electric Industries, Taiyo Yuden, TDK, Kioxia, Mitsui, Asahi Kasei, Nittobo, Shin-Etsu Chemical, and Shibaura. The executives discussed Japan's role in the AI supply chain.

Huang also connected Japan's manufacturing base with the development of robotics. "Japan has historically been very good at precision manufacturing and very large-scale manufacturing, but now we have AI," Huang said. "You can combine the two technologies and create robotics."

The Japanese government has launched a physical AI initiative involving manufacturers, industrial data providers, and technology companies. The program will develop open multimodal foundation models for robotics, digital twins, AI agents, and systems that interact with physical environments.

"With AI and robotics, you can augment the workers you have and increase national productivity," Huang said.

NVIDIA also introduced two edge computing modules based on its Thor architecture. The Jetson T3000 provides 865 FP4 teraflops of AI processing, 32GB of LPDDR5X memory, an eight-core Arm CPU, and 25-gigabit Ethernet connectivity.

The Jetson T2000 provides 400 FP4 teraflops and 16GB of memory for visual AI systems, autonomous mobile robots, industrial manipulators, and other edge applications. NVIDIA also introduced Cosmos 3 Edge, a four-billion-parameter model designed for on-device visual processing and action generation on Thor-based hardware.

Emulation support for the T3000 is scheduled for JetPack 7.2.1, while T2000 support is planned for a later release. Both modules are scheduled to become available in the first quarter of 2027.