Amid the rapid advancement of generative AI, we increasingly hear startups express a desire to build their own AI models strong in Japanese.

While overseas general-purpose models are indeed excellent, challenges remain in handling Japanese-specific expressions and industry-specific terminology. This is precisely why more startups are taking on the challenge of developing foundational models originating in Japan.

AWS has a specialized team for ML model training called the **GenAI Innovation Center (GenAIIC)**. The starting point for this initiative is exploring how we can leverage their expertise to support startups tackling foundational model development.

Developing foundational models is not easy. There are no guarantees of success. Still, we want to take on this challenge alongside startups driven by the ambition to bring Japan-born AI to the world.
We are looking for partners who can run alongside us, facing the same direction.


Challenges we want to solve through co-creation：
As the use of generative AI expands, we increasingly receive inquiries from startups expressing a desire to develop their own foundational models. Several common challenges underlie this trend.

- Support for Japanese and specialized domains
While general-purpose models from overseas perform well in many scenarios, there are still situations where they fall short when it comes to the subtle nuances of Japanese or specialized fields like medicine, law, and manufacturing. Many startups believe that models tailored to specific use cases could enable better service delivery.


- Data handling considerations
Some industries have companies that are cautious about sending highly confidential data to external APIs. There is also a growing need to control data handling by owning the model in-house.


- Access to Development Know-How
Training foundational models requires specialized knowledge: architecture design, hyperparameter tuning, distributed learning optimization, and more. Some express uncertainty about where to acquire the practical know-how that research papers alone cannot provide.


- Computational Resource Costs
Training large-scale models demands significant computational resources. For startups, this investment decision is far from straightforward.

The Technology We Seek in Startups
- Access to large-scale Japanese corpora
High-quality Japanese data in the hundreds of GB to TB range is essential for foundational model training. We require possession of data collected in-house, data obtainable from partners, or specialized data for specific domains (e.g., medical, legal, manufacturing).


- Foundational ML/NLP Knowledge and Development Capabilities
Teams must possess an understanding of Transformer architectures, experience with frameworks like PyTorch/JAX, and foundational knowledge of distributed learning. Having in-house ML engineers or researchers is desirable.


- Clear Use Cases and Domain Expertise
Instead of aiming for a "generic model," teams must have a concrete exit strategy, such as "solving XX in medical settings" or "streamlining YY in manufacturing." Deep expertise in a specific industry enables the development of differentiated models.

- Long-term development commitment of 6 months to 1 year
Foundational model development cannot be completed in the short term. We require commitment from management, a dedicated development team structure, and a medium-to-long-term funding plan.

Examples of Co-Creation Ideas

- Co-development of a Healthcare-Specialized Japanese LLM
Collaborating with healthcare startups to develop a model specialized for diagnostic support, medical record summarization, and medical paper analysis. Utilizing anonymized medical data for training, we aim for adoption by university hospitals and pharmaceutical companies. The GenAI IC team supports the construction of an ethical and safety evaluation framework for medical AI.

- Manufacturing DX-Specialized Multimodal LLM
Collaborating with manufacturing startups, we develop models for design drawing analysis, work procedure generation, and automated quality control reporting. Leveraging multimodal technology that integrates text and image understanding, we target deployment to small and medium-sized manufacturers following proof-of-concept trials at Toyota, Denso, and others.

- Legal & Contract-Specialized LLM
Partnering with legal tech firms, we develop models specialized for contract review, precedent case search, and legal risk analysis. Implemented technology capable of understanding contracts exceeding 100 pages, targeting adoption by major law firms and legal departments of financial institutions.

- Japanese × Asian Languages Bilingual LLM
Developed a model capable of high-precision processing of Japanese and English, Chinese, Korean, etc., in collaboration with a startup targeting global expansion. Aiming to sell it as a tool supporting Japanese companies' overseas expansion, leveraging AWS's global customer base.

Available resources

The following is a list of resources AWS may consider providing for co-creation projects. Actual offerings and scale will be determined through individual discussions based on project requirements, technical feasibility, progress, and other factors.


- Dedicated support from the AWS GenAI Innovation Center (GenAIIC) team
World-class ML experts provide end-to-end guidance from data pipeline construction to model design, training strategies, and troubleshooting. Through ad-hoc technical review sessions, they directly mentor architecture selection, hyperparameter tuning, and training stabilization techniques.


- Provision of large-scale training infrastructure
Includes AWS Trainium/Inferentia2 clusters and distributed training environments via Amazon SageMaker HyperPod. Access computational resources equivalent to thousands of GPUs.


- AWS Credit Application
Apply for AWS credits worth $1,000 to $100,000 based on project scale. Eliminate funding concerns and create an environment focused on technical development.


- Commercialization and Go-to-Market Support
We explore listing completed models on Amazon Bedrock, support sales on AWS Marketplace, and provide opportunities for introductions to enterprise customers. We accompany you not only through technical development but also to business success.

- Awareness and Branding Support
We support establishing your positioning as a "Japan-born AI model" through opportunities to speak at AWS-hosted events and case study features on the official AWS blog.


- Support from a Team Specialized in Large Language Model Training
Our expert team provides reliable support backed by proven track records and stable infrastructure, including support for initiatives like GENIAC led by the Ministry of Economy, Trade and Industry. For details, please refer to the following blog posts:
・Phase 2: https://aws.amazon.com/jp/blogs/news/geniac-cycle2-kick-off/
・Phase 3: https://aws.amazon.com/jp/blogs/news/geniac-cycle3-kick-off/
・Lessons Learned Through GENIAC: https://aws.amazon.com/jp/blogs/news/beyond-accelerators-lessons-from-building-foundation-models-on-aws-with-japans-geniac-program/