Taiwan is expanding fine-tuning functions for a national-level AI platform and moving to build next-generation computing infrastructure. It is a strategy to foster sovereign AI optimized for its own data and language environment and speed up industrial application of research results.

On Nov. 11 local time, Taiwanese media outlet IT Home reported that the National Center for High-performance Computing (NCHC) plans to enhance functions of its AI development platform, TAIWAN AI RAP, and expand AI computing resources.

Chang Chao-liang (張朝亮), head of the national network center, stressed at a recent AI technology results presentation that advancing sovereign AI requires securing sufficient computing power, in-house model development and capabilities to train on local data.

TAIWAN AI RAP is a national-level AI development platform that began pilot operations at the end of March last year. It provides computing resources and development tools to companies, startups and research institutions to support everything from AI model development to application in real services.

The number of platform users has nearly doubled over the past year. The national network center is expanding model fine-tuning services this year to help companies use their own data to develop AI models suited to specific industries.

On Nov. 6, it added API functions to Chameleon, its model fine-tuning service, and upgraded computing resources to Nvidia H200 GPUs.

It is also strengthening support for AI agent development. It plans to help companies test performance in a proof-of-concept environment before launching real services and connect via APIs with the national network center or private AI services.

The operating model is a B2B2C structure that links business-to-business transactions with consumer services. The national network center provides the technology platform, and private companies use it to develop services for end users.

It is also expanding computing infrastructure. It currently operates the Nano 4 and Nano 5 platforms and will additionally introduce AMD-based AI computing equipment. It has already completed procurement procedures for the equipment and is also reviewing the introduction of Nvidia's next-generation Vera Rubin platform.

The scope of sovereign AI development is expanding from language models to multimodal AI that understands images, voice and actions. The event also unveiled a system called TAIWAN GBA that combines YOLOv9-based object detection with action recognition technology to identify 12 actions, including falls and fights.

The national network center envisions building an independent AI ecosystem by combining its own computing infrastructure, the model research capabilities of Academia Sinica, and Taiwan's language and data resources.