A growing number of South Korean artificial intelligence startups are offering their AI models and services in on-premise environments, where software runs on a client's own servers rather than over the internet. The shift adds a new option for customers alongside internet-based software distribution, and such efforts are expected to expand as companies worry about leaks of confidential information through AI.
Twelve Labs is in talks with four South Korean organizations — large companies and state-run enterprises — to deploy its AI models in on-premise environments, venture industry sources said on the 21st. The company developed Marengo, an AI model that searches data by analyzing scenes and sounds in video, and Pegasus, a model that converts the scenes analyzed by Marengo into text.
Both models have been used primarily through application programming interface (API) calls, which require internet access to invoke the models' functions and link them to other software. Twelve Labs is now pursuing a new business line that would run the same process inside a client's own servers without any external internet connection. If the project goes through, it would mark the company's first deployment of Marengo and Pegasus on-premise in South Korea. Twelve Labs began building experience in the field after winning a contract to supply AI models in an on-premise environment to a U.S. government agency.
Other companies are also moving to distribute AI services on-premise. Flitto (300080.KQ), which has run its AI multilingual translation and interpretation service as software-as-a-service (SaaS) and through APIs, has started an on-premise business in earnest. The company won a 3.9 billion won ($2.7 million) contract this month to supply the service on-premise. Inquiries from other clients followed, prompting Flitto to change its business policy and broaden the ways it delivers the service.
MangoBoost recently agreed to supply LLM Boost, its software for optimizing AI model inference, to a defense contractor. The client asked for the software to be installed on its own servers, citing the strict security requirements of defense work, and MangoBoost decided to support an on-premise setup.
On-premise refers to installing and managing software on a company's own servers. Unlike uniform software distribution, it requires customized setup for each client, because server environments differ from company to company in specifications such as central processing units (CPUs) and graphics processing units (GPUs), and because each client has its own requirements for how the software is used. Suppliers must therefore carry out separate installation and configuration work, while clients bear the burden of maintaining the software installed on their servers.
Despite that inconvenience, the appeal of on-premise AI rests on concerns about data protection. Companies fear that AI will learn and memorize sensitive corporate information that could later be extracted by other users. Facing demand from companies that want to run AI services on their own servers with internet access blocked, AI service providers have offered on-premise deployment as a new option.
"Companies and public institutions that place a premium on security tend to prefer on-premise environments," said Lee Hye-min, head of Twelve Labs' Korean operations. "They want to manage sensitive information tied to corporate competitiveness or the public interest — such as confidential data on factories and production processes, or CCTV footage — inside a closed network rather than on an external cloud."