What the agreement provides for
The joint research laboratory between Nvidia and the Korea Advanced Institute of Science and Technology (KAIST) intends to release at least one million Korean-language training samples for artificial intelligence models during 2027. Kim Hyun-woo, the lab's head, said this while meeting journalists in Seoul on Tuesday. No precise release date has been set.
This is the first concrete research milestone announced since the July agreement between the U.S. company and the South Korean university to open the first joint AI laboratory in Asia between Nvidia and a university. The facility is located at KAIST's Kim Jaechul Graduate School of AI in Seoul and is supported by a five-year, $300 million collaboration dedicated to Korean-language AI and industrial applications.
Quality before quantity
According to Kim, Korea has invested heavily in developing models and computing infrastructure, but has devoted comparatively less attention to the high-quality training data needed to make those models competitive. For this reason, the lab will not work on another foundation model: it will focus on producing better Korean datasets and on how existing models use them.
The gap concerns training materials in particular: datasets for evaluating models in Korean exist, Kim explained, but high-quality ones remain limited compared with English and Chinese. Available datasets are often based on translated texts, while AI-generated Korean can contain unnatural expressions; producing large volumes of synthetic data without adequate filtering also risks degrading performance. "Generating the data is not the problem," Kim said: quality filtering and content curation require far more work.
The lab will develop techniques to generate realistic documents from limited information, then refine them for accuracy, diversity and duplication. Kim himself noted that the first million samples will not be enough to close the language gap.
From Nemotron to potential benefits for other languages
Nvidia's open Nemotron models will support the research, but the resulting datasets and training methods are intended for broader use by Korean developers: "Nemotron is just one of the trees in the forest," Kim said, adding that a healthy forest needs different species growing together.
A previous Nvidia initiative, Nemotron-Personas-Korea, had generated 7 million synthetic personas based on South Korean government statistics; according to Kim, the dataset was used in the development of LG AI Research's K-Exaone 2.5 and has recorded about 140,000 downloads. Asked whether making the Korean datasets public could also benefit foreign competitors, Kim replied that strengthening domestic research capabilities matters more than restricting access.
On timing and resources, Nvidia has begun recruiting full-time researchers in Korea for its global research organization, with the first hires expected by the end of the year. Of the $300 million collaboration, 250 million will come in the form of computing resources provided through Nvidia's Korean cloud partners over five years; the program will also support at least 10 KAIST researchers per year, including internships at Nvidia. Once the Korean data foundation is in place, the lab plans to expand to sector-specific datasets and agentic AI capable of performing complex tasks with limited human intervention. Industrial partnerships and intellectual property agreements remain under discussion. Kim lastly indicated that the research could benefit in the future other languages with little training data, such as Japanese and Indian languages, and that he would consider it a success if other Korean AI models widely used the methods developed.
