Nvidia and the Korea Advanced Institute of Science and Technology plan to release at least 1 million Korean-language AI training samples in 2027, The Korea Herald reported after the head of their joint research lab spoke to reporters in Seoul on 10 November. The planned public release is the lab’s first concrete research milestone since Nvidia and KAIST announced plans to establish it in July. The lab is backed by a five-year, $300 million collaboration.
Key points
- The lab plans to make at least 1 million Korean-language training samples public in 2027.
- Computing resources from Nvidia’s Korean cloud partners account for $250 million of the five-year, $300 million collaboration.
- Kim Hyun-woo says the lab will concentrate on data quality and training methods rather than develop another foundation model.
- The programme will support at least 10 KAIST researchers each year, including opportunities for internships at Nvidia.
Kim sets a 2027 target for Korean data
Kim Hyun-woo, an assistant professor at KAIST’s Kim Jaechul Graduate School of AI and head of the Nvidia-KAIST Joint AI Research Lab, said the samples would be available for others to use in training. The lab is based at the graduate school in Seoul. Its work focuses on Korean-language AI and applications for particular industries.
Kim said South Korea’s spending on models and computing infrastructure has outpaced its attention to the quality of data used to train them. Researchers have identified a shortage of high-quality Korean-language material as a constraint on the competitiveness of domestic models. According to Kim, training data of that kind is more limited in Korean than in English or Chinese, even though Korean-language datasets for evaluating models are available.
The lab intends to generate realistic documents from limited information and then check them for accuracy, diversity and duplication. Kim said existing collections often depend on translated text, while synthetic Korean can contain unnatural phrasing. Producing more synthetic material without adequate filtering could damage model performance, he said.
Kim cautioned that the initial 1 million samples would not close the language gap. He said the lab would work on improving how existing models learn from the data rather than build another foundation model. Nvidia’s open Nemotron models will support the research, while the resulting datasets and training methods are intended for use by other Korean AI developers.
Korean cloud partners provide $250 million in computing
Of the collaboration’s $300 million value over five years, $250 million is to take the form of computing resources supplied through Nvidia’s Korean cloud partners. The arrangement ties most of the stated value to access to computing rather than a payment described in cash. The programme will support at least 10 KAIST researchers annually and offer internship opportunities at Nvidia.
Once its Korean-language data work is established, the lab plans to develop industry-specific datasets and work on agentic AI, which carries out complex tasks with limited human intervention. Industry partnerships and intellectual property arrangements remain under discussion.
Kim said the work could eventually be useful for other languages with relatively little training material, including Japanese and Indian languages. He also said broader access to the Korean-language datasets mattered more to domestic research than limiting their use by foreign competitors.
Kim reports about 140,000 Nemotron-Personas-Korea downloads
An earlier Nvidia data initiative, Nemotron-Personas-Korea, generated 7 million synthetic personas using Korean government statistics. Kim said that dataset had recorded about 140,000 downloads and was used in developing LG AI Research’s K-Exaone 2.5. The new lab’s target concerns Korean-language training samples rather than personas.
The data project sits alongside Nvidia’s work on tools for Korean model developers. AI Affairs reported Nvidia’s release of Dynamo recipes for four South Korean open models. Kim said he would regard the lab’s work as successful if other Korean AI models made wide use of its data and training methods.
Nvidia has begun recruiting full-time researchers in Korea for its global research organisation. Its initial hires are expected by the end of 2026.
