LG, SK Telecom and Upstage Survive the Latest Cut in Korea's Sovereign AI Race

Three Korean companies have survived the latest round of the government's effort to build a homegrown artificial intelligence model, narrowing a national contest meant to produce a domestic answer to the leading systems from the United States and China. LG AI Research, SK Telecom and Upstage advanced past a second evaluation announced on Tuesday, while a rival was cut.

The winnowing has been steady. The Ministry of Science and ICT launched the project last year with five teams, trimmed the field to four after a first evaluation in January, and has now narrowed it to three. The eliminated firms are Motif Technologies, dropped in this round, and Naver Cloud and NC AI, both cut in January.

The models still standing

Each surviving team is fielding its own model. LG AI Research is competing with K-EXAONE 2.0, SK Telecom with A.X K2, and Upstage with Solar Open2, three systems that now carry Korea's hopes of building sovereign AI, models the country controls rather than rents from abroad.

The team that fell was not without merit. Motif Technologies entered with Motif 3, which scored 47 on the AAII benchmark and ranked tenth in the world, a strong technical showing that still was not enough to advance.

Why Motif was cut

The reason lay in how the models were judged. The second round scored teams on benchmarks worth 40 points, an expert assessment worth 35 and a user evaluation worth 25, a formula that weighs real-world usefulness alongside raw capability.

Second Vice Science Minister Ryu Je-myung explained the outcome for Motif directly, saying the company showed outstanding technological capabilities but received a lower evaluation than other firms in usability and applicability. In a contest that rewards models people can actually deploy, being brilliant on paper was not decisive.

A close race on the scoreboard

The average scores suggest the field is tight. Teams averaged 22.5 points on benchmarks, 28.8 on expert assessment and 17.6 on user evaluation, and the gaps between them were narrow, four points on benchmarks, 2.4 on the expert score and five on user evaluation. Small margins are deciding who stays in.

The compute behind it

The state is backing the effort with hardware as well as money. Teams drew on roughly 768 Nvidia B200 graphics processors in the first half of the year, and the allocation is set to rise to about 1,000 in the second half, a supply of the scarce chips that training a competitive model demands.

The prize at the end

The stakes reach beyond a single grant. A third evaluation is due in the second half of the year, and the government's aim is to end up with a model that can stand against the best from the United States and China, part of a broader push to make Korea one of the world's top three AI powers. For the three survivors, the next cut will decide who gets to carry that ambition forward.