Companies developing the government's sovereign AI foundation models shared examples of how they use Nvidia platforms and software. Nvidia also unveiled inference optimisation technology tailored to Korean AI models and is expanding cooperation in the sovereign AI ecosystem.

Researchers from LG AI Research, SK Telecom and Upstage, three companies that advanced to the third round of the sovereign foundation model project, took part in the 'K-AI Frontier Builders' panel discussion at 'Nvidia AI Day Seoul 2026' held at COEX in Samseong-dong, Seoul, on Nov. 9.

Nvidia's AI inference optimisation software Dynamo was cited as a key technology. It distributes AI inference tasks efficiently across multiple GPU servers and reuses computation results to increase processing speed and resource utilisation.

Upstage said it is improving operational efficiency of AI services by adopting Dynamo. It said it saw benefits in managing the KV cache, which temporarily stores information previously computed by AI to reduce duplicate calculations.

Song In-seo (송인서), LLM tech lead at Upstage, said, "We gained many advantages in areas such as KV cache management as we introduced Dynamo." He said, "We were able to significantly improve throughput with fewer resources."

LG AI Research also shared its experience cooperating with Nvidia on inference optimisation technology. It highlighted the importance of distributing user requests to the appropriate servers in large-scale AI services.

Cho Hyun-jik (조현직), head of the Exaone Lab team at LG AI Research, said, "What matters is not a single server, but which server to route the current query to." He said, "As services move toward large-scale deployment, routing capabilities will become extremely important."

He added that technologies that orchestrate multiple models, rather than just one model, to respond while maintaining the same quality of answers would become very promising.

Sovereign foundation model developers are using Nvidia technology not only for inference but also for AI training data processing and the model training process.

SK Telecom said it significantly shortened the data processing period by using Nvidia's NeMo Curator, which removes duplicate documents in AI training data.

Yang Hyun-ho (양현호), a researcher at SKT, said, "It used to take 2 to 3 weeks to process duplicate documents, but GPU acceleration with NeMo Curator reduced it to about 3 days, which had a big effect."

Yang also highlighted the impact of training efficiency on costs in developing large-scale AI models. He said, "Even if model training speed increases by only 10 percent, that is about 6 days faster in terms of GPU usage time, and 600 million won in money."

LG AI Research also introduced its use of Nvidia's Megatron Bridge, which supports training and conversion of large AI models.

Cho said, "After moving to Megatron Bridge, it was very convenient for affiliates to quickly try training Exaone models." He said, "It contributed to improving the overall efficiency of affiliates."

Upstage is also using Nvidia's NVFP4 quantisation technology, which improves processing efficiency by lowering computational precision of AI models. Song said model performance can decline during quantisation, so the company maintains higher precision for accuracy-sensitive parts.

Nvidia unveiled Dynamo optimisation technology tailored to four Korean AI models: SKT's A.X K2, LG AI Research's K-Exaone 2.0, Upstage's Solar Open 2 and Motif Technologies' Motif-3.

According to Nvidia's own test results, after applying the optimisation technology in a B200 GPU environment, output throughput per GPU increased versus the previous level by 1.72 times for A.X K2, 1.58 times for K-Exaone 2.0 and 1.43 times for Solar Open 2. Motif-3 improved 3.67 times in a separate distributed inference environment.

Separately, Nvidia AI Day Seoul 2026 also saw Nvidia, SK Telecom and Rebellions sign a three-party memorandum of understanding for cooperation on NVLink Fusion.

The Ministry of Science and ICT congratulated the parties on signing the agreement and jointly held an 'AI Startup Partnership Forum' with Nvidia. The agreement was signed as part of cooperation linking Korea-made neural processing units that the ministry and Nvidia have discussed. The partnership forum was set up to support overseas expansion by domestic AI startups.