Nvidia tools cut Korean AI data processing from weeks to days

Accelerating Sovereign AI

Korean companies involved in the government's sovereign AI project are increasingly relying on Nvidia's hardware and software ecosystem. These developers are using Nvidia technology across the entire lifecycle of AI development, from the initial processing of training data to the final deployment of models. This cooperation aims to build robust, locally-tailored AI foundation models that can serve the specific needs of the Korean market.

During a panel at the Nvidia AI Day Seoul 2026, researchers from LG AI Research, SK Telecom, and Upstage discussed their specific use cases. They highlighted how these tools provide the necessary infrastructure to manage the massive scale required for modern AI. By integrating Nvidia's platforms, these companies can better manage the complexities of large-scale AI services and ensure high-quality outputs.

The collaboration is also being supported at a policy level. The Ministry of Science and ICT participated in the event, which included a forum to help domestic AI startups expand into overseas markets. This suggests a coordinated effort to combine Korean innovation with Nvidia's global technological leadership to create a competitive sovereign AI landscape.

Efficiency in Data and Training

One of the primary bottlenecks in creating large-scale AI models is the preparation of training data. SK Telecom reported that using Nvidia's NeMo Curator, a tool designed to remove duplicate documents from training sets, significantly shortened their workflow. According to researcher Yang Hyun-ho, the process used to take between 2 and 3 weeks but was reduced to about 3 days with GPU acceleration.

The financial implications of these efficiency gains are substantial for developers. Yang noted that even a small 10 percent increase in training speed can save approximately 6 days of GPU usage time. For a large-scale project, this translates to a cost saving of roughly 600 million won. These types of optimizations are critical for making the development of sovereign models economically viable for private and public entities.

LG AI Research also highlighted the use of Nvidia's Megatron Bridge to streamline the training and conversion of large models. Cho Hyun-jik, head of the Exaone Lab team, stated that the tool made it very convenient for affiliates to quickly test Exaone models. By improving the overall efficiency of these internal workflows, the company can iterate on its AI products more rapidly without incurring unnecessary overhead.

Optimizing Inference and Routing

Once a model is trained, the challenge shifts to inference, which is the process of the AI generating responses for users. Upstage is currently using Nvidia's Dynamo software to improve the operational efficiency of its AI services. Dynamo works by distributing inference tasks across multiple GPU servers and reusing previous computation results to increase processing speed and resource utilization.

A specific area of improvement for Upstage involves managing the KV cache. This is a temporary storage area for information previously computed by the AI, which helps avoid duplicate calculations. Song In-seo, the LLM tech lead at Upstage, noted that the company gained many advantages in KV cache management by introducing Dynamo. This allows them to achieve significantly higher throughput while using fewer physical resources.

LG AI Research emphasized that as AI services scale, the focus must shift from single-server performance to intelligent routing. Cho Hyun-jik explained that it is vital to know which specific server to route a user's query to in a large-scale environment. He suggested that technologies capable of orchestrating multiple models to provide consistent quality will be the next major frontier for large-scale AI deployment.

Quantization and Performance

To further improve speed, developers often use quantization, which is the process of lowering the computational precision of an AI model. Upstage utilizes Nvidia's NVFP4 quantization technology to make their models run more efficiently. However, because reducing precision can sometimes cause a decline in model performance, the company carefully maintains higher precision for parts of the model that are sensitive to accuracy.

Nvidia also conducted specific tests on four Korean AI models to measure the impact of its optimization technology in a B200 GPU environment. The results showed varying levels of improvement in output throughput per GPU. For example, SKT's A.X K2 saw a 1.72x increase, while LG AI Research's K-Exaone 2.0 saw a 1.58x increase, and Upstage's Solar Open 2 saw a 1.43x increase.

The most significant jump was seen in Motif Technologies' Motif-3, which improved by 3.67x in a separate distributed inference environment. These diverse results demonstrate how different optimization techniques can be tuned to specific model architectures. By leveraging these tools, Korean developers can maximize the hardware capabilities of Nvidia's latest chips to serve a wider audience.

Future Hardware Partnerships

The relationship between Nvidia and the Korean AI ecosystem extends beyond software to hardware integration. During the Seoul event, Nvidia, SK Telecom, and Rebellions signed a three-party memorandum of understanding regarding NVLink Fusion. This agreement is intended to link Korea-made neural processing units (NPUs) with Nvidia's existing infrastructure. This move is part of a broader discussion between the Korean government and Nvidia regarding the integration of domestic hardware.

The goal of such partnerships is to create a more interconnected AI ecosystem where different types of processors can work together seamlessly. By linking domestic NPUs with Nvidia's technology, Korean developers may gain more flexibility in how they deploy their sovereign models. This could reduce reliance on a single hardware provider while still benefiting from Nvidia's software and interconnect standards.

Looking ahead, the focus will likely remain on balancing high-performance computing with cost-effective scaling. As the sovereign AI project matures, the ability to orchestrate multiple models and optimize data pipelines will be the primary differentiator. The ongoing cooperation between the Korean government, domestic tech giants, and Nvidia suggests a long-term strategy to establish Korea as a significant hub for AI development.