South Korea's AI Strategy: Challenging OpenAI and Google

South Korea's Push for Sovereign AI

Companies throughout South Korea, ranging from established technology leaders to emerging startups, are actively engaged in the development of large language models. These models are specifically designed to understand and process the nuances of the Korean language and culture.

This initiative is positioning these Korean entities to directly challenge global AI powerhouses such as OpenAI and Google in the rapidly evolving artificial intelligence landscape.

National AI Investment

A significant national commitment was made last month with the launch of a substantial sovereign AI program. ₩530 billion (approximately $390 million USD) has been allocated to five domestic companies for the construction of large-scale foundational models.

This investment highlights Seoul’s strategic objective to lessen its dependence on AI technologies originating from other countries.

Strengthening Security and Data Control

The government anticipates that fostering domestic AI capabilities will bolster national security interests. It also aims to provide greater oversight and control regarding data management within the context of the AI era.

The Selected Companies

The Ministry of Science and ICT has chosen the following organizations to participate in this competitive program: LG AI Research, SK Telecom, Naver Cloud, NC AI, and the innovative startup, Upstage.

A Phased Approach to Funding

Progress will be rigorously evaluated by the government on a semi-annual basis. Companies demonstrating insufficient advancement will be removed from the program.

Funding will be continuously directed towards the leading contenders, ultimately narrowing the field to just two entities that will spearhead South Korea’s sovereign AI efforts.

Unique Strengths in the Competition

Each participating company possesses distinct strengths that they intend to leverage in South Korea’s AI development race. TechCrunch engaged with several of the selected organizations to understand their strategies for competing with established players like OpenAI, Google, and Anthropic within the Korean market.

NC AI chose not to provide a statement at this time.

LG AI Research: Exaone

LG AI Research, the research and development division of the LG Group based in South Korea, presents Exaone 4.0, a sophisticated AI model employing hybrid reasoning. This newest iteration integrates extensive language understanding capabilities with the enhanced reasoning functionalities initially featured in the company's previous Exaone Deep model.

Exaone 4.0 Performance and Strategy

Currently, Exaone 4.0 (32B) demonstrates competitive performance against other models on the Intelligence Index benchmark developed by Artificial Analysis, mirroring the results achieved by Upstage’s Solar Pro2. LG intends to further enhance its position by leveraging its extensive access to practical data from diverse industries.

These industries span biotechnology, cutting-edge materials science, and advanced manufacturing processes. Rather than simply increasing model size, LG prioritizes intelligent data refinement before utilizing it for training purposes.

Focus on Practical Value

The company aims to optimize the entire AI development process, enabling its AI to provide tangible, real-world benefits that surpass the capabilities of broadly applicable models. Co-head Honglak Lee articulated this core philosophy to TechCrunch.

Iterative Improvement Through APIs

LG is refining its models through a common strategy: providing access via Application Programming Interfaces (APIs). The resulting real-world data generated by API users is then utilized to further train and improve the model’s performance.

As LG’s AI models become more advanced, its partners are empowered to deliver superior services, fostering increased economic value and generating even more comprehensive data sets.

Efficiency and Specialization

Instead of competing through massive investments in Graphics Processing Units (GPUs), LG AI Research is concentrating on maximizing efficiency and extracting optimal performance from existing hardware. A key focus is the development of AI models tailored to specific industry needs.

The objective is not to exceed global competitors in spending, but to surpass them through the creation of highly effective and resource-efficient AI solutions.

SK Telecom's A.X: A New AI Agent

SK Telecom (SKT), a leading telecommunications company in South Korea, initially introduced its personal AI agent service, A., in late 2023. This July saw the release of its latest large language model, A.X.

A.X 4.0 is constructed utilizing the Qwen 2.5 open-source model originating from Alibaba Cloud. It is available in two distinct versions: a substantial 72-billion-parameter model and a more streamlined 7-billion-parameter model.

Performance and Open Source Initiatives

According to SK Telecom, A.X 4.0 demonstrates a 33% improvement in processing efficiency for Korean language inputs compared to OpenAI’s GPT-4o. This highlights a significant advantage in handling the nuances of the Korean language. (Comparative data for OpenAI’s GPT 5.0 is currently unavailable.)

Earlier this summer, SKT also made its A.X 3.1 models publicly available through open-source licensing. The A. service itself provides functionalities such as AI-powered call summaries and automated note-taking.

As of August 2024, the service has attracted approximately 10 million users.

Leveraging Telecom Infrastructure

A key differentiator for SK Telecom lies in its versatile access to data derived from its extensive telecom network. This encompasses information related to navigation services and ride-hailing applications.

Taeyoon Kim, head of the foundation model office at SK Telecom, explained to TechCrunch: “SK Telecom functions as a connector between pioneering model research and practical applications. Through our telecom infrastructure, a large user base, and established services like A., we integrate AI into daily life across sectors like customer support, transportation, and manufacturing.”

Investment in AI Infrastructure and Partnerships

SK Telecom is actively investing in the necessary AI infrastructure. This includes utilizing GPUaaS, South Korea’s largest GPU-based service, and the construction of a new, large-scale AI data center in collaboration with AWS.

The company is addressing any potential gaps through strategic partnerships.

Kim further stated: “We are developing a comprehensive ecosystem in collaboration with Korean AI chipmaker Rebellions. We are also establishing secure data partnerships with governmental bodies and academic institutions, and cultivating a global research network.”

“This includes collaborative efforts, such as our partnership with MIT (MGAIC), which focuses on applying foundation models to advancements in manufacturing, as well as innovations in battery and semiconductor technologies.”

Naver Cloud: HyperCLOVA X

In 2021, Naver Cloud, the cloud division of the prominent South Korean internet firm Naver, initially launched its large language model, HyperClova. A subsequent, enhanced iteration, HyperCLOVA X, was revealed two years later.

Alongside HyperCLOVA X, Naver introduced new offerings leveraging this technology. These include CLOVA X, an AI chatbot, and Cue, a generative AI search engine.

Cue is designed to compete directly with Microsoft’s CoPilot-integrated Bing and Google’s AI Overview features.

Furthermore, this year saw the unveiling of HyperCLOVE X Think, Naver’s multimodal reasoning AI model.

The Role of LLMs at Naver

According to a Naver representative, the core strength of large language models lies in their ability to function as “connectors.” They bridge the gap between established systems and isolated services, thereby enhancing overall utility.

Naver distinguishes itself as the sole Korean company – and one of a limited number globally – possessing a truly comprehensive “AI full stack.”

Building the AI Ecosystem

The company developed the HyperCLOVA X model independently and manages the extensive infrastructure required to support it.

This includes large-scale data centers, cloud services, AI platforms, applications, and direct consumer services, all bringing the technology to fruition, as explained by the spokesperson.

Integration into Core Services

Mirroring Google’s approach, but specifically optimized for the South Korean market, Naver is integrating its AI into fundamental services.

These services encompass search, shopping, maps, and financial applications.

A key advantage for Naver is its access to real-world data.

For example, its AI Shopping Guide provides recommendations based on actual consumer purchasing behavior.

Additional Offerings

• CLOVA Studio: Enables businesses to create customized generative AI solutions.
• CLOVA Carecall: An AI-driven check-in service designed for elderly individuals living independently.

Competition and Future Strategy

The Naver spokesperson asserts that surpassing global AI leaders like OpenAI and Google depends on two critical factors.

These are refining the model development “recipe” and securing the necessary investment for scaling operations.

However, the company prioritizes sophistication over sheer size, maintaining that its AI is already globally competitive at similar scales.

HyperCLOVA X represents a significant step in Naver’s AI strategy.

Upstage and the Solar Pro 2 Model

Within this project, Upstage stands as the sole startup contender. The company’s Solar Pro 2 model, initially released in July of the previous year, achieved a significant milestone. It became the first Korean-developed model to be classified as a frontier model by Artificial Analysis.

This recognition positions Upstage alongside industry giants such as OpenAI, Google, Meta, and Anthropic, as noted by Soon-il Kwon, Executive Vice President at Upstage.

A key advantage of Solar Pro 2 lies in its efficiency. Despite possessing only 31 billion parameters – considerably fewer than the 100 to 200 billion found in many competing frontier models – it delivers superior performance for users in South Korea.

Furthermore, it offers a more economical solution, as explained by Kwon in a discussion with TechCrunch.

Solar Pro 2 has demonstrated its capabilities by exceeding the performance of established global models on prominent Korean benchmarks. Upstage’s objective with this current project is to elevate Korean language performance to 105% of the prevailing global standard, according to Kwon.

The company intends to distinguish itself through a focus on tangible business applications, rather than solely pursuing benchmark scores.

Industry-Specific Model Development

Upstage is actively developing specialized AI models tailored for sectors including finance, law, and medicine. This strategic approach aims to foster a Korean AI ecosystem spearheaded by startups designed from the ground up for artificial intelligence.

This initiative underscores Upstage’s commitment to building a robust and innovative AI landscape within Korea.