South Korea is trying to make its AI future less dependent on foreign systems. From major technology groups to a startup, local players are building large language models designed around Korean language, culture, data, and business needs.
Last month, the nation launched its most ambitious sovereign AI initiative to date. The plan pledges ₩530 billion, (about $390 million), to five local companies developing large-scale foundational models, with the goal of creating stronger domestic alternatives to systems from OpenAI, Google, Anthropic and the rest.
A sovereign AI race with staged funding
The organizations selected by the Ministry of Science and ICT are LG AI Research, SK Telecom, Naver Cloud, NC AI, and Upstage. NC AI declined to comment to TechCrunch.
The structure of the program is competitive. Every six months, the government will review the first cohort’s progress, remove underperformers, and keep backing the companies that appear strongest. In the end, just two are expected to remain as the leaders of the country’s sovereign AI drive.
The logic behind the initiative is clear. Seoul wants to cut reliance on foreign AI technologies, strengthen national security, and maintain tighter control over data in the AI era. That does not mean every company is pursuing the same strategy. Each selected player is coming into the race with a different asset: industry data, telecom infrastructure, search and cloud services, startup focus, or specialized model design.
LG AI Research is betting on efficiency and industry data
LG AI Research, the R&D unit of South Korean giant LG Group, is offering Exaone 4.0, a hybrid reasoning AI model. The latest version combines broad language processing with reasoning features introduced earlier in Exaone Deep.
Exaone 4.0 (32B) already scores reasonably well against competitors on Artificial Analysis’s Intelligence Index benchmark. LG’s plan is to improve from there by drawing on real-world industry data across areas such as biotech, advanced materials, and manufacturing.
The company is not framing the race as a simple contest over scale. Instead, it is focusing on making the data and training process more useful before information reaches the model. Co-head Honglak Lee described this as the company’s core method, telling TechCrunch: “This is our fundamental approach,”
LG is also using a familiar improvement loop. It offers models through APIs, then uses real-world data generated by service users to train and improve the model. Lee said: “As LG’s models improve, our partners can deliver better services, which in turn generate greater economic value and even richer data,”
The company’s position is that a more efficient model can still compete if it is tuned to practical industry needs. Rather than chasing massive GPU clusters, LG AI Research is trying to get more performance out of every chip while building industry-specific models.
SK Telecom brings AI into daily services
SK Telecom has a different advantage: it already operates services and infrastructure that connect directly to users. The company launched its personal AI agent A. service in late 2023 and rolled out its large language model, A.X, this July.
A.X 4.0 is built on top of Alibaba Cloud’s Chinese open source model, Qwen 2.5. It comes in two versions: a 72-billion-parameter model and a lighter 7B model.
SK says A.X 4.0 processes Korean inputs about 33% more efficiently than GPT-4o did. The source notes that OpenAI’s GPT 5.0 comparison data is not available. SKT also open sourced its A.X 3.1 models earlier this summer.
The company’s A. service includes AI call summaries and auto-generated notes. As of August 2025, it had about 10 million subscribers.
Taeyoon Kim, head of the foundation model office at SK Telecom, told TechCrunch that SK Telecom’s role is to connect model research with real-world impact. He pointed to telecom infrastructure, a large user base, and services such as A. as ways to bring AI into customer service, mobility, and manufacturing.
SK Telecom is also investing in infrastructure. It uses GPUaaS, South Korea’s largest GPU-based service, and is building a new hyperscale AI data center with AWS. Kim also cited work with Korean AI chipmaker Rebellions, trusted data partnerships through government and universities, and a global research network that includes collaboration with MIT (MGAIC).
Naver Cloud is building from model to consumer product
Naver Cloud, the cloud services arm of South Korea’s leading internet company, has been developing its own large language model path for several years. It introduced HyperClova in 2021 and later unveiled HyperCLOVA X, along with products powered by the technology.
Those products include CLOVA X, an AI chatbot, and Cue, a generative AI-driven search engine positioned against Microsoft’s CoPilot-enhanced Bing and Google’s AI Overview. This year, Naver Cloud also unveiled HyperCLOVE X Think, a multimodal reasoning AI model.
A Naver spokesperson told TechCrunch that LLMs can work as “connectors” between legacy systems and siloed services. The company’s broader claim is that it has an “AI full stack,” because it built HyperCLOVA X from scratch and also runs data centers, cloud services, AI platforms, applications, and consumer services.
Naver is embedding AI into services such as search, shopping, maps, and finance. Its AI Shopping Guide offers recommendations based on what people actually want to buy. Other services include CLOVA Studio, which helps businesses build custom generative AI, and CLOVA Carecall, an AI-powered check-in service for seniors living alone.
The company’s view is that beating global AI giants depends on perfecting its model “recipe” and securing the capital to scale. Even so, it is emphasizing sophistication rather than only size, arguing that its AI is already globally competitive at comparable scales.
Upstage shows how a startup can compete
Upstage is the only startup in the project. Its Solar Pro 2 model, launched last July, was the first Korean model recognized as a frontier model by Artificial Analysis, according to Soon-il Kwon, executive vice president at Upstage.
Kwon told TechCrunch that while most frontier models have 100 billion to 200 billion parameters, Solar Pro 2 has just 31 billion and performs better for South Koreans while being more cost-effective. He said: “Solar Pro 2 has outperformed global models on major Korean benchmarks. With this project, Upstage aims to achieve a Korean language performance of 105% of the global standard,”
Upstage is also trying to define success beyond benchmarks. The company is developing specialized models for industries such as finance, law, and medicine, while pushing for a Korean AI ecosystem led by AI-native startups.
Taken together, the selected companies show the shape of South Korea’s homegrown AI strategy. The country is not relying on one model or one company. It is creating a contest among different approaches, then narrowing the field as results come in.
The central bet is that local language strength, trusted data, practical services, and efficient models can matter as much as raw scale. If the plan works, South Korea’s sovereign AI initiative could give the country more control over the systems that increasingly shape business, public services, and everyday digital life.