SK telecom, LG clash in korea’s AI model race

The gist

South Korea’s AI giants, SK Telecom and LG, are locked in a high-stakes showdown to build the nation’s first truly homegrown AI powerhouses—no foreign weights allowed.

What to know

  • SK Telecom’s new A.X K2 (688B parameters) and LG’s K-EXAONE 2.0 (750B parameters) models went head-to-head in July 2026, each boasting unique architectures and massive scale.
  • A $683B, 400-member AI Data Center Alliance aims to deliver 18.4 GW of domestic AI data center power by 2035, using local chips to cut reliance on foreign hardware.
  • The government’s ‘AI for All’ initiative launches a free nationwide chatbot in December 2024, powered by Korean models and backed by a hefty GPU allocation for domestic players.

Industrial AI Consortium Expands

SK Telecom’s Sovereign AI Foundation Model Consortium now unites B2B and manufacturing powerhouses to drive AI deep into Korea’s industrial core, signaling a national push for technological self-reliance.

In July 2026, SK Telecom significantly advanced South Korea’s sovereign AI ambitions by expanding its Sovereign AI Foundation Model Consortium to include SK AX and Technomatrix, aligning closely with the government-backed Sovereign AI Foundation Model Project. This strategic inclusion brought together SK AX’s expertise in B2B AI transformation—leveraging its A.X model platform for workplace productivity and AI agent development—and Technomatrix’s specialized capabilities in AI/ML/FM-Ops, crucial for accelerating manufacturing AI adoption and ensuring robust model operations in real-world industrial settings. Together, these partnerships underscore a concerted effort to embed AI deeply into industrial and manufacturing sectors, reflecting a broader national drive for technological self-reliance and innovation.

Building on its foundational work, the consortium has already developed the 'A.X K2' foundation model, positioning itself as a versatile platform that supports model creation, research, data integration, and service expansion. With ambitions extending beyond manufacturing, the consortium plans to broaden AI applications across diverse sectors including public services, defense, gaming, mobility, search, cybersecurity, and biotechnology. This expansive vision demonstrates a strategic evolution from a focused industrial transformation initiative to a comprehensive AI ecosystem, aiming to embed sovereign AI capabilities across South Korea’s critical economic and technological domains.

Sources

Tech Titans’ Model Showdown

LG and SK Telecom’s dueling 700B-parameter models reveal radically different architectures and training strategies, with citizen juries and real-world use cases shaping which AI will define Korea’s digital future.

In a pivotal moment for South Korea’s sovereign AI ambitions, SK Telecom and LG AI Research unveiled massive open-weight foundation models within just 48 hours in late July 2026, signaling the competition’s critical elimination phase. SK Telecom’s A.X K2, a 688-billion-parameter model, and LG’s K-EXAONE 2.0, boasting 750 billion parameters, both leverage a Mixture of Experts architecture with 256 specialized modules, enabling efficient inference by activating only a subset of parameters per token. Despite their similar scale, the models diverge technically: LG’s hybrid attention system supports an unprecedented 262,144-token context window, while SK Telecom’s proprietary Sparse Gated Attention selectively focuses on the top 2,048 relevant positions, optimizing memory use and accuracy.

The competing teams adopted contrasting training philosophies that reflect their strategic priorities. LG AI Research upcycled its existing 236-billion-parameter EXAONE model through continual pretraining and stabilization techniques to reach 750 billion parameters, emphasizing scale and benchmark dominance. Conversely, SK Telecom built A.X K2 from the ground up using the less common FP8 precision format and proprietary Sparse Gated Attention, training on an immense 8.2 trillion tokens to optimize industrial deployment and operational efficiency. These divergent approaches underscore the dynamic innovation landscape within South Korea’s sovereign AI ecosystem.

South Korea’s AI competition uniquely integrates a rigorous evaluation process that blends technical benchmarks with real-world applicability, featuring a panel of 200 citizen evaluators whose demographics mirror the nation’s population. This inclusive assessment, conducted in August 2026, shifts focus from pure performance metrics to industrial relevance and ecosystem openness, with the Ministry of Science and ICT and NIPA synthesizing citizen, expert, and benchmark feedback to determine which models will shape the country’s AI future. The four finalists—LG AI Research, SK Telecom, Upstage, and Motif Technologies—each showcase distinct innovations, from LG’s scale and benchmark prowess to SK Telecom’s efficiency and industrial focus, reflecting a multifaceted approach to sovereign AI development.

South Korea’s steadfast commitment to AI independence is exemplified by the strict prohibition against using frozen weights from foreign models, a rule that led to Naver Cloud’s elimination in early 2026 for embedding Alibaba’s Qwen encoder weights. This no-foreign-weights mandate has earned all four remaining sovereign AI teams—LG, SK Telecom, Upstage, and Motif—global recognition as 'Notable Models' by Epoch AI, positioning South Korea alongside the US and China as a leading nation in truly indigenous AI development. The competition’s structure, supported by a $350 million government budget and biannual evaluations that reallocate resources to top performers, fosters a fiercely competitive environment driving rapid innovation and national technological sovereignty.

Mega AI Alliance Faces Gridlock

Korea’s $683B, 400-member AI Data Center Alliance is racing to build global-scale infrastructure, but a looming power grid bottleneck threatens to derail the nation’s AI ambitions.

Launched on July 29, 2026, the AI Data Center (AIDC) Alliance unites over 400 organizations including SK Telecom, Naver Cloud, LG Electronics, Kakao, and Samsung SDS in a $683 billion endeavor to build 18.4 GW of AI data center capacity by 2035. Co-chaired by Deputy Prime Minister Bae Kyung-hoon and SK Telecom CEO Jung Jae-heon, the alliance adopts a 'one team' governance model reminiscent of Korea's historic export campaigns, aiming to cement the nation’s status as a global AI infrastructure powerhouse through a comprehensive regulatory roadmap and special legislation slated for early 2025.

Structured into three core divisions—Demand led by Naver Cloud, Supply by LG Electronics, and Foundation by Kakao—plus an Export Industrialization Task Force targeting markets like the UAE, the alliance orchestrates end-to-end development from construction and ICT to regulatory reform and workforce cultivation. This holistic approach supports Korea’s ambition to not only build massive AI data center capacity but also export integrated AI infrastructure solutions globally, positioning data centers as future 'export factories' supplying intelligence worldwide.

Despite its bold vision, the alliance faces a critical bottleneck in South Korea’s power grid capacity, which remains unresolved and threatens to constrain the ambitious AI data center build-out. Government pledges to address power, land, and permitting challenges underscore the urgency, as the alliance’s success hinges on overcoming this infrastructural hurdle to realize its full potential.

A strategic pillar of the alliance is the integration of domestic AI hardware, notably FuriosaAI’s RNGD neural processing unit, which aims to supply 30-40% of NPUs in future AI data centers. This push for homegrown AI chips reflects a national strategy to reduce reliance on foreign suppliers like Nvidia, embedding Korea’s sovereign innovation into the heart of its AI infrastructure mega project and enhancing its competitive edge in the global AI ecosystem.

Nationwide AI Access Push

Korea’s government is launching a free, unlimited AI chatbot for every citizen, backed by strict domestic tech mandates and massive GPU allocations to break foreign dependence and foster local innovation.

South Korea’s ambitious 'AI for All' initiative is set to democratize AI access by launching a free, nationwide chatbot service with unlimited usage, beginning a beta rollout in September 2024 and full public availability by December. This bold move treats AI as essential public infrastructure, aiming to empower every citizen with AI capabilities while deliberately minimizing reliance on foreign AI models, which will be allowed only in limited capacity without government support to bolster domestic innovation.

To operationalize this vision, the government has allocated 512 Nvidia B200 GPUs among three major consortia led by SK Telecom, KT, and Kakao, mandating that at least 80% of AI model usage prioritize domestic technologies. These consortia, which include telecom giants and platform leaders like LG Uplus and Kakao, are tasked with deploying multi-vendor, multi-channel AI services that integrate proprietary Korean models such as SKT’s AX K2, KT’s Mi:dm K Pro 2.5, and Kakao’s Kanana and EXAONE, thereby reducing vendor lock-in and enhancing accessibility across popular platforms like KakaoTalk and Daum.

Beyond basic chatbot functions, the 'AI for All Project' is accelerating toward a transformative one-agent-per-person era by subsidizing AI agent use and fostering cross-industry consortiums to develop agentic AI capable of executing real-world tasks. SK Telecom, leading a consortium of about 20 organizations spanning finance, healthcare, education, and public services, plans to implement an 'intelligence improvement loop' that leverages real user data from its AI service 'A.' with over 10 million monthly users to continuously enhance model performance, while emphasizing safety, trust, and cost-efficient service through advanced architectures and rigorous monitoring.

Sovereign AI in Critical Sectors

SK Telecom is embedding homegrown AI into manufacturing, defense, and cybersecurity, forging industry-spanning alliances and feedback loops to ensure Korean models power real-world applications.

SK Telecom is spearheading the development of sovereign AI applications tailored for critical sectors such as manufacturing, defense, legal, and tax. Their proprietary A.X K1 and A.X K2 foundation models have been deployed in diverse industrial contexts, including AI agents at steelmaker KG Steel and auto parts manufacturer Conec, as well as quantized AI models co-developed with the Ministry of National Defense for all military branches. This targeted approach underscores SKT's commitment to embedding sovereign AI deeply into South Korea’s industrial and defense infrastructure.

To unify AI development and deployment, SK Telecom leads a consortium that integrates model creation, data curation, domestic AI semiconductor utilization, and service verification into a cohesive ecosystem. This consortium spans about 20 organizations across finance, mobility, media, education, healthcare, and public sectors, orchestrating AI agents capable of multi-stage execution processes that enhance end-to-end success rates. Central to this effort is an 'intelligence improvement loop' championed by SKT’s Ryu Kyung-sang, which rapidly refines AI models through real-world feedback, ensuring continuous enhancement of AI's execution capabilities in industrial and public applications.

Recognizing the critical importance of security and reliability, SK Telecom has formed a specialized consortium with Upstage, nine leading cybersecurity firms—including AhnLab, SK Shieldus, and RaonSecure—and academic institutions like Korea University and Soongsil University. This coalition aims to develop a domestic AI foundation model tailored to South Korea’s unique cyber threat landscape, automating tasks from threat detection to response amid escalating workloads that see up to 10,000 daily alerts. The initiative not only addresses a shortage of cybersecurity professionals but also seeks to build a globally competitive security AI ecosystem fortified by rigorous independent evaluation and workforce training facilitated by KISIA.

In the legal tech domain, Law&Company stands out as the sole specialized firm within South Korea’s sovereign AI foundation model consortium, driving the integration of AI into legal workflows. Through strategic acquisitions like LawTime Biztech and investments in AI infrastructure, the company has expanded its offerings across platforms such as LawTalk and SaaS solutions including BigCase and SuperLawyer. This momentum reflects a broader national push under South Korea’s AI initiatives to embed sovereign AI capabilities across specialized industries, fostering cross-sector collaboration and innovation.

Blueprint for Telecom AI Sovereignty

SK Telecom’s hands-on, cost-conscious approach to building Korean-language AI models offers a practical roadmap for telcos seeking national-scale AI without sacrificing operational efficiency.

SK Telecom’s strategic vision for AI sovereignty is anchored in its development of proprietary AI models tailored specifically for the Korean language and domestic industries, exemplified by its A.X model family, including the 519 billion-parameter A.X K1 and the 688 billion-parameter multi-modal A.X K2. This initiative, part of Korea’s Sovereign AI Foundation Model Project since 2018, positions SK Telecom not only as a national-scale AI model builder but also as a pivotal contributor to Korea’s technological independence. The operator’s recent MOU with the Ministry of National Defense to develop specialized AI models for defense applications and ongoing validation in manufacturing sectors underscore a deliberate effort to embed AI sovereignty within critical national infrastructures, reflecting a broader strategic alignment with Korea’s national priorities.

Balancing AI sovereignty with cost efficiency and operational scalability is central to SK Telecom’s blueprint, offering valuable lessons for other telecommunications operators aiming to carve out their own AI capabilities. Taeyoon Kim, Head of Foundation Model at SK Telecom, emphasizes the importance of telcos clearly defining which AI capabilities are essential to their services and gaining hands-on experience across the full model lifecycle—from training to deployment. This pragmatic approach encourages operators to strategically leverage open models where appropriate while focusing proprietary efforts on core competencies, thereby optimizing resource allocation and ensuring scalable, real-world AI deployments tailored to specific service requirements.

Sources

Get the stories behind the trends

Deep-dive reporting and the weekly brief, in your inbox.