General-Purpose Physical AI Models & Computing Platforms

The Sovereign AI Imperative

In the global AI landscape of 2026, a stark asymmetry defines the competitive terrain: the United States dominates foundation model development, with OpenAI, Google DeepMind, Anthropic, and Meta collectively controlling the most powerful general-purpose AI systems in the world. China has established a parallel ecosystem anchored by Baidu, Alibaba, ByteDance, and DeepSeek. For every other nation—including technologically advanced economies like South Korea—the question of AI sovereignty has become an urgent strategic priority.

K-Moonshot Mission 7 confronts this asymmetry directly. Korea's approach to sovereign AI is not merely about building models that match US or Chinese capabilities parameter-for-parameter; it is about developing AI systems that are deeply optimized for the Korean language, culture, and industrial context, while establishing the domestic computing infrastructure necessary to train and deploy these systems without dependence on foreign cloud providers. The mission's scope encompasses both the "intelligence" layer (foundation models) and the "infrastructure" layer (GPU clusters, national computing centers), recognizing that sovereignty requires control over both.

The Five Sovereign AI Consortia

The Ministry of Science and ICT has selected five consortia to lead Korea's sovereign AI model development, each anchored by a major Korean technology company and supported by academic institutions, startups, and government research labs. This consortium structure—rather than a single national champion model—reflects a deliberate strategy to maintain competitive diversity while concentrating resources above the threshold of global relevance.

The Five Consortia

Consortium Lead	Focus Area	Key Model(s)	Distinguishing Capability
LG AI Research	Industrial & Enterprise AI	EXAONE 3.5	Manufacturing, materials science, B2B applications
SK Telecom	Telecommunications & Infrastructure AI	519B parameter model	Largest Korean model, telco-optimized
Naver Cloud	Consumer & Language AI	HyperCLOVA X	Korean language supremacy, omnimodal
NCSoft	Interactive & Generative AI	VARCO	Game AI, simulation, digital worlds
Upstage	Document & Enterprise AI	Solar	Small-model efficiency, document understanding

The $381 million in government funding is distributed across these consortia as catalytic capital, with the expectation that private-sector co-investment will multiply the total spending several-fold. Each consortium is required to open-source certain model components, contribute to shared evaluation benchmarks, and collaborate on safety and alignment research—conditions designed to prevent the fragmentation of Korea's AI ecosystem into proprietary silos.

Naver HyperCLOVA X: Korea's Flagship Language Model

Naver, Korea's dominant internet platform, has invested heavily in HyperCLOVA X as the country's most prominent foundation model. The model's defining characteristic is its unprecedented depth of Korean language understanding, trained on a dataset containing 6,500 times more Korean-language data than GPT-4. This data advantage—drawn from Naver's search engine, news platform, online encyclopedia (Naver Knowledge iN), and e-commerce ecosystem—creates a model that understands Korean linguistic nuance, cultural context, and domain-specific terminology at a level that no foreign model can match.

CES 2026: Omnimodal Capabilities

At CES 2026 in Las Vegas, Naver unveiled the latest iteration of HyperCLOVA X with omnimodal capabilities—the ability to process and generate across text, images, audio, video, and code within a single unified model. This omnimodal architecture represents a technical advance beyond the multimodal models that characterized the 2024-2025 generation, enabling seamless cross-modal reasoning that is essential for physical AI applications. For example, the omnimodal HyperCLOVA X can analyze a video of a manufacturing process, generate text instructions for correction, and synthesize an audio explanation—all within a single inference pass.

Naver's strategy positions HyperCLOVA X not as a competitor to GPT-5 or Gemini in the English-language consumer market, but as the sovereign AI backbone for Korea's enterprise and government sectors. Naver Cloud, the company's cloud computing division, offers HyperCLOVA X through its CLOVA Studio platform, enabling Korean enterprises to fine-tune and deploy the model on domestic infrastructure—a critical requirement for government agencies and regulated industries that cannot rely on foreign cloud providers for sensitive workloads.

SK Telecom: Korea's Largest Model

SK Telecom, Korea's largest telecommunications company, has taken a different strategic approach: building the largest model in Korea by parameter count. The company's current flagship model stands at 519 billion parameters—the largest language model developed by a Korean organization—with plans to upgrade to a model exceeding one trillion parameters by the second half of 2026.

The Trillion-Parameter Ambition

SK Telecom's push toward a trillion-parameter model is driven by both technical and strategic logic. At the technical level, scaling laws in large language models have consistently shown that larger models, trained on sufficient data, outperform smaller ones across a wide range of tasks. SK Telecom's leadership has stated that achieving parity with frontier US models requires competing at scale, not merely optimizing smaller models for efficiency.

Strategically, SK Telecom's model development is tightly integrated with the company's telecommunications infrastructure. The model is designed to power next-generation customer service, network optimization, and enterprise applications across SK Telecom's 30+ million subscriber base. The company has also partnered with SK Hynix, its sister company within the SK Group and the world's second-largest memory chip manufacturer, to co-optimize the model's architecture for deployment on High Bandwidth Memory (HBM)-equipped inference servers—a hardware-software integration that leverages SK Group's unique position spanning both AI and semiconductor manufacturing.

SK Telecom's AI subsidiary, SK Telecom AI, has also invested in partnerships with global AI companies including Anthropic, taking a strategic stake in the US AI safety company. This dual approach—building sovereign Korean models while maintaining relationships with frontier US developers—reflects a pragmatic hedging strategy common among Korea's major technology groups.

LG EXAONE: The Enterprise and Industrial AI Contender

LG AI Research, the artificial intelligence division of LG Group, has developed EXAONE (Expert AI for Everyone) as a foundation model specifically optimized for enterprise and industrial applications. In independent evaluations, EXAONE has been ranked 7th worldwide among foundation models—a remarkable achievement for a model developed outside the US-China duopoly and a testament to LG AI Research's engineering depth.

K-EXAONE and the MWC 2026 Push

At Mobile World Congress (MWC) 2026 in Barcelona, LG AI Research launched K-EXAONE, a Korean-language-optimized version of the EXAONE model specifically designed for domestic enterprise deployment. K-EXAONE integrates domain-specific fine-tuning for manufacturing, chemical engineering, and materials science—sectors where LG Group has deep operational expertise through LG Chem, LG Energy Solution, and LG Electronics' manufacturing operations.

LG's industrial AI strategy is distinctive within the Korean sovereign AI landscape. While Naver focuses on consumer-facing language applications and SK Telecom targets telecommunications infrastructure, LG EXAONE is purpose-built for the physical world: optimizing factory production lines, predicting battery degradation patterns, designing new materials, and automating quality inspection. This positioning directly connects Mission 7 to Mission 6 (Humanoid Robots), as LG's KAPEX humanoid platform will rely on EXAONE-derived models for real-world reasoning and manipulation.

Kakao: The Consumer AI Platform

Kakao, the operator of KakaoTalk (Korea's dominant messaging platform with over 48 million monthly active users), has pursued a distinctive AI strategy that combines domestic model development with strategic international partnerships. Kakao's Kanana AI agent represents the company's approach to deploying AI within the conversational commerce ecosystem that KakaoTalk enables.

The OpenAI Partnership

In a move that highlights the tension between sovereign AI ambitions and pragmatic access to frontier capabilities, Kakao has entered a partnership with OpenAI to integrate GPT-family models into KakaoTalk. This partnership gives KakaoTalk's massive user base access to state-of-the-art conversational AI while allowing Kakao to focus its own development resources on Korean-language optimization and platform-specific applications rather than attempting to build a frontier model from scratch.

The Kakao-OpenAI relationship illustrates a broader strategic question facing Korea's AI ecosystem: whether sovereign AI requires building every layer domestically, or whether selective partnerships with foreign frontier developers—combined with sovereign control over data, infrastructure, and application layers—can achieve the same strategic objectives more efficiently. Korea's approach, as evidenced by the five-consortium structure, accommodates both philosophies.

GPU Infrastructure: The Computational Foundation

Foundation model development is ultimately constrained by computational infrastructure. Training a frontier model requires tens of thousands of high-end GPUs operating in concert for weeks or months, consuming megawatts of electricity. Korea's sovereign AI ambitions are therefore inseparable from its national GPU infrastructure strategy.

Current State and Scaling Plan

As of early 2026, Korea has deployed over 50,000 GPUs across government-funded computing centers and private-sector facilities. The government's phased scaling plan commits to distributing 4,000 GPUs starting in March 2026, scaling to 52,000 by 2028, and reaching the full 260,000 target by 2030. These deployments are concentrated at the National AI Computing Center and at facilities operated by Naver, SK Telecom, and KT (Korea Telecom).

The GPU procurement programme relies heavily on NVIDIA, whose H100 and successor Blackwell-architecture GPUs represent the global standard for AI training and inference workloads. Korea's relationship with NVIDIA is strategically significant: NVIDIA CEO Jensen Huang has made multiple visits to Korea to strengthen partnerships with Samsung (which fabricates certain NVIDIA chip components) and SK Hynix (which produces the HBM chips essential to NVIDIA's GPU architecture). This supply chain interdependence gives Korea leverage in securing GPU allocation even during periods of global GPU shortage.

The National AI Computing Center

The National AI Computing Center, operated under the auspices of MSIT and the Institute for Information & Communications Technology Planning & Evaluation (IITP), serves as the anchor facility for government-funded AI research. The center provides computing resources to universities, research institutes, and startups that lack the capital to build their own GPU clusters—democratizing access to the computational resources required for serious AI research.

The center operates on a prioritized allocation model: sovereign AI consortium members receive guaranteed capacity for training runs, while remaining capacity is allocated through a competitive proposal process to academic researchers and startups. This structure ensures that Korea's most strategically important AI development efforts are not bottlenecked by compute availability, while still fostering a broad ecosystem of innovation.

From Language Models to Physical AI

Mission 7's full title—"General-Purpose Physical AI Models and Computing Platforms"—points to an ambition that extends beyond text-based language models into the domain of AI systems that understand and interact with the physical world. This "Physical AI" dimension connects Mission 7 directly to Korea's broader industrial strategy and to multiple other K-Moonshot missions.

Physical AI models must process sensor data from cameras, LIDAR, tactile sensors, and inertial measurement units; reason about spatial relationships, object properties, and physical dynamics; and generate motor commands that produce precise, safe interactions with the real world. These capabilities require foundation model architectures that are fundamentally different from text-only language models—multimodal or omnimodal systems that can fuse information across modalities and reason in the physical domain.

Application Domains

• Humanoid robotics: Foundation models for Mission 6's humanoid robots, enabling dexterous manipulation, navigation in unstructured environments, and human-robot collaboration
• Autonomous vehicles: Physical AI for Hyundai's autonomous driving programme, integrating perception, prediction, and planning in a single model architecture
• Smart manufacturing: AI models that optimize factory operations in real time, predict equipment failures, and coordinate multi-robot production cells
• Drug discovery: AI systems that model molecular dynamics and protein interactions for Mission 1's drug development acceleration
• Materials science: Foundation models for discovering and optimizing new materials, connecting to Mission 9's rare earth strategy

Korea's competitive advantage in physical AI derives from the proximity of its AI developers to its manufacturing base. Unlike Silicon Valley AI companies that must partner with external manufacturers to deploy physical AI, Korean conglomerates like Hyundai, Samsung, and LG operate massive production facilities where physical AI models can be trained, tested, and refined in real industrial environments—a closed-loop development cycle that accelerates iteration.

The Data Sovereignty Dimension

Underlying the entire sovereign AI model strategy is the question of data sovereignty. Korean-language training data of sufficient quality and scale can only be generated within Korea's digital ecosystem—Naver's search and commerce platforms, Kakao's messaging and payment systems, government databases, academic repositories, and industrial operational data. Control over this data and the models trained on it represents a form of digital sovereignty that Korea's policymakers view as strategically equivalent to semiconductor manufacturing capability or energy security.

Korea's Personal Information Protection Act (PIPA) and related data governance frameworks are being updated to balance two objectives: protecting individual privacy while enabling the responsible use of Korean-language data for sovereign AI training. The government has established data trusts and anonymization protocols that allow model developers to access large-scale datasets without compromising personal information—a framework that must walk the fine line between enabling AI development and maintaining the trust of Korea's digitally sophisticated citizenry.

Risk Assessment

• Scale gap with frontier models: Despite impressive progress, Korean models remain smaller and less capable than the latest offerings from OpenAI, Google, and Anthropic. The 519-billion-parameter SK Telecom model is large by Korean standards but modest relative to frontier US and Chinese models reportedly exceeding several trillion parameters. Closing this gap requires sustained investment at a scale that may test even Korea's substantial commitments.
• GPU supply dependence: Korea's 260,000 GPU target depends almost entirely on NVIDIA, creating a single-supplier risk. Any disruption to NVIDIA supply—whether from export controls, production constraints, or geopolitical events—would directly impact Korea's sovereign AI timeline. This risk motivates Mission 11's AI accelerator chip development, but domestic alternatives remain years from production at scale.
• Talent constraints: Building and operating frontier AI systems requires a specialized workforce that Korea does not yet possess in sufficient numbers. The country produces approximately 2,500 AI-related PhD graduates annually—a significant number but insufficient for the ambitions of five sovereign AI consortia, dozens of startups, and major corporate research labs competing for the same talent pool. Mission 10's AI talent development addresses this constraint but results will take years to materialize.
• Fragmentation risk: Five separate sovereign AI consortia may dilute Korea's resources rather than concentrate them. If each consortium pursues independent architectures, training pipelines, and evaluation frameworks, the country could end up with five mediocre models rather than one or two globally competitive ones. The MSIT consortium governance structure includes provisions to mitigate this risk, but coordination across competing corporations is inherently difficult.
• Open-source competition: Meta's Llama, Alibaba's Qwen, and other open-source model families are rapidly improving and are available at zero licensing cost. Korean enterprises may choose to fine-tune these open-source models for Korean-language tasks rather than adopt commercially licensed sovereign Korean models, undermining the business case for domestic model development.

Analytical Assessment

Mission 7 addresses what is arguably the most strategically consequential challenge within K-Moonshot: ensuring that Korea possesses sovereign capability in the foundational technology—artificial intelligence—that underpins all other missions. Without competitive AI models and the computing infrastructure to train and deploy them, Korea's ambitions in robotics, drug discovery, quantum computing, and every other mission domain are ultimately dependent on foreign technology providers.

Korea's five-consortium approach is a pragmatic response to the reality that no single Korean company can match the resources of OpenAI ($13+ billion in capital) or Google DeepMind (backed by Alphabet's $300+ billion annual revenue). By distributing sovereign AI development across five entities with distinct specializations, Korea creates multiple pathways to success while maintaining competitive dynamics that prevent complacency.

The 260,000 GPU target, if achieved, would give Korea one of the largest national AI computing infrastructures outside the US and China. Combined with Korea's semiconductor manufacturing capabilities—Samsung's foundry operations and SK Hynix's HBM dominance—this infrastructure creates a vertically integrated AI capability stack that few nations can replicate. The critical variable is execution: converting budgetary commitments into operational computing capacity, trained models, and deployed applications at the pace that the global AI race demands. The 2030 midpoint assessment will reveal whether Korea's sovereign AI strategy has achieved its objectives or whether the gap with frontier developers has widened despite substantial investment.