NVIDIA Expands Korea Sovereign AI Initiative With MSIT, Invests in Point2, and Releases Dynamo Deployment Recipes for Local Foundation Models

South Korea’s Ministry of Science and ICT (MSIT) and NVIDIA have expanded their sovereign AI collaboration through infrastructure, model deployment, and investment initiatives unveiled at the MSIT-NVIDIA AI Startup Partnership Forum at COEX in Seoul. During the forum, Ryu Je-myung, South Korea’s Vice Minister of Science and ICT, met with NVIDIA leadership and over 20 domestic AI enterprises across semiconductors, foundation models, physical AI, and infrastructure. NVIDIA announced a direct investment in Korean interconnect developer Point2 Technology, alongside planned future investments in agent lab Motif Technologies, physical robotics startup RLWRLD, and foundation model builder Upstage.

The announcements reflect an evolving approach to sovereign AI. Rather than treating sovereignty purely as a mandate for domestic GPU procurement, the initiative spans physical interconnects, open foundation models, distributed inference serving, and runtime agent sandboxing. Supported by MSIT’s “AI for All” initiative, the collaboration lays groundwork for Korean accelerators to explore participation in the NVIDIA NVLink Fusion ecosystem, while providing preconfigured serving recipes for four Korean open-weight models. By linking foreign accelerated compute to domestic model development, MSIT and NVIDIA are outlining a pragmatic path for regional technological autonomy.

Interconnects and Exploratory Silicon Collaboration

At the infrastructure layer, NVIDIA’s investment in Point2 Technology targets critical data center scaling bottlenecks, including interconnect reach, thermal load, and energy consumption. Point2 designs radio frequency-based (RF) interconnects that move data between chips and server racks, offering an alternative to copper cables and optical transceivers for high-density scale-up clusters.

Alongside this startup investment, MSIT is backing collaboration between NVIDIA, telecommunications provider SK Telecom, and AI chipmaker Rebellions. The project establishes groundwork for Korean neural processing units (NPUs) and XPUs to explore technical participation in NVIDIA’s NVLink Fusion ecosystem. Rather than representing completed integration or guaranteeing an end to proprietary lock-in, the initiative provides an exploratory channel to evaluate how domestic hardware can interface with multi-node GPU clusters.

Distributed Serving Recipes for Domestic Open Models

On the software layer, NVIDIA released Dynamo-optimized deployment recipes for four Korean open models: SK Telecom’s A.X K2, LG AI Research’s K-EXAONE 2.0, Upstage’s Solar Open 2, and Motif Technologies’ Motif-3. NVIDIA Dynamo is an open-source, distributed inference framework designed to deploy models across multi-node, multi-GPU environments at data center scale.

Optimized for NVIDIA B200 GPUs, NVIDIA states these Dynamo serving recipes deliver up to 2x performance out of the box by taking guesswork out of cluster deployment. While the announcement does not provide workload baselines or benchmark parameters for that metric, preconfigured deployment templates aim to simplify how domestic enterprises transition local open models into operational environments.

Enterprise Agent Governance and OpenShell Runtime

As Korean enterprises build autonomous workflow agents, runtime safety has emerged as a primary requirement. Hyundai Motor Group has begun exploring NVIDIA NemoClaw blueprints alongside the NVIDIA OpenShell runtime. Operating within NVIDIA’s Open Agent Safety Platform, OpenShell provides sandboxed execution and policy-based controls to govern tool access, protect corporate data, and monitor autonomous transactions.

Concurrently, cybersecurity provider AIM Intelligence is integrating its Starfort agent control platform with OpenShell. The integration aims to combine OpenShell’s runtime access controls with Starfort’s detection of malicious instructions and data leakage. For public sector workflows, AIM Intelligence is considering a preliminary proof of concept in South Korea’s public energy utility sector. As analyzed in our report on NVIDIA Sentry and OpenShell runtime enforcement, placing policy verification outside of prompt context is essential for enterprise security.

To ground models in Korean contexts, LG AI Research utilized NVIDIA Nemotron datasets, including Nemotron-Personas-Korea, in developing K-EXAONE-2.5, for which LG is opening API access. The dataset provides 7 million fully synthetic Korean personas grounded in official government statistics, offering localized synthetic data for model training and evaluation.

Physical AI and Industrial Simulation

The initiative extends into Korea’s physical AI manufacturing ecosystem. In semiconductor engineering, Samsung Electronics is collaborating with NVIDIA across technology computer-aided design, using NVIDIA cuISS and cuDSS to accelerate sparse linear system computations, while exploring GPU acceleration for optical and process simulations. In a separate thermal and mechanical analysis workflow using NVIDIA PhysicsNeMo, NVIDIA reported that one AI homogenization case combining AI-predicted material properties, coarser grids, and GPU acceleration achieved a roughly 1,600x speedup, reducing solve time from 18 hours for a fine-grid CPU solve to 41 seconds on a coarse GPU solve.

In robotics, LG Electronics demonstrated Isaac GR00T-powered manipulation using its CLOiD robot in an active manufacturing facility, combining Cosmos data augmentation, GR00T training, and Jetson AGX Thor onboard compute. In mobility, Kakao Mobility plans to develop a Level 4 autonomous driving system utilizing NVIDIA Hyperion, DRIVE AGX Thor, and NVIDIA Halos on Kia purpose-built vehicles.

Architectural Matrix: South Korea’s Sovereign AI Deployments

Layer	Key Domestic Partners	NVIDIA Technology Component	Operational Consequence
Interconnects	Point2 Technology, Rebellions, SK Telecom	NVLink Fusion, Point2 Direct Investment	RF chip-to-rack interconnect investment; exploratory groundwork for Korean NPU/XPU participation in NVLink Fusion
Foundation Models	Upstage, SK Telecom, LG AI Research, Motif	Dynamo B200 Deployment Recipes	NVIDIA states preconfigured recipes deliver up to 2x performance out of the box on B200 GPUs across four domestic open models
Agent Governance	Hyundai Motor Group, AIM Intelligence	OpenShell Runtime, NemoClaw Blueprints	Sandboxed agent tool execution; planned Starfort detection of malicious instructions and data leakage; preliminary public energy PoC considered
Physical AI	Samsung Electronics, LG Electronics, Kakao Mobility	Isaac GR00T, DRIVE AGX Thor, PhysicsNeMo	PhysicsNeMo thermal/mechanical homogenization speedup (~1,600x coarse GPU vs fine CPU, NVIDIA report); cuISS/cuDSS sparse linear solve exploration; LG CLOiD factory trial; planned Kakao L4 AV

What This Means for Sovereign AI Strategy

South Korea’s coordinated engagement with NVIDIA offers a pragmatic counterweight to binary sovereign AI strategies. By backing domestic interconnect developers like Point2 while establishing exploratory channels for domestic accelerators in NVLink Fusion, the initiative seeks to preserve a seat for Korean hardware in advanced cluster architectures.

Simultaneously, preconfigured Dynamo deployment recipes and culturally grounded synthetic datasets provide domestic enterprises with practical tooling to operationalize open foundation models. While elements such as NPU interconnect integration, Level 4 mobility, and utility agent deployments remain exploratory or planned, the initiative demonstrates how regional economies can combine domestic capabilities with global accelerated computing.