NVIDIA bundles chips, open models and startup cash into one Korea sovereign AI push

What NVIDIA announced

NVIDIA and Korea's Ministry of Science and ICT hosted an AI Startup Partnership Forum at COEX, where ministry vice minister Ryu Je-myung met NVIDIA leaders and more than 20 Korean startups working on foundation models, physical AI, AI infrastructure and AI semiconductors. At the forum NVIDIA announced an investment in Point2 Technology and said it plans further investments in Motif Technologies, RLWRLD and Upstage. The four companies span the stack NVIDIA says it wants to build in Korea: Point2 builds radio-frequency interconnects that move data between chips and racks in AI factories, Motif is a Seoul lab building foundation models for agents, RLWRLD builds robotics foundation models, and Upstage builds Solar language models and document AI tools. m:…3716

The investments sit on top of a startup program that is already large. NVIDIA says its Inception program has more than 1,000 Korea-based members, and that the program offers preferred pricing, technical expertise and introductions to investors. That matters because the forum was not a one-off showcase: the ministry co-hosted it, and the companies in the room cover the layers a sovereign AI stack needs, from chips and interconnects up to models and agents. The pattern is a bundle — government backing, startup funding and a pipeline of local builders — rather than a single deal. m:…3716

Open models get a production path

The models themselves got the most concrete help. NVIDIA is releasing Dynamo-optimized deployment recipes for four Korean open models. Dynamo is NVIDIA's open-source framework for serving inference across many machines in a data center, and the recipes are tuned for NVIDIA B200 GPUs. NVIDIA says they deliver up to twice the performance out of the box. The problem they solve is ordinary and expensive: once a model grows beyond one GPU it has to be split across several, and routing requests and coordinating communication between them takes careful work that every team would otherwise repeat. m:…8956

That is the open-model half of the sovereign AI idea. Korea's leading models are open, which means any team can inspect and use them, but open weights alone do not put a model into production. The recipes close that gap for the four models, and they tie the models to NVIDIA's hardware and serving software at the same time. The implication for an executive weighing where to build is that the local models now come with a supported path to run at scale, not just a download link. m:…8956

Sovereign AI, from chips to skills

The wider collaboration with the ministry covers AI semiconductors, data center infrastructure, talent, startups, models, agents and security, NVIDIA says. One piece opens the door for Korean NPU and XPU companies to explore joining NVIDIA's NVLink Fusion ecosystem, which would let local chipmakers connect their designs to NVIDIA's interconnect technology. Another is a collaboration with SK Telecom and Rebellions, supported by the ministry, to advance sovereign AI infrastructure for local models and workloads. Through Korea's AI for All initiative, NVIDIA will work with three selected consortia, and it is already working with four Korean foundation model builders to optimize and scale inference with Dynamo. m:…8955

Skills are part of the same package. NVIDIA says the collaboration includes training and certification programs, hackathons and technical seminars aimed at strengthening GPU use in Korea. Sovereign AI programs often start with hardware and stop there; this one pairs the hardware with people who can run it and with local models that give them something to run. That combination is what makes the announcement more than a procurement story. m:…8955

Agents get a guardrail layer

The agent layer got its own announcements. Hyundai Motor Group is exploring NVIDIA's NemoClaw blueprints and the OpenShell runtime, while AIM Intelligence is integrating its Starfort platform with OpenShell. NVIDIA describes OpenShell as a secure runtime with sandboxed execution and policy-based controls, and the company says the goal is to combine those runtime controls with Starfort's detection of malicious instructions and data leakage. LG AI Research is opening API access for K-EXAONE-2.5, a model developed using NVIDIA's Nemotron datasets, and NVIDIA released Nemotron-Personas-Korea, an open dataset of 7 million fully synthetic Korean personas grounded in official Korean statistics. m:…8954

Taken together, the announcements read as one deal with four parts: money for startups, a production path for open models, infrastructure and skills under a government partnership, and a security layer for agents. Each part on its own would be a routine vendor announcement. The reason it matters to an AI executive is that a buyer or builder in Korea no longer has to assemble those parts separately — and a competitor selling into Korea now faces a stack that arrives pre-assembled, with the ministry's name on it. m:…3716 m:…8956 m:…8955 m:…8954