NVIDIA and South Korea expand sovereign AI push

What changed?

On November 8, NVIDIA published three blog posts on deeper ties with South Korea. The umbrella is an expanded collaboration with Korea's Ministry of Science and ICT (MSIT) (NVIDIA Blog), covering semiconductors and data centers, talent, startups, models, agents and security.

It fits what we covered in South Korea's record AI budget: Seoul is serious about building its own sovereign AI stack, and NVIDIA is now adding the software layer.

The three announcements

• Hardware: the deal opens a path for Korean NPU and XPU makers to explore the NVLink Fusion ecosystem. A separate collaboration with SK Telecom and Rebellions, backed by MSIT, targets infrastructure for Korean models and workloads.
• Inference: NVIDIA is releasing Dynamo optimized deployment recipes for four Korean open models (NVIDIA Blog): SK Telecom's A.X K2, LG AI Research's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif Technologies' Motif-3. Tuned for B200 GPUs, they deliver up to 2x performance out of the box, according to NVIDIA.
• Agent security: Hyundai Motor Group is exploring NemoClaw blueprints and the OpenShell runtime, while AIM Intelligence is integrating its Starfort platform with OpenShell (NVIDIA Blog).

Models and data

LG AI Research is opening API access to K-EXAONE-2.5, trained with NVIDIA Nemotron datasets including Nemotron-Personas-Korea. That open dataset holds 7 million fully synthetic Korean personas grounded in official Korean statistics, used to anchor training and evaluation data in Korean contexts.

Under Korea's AI for All initiative, NVIDIA will work with three selected consortia and four leading Korean foundation model builders on scaling inference with Dynamo. It also plans training, certification, hackathons and technical seminars to improve GPU utilization.

Why it matters

Models rarely fit on a single GPU anymore, and splitting them across GPUs and nodes takes a lot of manual tuning for routing and communication. Dynamo is an open source, distributed inference serving framework, and ready made recipes shortcut exactly that step. If you want to put a Korean open model into production, this saves real time.

The security piece builds on OpenShell, which we covered with the OpenShell reference architecture. AIM Intelligence aims to pair OpenShell's runtime access controls with Starfort's detection of malicious instructions and data leakage, and is considering a preliminary proof of concept in the public energy sector.

What to watch

The numbers are NVIDIA's own: the 2x figure applies to the B200 tuned recipes, and we have not seen independent benchmarks. Teams running LG, Upstage or SK Telecom models should compare against their own deployments. Hyundai and AIM Intelligence are still at the exploration and integration stage, not production rollout.