Korea's sovereign AI push now reaches agent sandboxes and local data. Read each announcement by its verb tense
NVIDIA's Korean announcements add agent security runtimes, a synthetic Korean persona dataset and a plan to pair Rebellions NPUs with NVIDIA GPUs by 2028. Most items are plans; the scope is the news.
Listen to this article · 4 min
AI-generated narration of the full article.

Sovereign AI used to mean a national data center and, perhaps, a national model. NVIDIA’s announcements in South Korea on November 8 describe something wider: local chips, local training data and a local layer of agent security, all connected to a global platform.
The useful way to read them is by verb tense. Very little here is shipped. The scope is what is new.
Agent security as part of the national stack
NVIDIA presented OpenShell, a secure runtime with sandboxed execution and policy-based controls for agents’ access to sensitive data and enterprise tools, as a key part of its Open Agent Safety Platform. Around it:
- Hyundai Motor Group is exploring NVIDIA’s NemoClaw blueprints and the OpenShell runtime.
- AIM Intelligence is integrating its Starfort agent control platform with OpenShell, combining runtime access controls with detection of malicious instructions and data leakage, and is considering a preliminary proof of concept for the public energy sector.
- LG AI Research is opening API access to K-EXAONE-2.5, a model developed with NVIDIA’s Nemotron datasets, including Nemotron-Personas-Korea, an open set of 7 million synthetic Korean personas grounded in official Korean statistics.
The dataset deserves attention. Evaluating an assistant for Korean citizens with American test personas measures the wrong population. A public, statistically grounded persona set lets Korean developers test for their own demographics, and that kind of evaluation data is a sovereignty asset as real as a data center.
Local chips inside a global interconnect
The second announcement is further out. Rebellions, a Korean chip designer, plans to connect its R200 neural processors to NVIDIA’s platform through NVLink Fusion and build rack-scale systems on NVIDIA’s MGX architecture. SK Telecom plans to deploy them in its AI data center, with support from Korea’s Ministry of Science and ICT. The design splits inference: NVIDIA GPUs handle prefill, the stage that processes the prompt, and Rebellions NPUs handle decode, the stage that generates tokens. R200 systems are expected to be ready for deployment in 2028.
That is a different model of sovereignty from building everything at home. The country contributes the chip that runs the high-volume, power-hungry part of inference, and connects it to a global ecosystem for the rest. It reduces dependence without pretending to independence.
What enterprises outside Korea can take from it
Few companies will buy Rebellions racks. The pattern still applies to any organization whose regulator, government customer or board asks where its AI runs and who controls it.
Sovereignty is a stack, not a location. We argued that sovereign AI now reaches the data center floor. Korea’s announcements add two layers above it: the runtime that constrains agents, and the evaluation data that reflects local users. Ask vendors about all of them, not only the region.
Agent security is becoming infrastructure. A sandboxed runtime with policy controls, offered as part of a national platform, follows the same logic as containment for agents: the boundary belongs in the runtime, not in the prompt.
Discount the future tense. Exploring, considering and planning are early stages. A procurement plan should rest on what is deployable now, with 2028 hardware treated as an option, not a commitment.