$ today_in_10_seconds
NVIDIA used AI Day Seoul to argue that enterprise agents only ship when security ships with them. Hyundai is testing NemoClaw blueprints on NVIDIA's sandboxed OpenShell runtime, and a Korean security startup is building its agent controls on that same runtime. Everything else NVIDIA announced in Seoul today fills in the stack around those runtimes: open models, inference recipes, and startup money.
$ lead_story
Hyundai and AIM Intelligence Bet on Sandboxed Agents
Hyundai Motor Group is exploring NVIDIA NemoClaw blueprints and the OpenShell runtime, while AIM Intelligence is integrating its Starfort agent-control platform with OpenShell to secure agents for employee productivity and enterprise workflows. OpenShell is a core part of NVIDIA's Open Agent Safety Platform, giving agents a secure runtime with sandboxed execution and policy-based controls. On the model side, LG AI Research is opening API access to K-EXAONE-2.5, built using Nemotron datasets including Nemotron-Personas-Korea. For builders, the takeaway is that large enterprises are starting to buy the agent runtime (where tool calls execute and which policies gate them) instead of having each team build its own.
// why_it_matters: The fight for enterprise agents is moving from the model to the sandbox it runs in.
$ also_worth_attention
NVIDIA AI Day Seoul: Agents, Claws and Physical AI
Today's COEX program covers physical AI, agentic AI and claws, and AI infrastructure, with claws pitched as agents that run continuous tasks at a high level of autonomy. Keynotes from VPs Deepu Talla and Jonathan Cohen lay out NVIDIA's roadmap for edge robotics and enterprise agent frameworks, and tomorrow's paid DLI program includes two agentic AI workshops. Naming a category of always-on agents is how a vendor sells the stack underneath them, so expect the claw label to spread.
link →NVIDIA Dynamo Recipes Bring Korea's Open Models to Production
The deployment recipes are tuned for B200 GPUs and claim up to 2x performance out of the box for Korean open models. They back the government's AI for All initiative, which supports three selected consortia and four Korean foundation-model developers. Agents call the model many times per task, so serving efficiency largely determines what each agent run costs.
link →NVIDIA Plans New Investments in Korean Startups
One named company is Motif Technologies, a Seoul lab building foundation models for agents that handle complex tasks, including the open-weight Motif-3. NVIDIA now counts more than 1,000 Korea-based Inception members. NVIDIA is funding the model layer that will run on its own runtime layer, which looks like a deliberate flywheel.
link →
$ bigger_picture
The trend behind the headlines
Read today's posts together and you get NVIDIA's full-stack template for nations. It covers chips (including SK Telecom and Rebellions integrating NVLink Fusion chiplets into domestic NPU packages for rack-scale systems), open models, inference recipes, a secure agent runtime, and startup capital. Sovereign AI used to mean owning the model. With agents, it increasingly means owning the policy layer that decides what those agents can touch. An open runtime like OpenShell is both a gift and a hook: enterprises get security primitives for free, and NVIDIA becomes the default substrate. If you're building agent security or governance, note what AIM Intelligence did: it built on the platform instead of competing with it.
NVIDIA OpenShell
It's a secure runtime for agents with sandboxed execution and policy-based controls, and it's now attracting both an automaker and a third-party security vendor. It's worth evaluating if your agents execute code or touch internal systems and you'd rather not build isolation and policy enforcement yourself.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime