NVIDIA and Korea Expand Sovereign AI: NVLink Fusion, Samsung, LG, Kakao and OpenShell

NVIDIA and South Korea's Ministry of Science and ICT (MSIT) widened their sovereign AI collaboration on November 8, 2026, the day before NVIDIA AI Day Seoul opened. NVIDIA put it out as four blog posts. Together they cover Korean chips, Korean models, physical AI at Samsung, Kakao Mobility and LG, and agent security with OpenShell. Most of it is a plan or a pilot, not a finished product, so this post separates what is announced from what is shipping.

If you want only the deployment side, our companion post covers the Dynamo recipes for four Korean open models.

TL;DR: what is in the package

What did NVIDIA and MSIT actually agree?

The core post, "From Chips to Skills, NVIDIA and Korea Expand Sovereign AI Collaboration", is dated November 8 and written by Jesse Clayton. It says the expanded work with MSIT spans Korea's sovereign AI ecosystem: AI chips, data center infrastructure, talent, startups, models, agents and security.

Four concrete pieces stand out.

• NVLink Fusion preparation. The collaboration "prepares" Korean NPU and XPU companies to explore joining the NVLink Fusion ecosystem. The wording is "explore", not "join".
• Training. Planned activities include NVIDIA training and certification programs, hackathons and technical seminars aimed at improving GPU utilization.
• AI for All. NVIDIA will work with three selected consortia and four leading Korean foundation model builders. They will use the NVIDIA Dynamo inference software to optimize and scale inference.
• SK Telecom and Rebellions. A separate collaboration, supported by MSIT, aims to build sovereign AI infrastructure for Korean models and workloads using NVLink Fusion. NVIDIA links to a press release for those details, which we did not find terms for.

The pattern is familiar from other national deals: a ministry, a local telecom, a local chip start-up and a list of named models. If the idea of "sovereign" AI is new to you, our explainer on what sovereign AI means and its layers is the place to start.

Why NVLink Fusion for Korean NPUs is the strategic part

NVLink is the fast interconnect NVIDIA uses to join GPUs into a single large system. NVLink Fusion opens that fabric to partner silicon. For a country with strong memory makers and a growing group of custom accelerator companies, the pitch is simple: build your own NPU, but keep it compatible with the system most AI clusters already use.

Rebellions is the Korean NPU company named in the SK Telecom project. Our earlier note on the cancelled visit by Madison Huang records that a Rebellions headquarters stop was dropped from her schedule, so this announcement fills in some of what that visit would have covered. See Madison Huang cancels the AI Day Seoul trip.

The same NVLink Fusion idea showed up outside Korea earlier this year. We covered how Apple's AI server plans connect to NVLink Fusion. The common thread is that NVIDIA wants its interconnect to be the standard even when the compute silicon is someone else's.

There is a real trade-off here. Korean chip makers gain a path into existing NVIDIA-based data centers. They also tie their designs to a vendor that already dominates the market. NVIDIA's post does not discuss that tension, and it does not give any dates for when a Korean NPU might ship on NVLink Fusion.

Physical AI: Samsung, Kakao Mobility and LG

The second post, "Korea Advances Physical AI Across Industries With NVIDIA", says NVIDIA is helping Korea's physical AI ecosystem move "from development into real-world production". It includes no executive quotes. It lists three company stories.

Samsung Electronics: chip design and simulation

Samsung is expanding its collaboration with NVIDIA on semiconductor innovation using accelerated computing and AI physics. NVIDIA names several libraries:

• cuISS and cuDSS speed up sparse linear system computations in technology computer-aided design (TCAD).
• GPU acceleration extends to process and optical simulations.
• cuEST helped Samsung's sMOL platform run selected quantum chemical calculations up to 5x faster end to end.
• cuEquivariance is being explored for tensor operations in atomistic AI models.
• PhysicsNeMo supports chip-scale thermal and mechanical analysis. In one AI homogenization case, simulation time fell from about 18 hours (a fine-grid solve on a 32-core CPU) to 41 seconds (a GPU-accelerated coarse-grid solve), roughly 1,600x.

Read these numbers as NVIDIA's. The posts do not publish the benchmark setup beyond the one-line description, and a coarse-grid solve is not the same computation as a fine-grid solve. That is the nature of the trick: AI homogenization trades exactness for speed. The result is plausible for that specific case and should not be generalized to all chip simulation.

NVIDIA also says its Agent Toolkit could eventually connect these tools so engineering agents can run analyses and compare designs. "Could eventually" is a roadmap, not a feature.

Kakao Mobility: a level 4 driving plan

Kakao Mobility plans to build a level 4 autonomous driving system on NVIDIA's automotive stack. The pieces named are NVIDIA Hyperion, a reference platform for sensors and compute powered by DRIVE AGX Thor, and NVIDIA Halos for safety across development and validation. Kakao will combine its own driving software and its experience operating in Seoul's Gangnam district. Initial development centers on Kia purpose-built vehicles, with the goal of supporting future level 4 mobility services in Korea.

Again the verbs are "plans to" and "goal". No launch date, fleet size or road permit is given.

LG Electronics: CLOiD and Isaac GR00T

LG demonstrated Isaac GR00T-powered robotic manipulation and manufacturing tasks with its CLOiD robot in an actual manufacturing environment. NVIDIA describes a physical AI data flywheel with these steps: real-world data collection, Cosmos-based data augmentation, GR00T model training, Jetson AGX Thor onboard computing, and simulation-based evaluation. LG teams are refining a repeatable way to train, evaluate and improve manufacturing robots.

This is the kind of workflow our readers see across robotics now. For a related NVIDIA release on robot data, see our post on the NVIDIA Flora open models and robot data engine.

Robot arm gently holding a leaf, illustrating physical AI manipulation in the NVIDIA Korea package

Agents: OpenShell, NemoClaw and K-EXAONE-2.5

The third post, "Korea's AI Ecosystem Advances and Secures Agentic AI With NVIDIA", was written by Nicola Sessions. It covers three things.

Hyundai Motor Group is exploring NVIDIA NemoClaw blueprints and the NVIDIA OpenShell runtime. NVIDIA calls OpenShell "a key component of the NVIDIA Open Agent Safety Platform". It offers sandboxed execution and policy-based controls for protecting sensitive data, managing access to enterprise tools and governing autonomous agent actions. Our explainer on the Open Agent Safety Platform and OpenShell covers the design, and the security record of NemoClaw is worth reading too: see the NemoClaw CVE-2026-65105 takeover issue.

AIM Intelligence, a Korean AI security company, is integrating its Starfort agent control platform with OpenShell. The aim is to pair OpenShell's runtime access controls with Starfort's detection of malicious instructions and data leakage. The company is considering a preliminary proof of concept for the public energy sector, with possible uses across enterprise, industrial and robotics settings.

LG AI Research is opening API access to K-EXAONE-2.5, a model developed using NVIDIA Nemotron datasets, including Nemotron-Personas-Korea. That dataset is open and holds 7 million fully synthetic Korean personas grounded in official Korean statistics. Model builders use it to ground training data and evaluation in Korean contexts. Earlier Nemotron persona sets for Europe are covered in our posts on UK and Germany personas and France and Belgium personas.

Agent sandbox test illustrating runtime isolation for NVIDIA OpenShell Korea agents

What is announced, what is shipping, what is only a claim

Context: why this lands on the eve of AI Day Seoul

The timing matters. NVIDIA AI Day Seoul runs at COEX on November 9 and 10, according to Korean reports we cited when Madison Huang canceled her trip. Korea's government has also pledged heavy spending: our post on the 2027 budget pledge of about $15.97 billion shows the same priority list of AI, chips and physical AI. The week also included an unannounced visit by Sam Altman, covered in what is reported about Altman's Korea trip and the HBM talks.

For a wider picture of Korea's domestic models, our earlier overview of the South Korea AI ecosystem and sovereign HyperCLOVA push is the background. NVIDIA's posts fit that picture: they plug NVIDIA software and interconnects into models and companies Korea already champions.

NVIDIA is also in a hard news week at home. Its chief executive is the subject of a board review tied to a report about a suspected Chinese intelligence operation. Nothing in the Korea posts refers to it. Our running coverage is in the Jensen Huang and Megaspeed claimed-versus-verified post, which carries dated updates.

What it means for builders

• If you build Korean-language systems, the open Nemotron-Personas-Korea dataset and the deployment recipes for Korean models are the immediately usable parts.
• If you run agents in the enterprise, watch OpenShell integrations. A sandbox and policy layer is becoming the default expectation, whatever vendor you choose. A neutral place to start thinking about agent risk is AgentBeam, the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.
• If you work in chips or robotics, treat the Samsung and LG items as case studies. Ask for the baseline: which solver, which hardware, which accuracy trade-off.
• If you track geopolitics, the NVLink Fusion move is the one to follow. It shows NVIDIA offering a controlled route for allied chip designs to coexist with its platform.

How to verify and what to watch

1. Read the primary posts yourself: the sovereign AI post, the physical AI post, the agentic AI post and the Dynamo recipes post.
2. Look for the SK Telecom and Rebellions press release, which NVIDIA links but which we did not find terms for.
3. Watch for statements from MSIT, Samsung, Kakao Mobility, LG and Hyundai. Their own announcements will show which items are firm.
4. Check whether any NPU maker publishes an NVLink Fusion design, a date or a customer. That would turn "explore" into a product.

Related reading

• Dynamo recipes for Korea's open models
• What is sovereign AI?
• South Korea's $15.97B 2027 AI budget pledge
• Madison Huang cancels the AI Day Seoul trip
• NVIDIA Open Agent Safety Platform and OpenShell
• Apple's AI server and NVLink Fusion
• South Korea's AI ecosystem and HyperCLOVA

Details are accurate as of November 9, 2026. The posts we cite are NVIDIA's own and describe plans that may change.