NVIDIA widens Korea AI push with Samsung, LG, Hyundai and Kakao Mobility

NVIDIA is publishing a set of posts on Korea’s AI ecosystem. They describe how Samsung Electronics, LG Electronics, Kakao Mobility, Hyundai Motor Group and the security firm AIM Intelligence are using its software and hardware for chip engineering, factory robots, self-driving development and agent security. The detail most relevant to API builders is that LG AI Research is opening application programming interface (API) access for its K-EXAONE-2.5 model, which NVIDIA says was developed using its Nemotron datasets, including a new Korean persona dataset.

Both posts list two related NVIDIA posts in the same series: “From Chips to Skills, NVIDIA and Korea Expand Sovereign AI Collaboration” and “NVIDIA Dynamo Recipes Bring Korea’s Leading Open Models to Production”. Their full text was not available for this report. Who the collaboration involves, which models the Dynamo recipes cover and what they contain are therefore not confirmed here. This article reports only what the readable posts state.

What NVIDIA’s Korea posts confirm

The readable posts cover two themes. The first is “physical AI”, meaning AI that senses and acts in the real world: robots, vehicles and simulations of physical processes. NVIDIA says Korean industry leaders are applying accelerated computing, AI and simulation to semiconductor manufacturing, robotics and autonomous vehicles. It says these efforts help move physical AI from development into real-world production. The second theme is agentic AI: models that take actions on a user’s behalf, and the security controls around them.

The posts carry a date of 9 November 2026 in their headers, while the related-news lists on the same pages date the series to 8 November 2026. The gap is probably a time-zone difference, but the sources do not say.

Partner	Area	NVIDIA technology named	Status as described
Samsung Electronics	Semiconductor simulation and materials research	cuISS, cuDSS, cuEST, cuEquivariance, PhysicsNeMo	Expanding collaboration; some results reported
Kakao Mobility	Level 4 autonomous driving	Hyperion, DRIVE AGX Thor, Halos	Planned
LG Electronics	Factory robots (CLOiD)	Isaac GR00T, Cosmos, Jetson AGX Thor	Demonstrated in a manufacturing environment
Hyundai Motor Group	Secure AI agents	NemoClaw blueprints, OpenShell	Exploring
AIM Intelligence	Agent security	OpenShell	Integrating its Starfort platform
LG AI Research	Language model K-EXAONE-2.5	Nemotron datasets, including Nemotron-Personas-Korea	Opening API access

K-EXAONE-2.5 and a Korean persona dataset

LG AI Research is opening API access for K-EXAONE-2.5, which NVIDIA says was developed using Nemotron datasets, including Nemotron-Personas-Korea. NVIDIA describes Nemotron as a family of open models, datasets and techniques for building specialised, efficient AI agents.

Nemotron-Personas-Korea is described as an open dataset of 7 million fully synthetic Korean personas, grounded in official Korean statistics. A synthetic persona is a generated profile of a fictional person. Datasets like this let model builders write training examples and test cases that reflect a real population’s mix of ages, regions and occupations without using real people’s data. NVIDIA says builders use this dataset to ground training data and evaluation in Korean contexts.

For developers this is the most concrete item in the announcements, though important details are missing. NVIDIA’s post gives no pricing, rate limits, context window, regions or benchmark results for K-EXAONE-2.5. It also does not say whether the model’s weights are open. Teams that need a Korean-language model should wait for LG AI Research’s own documentation before planning around it.

Secure agents: Hyundai, AIM Intelligence and OpenShell

The agent-security post opens with NVIDIA’s framing: as AI agents take on more work, they need capable open models and security that keeps pace with them. Hyundai Motor Group is exploring NVIDIA NemoClaw blueprints and the NVIDIA OpenShell runtime. “Exploring” is NVIDIA’s word, and it signals evaluation, not deployment.

NVIDIA calls OpenShell a key part of its Open Agent Safety Platform. It describes OpenShell as a secure runtime with sandboxed execution and policy-based controls that protect sensitive data, manage access to enterprise tools and govern the actions agents take on their own. In practice, the agent runs in an isolated environment, and rules decide which tools and data it can touch.

AIM Intelligence, which NVIDIA describes as a Korean AI security company, is integrating its Starfort agent control platform with OpenShell. The goal is to combine OpenShell’s runtime access controls with Starfort’s detection of malicious instructions and data leakage. “Malicious instructions” covers attacks such as prompt injection, where hostile text in a document or web page tries to redirect an agent. NVIDIA says the integration targets secure agents for employee productivity and enterprise workflows. AIM Intelligence is also considering a preliminary proof of concept for the public energy sector, with possible uses in enterprise, industrial and robotics settings. That proof of concept is described as under consideration, not under way.

Physical AI: Samsung, LG Electronics and Kakao Mobility

Samsung’s simulation speedups

Samsung Electronics is expanding its collaboration with NVIDIA across materials research, process development and device engineering. In technology computer-aided design, the software used to model how chips are made and how they behave, Samsung is using NVIDIA cuISS and cuDSS to speed up calculations involving sparse linear systems. The work also covers GPU acceleration for process and optical simulations.

NVIDIA reports two specific results:

• Samsung’s sMOL materials platform has run selected quantum chemical calculations up to 5x faster end to end with NVIDIA cuEST.
• In one “AI homogenization” case for chip-scale thermal and mechanical analysis built with PhysicsNeMo, NVIDIA reports a speedup of roughly 1,600x. Simulation time fell from about 18 hours for a fine-grid solve on a 32-core CPU to 41 seconds for a GPU-accelerated coarse-grid solve.

The 1,600x figure needs careful reading. NVIDIA attributes it to three changes at once: AI-predicted effective material properties, coarser grids and GPU acceleration. It is not a like-for-like comparison of the same calculation on a CPU and a GPU, and the post does not say how close the coarse-grid answer comes to the fine-grid one. Samsung is also exploring NVIDIA cuEquivariance to speed up tensor operations in training and inference for atomistic AI models. NVIDIA adds that its Agent Toolkit “could” connect these simulation tools and AI models so that future engineering agents run analyses and compare design options. That is a stated possibility, not a product.

LG Electronics’ CLOiD robot

NVIDIA and LG Electronics demonstrated Isaac GR00T-powered robotic manipulation and manufacturing tasks using LG’s CLOiD robot in a real manufacturing environment. The teams built what NVIDIA calls a physical AI data flywheel: real-world data collection, data augmentation based on NVIDIA Cosmos, GR00T model training, onboard computing on Jetson AGX Thor and simulation-based evaluation. NVIDIA says the teams are using this work to refine a repeatable process for training, evaluating and improving manufacturing robots. The post reports no task success rates or throughput figures.

Kakao Mobility’s level 4 plans

Kakao Mobility plans to build on NVIDIA’s vehicle computing and safety technology to develop a level 4 autonomous driving system, meaning a vehicle that drives itself without a human ready to take over, within a defined operating area. The work combines Kakao Mobility’s own driving software and its experience operating in Seoul’s Gangnam district with NVIDIA Hyperion, a reference platform for vehicle sensors and compute powered by DRIVE AGX Thor. NVIDIA Halos supports safety during development and validation. Initial development is planned around Kia purpose-built vehicles and aims to support future level 4 mobility services in Korea. The post gives no timeline.

How this fits NVIDIA’s earlier Korea commitments

These posts follow a series of earlier Korea announcements. A LinkedIn post by Erudite Asia from November 2025 summarised NVIDIA’s partnership with the South Korean government and companies including Samsung Electronics, SK Hynix and Hyundai Motor. According to that post, NVIDIA would supply around 260,000 GPUs. It said 50,000 would support a government initiative to build a national AI cloud computing centre, Samsung and SK Hynix would each receive 50,000, and Hyundai would use 50,000 Blackwell GPUs for self-driving cars, smart factories and robotics. A second item in the same feed said the first deployment phase would include about 13,000 Blackwell GPUs, and that NVIDIA planned a joint research centre and AI training programmes for Korea’s workforce. These figures come from a social media summary of news coverage, not from NVIDIA directly.

In July 2026, Vladimir Troy wrote on LinkedIn that South Korean President Lee Jae Myung joined NVIDIA CEO Jensen Huang and other technology, industry and research leaders at an AI summit in San Francisco. The discussion covered Korea’s AI plans across infrastructure, physical AI and research. Earlier, at NVIDIA’s GTC 2025 conference, Naver Cloud said it would work with NVIDIA on large language models, infrastructure and applications for Southeast Asian markets, as part of a sovereign AI push built around its HyperCLOVA X models, CHOSUNBIZ reported. “Sovereign AI” here means a country or company building AI capability on infrastructure, data and models it controls.

The November posts shift from those hardware and infrastructure commitments to named software tools and specific workloads. The “From Chips to Skills” title suggests the sovereign AI post also covers training and skills, but its contents could not be checked for this report.

What it means for builders, and the open questions

For teams building on model APIs, three points stand out.

• A new Korean-language API option. K-EXAONE-2.5 API access from LG AI Research is the most direct change for API users. Terms, pricing and capabilities are not stated in NVIDIA’s post.
• An open Korean dataset. Nemotron-Personas-Korea is described as open, so teams fine-tuning or evaluating models for Korean users may be able to use its 7 million synthetic personas. NVIDIA’s post does not state the licence.
• Agent sandboxing is becoming a product category. OpenShell’s sandboxing and policy controls, and AIM Intelligence’s added detection of malicious instructions and data leakage, address problems any team running tool-using agents faces.

The limits are significant. Every figure here comes from NVIDIA’s own blog and has not been independently tested. Several items are plans or explorations, not deployments: Kakao Mobility’s level 4 system, Hyundai’s OpenShell evaluation and AIM Intelligence’s energy-sector proof of concept. The Dynamo recipes post, which the series title says brings “Korea’s Leading Open Models to Production”, could not be read for this report. Which models it covers, which hardware it targets and what performance it claims remain unconfirmed. The public ai-dynamo/dynamo code repository does include recipe files that are actively maintained, as a recent commit changing their inputs shows. That commit says nothing about Korean models.

What to watch next

• The full text of NVIDIA’s “From Chips to Skills” post, to see who the sovereign AI collaboration involves and what it commits to.
• The Dynamo recipes post: which Korean open models it covers, and whether it includes reproducible serving configurations and throughput numbers.
• LG AI Research’s own documentation for K-EXAONE-2.5 API access, including pricing, limits and any published evaluations.
• Whether Hyundai Motor Group moves from exploring OpenShell to deploying it, and whether AIM Intelligence’s energy-sector proof of concept goes ahead.
• Independent results for Samsung’s simulation speedups and a timeline for Kakao Mobility’s level 4 vehicles.

For related NVIDIA coverage, see the report on NVIDIA’s Dory 4B model, which reuses its layers to trade compute for accuracy.