Nvidia used its AI Day Seoul event on November 9, 2026 to announce a set of Korea deals, and the biggest one is about chips that are not Nvidia’s. Rebellions, a Korean chip startup, will connect its next-generation R200 inference processors to Nvidia’s rack platform through NVLink Fusion, and SK Telecom plans to run those systems in its AI data center. Nvidia posted four blog announcements at 05:00 UTC and a joint press release the same morning, Seoul time.
The event opened at COEX in Seoul with more than 2,000 developers, startups and researchers, by Nvidia’s count. It also landed a day after Nvidia’s board opened an independent investigation into reports about Jensen Huang, which shaped the coverage in Korea. Below: what was announced, what it changes for builders, and what was not announced.
On this page · 8 sectionsOpen
Nvidia announced a set of Korea deals at AI Day Seoul on November 9, 2026, at COEX, with more than 2,000 attendees by its count.
Rebellions will connect its R200 NPUs to Nvidia racks through NVLink Fusion, SK Telecom will deploy them, and the systems are expected in 2028.
The plan splits inference: Nvidia GPUs read the prompt (prefill) and Rebellions NPUs write the answer (decode).
Nvidia released Dynamo recipes for four Korean open models on B200 GPUs and claims up to 2x performance; Hyundai is exploring OpenShell.
No new GPU order or dollar figure was announced; Korea’s ministry expects contracted Nvidia GPU deliveries to continue.
§ 01What Nvidia announced in Seoul
The announcements fall into five groups. None of them includes a new GPU order or a dollar figure.
| Area | Who | What was announced |
|---|---|---|
| Sovereign AI | Korea’s Ministry of Science and ICT | Wider collaboration on chips, data centers, training, startups, models and security |
| Korean chips | Rebellions and SK Telecom | R200 NPUs joined to Nvidia racks through NVLink Fusion and MGX |
| Open models | SK Telecom, LG AI Research, Upstage, Motif | Dynamo deployment recipes for four Korean open models |
| Physical AI | Samsung, Kakao Mobility, LG Electronics | Chip simulation, level 4 driving work, factory robot training |
| Secure agents | Hyundai Motor Group, AIM Intelligence, LG AI Research | OpenShell and NemoClaw work, K-EXAONE-2.5 API access |
§ 02Korean chips inside Nvidia’s racks
NVLink Fusion is Nvidia’s way of letting other companies’ processors sit in its rack-scale systems and talk to its GPUs over its high-speed links. Rebellions will connect its R200 neural processing units this way and build on Nvidia’s MGX rack design, which gives it a ready supply chain of manufacturers and parts.
The plan splits the work of answering a prompt between two kinds of chips. Nvidia GPUs handle prefill, reading the prompt; Rebellions NPUs handle decode, writing the answer one token at a time. The companies say this should improve throughput and power efficiency. SK Telecom plans to deploy the systems with its own models, including A.X K2, and its public AI services.
“Korea has the talent, technology and industrial ambition to become a world leader in AI,” Jensen Huang said in the release. Korea’s science minister, Bae Kyung-hoon, said the government will support the work across areas including physical AI and “securing GPU supply.”
The timing is the catch: the release says Rebellions’ R200 systems are expected to be ready for deployment in 2028. This is a plan for the next rack generation, not hardware anyone can rent now.
§ 03Open models and Dynamo recipes
Nvidia is releasing Dynamo deployment recipes for four Korean open models: A.X K2 from SK Telecom, K-EXAONE 2.0 from LG AI Research, Solar Open 2 from Upstage and Motif-3 from Motif Technologies. Dynamo is Nvidia’s open source inference server for models too large for one GPU. The recipes are tuned for B200 GPUs, and Nvidia says they deliver “up to 2x performance out of the box.”
Separately, through Korea’s AI for All program, Nvidia will work with three selected consortia and four Korean foundation model builders on scaling inference with Dynamo. LG AI Research is opening API access to K-EXAONE-2.5, trained in part on Nemotron-Personas-Korea, an open dataset of 7 million synthetic Korean personas.
§ 04Physical AI: chips, cars and robots
The physical AI post names three projects, each with a number or a concrete step.
| Company | Work | Reported result or plan |
|---|---|---|
| Samsung Electronics | Quantum chemistry on its sMOL platform | Up to 5x faster end to end |
| Samsung Electronics | Chip thermal simulation with PhysicsNeMo | About 18 hours on a 32-core CPU down to 41 seconds |
| Kakao Mobility | Level 4 autonomous driving | Hyperion, DRIVE AGX Thor and Halos, starting on Kia vehicles |
| LG Electronics | CLOiD robot on a real factory floor | GR00T-trained manipulation, Jetson AGX Thor on board |
Nvidia attributes the Samsung result, roughly 1,600x, to three things together: AI-predicted material properties, a coarser grid and GPU acceleration. It is one case, not a general speedup for chip simulation.
§ 05Secure agents
Hyundai Motor Group is exploring Nvidia’s NemoClaw blueprints and the OpenShell runtime, a sandbox with policy controls for what an agent can touch. AIM Intelligence, a Korean AI security company, is connecting its Starfort platform to OpenShell to add detection of malicious instructions and data leakage. Our earlier look at OpenShell is on the Nvidia OpenShell reference architecture page.
§ 06What was not announced
No new GPU order, dollar commitment or delivery date came out of the event. Reuters reported that Korea’s deputy science minister expects no disruption to contracted Nvidia GPU deliveries despite the company’s current issues.
The Huang investigation was visible at the edges. Madison Huang, Jensen Huang’s daughter and an Nvidia executive, had planned to visit Rebellions and hold a session with Kakao Mobility; Rebellions said those events were cancelled due to “unavoidable internal circumstances at Nvidia.” The board’s investigation is covered in our Nvidia board investigation post.
§ 07What we are watching
- Rebellions R200 milestones before the 2028 deployment date.
- Other Korean NPU makers joining NVLink Fusion, which the ministry says it will support.
- A GPU number. Any new order or delivery figure from the ministry or Nvidia.
§ 08Sources
- Nvidia, AI Day Seoul live updates
- Nvidia, From Chips to Skills, NVIDIA and Korea Expand Sovereign AI Collaboration, November 9, 2026
- Nvidia, Korea Advances Physical AI Across Industries With NVIDIA, November 9, 2026
- Nvidia, NVIDIA Dynamo Recipes Bring Korea’s Leading Open Models to Production, November 9, 2026
- Nvidia, Korea’s AI Ecosystem Advances and Secures Agentic AI With NVIDIA, November 9, 2026
- NVIDIA, Rebellions and SK Telecom press release, via GlobeNewswire on Yahoo Finance, November 9, 2026
- Reuters, South Korea says Nvidia issues won’t affect contracted GPU supplies, November 9, 2026
Q1What is NVLink Fusion?
Nvidia’s way of letting other companies’ processors join its rack-scale systems and connect to its GPUs over NVLink.
Q2Which Korean models got Dynamo recipes?
A.X K2 from SK Telecom, K-EXAONE 2.0 from LG AI Research, Solar Open 2 from Upstage and Motif-3 from Motif Technologies.
Q3What was Samsung's 1,600x result?
A chip thermal simulation that took about 18 hours on a 32-core CPU ran in 41 seconds with AI-predicted properties, a coarser grid and GPUs. It is one case.
Q4Did Jensen Huang attend?
The release carries his quote. His daughter Madison Huang cancelled planned visits to Rebellions and Kakao Mobility.
Q5Is this tied to the Huang investigation?
No. The announcements are commercial. The investigation only affected the visit schedule, per Rebellions and Reuters.