NVIDIA expands Korean sovereign AI work with Rebellions racks

By Rebecca Stone

NVIDIA said on November 9, 2026 that Rebellions’ Rebel R200 neural processing units will connect through NVLink Fusion, with the resulting rack systems expected to be ready for deployment in 2028.

The announcement expanded NVIDIA’s collaboration with Korea’s Ministry of Science and ICT, or MSIT, across “AI chips, data centers, models, skills and agent security.” SK Telecom also plans to deploy the systems in its AI data center, linking them with proprietary models and public-facing services under AI for All.

The agreement gives South Korea’s sovereign AI effort a production path that combines Rebellions NPUs with NVIDIA’s GPU, networking and software stack. It also creates a concrete deployment test for SK Telecom, while MSIT’s backing opens broader work between NVIDIA and Korean NPU companies.

What does the expanded collaboration cover?

The joint announcement divides the program into five areas: AI chips, data centers, models, skills and agent security. Its practical center is a mixed-processor rack built around Rebellions silicon and NVIDIA infrastructure.

• AI chips and data centers: Rebellions will connect Rebel R200 NPUs through NVLink Fusion and use NVIDIA’s MGX architecture for rack-scale development.
• Models and serving: NVIDIA will use Dynamo with three selected consortia and four Korean foundation-model builders under AI for All.
• Skills: The program adds NVIDIA training, certification, hackathons and technical seminars focused on GPU utilization.
• Agent security: Hyundai Motor Group and AIM Intelligence will explore NemoClaw, OpenShell and Starfort in separate workstreams.

How will the GPUs and NPUs split inference?

Under the plan, Rebellions will connect its next-generation Rebel R200 NPUs to NVIDIA’s platform through NVLink Fusion. The company also plans to use the MGX architecture to develop rack-scale systems that SK Telecom expects to deploy in its AI data center.

Rebellions said the R200 is “designed for low-latency, energy-efficient inference on frontier models,” including models with multiple trillions of parameters. The announcement places the chip at the inference stage for some of the largest models in the collaboration.

NVIDIA says NVLink Fusion lets chipmakers combine specialized silicon with its high-bandwidth, low-latency scale-up networking and established infrastructure ecosystem. In the announced configuration, Rebellions NPUs join NVIDIA GPUs in a partitioned system rather than handling every stage of model serving.

The companies plan to use “disaggregated inference,” assigning different stages of AI response generation to the processor architecture best suited to each task. NVIDIA GPUs would handle prefill, while Rebellions NPUs would handle decode. The stated goal is higher inference throughput and better power efficiency.

MGX supplies the mechanical and operational framework for that mix. According to the joint announcement, the architecture gives GPU and NPU systems a common rack footprint and shared power-delivery and cooling elements. The design also supports data-center planning alongside Rebellions’ processor and system development.

The announcement adds that the MGX ecosystem gives Rebellions access to original design manufacturers and component suppliers. It also identifies the shared MGX design as a way to coordinate data-center planning with the remaining processor and system work.

Which Korean models will NVIDIA optimize?

Through AI for All, NVIDIA will work with three selected consortia and four Korean foundation-model builders. The company will use Dynamo, its open-source distributed inference-serving framework, to optimize and scale inference for multi-node data-center deployments.

Dynamo coordinates request routing and communication when a model grows beyond one GPU and must be split across several GPUs. NVIDIA’s deployment recipes apply that framework to four named Korean models:

• SK Telecom: A.X K2.
• LG AI Research: K-EXAONE 2.0.
• Upstage: Solar Open 2.
• Motif Technologies: Motif-3.

NVIDIA reports that the recipes, optimized for B200 GPUs, deliver up to 2x out-of-the-box performance. The company says the recipes are intended to help developers move the Korean open models into production and support broadly accessible AI for All services.

The model work is ahead of the hardware schedule. The joint announcement says model-serving recipes are being released for Korean developers now, while the R200-based racks still require processor integration and system development before SK Telecom’s planned deployment.

What does the training program add?

The workforce component extends the collaboration beyond hardware and software. NVIDIA will provide training and certification programs, hackathons and technical seminars intended to strengthen GPU utilization and the operational skills needed to run GPU-based infrastructure.

The announcement presents those programs as an operational extension of the model and rack work. It places workforce preparation alongside the disaggregated inference design, where NVIDIA GPUs perform prefill and Rebellions NPUs perform decode.

What does the agent-security work cover?

Hyundai Motor Group is exploring NemoClaw blueprints and the OpenShell runtime. The company’s work forms a separate track from the Rebellions rack project, but it brings agent tooling into the same expanded Korean collaboration.

The joint announcement says OpenShell forms part of the NVIDIA Open Agent Safety Platform. It provides sandboxed execution and policy-based controls for sensitive data, enterprise-tool access and autonomous agent actions.

AIM Intelligence is integrating its Starfort agent-control platform with OpenShell for employee-productivity and enterprise-workflow agents. The planned integration combines OpenShell’s runtime access controls with Starfort’s detection of malicious instructions and data leakage.

AIM Intelligence is also considering a preliminary proof of concept for the public energy sector. The announcement describes that effort as preliminary and attaches no deployment date.

What has to happen before 2028?

The announcement identifies a split between software and hardware readiness. Model-serving recipes can move directly into developer testing, while the mixed GPU-NPU racks require further processor integration, system development and a planned installation at SK Telecom’s AI data center.

MSIT’s support begins what the announcement calls broader work between NVIDIA and Korean NPU companies. The immediate test is a staged architecture in which Rebellions NPUs decode requests after NVIDIA GPUs perform prefill, with MGX providing the shared rack and cooling framework.

For SK Telecom, the systems must move from shared MGX specifications to working racks tied to its proprietary models and public-facing services. For model developers, Dynamo and the B200 recipes provide an earlier production route through AI for All. The next checkpoint is the R200 integration itself: without it, the 2028 date remains a deployment target rather than an operating system.