NVIDIA, Rebellions, and SK Telecom announced a plan to build sovereign AI inference infrastructure in South Korea around Rebellions’ next-generation R200 neural processing unit, connected inside the rack with NVIDIA NVLink Fusion and built on the NVIDIA MGX rack architecture. SK Telecom will deploy the systems in its AI data center alongside its models and services. The three companies expect R200 systems to be ready for deployment in 2028.
NVLink Fusion Connects R200 NPUs
Rebellions is designing the R200 for low-latency, energy-efficient inference on frontier models, including multitrillion-parameter mixture-of-experts models. NVLink Fusion is NVIDIA’s program for bringing custom silicon onto its scale-up fabric: a chipmaker integrates NVLink into its processor and gains access to NVIDIA’s NVLink switching, its high-bandwidth, low-latency networking, and the surrounding infrastructure ecosystem. Rebellions plans to use it to link R200 NPUs within a rack over NVLink scale-up networking and to tie those processors into the NVIDIA platform as part of a rack-scale system.
The companies have not disclosed R200 specifications, the NVLink bandwidth per NPU, or any performance results for the planned configuration. Rebellions lists the Rebel100 as its current silicon, and the R200 announcement says nothing about memory, process node, or power, so the step between the two generations is unknown.
MGX Rack Architecture Supports System Development
NVIDIAs MGX rack architecture and reference designs give Rebellions a starting point for the system, along with access to the original design manufacturers and component suppliers already building MGX hardware. MGX defines the shared pieces of a rack, including footprint, power delivery, and cooling, so GPU systems and NPU systems can occupy the same data center design. SK Telecom plans to deploy the R200 systems in its AI data center on the standardized MGX rack, which lets it plan power and cooling before the processor is final.
GPUs for Prefill, NPUs for Decode
The most interesting technical detail is the inference split the three companies plan to develop. Inference has two phases with different hardware demands: prefill, which processes the prompt and is typically compute-bound, and decode, which generates tokens one at a time and is typically bound by memory bandwidth and latency. The proposed configuration runs prefill on NVIDIA GPUs and decode on Rebellions NPUs, with the two connected over the NVLink fabric.
The stated goals are higher inference throughput and better power efficiency. The benefit of handing decode to an NPU remains a design intent until systems exist, but the logic makes sense on paper. Disaggregated prefill and decode is already common practice across GPU pools, including in NVIDIA’s Dynamo serving framework; splitting the two phases across different processor architectures over NVLink is the new part.
Rebellions cofounder and CEO Sunghyun Park said the collaboration combines the company’s AI silicon and system design with NVIDIA’s interconnect and rack-scale capabilities and SK Telecom’s models and data center operations, with the aim of low-latency, energy-efficient inference for frontier models at scale.
SK Telecom Connects Models and Services
SK Telecom brings the models and the customers. It plans to run its proprietary foundation models, including A.X K2, on the infrastructure, along with public-facing services developed through the Korean government’s AI for All project. SK Telecom’s data center operations, models, and services round out what the companies describe as a full-stack sovereign AI strategy, and the work builds on the broader SK and NVIDIA partnership floated at the Korea AI Summit in July.
Korea’s Ministry of Science and ICT is backing the collaboration between NVIDIA and Korean NPU companies, starting with Rebellions and SK Telecom, and described wider cooperation covering physical AI and GPU supply. The aim is AI services tuned to Korea’s language, industries, and priorities across the public and private sectors.
Verfügbarkeit
Rebellions R200 systems are expected to be ready for deployment in 2028. There is no sampling schedule, early-access program, general availability date, or pricing.
Amazon