On November 9, 2026, at NVIDIA AI Day Seoul, NVIDIA, SK Telecom, and chipmaker Rebellions announced a strategic sovereign AI infrastructure partnership backed by South Korea’s Ministry of Science and ICT. The collaboration combines Rebellions’ Rebel R200 NPUs with NVIDIA’s NVLink Fusion and MGX rack architectures for disaggregated GPU-NPU inference, providing research institutions and university computing facilities a blueprint for power-efficient, cost-effective academic frontier model deployment.
Three-Way Agreement Signed at NVIDIA AI Day Seoul
On November 9, 2026, at the NVIDIA AI Day Seoul conference at the COEX convention centre in Gangnam District, Seoul, leadership from NVIDIA, SK Telecom (SKT), and South Korean AI semiconductor startup Rebellions signed a strategic memorandum of understanding to build sovereign artificial intelligence infrastructure. Attendees included Rebellions Chief Executive Officer Sunghyun Park, SK Telecom Chief Technology Officer Chung Suk-geun, Ashish Karandikar, Vice President of NVIDIA’s NVLink Fusion division, and SK Telecom President and CEO Jung Jai-hun. The initiative is backed by South Korea’s Ministry of Science and ICT (MSIT) under Deputy Prime Minister and Minister Bae Kyung-hoon.
Under the agreement, Rebellions will connect its next-generation Rebel R200 neural processing units (NPUs) to NVIDIA’s computing platform via the NVIDIA NVLink Fusion interconnect. Rebellions will develop rack-scale inference systems based on the NVIDIA MGX modular server architecture, while SK Telecom will deploy and validate the hardware across its domestic AI data centres. SK Telecom plans to pair the heterogeneous systems with its proprietary foundation models, including A.X K2, and public-facing services under the government’s ‘AI for All’ project. The partners target commercial deployment of the integrated Rebel R200 systems in 2028.
Technical Specifications and Disaggregated Inference Architecture
The partnership targets the operational divergence between prompt processing and token generation in frontier Mixture-of-Experts (MoE) models. Large language model inference divides into two computational phases: the compute-bound prefill stage and the memory-bandwidth-bound decode stage. During prefill, the system evaluates input prompts in parallel using dense General Matrix Multiply (GEMM) operations, fully utilizing GPU tensor cores. During decode, the model generates output tokens sequentially. Each step requires retrieving past Key-Value (KV) cache tensors from memory, leaving high-throughput GPU floating-point units underutilized while waiting for memory transfers.
The consortium’s architecture implements heterogeneous disaggregated inference. NVIDIA GPUs execute the compute-heavy prefill stage, while Rebellions’ Rebel R200 NPUs run the decode stage. Connecting these processors is NVIDIA NVLink Fusion, a scale-up interconnect allowing custom third-party ASICs and XPUs to communicate directly with NVIDIA GPUs and CPUs at NVLink bandwidth within a server rack, eliminating PCIe and external network latency bottlenecks during KV-cache handoffs.
Physical deployment relies on the NVIDIA MGX modular server architecture. MGX provides standardized rack footprints, liquid cooling paths, and power distribution compatible with Open Compute Project (OCP) standards, giving Rebellions access to established original design manufacturer (ODM) supply chains.
| Inference Dimension | Prefill Stage (NVIDIA GPUs) | Decode Stage (Rebellions Rebel R200 NPUs) |
|---|---|---|
| Primary Bottleneck | Compute-bound (dense matrix multiplications) | Memory-bandwidth-bound (sequential KV-cache fetches) |
| Workload Characteristics | High arithmetic intensity, parallel tensor-core execution | Low arithmetic intensity, token latency sensitivity |
| Hardware Profile | NVIDIA Hopper/Blackwell MGX Server GPUs | Rebellions Rebel R200 specialized inference NPUs |
| Interconnect Layer | NVIDIA NVLink intra-node cluster fabric | NVIDIA NVLink Fusion scale-up heterogeneous bus |
| Operational Benefit | Minimized time-to-first-token (TTFT) across long contexts | Higher token throughput per watt and lower operating TCO |
Context: Sovereign AI Ambitions and Rebellions’ Silicon Evolution
This tripartite alliance follows SK Telecom’s June 2026 agreement with NVIDIA to construct gigawatt-scale AI factories in South Korea using NVIDIA’s DSX modular architecture, with an initial facility planned for 2027. To mitigate dependency on foreign hardware, South Korea’s MSIT has funded domestic semiconductor development to cultivate sovereign infrastructure capable of processing national data domestically.
Founded in 2020, Rebellions previously developed the Atom NPU and its flagship Rebel R100, which featured a chiplet architecture using fifth-generation high-bandwidth memory (HBM3E). The upcoming Rebel R200 expands on this foundation for multi-trillion-parameter MoE architectures. Backed by Arm, Samsung Electronics, SK Hynix, SK Telecom, KT, Aramco’s Wa’ed Ventures, and Mirae Asset Group, Rebellions avoids the isolation of proprietary host buses by linking into NVLink Fusion, gaining direct entry into mainstream data centre racks.
Executive and Ministerial Reaction
Consortium leaders and government representatives emphasized the strategic significance of the agreement:
- Jensen Huang, founder and CEO of NVIDIA: “Sovereign AI gives nations the power to turn their language, culture and expertise into intelligence. With NVLink Fusion, Korea’s chipmakers can connect their custom ASICs to NVIDIA’s AI factory platform, tightly coupling their processors with NVIDIA compute to run AI workloads.”
- Sunghyun Park, cofounder and CEO of Rebellions: “Today’s announcement marks a turning point, opening a new chapter for sovereign AI by connecting Korea’s AI silicon and models with global systems technology. By integrating Rebellions’ purpose-built silicon and system design with NVIDIA’s NVLink Fusion… we will deliver low-latency and energy-efficient inference for frontier models at scale.”
- Jung Jai-hun, President and CEO of SK Telecom: “By linking SKT’s AI data centers, proprietary AI models including A.X K2, and public-facing AI services developed under the ‘AI for All’ project, we will strengthen our full-stack sovereign AI competitiveness and contribute to the growth of Korea’s AI ecosystem.”
- Bae Kyung-hoon, South Korea Deputy Prime Minister and Minister of Science and ICT: “This collaboration with NVIDIA marks a significant step forward, as it enables Korean-developed NPUs to enter the computing infrastructure market alongside NVIDIA, a global leader in AI computing, through NVLink Fusion. The government will actively support this collaboration to deliver tangible results.”
Implications for University Computing Centres and AI Research Facilities
The disaggregated GPU-NPU model directly impacts academic high-performance computing (HPC) centres, national laboratories, and university research computing divisions. Academic facilities face strict power and cooling ceilings. Running monolithic GPU clusters for conversational models wastes energy on memory-bound decoding loops. Offloading token generation to energy-optimized NPUs advances institutional goals for sustainable HPC by reducing power consumption per token while maintaining low inference latency.
For AI systems researchers, NVLink Fusion offers an accessible environment for investigating heterogeneous memory hierarchies, KV-cache transfers, and mixed-precision execution without proprietary interconnect barriers. As research funding agencies increasingly demand environmental audits, calculating a rigorous compute FLOPs estimate alongside token-per-watt throughput will be vital for grant compliance.
From an institutional governance standpoint, sovereign AI infrastructure allows universities to safeguard sensitive research assets, including clinical health records, genomic datasets, and proprietary patent disclosures. Serving models on domestic sovereign clouds complies with strict data protection mandates and aligns with broader national data infrastructure standards. In addition, researchers deploying split architectures must record firmware and runtime configurations carefully to guarantee a reproducible AI experiment across heterogeneous clusters.
Development Milestones and 2028 Rollout Roadmap
The consortium has outlined a four-stage roadmap toward production readiness:
- 2026–2027: Silicon Co-Design and Interconnect Validation. Rebellions and NVIDIA will finalize the Rebel R200 physical design and validate NVLink Fusion signaling inside prototype NVIDIA MGX chassis.
- 2027: Testbed Deployment at SK Telecom. SK Telecom will deploy prototype MGX systems within its domestic AI data centres to evaluate power efficiency, thermal stability, and performance on A.X K2 models.
- 2028: Commercial Sovereign AI Launch. Commercial deployment of Rebel R200 rack systems will begin in SK Telecom data centres, supporting public services under ‘AI for All’ and offering compute to universities and enterprises.
- Long-Term International Expansion. Following domestic validation, the consortium plans to market the MGX sovereign AI infrastructure package to global telecom operators and sovereign cloud providers.
Sources
- NVIDIA Newsroom: NVIDIA, Rebellions and SK Telecom Announce Collaboration to Advance Sovereign AI Infrastructure With NVIDIA NVLink Fusion (Published 2026-11-08 / 2026-11-09).
- Evertiq: Nvidia, Rebellions, SK Telecom to advance sovereign AI infrastructure (Published 2026-11-09).
- Unite.AI: Rebellions to Use NVIDIA NVLink Fusion for Korean Sovereign AI (Published 2026-11-09).
- The Elec: Rebellions, Nvidia to Build Sovereign AI Infrastructure in Korea (Published 2026-11-10).
- Maeil Business Newspaper (MK): SK Telecom, NVIDIA and Rebellions’ ‘NVLink Alliance’ (Published 2026-11-09).







