NVIDIA, South Korean AI chipmaker Rebellions and telecommunications operator SK Telecom are collaborating to develop sovereign AI infrastructure that combines NVIDIA GPUs with Korean-designed neural processing units (NPUs). The partnership will use NVIDIA NVLink Fusion and MGX rack architecture to build integrated AI computing systems, with a focus on improving inference performance, energy efficiency and domestic AI capabilities.
South Korea is moving to strengthen its domestic artificial intelligence infrastructure through a collaboration between NVIDIA, Rebellions and SK Telecom. The companies plan to integrate Korean-designed AI processors with NVIDIA’s computing and networking technologies, creating a heterogeneous infrastructure designed to support large-scale AI inference and locally developed AI services.
The collaboration brings together three distinct capabilities: NVIDIA’s AI infrastructure and high-speed interconnect technology, Rebellions’ custom AI silicon, and SK Telecom’s data center operations, foundation models and consumer-facing services.
At the center of the agreement is NVIDIA NVLink Fusion, which will connect Rebellions’ next-generation Rebel R200 neural processing units to NVIDIA’s infrastructure platform. The companies also plan to use NVIDIA MGX rack architecture to develop integrated GPU-NPU systems for deployment in SK Telecom’s AI data centers.
Supported by South Korea’s Ministry of Science and ICT (MSIT), the initiative reflects a broader effort to develop sovereign AI infrastructure that can support the country’s language, industrial requirements and public-sector priorities.
Integrating Korean AI Chips With NVIDIA Infrastructure
Rebellions is developing its R200 NPU for low-latency, energy-efficient inference workloads, including large mixture-of-experts (MoE) models. Through NVLink Fusion, the company intends to connect its processors to NVIDIA’s high-bandwidth, low-latency scale-up networking architecture.
This approach could give specialized AI chipmakers access to an established infrastructure ecosystem without requiring them to independently develop every component of a rack-scale computing environment.
NVIDIA MGX provides standardized rack architecture and reference designs covering infrastructure elements such as power delivery, cooling and physical rack configurations. Its ecosystem also connects developers with original design manufacturers and component suppliers.
For Rebellions, this integration offers a route to develop systems around its own silicon while leveraging NVIDIA’s broader infrastructure technologies.
The arrangement is strategically significant because AI infrastructure is becoming increasingly heterogeneous. Rather than relying exclusively on GPUs for every workload, data center operators are exploring combinations of specialized processors designed for different stages of AI computation.
Disaggregated Inference Targets Performance and Efficiency
One of the most technically significant elements of the collaboration is the planned development of disaggregated inference.
AI inference involves several computational stages. During the prefill stage, a model processes the initial input and prepares the context needed to generate a response. During decode, the model generates output tokens sequentially.
NVIDIA and Rebellions plan to distribute these workloads across different processor architectures. NVIDIA GPUs would handle prefill, while Rebellions NPUs would perform decode.
The objective is to assign each stage to the hardware best suited to its computational requirements, potentially improving inference throughput and power efficiency.
This architecture could become relevant as enterprises deploy increasingly large foundation models and AI agents that require continuous inference capacity.
Mixture-of-experts models add another dimension to the infrastructure challenge because their architecture activates selected expert networks rather than every model component for each input. Supporting these workloads efficiently requires coordinated computing, memory and networking resources.
However, the expected performance and efficiency improvements will need to be established through system-level testing and real-world deployments. The announcement does not provide comparative benchmarks, throughput figures or quantified energy savings.
SK Telecom Connects Infrastructure With AI Services
SK Telecom will provide the operational environment for the integrated systems, drawing on its experience building and managing AI data centers.
The telecommunications company also plans to connect the infrastructure with its proprietary foundation models, including A.X K2, and public-facing AI services developed under South Korea’s AI for All initiative.
This gives the collaboration a broader scope than semiconductor integration alone. Its objective is to connect domestic AI processors, data center infrastructure, locally developed models and end-user services within a single ecosystem.
For SK Telecom, the project supports its full-stack sovereign AI strategy, extending its role from telecommunications and data center operations into AI infrastructure and services.
For South Korea, the partnership could provide an additional development pathway for domestic semiconductor companies seeking to participate in the global AI computing market.
Sovereign AI Meets an Open Infrastructure Ecosystem
The agreement also highlights an important strategic distinction in the evolving AI infrastructure market.
Sovereign AI does not necessarily require every component of an AI computing system to be developed domestically. Instead, countries can combine locally developed processors, models, datasets and operational capabilities with international technology platforms.
NVIDIA’s NVLink Fusion is relevant to this model because it enables custom processors to connect with NVIDIA’s broader infrastructure ecosystem.
Major technology companies, including Microsoft, Google and Amazon, are expanding their AI cloud and computing infrastructure capabilities. NVIDIA has established a significant position through its accelerated computing ecosystem, while specialized semiconductor developers are exploring opportunities to provide alternative processors for selected workloads.
Rebellions’ participation demonstrates how domestic AI chipmakers can seek to integrate their technology into established infrastructure platforms rather than compete across the entire computing stack.
South Korea’s Ministry of Science and ICT is supporting the collaboration as part of efforts to strengthen national AI computing capabilities. The government has also indicated interest in extending cooperation with NVIDIA across areas such as physical AI and GPU supply.
Development Timeline and Industry Implications
The collaboration is structured around future system development and deployment rather than an immediately completed infrastructure rollout. SK Telecom plans to deploy and validate the integrated systems in its AI data center, while Rebellions continues developing the R200 platform. NVIDIA Korea’s announcement indicates that R200 systems are expected to be ready for deployment in 2028.
The longer-term opportunity lies in establishing a repeatable architecture through which domestic AI processors can operate alongside globally established computing infrastructure.
For enterprise AI, this could provide additional options for inference optimization, processor selection and data center planning. For semiconductor developers, it creates opportunities to access established system ecosystems while retaining differentiated silicon designs.
The partnership also illustrates how the next phase of AI infrastructure competition may increasingly involve heterogeneous computing rather than a single processor architecture.
Ultimately, its success will depend on whether the companies can demonstrate reliable interoperability, commercially viable inference performance and meaningful energy-efficiency improvements at production scale.
Market Landscape
The AI infrastructure market is moving beyond GPU-only architectures toward heterogeneous computing environments combining GPUs, NPUs, custom ASICs and specialized accelerators.
NVIDIA’s NVLink Fusion strategy creates an opportunity for third-party chipmakers to integrate custom silicon into its broader infrastructure ecosystem. This could expand the role of specialized processors while maintaining compatibility with established rack-scale computing technologies.
South Korea’s initiative also reflects the growing importance of sovereign AI infrastructure. Governments and telecommunications operators are seeking greater control over domestic computing capacity, AI models, semiconductor development and locally relevant services.
The competitive landscape includes NVIDIA, AMD, Intel, specialized AI chip developers and hyperscale cloud providers. A key differentiator will be the ability to deliver efficient inference infrastructure that supports increasingly complex foundation models and autonomous AI workloads.
Top Insights
- Heterogeneous AI computing: NVIDIA and Rebellions plan to combine GPUs and Korean-designed NPUs to distribute inference workloads across specialized processors.
- Inference optimization: The planned prefill-and-decode separation could improve throughput and power efficiency, subject to real-world system validation.
- Sovereign AI infrastructure: SK Telecom will connect domestic AI processors with its data centers, proprietary models and public-facing AI services.
- Custom silicon integration: NVLink Fusion provides Rebellions with a pathway to connect its R200 processors to NVIDIA’s rack-scale infrastructure ecosystem.
- Long-term deployment: R200 systems are expected to be ready for deployment in 2028, making system validation and commercialization important milestones.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
