The race for artificial intelligence is shifting from a battle of raw training power to a quest for efficient, localized execution. In a move to decentralize AI control and reduce the massive energy overhead of large-scale models, Rebellions has announced a strategic collaboration with SK Telecom (SKT) and Arm to build specialized AI inference infrastructure targeting sovereign AI and telecommunications-focused data centers.
This partnership aims to create a high-performance server ecosystem that allows nations and enterprises to own and operate their AI capabilities independently—a concept known as “sovereign AI.” By combining Arm’s first-ever designed data center CPU with Rebellions’ purpose-built AI chips, the trio is attempting to break the industry’s heavy reliance on general-purpose GPUs, which are often costly and energy-intensive to maintain at scale.
The collaboration focuses specifically on the “inference” stage of AI—the process where a trained model actually generates a response to a user query. While training requires immense bursts of power, inference happens billions of times a day across global networks. For telecommunications providers and governments, the ability to run these models on energy-efficient, air-cooled hardware within their own borders is becoming a matter of both economic and national security.
The technical heart of this initiative is the integration of the Arm AGI CPU, built on the Arm Neoverse CSS V3 architecture, with Rebellions’ RebelCard™ accelerator. This hardware pairing is designed to handle the complex orchestration of workloads across memory and networking while maximizing the throughput of AI calculations.
Engineering the ‘Sovereign’ Stack
For many governments, the primary hurdle to AI adoption is the “black box” nature of cloud-based AI, where data often leaves national borders to be processed by a handful of global providers. The Rebellions-Arm-SKT alliance is positioning its infrastructure as a way for the public sector and telcos to maintain a closed loop of data and compute.
The partnership is not merely a hardware assembly; it involves the co-development of the entire software stack, including firmware. This ensures that the hardware is optimized for the specific types of large-scale data processing required by telecom networks. A key part of the validation process will involve running SK Telecom’s proprietary foundation model, A.X K1, on these modern servers to test stability and performance in a live data center environment.
According to Jinwook Oh, CTO of Rebellions, the goal is to establish a “one-team” collaboration that serves as a precedent for AI-specialized infrastructure. By pairing the RebelCard with full-stack software, the company aims to provide a core pillar for next-generation data centers that prioritize power efficiency without sacrificing the ability to run ultra-large multimodal models.
The RebelCard and the Shift to NPU Architecture
At the center of the hardware strategy is the RebelCard™, which utilizes the “Rebel 100” semiconductor. Unlike traditional GPUs that were originally designed for graphics and later adapted for AI training, the Rebel 100 is a purpose-built inference accelerator. It employs a chiplet architecture, integrating four NPU (Neural Processing Unit) chiplets with 5th-generation High Bandwidth Memory (HBM3E).
This design is specifically tuned for Mixture of Experts (MoE) models—a type of AI architecture that only activates a fraction of its parameters for any given task, significantly reducing the computational load. By utilizing high-speed chip-to-chip communication, the RebelCard aims to match the performance of flagship GPUs while offering superior power efficiency and the ability to operate using standard air-cooling, which drastically lowers the cost of data center operations.
| Feature | Specification / Detail |
|---|---|
| Architecture | 4 NPU Chiplets |
| Memory | 5th-Gen HBM3E |
| Optimization | Multimodal & Mixture of Experts (MoE) |
| Cooling | Air-cooled operation support |
| Primary Use | AI Inference (Sovereign AI/Telecom) |
Strategic Implications for the Asian Market
While the technology has global applications, the partners are eyeing a strong foothold in Asia. The region is seeing a surge in demand for independent AI infrastructure as countries seek to develop AI that reflects their own languages, cultures, and regulatory requirements without relying on Silicon Valley-centric clouds.
For SK Telecom, the move is about vertical integration. By combining its own foundation model, A.X K1, with an optimized hardware layer, the company can offer a “full package” to other global telcos. This reduces the latency of AI services and ensures that the infrastructure is scaled specifically for the high-volume, real-time data processing that characterizes telecommunications.
Eddie Ramirez, vice president of travel-to-market for Arm’s Cloud AI Business Unit, noted that CPUs are critical for coordinating workloads across accelerators and networking. The Arm AGI CPU is designed specifically to meet the demands of these large-scale deployments, providing the necessary “glue” to make the RebelCard’s processing power accessible and scalable.
What Comes Next
The immediate next step for the alliance is the technical validation phase within SK Telecom’s live data center environments. Once the stability of the hardware-software integration is proven—specifically regarding the A.X K1 model’s performance—the group plans to move toward broader commercial deployment.
The long-term goal is to supply customized, stability-proven solutions to public sectors and global telecommunications companies that require autonomous AI infrastructure. This will likely involve a rollout of “sovereign-ready” server racks that can be deployed in regional data centers with minimal reconfiguration.
The success of this venture will depend on whether the Rebel 100 can maintain its performance edge as AI models continue to evolve in size and complexity. The partners are now moving toward a validation checkpoint to confirm that the air-cooled, NPU-driven approach can truly displace the GPU-heavy status quo in the inference market.
This article is intended for informational purposes only and does not constitute financial or investment advice.
We invite you to share your thoughts on the rise of sovereign AI and the shift toward NPU-based infrastructure in the comments below.
