SEOUL, July 23, 2026 — Rebellions, a global leader in AI inference infrastructure, announced that it has successfully run SK Telecom’s sovereign AI foundation model ‘A.X K1’ (hereinafter K1) on its own server infrastructure, and demonstrated an agent service capable of handling concurrent requests from multiple users on top of it. The achievement — pairing a domestic large-scale model with a high-performance domestic NPU to efficiently process complex, large-scale computation — lays the groundwork for commercializing real-world AI services built on Korean technology.

The demonstration centers on maximizing “infrastructure efficiency” for large-scale AI operations through joint optimization between the two companies. Rebellions stably ran Korea’s largest large language model, with over 500 billion (500B) parameters, on RebelServer — its high-performance AI semiconductor product — delivering exceptional serving1 efficiency. The result: infrastructure capability on par with global high-performance GPU servers, achieved through a single-server configuration alone. Because large-scale model computation is carried out within a single server without waste, the burden of data center build-out and operation is substantially reduced. Applied at scale, this strength translates directly into service-level competitiveness.

This level of infrastructure efficiency stems from the structural strengths of K1 working in tandem with Rebellions’ software technology. K1 is built on an MoE2 architecture — widely adopted globally for its efficiency — which allows it to maintain efficiency even at the 500B-parameter scale. On top of supporting such MoE models, Rebellions optimized its dedicated software to fully bring out K1’s characteristics in real service environments. In particular, by implementing a distributed processing technique that partitions and parallelizes different types of operations across the chip in the most efficient way, Rebellions enabled a single server to handle a large number of concurrent user requests in real time without bottlenecks.

The demonstration went beyond simply running the model, extending into a “Map-based agent service” demo to prove real-world applicability. The core achievement was successfully handling one of the toughest technical hurdles for consumer-facing services: concurrent multi-user requests. When multiple users log in simultaneously and ask different everyday questions such as “which pharmacies are open this weekend,” A.X K1 processes each in real time without bottlenecks and instantly displays the results on a map. It is a representative example showing that a combination of purely domestic AI technologies can reliably support real-world public AI agent services used by many people at once.

“This demonstration is a tangible example of sovereign AI, showing that a domestic large-scale model and a domestic NPU alone can operate a large-scale AI service, and it is meaningful in that it validates both infrastructure efficiency and real-world viability at once,” said Sungwon Park, CEO of Rebellions. “Going forward, we will continue to work with SK Telecom to build a Korean AI infrastructure that lets everyone experience AI in their daily lives.”

Rebellions is a member of the SKT consortium in the Ministry of Science and ICT’s Sovereign AI Foundation Model program, where it is validating domestic NPU-based infrastructure and helping expand the reach of Korean AI models across industries and services. The company has also been stably operating high-traffic real-time AI services such as A. Phone Call Summary for approximately one year. Building on this experience, Rebellions plans to contribute as a domestic NPU infrastructure partner to building the foundation for AI that “anyone can access without barriers” — a principle emphasized by the government through its “Three Grand Mega-Projects for Korea’s Great Leap” and “AI for All” initiatives.

About Rebellions

Rebellions builds what the AI era needs: purpose-built inference accelerators delivering the performance, efficiency, and supply chain resilience that enterprises and nations demand. Its flagship Rebel100™ delivers the best performance per dollar per watt – built for inference from the ground up, not retrofitted from training – with full-stack software built entirely on open source and open standards. With proven commercial deployments already live across enterprises and governments, and a chiplet architecture built for the most demanding AI workloads, Rebellions gives every organization and nation the ability to own and control their AI, not just access it. Backed by Aramco, Arm, Kindred Ventures, KT, Mirae Asset Group, Samsung, SK Hynix, and SK Telecom, Rebellions is headquartered in South Korea with operations in the United States. Learn more by visiting www.rebellions.ai, on LinkedIn at www.linkedin.com/company/rebellions-ai, and on X at twitter.com/rebellions_inc.