[Edaily Reporter Shin Ha-yeon ] #Acryl Inc., a company specializing in Artificial Intelligence Transformation (AX) infrastructure, has unveiled its “Sovereign AX” strategy, which integrates NVIDIA graphics processing units (GPUs) and domestically developed neural processing units (NPUs) into a single resource.

Acryl Inc. announced on the 24th that it participated in the “2026 OCP Korea Tech Day,” held on the 21st at COEX in Samseong-dong, Seoul, where it introduced its Sovereign AX strategy designed to enhance investment efficiency and operational autonomy in AI data centers.

Sovereign AX is a strategy that connects the various layers required for implementing AI services—including infrastructure, models, and operations—via open interfaces to reduce dependence on specific vendors. It focuses on enabling companies and institutions to minimize their reliance on external APIs or individual suppliers and to operate AI infrastructure and services tailored to their own environments.

The core solution is “GPUBASE,” a GPU cluster optimization platform developed in-house by Acryl Inc. According to the company’s own proof-of-concept results, GPUBASE achieved an effective GPU utilization rate of 93% and a network path utilization rate of 98%. The company explains that applying it to a cluster of 100 NVIDIA H100 GPUs could save approximately 2.1 billion won annually. The solution is deployed as software on existing servers without requiring the purchase of additional GPUs or switches, nor any changes to training code.

In particular, it enables the management of not only GPUs from different generations—ranging from the NVIDIA A100 to the B200—but also various AI accelerators such as AMD, Google TPU, Rebellion ATOM, and FuriosaAI RNGD, all under a single scheduler, quota, and billing system. The company explains that this allows for the diversification of the supply chain without a complete system overhaul, as domestic NPUs can be added gradually while maintaining existing NVIDIA infrastructure.

Acryl Inc. also conducted technical demonstrations of the technology in both cloud and private network environments. Large-scale demonstrations were conducted using more than 500 GPUs each on Amazon Web Services (AWS) and Google Cloud Platform (GCP), and more than 200 GPUs on Microsoft Azure. In South Korea, the system was operated on the closed networks of EPC companies and in network-segregated environments of government and local authorities. The AI stack has also been operated in closed-network environments at 18 medical institutions—including 11 tertiary general hospitals—with a cumulative total of more than 10,000 hospital beds.

During a technical session held that day, Yeom Ik-jun, CTO of Acryl Inc., highlighted the issues of inter-node communication latency and network path congestion, which reduce GPU utilization rates. “Networks are not a performance issue but a cost issue,” said CTO Yeom. “Reducing communication latency allows more GPUs to operate on the same hardware, thereby lowering the budget required for expansion.”

Acryl Inc. holds a total of more than 83 patents, including 22 related to GPUBASE.

Park Oe-jin, CEO of Acryl Inc., said, “AI sovereignty does not depend on how many GPUs one has secured, but rather on how efficiently one can control and operate the entire infrastructure, models, and operations,” adding, “Based on technology that connects these three layers and empirical data, Acryl Inc. will contribute to building an open, vendor-agnostic ‘Sovereign AX’ ecosystem.”