[Edaily Reporter Han Kwangbeom ] The NPU developed by AI semiconductor company Rebellion has been expanded to support the core functions of four major commercial AI services on SKTelecom(017670).
Rebellion announced on the 21st that its NPU has been integrated into four of KTCorporation’s core services, starting with A.dot Call Summary, followed by A.dot’s voice synthesis feature, the AI Customer Center (AICC), and the “ScamGuard” phishing text message detection feature.
The two companies began testing in June of last year to integrate a domestically developed NPU into SKT’s major AI services. Subsequently, in December of the same year, a server based on Rebellion’s AI semiconductor “ATOM-Max” was first commercially deployed for the “relationship inference feature” of A.dot Call Summary and has been operating stably for over half a year.
Starting in June of this year, Rebellion has sequentially expanded the scope of application to include: △ the “one-line summary feature” of A.dot Call Summary; △ A.dot’s “voice synthesis feature (streaming TTS)”; △ the AICC “consultation post-processing and search system”; and △ ScamGuard’s “phishing text message detection feature.”
Currently, the real-time AI traffic processed by Rebellion’s infrastructure across these four services amounts to approximately 14 million requests per day, exceeding 4 billion tokens daily. This scale is comparable to large-scale commercial AI operations by global companies.
From a technical perspective, leveraging low-latency, high-performance computing capabilities, the platform stably runs highly complex real-time services such as streaming TTS, which synthesizes speech at the sentence level simultaneously with text generation. By supporting a global-standard AI development and deployment environment, it enables ScamGuard to immediately apply detection models for new and variant scam texts without service interruptions.
By adopting the Rebellion NPU, SKT has optimized computing resources by dynamically allocating GPUs and NPUs based on the size and performance characteristics of AI models. Furthermore, the company has enhanced infrastructure operational efficiency by configuring an “NPU Farm”—a cluster comprising dozens of NPUs and hundreds of cards—allowing multiple in-house services to share the same resources.
Park Seong-hyun, CEO of Rebellion, stated, “The past year has marked the transition of domestically produced NPUs from ‘potential’ to ‘proven infrastructure.’ The fact that four services used daily by the public are running stably on Rebellion’s infrastructure is the clearest evidence of this.” He added, “Moving forward, we will continue to expand the reach of AI services both domestically and internationally based on our track record of large-scale commercial operations.”
Park Byeong-gwan, Head of Core Platform at SKT Corporation, said, “As demand for AI computing grows, we are actively adopting and utilizing not only GPUs but also Rebellion’s NPUs. By selectively utilizing and reallocating GPUs and NPUs according to the specific characteristics of each AI service task, we have improved the efficiency of AI resource utilization across the entire company,” he said. “Going forward, we plan to expand our NPU farm to maximize synergies from using domestically produced NPU semiconductors and respond agilely to AI computing demands both internally and externally.”
Following its success in powering SKT’s AI foundation model “A.X K1” and now with the expansion of this commercial service, Rebellion plans to extend its supply of NPU infrastructure to various industries, including finance, telecommunications, and the public sector.
