[Edaily Reporter Kim Hyun-ah ] South Korean AI semiconductor company Rebellion is expanding the commercialization of its AI semiconductor technology by applying domestically developed neural processing units (NPUs) to the public safety sector.

Rebellion (CEO Park Seong-hyun) announced on the 3rd that it will supply server-based NPU infrastructure to the CCTV monitoring centers of South Gyeongsang Province and Ulsan Metropolitan City as part of the “Transition to Public AI CCTV Based on Domestically Produced NPU” project led by the Ministry of Science and ICT and the National Information Society Agency (NIA).

This project marks the first instance of applying domestically produced AI semiconductors to public CCTV monitoring systems. Rebellion has secured contracts for server infrastructure projects with two organizations—South Gyeongsang Province and Ulsan Metropolitan City—and plans to begin supplying its high-performance server-type NPU, “ATOM-Max,” starting this August.

This AI CCTV monitoring system is designed to go beyond simple video detection and enable AI-based precision analysis. The system first detects basic objects—such as people and vehicles—using cameras and edge devices, and then an AI model on the monitoring center’s server performs additional analysis of the video to identify emergency situations such as fires and crimes.

Rebellion is responsible for the server-side AI processing in this process. The ATOM-Max server runs a large-scale video language model (VLM) with over 30B (30 billion parameters) at the monitoring center to analyze CCTV footage, and it even supports situation descriptions and confidence assessments.

Large-scale VLMs require extensive memory and high computational performance, meaning that existing small-scale edge AI chips alone have limitations in performing precise analysis and minimizing false positives and missed detections. Rebellion plans to stably implement a large-scale AI inference environment using its server-based NPU.

In particular, the Rebellion NPU officially supports “vLLM,” an AI model execution platform widely used by developers. As a result, public institutions can adopt the domestically developed NPU without significantly altering their existing AI development environments, thereby reducing system deployment costs and dependence on specific technologies.

It is also capable of operating independently within closed-network environments isolated from the external internet, and because it facilitates easy replacement and deployment of AI models, it is considered well-suited for the public sector, where security is paramount.

Through this project, Rebellion aims to expand its technology for running super-large AI models into actual public service settings.

Previously, Rebellion successfully ran the ultra-large AI model “A.X K1” from SKTelecom(017670)on a single server using its proprietary AI server, “RebelServer.” A.X K1 is a model with over 500 billion parameters; Rebellion optimized its software for the Mixed Expert (MoE) architecture and utilized parallel NPU computations to ensure real-time processing performance.

This public AI CCTV project is particularly significant as it marks the first instance of a domestically produced NPU moving beyond the research and validation stages to be applied in actual public infrastructure. As the government pushes for self-reliance in AI semiconductors and the expansion of domestically produced AI infrastructure in the public sector, domestic NPU companies are expected to accelerate their entry into the market.

Park Seong-hyun, CEO of Rebellion, stated, “It is symbolic and meaningful to have been part of the government’s initiative to expand the use of domestically produced NPUs in the public sector from the very beginning,” adding, “We will create a turning point where domestic NPU technology establishes itself as on-site infrastructure that safeguards public safety, moving beyond the demonstration of hyper-large AI models.”

Meanwhile, Rebellion, an AI semiconductor company founded in 2020, is developing inference-specialized NPUs for data centers, the software that powers them, and AI infrastructure solutions at the server and rack levels.

Its in-house developed “ATOM” series of NPUs is being applied to AI services at domestic data centers and by telecommunications and cloud companies, and the company is currently pursuing entry into the global market with its HBM3E-based high-performance NPU, the “Rebel100.”