"The core of an AIDC (AI data center) is increasing the number of tokens processed per unit of power. That is why securing the latest GPUs (graphics processing units) is important."
Choi Jung-jin, head of the Evangelist Group at Samsung SDS, emphasized this point at the "2026 AI Infrastructure Frontier Conference" held in Hall B of COEX in Samsung-dong, Seoul. Choi delivered a presentation on the topic of "Enterprise AI Infrastructure with Samsung Cloud Platform (SCP)."
The latest GPUs are needed to increase the number of tokens processed per unit of power. Choi said, "As NVIDIA GPU generations advance, prices typically double while performance improves by about threefold," adding, "Since performance per unit of power also improves, introducing the latest accelerators early is an efficient way to operate."
Samsung SDS provides services such as GPUaaS (a service for renting graphics processing units) and NPUaaS (a service for renting neural processing units) on a serverless basis. While purchasing and using accelerators like GPUs and NPUs incurs significant time and cost each time a new accelerator is released, serverless services are economical because they utilize the latest accelerators secured by Samsung SDS.
Samsung SDS launched B300 GPUaaS in March, the first of its kind domestically, and began NPUaaS based on FuriosaAI's second-generation NPU (neural processing unit), "Renegade (RNGD)," in July, expanding its serverless services. Samsung SDS provides these serverless services through SCP.
Serverless services also offer the advantage of allowing companies to verify AI utilization strategies and potential before making large-scale investments. Users can select based on specific needs, such as using GPUs for AI training and NPUs for tasks like inference.
Choi explained, "Rather than purchasing GPUs from the start, finding usage methods or conducting a PoC (proof of concept) via serverless first enables more efficient business operations," adding, "If you want to apply GPUs across your entire organization, creating a cluster with GPUs and combining them with NPUs is more efficient."
AIDCs are also operated with energy efficiency in mind. Advantages include △cooling and facility design capabilities, △PUE (Power Usage Effectiveness) improvement, △global-level operational certifications, and △stable power and facility structures.
Samsung SDS has filed four domestic patents for cooling technologies in ultra-high-density computing environments. The Dongtan data center achieved the highest level domestically by lowering its PUE to 1.19. A PUE closer to 1 indicates higher efficiency.
In addition, the company features seismic-resistant design capable of withstanding a magnitude 9 earthquake, and dual power supply and facility structures including redundant cooling equipment and intake pipelines.
It has also obtained certifications such as ISO 27001, 22301, 50001, and ISMS-P. Furthermore, Samsung SDS designs and builds AIDCs by combining NVIDIA's DSX with its own data center infrastructure management program (DCIM). AI predicts in real time and proactively responds to gaps between DSX and DCIM.
Samsung SDS also offers a simple product called Simple AI Training & Inference. Simple AI Training is a product that allows customers to train LLMs without directly configuring the infrastructure themselves. Models trained through this service, or open-weight models pre-built by Samsung SDS, can be used with Simple AI Inference.
in TECH
