Korean IT service corporations have begun adopting domestic Neural Processing Units (NPU) to replace foreign Graphics Processing Units (GPU). The goal is to reduce dependence on high-performance but expensive GPUs and expand the use of NPUs specialized for artificial intelligence (AI) inference to boost the price competitiveness and efficiency of AX (AI transformation) businesses. In line with the government's "K-Nvidia promotion" policy, they also appear to be seeking an advantageous position in winning public-sector AX and cloud contracts.
On the 7th, according to the industry, major IT service corporations such as Samsung SDS, LG CNS, POSCO DX, and Lotte Innovate have recently teamed up with homegrown semiconductor companies to accelerate the verification and adoption of domestic NPUs. Moves are underway to convert existing Nvidia GPU-based cloud services and AI systems to hybrid infrastructure applying NPUs.
On Apr. 2, Samsung SDS said it will offer RNGD, the second-generation NPU developed by AI Semiconductor company FuriosaAI, as an NPU-as-a-Service (NPUaaS) on its Samsung Cloud Platform (SCP) starting in July this year. This is the first case of a domestic cloud company commercializing a homegrown NPU as a subscription service. Executive Director Lee Joo-pyeong of Samsung SDS said, "We are preparing NPUaaS so that customers can use RNGD on a subscription basis in units of 1, 2, 4, or 8 cards, as needed."
POSCO DX also signed a business agreement with AI Semiconductor startup Mobilint on the same day to implement NPU-based AX. The plan is to apply Mobilint's NPU to POSCO DX's industrial control systems to reduce AI infrastructure expense and realize an "intelligent factory" capable of instant analysis and control on site. Lotte Innovate also signed a business agreement the same day with AI Semiconductor company DeepX on using NPUs. The goal is mass production in the second half of the year. Lotte Innovate will equip its solution using intelligent CCTVs with DeepX's low-power, high-efficiency NPU, and DeepX will support development so that Lotte Innovate's in-house vision AI model can operate smoothly in an NPU environment.
LG CNS also decided to develop NPU-based services with FuriosaAI. It will apply FuriosaAI's NPU to the infrastructure running its in-house, enterprise agentic AI platform AgenticWork to conduct technology validation, and it plans to verify performance-optimization technology for NPU-based GPUaaS (GPU as a Service).
IT service corporations cited expense reduction as the reason they are actively pursuing the introduction and commercialization of domestic NPUs. An NPU is a semiconductor specialized for AI inference. Because it processes only the required computations, it can perform the same AI inference tasks with less power than a GPU, and it also has the advantage of being cheaper. GPUs excel at processing massive amounts of data and are widely used for AI training, but they are relatively expensive.
With the recent spread of AI services, GPU costs and power burdens are mounting, and the industry says high-efficiency NPUs have emerged as an effective alternative. Whereas training AI models used to be the focus, the center of gravity is shifting toward boosting productivity by applying trained models to real work and services, driving a surge in inference demand. From a corporation's perspective, using NPUs specialized for inference alongside GPUs can reduce infrastructure expense, making a shift to hybrid infrastructure necessary. A source in the cloud industry said, "Introducing NPUs can improve power efficiency and cost competitiveness, which helps optimize infrastructure."
NPUs are also seen as suitable for implementing "Edge AI," which does not pass through AI data centers or servers. Because data can be processed internally or on site without sending it to an external cloud, the risk of leakage of sensitive industrial data or corporate information can be reduced. This feature is expected to be a strength in sectors where data security is critical, such as public institutions, finance, and manufacturing.
The government's ongoing "K-Nvidia promotion project" is also seen as influencing corporations' decisions to adopt NPUs. The government said it will invest 50 trillion won in AI and semiconductors over the next five years to reduce reliance on GPUs and build an AI Semiconductor ecosystem centered on domestic NPUs. Korean IT service and cloud corporations are expanding AX contract wins in the public sector, and their alignment with the government's push to spread domestic NPUs is interpreted as an attempt to secure a competitive edge in the public procurement market.
However, some note that there are practical limits to converting a corporation's AI systems or infrastructure to NPUs. A source at an IT service company said, "Considering performance and general-purpose use, it is difficult to fully switch GPU-based AI systems to NPUs, and we are pursuing ways to replace certain areas, such as on-site Edge AI applications, with NPUs." Industry sources also said the barriers to NPU adoption felt in the field remain high.
While Nvidia's GPU monopoly in the AI chip market owes something to hardware performance, the GPU-specialized software platform CUDA has also played a major role. On top of that, the full-stack AI ecosystem includes the power-optimization network platform NVLink, leading to a "lock-in" effect that leaves users little choice but to use GPUs. Industry sources said that domestic NPUs also need to establish full-stack infrastructure spanning hardware and software if Korea's AI industry is to build self-sufficiency.
Lee Jin-ho, a professor in the Department of Electrical and Computer Engineering at Seoul National University, said at a recent forum, "Models optimized for NPU characteristics must be developed," adding, "High-bandwidth, low-latency networks among NPUs and compute-communication optimization technologies are also needed."