[Issue] Why Is Gwangju Pursuing an 'NPU Computing Center' Again?

Alternative Initiative Launched After Failure to Attract National AI Computing Center
Strategy in Motion to Demonstrate Domestic Semiconductors and Secure AI Sovereignty
Lawmakers Urge Inclusion in Main Budget
Industry Sees It as "First Stage for Domestic Chip Validation"
Gwangju City: "Parallel Use of GPU and NPU, Establishing a Hub for Efficient Demonstration"

After failing to attract the National AI Computing Center, Gwangju has taken action once again. This time, the focus is on an "NPU (Neural Processing Unit) Computing Center." Rather than being a simple replacement project, this is a strategic renewed effort to demonstrate domestically developed AI semiconductors and secure AI sovereignty. Lawmakers are calling for budget allocations, while the industry expects this initiative to serve as the first major testbed for domestic chip validation.

NPU: A Dedicated Processor Specialized for AI Inference

RNGD server equipped with eight 2nd generation NPU RNGD (Renegade) cards from FuriosaAI. Provided by FuriosaAI

An NPU (Neural Processing Unit) is a specialized semiconductor designed to rapidly process the thinking and decision-making tasks of artificial intelligence. In the past, CPUs (Central Processing Units) or GPUs (Graphics Processing Units) handled these computations, but they faced limitations in terms of speed and power efficiency.

An official from a domestic AI semiconductor company explained, "NPUs are dedicated processors specialized for inference operations in artificial intelligence, especially deep learning models, and they offer far greater efficiency compared to GPUs, which perform general-purpose graphic computations." The official added, "If a GPU is like a 'multi-purpose engine' that quickly handles a variety of tasks, an NPU is closer to a 'specialized engine' optimized for AI computations."

NPUs can perform multiple computations simultaneously, much like the human brain. This allows them to process the same tasks faster and with less power consumption. AI features that operate within devices-such as facial recognition on smartphones, voice captions, and real-time translation-are all enabled by NPUs. They are also the core technology behind "on-device AI," which continues to function even when there is no internet connection.

While GPUs are used to "train" AI through large-scale computations, NPUs are responsible for "inference" with 10 to 100 times greater efficiency than GPUs. According to the city of Gwangju, "For inference services, a computing center with 80% NPU and 20% GPU is the most efficient configuration."

'AI Demonstration City Gwangju'... Shifting Focus to NPU

Kang Gijeong, Mayor of Gwangju Metropolitan City, delivered the keynote speech on the theme "South Korea's Top 3 AI Demonstration City, Gwangju" at the 10th Seoul Future Conference held at the Shilla Hotel in Seoul on the morning of the 5th. Provided by Gwangju City

Gwangju officially proposed the "NPU Computing Center" on November 5. Kang Gijeong, Mayor of Gwangju, announced the "AI Demonstration City Gwangju" strategy at the 2025 Seoul Future Conference held at the Shilla Hotel in Seoul, and suggested to the government the establishment of a national NPU-dedicated computing center.

Mayor Kang stated, "It is necessary for building a domestic AI semiconductor ecosystem and for South Korea to become one of the top three AI powerhouses," adding, "We must support both the mass production systems of companies and the training of talent at universities and research institutions." Gwangju has already signed agreements with multiple fabless companies and is conducting the first phase of a demonstration project (2023-2024). Starting this year, the city is accelerating the commercialization of domestic NPUs by launching a second-phase advancement project worth 40 billion won (2025-2027).

On that day, Mayor Kang directly proposed the establishment plan to Vice Prime Minister and Minister of Science and ICT Bae Kyunghoon. He then presented four key strategies: ▲ National AI Research Institute ▲ AI+Mobility New City ▲ Mega Sandbox-type Cluster, and others, stating, "Although Gwangju missed out on the National AI Computing Center, we will lead South Korea to become a top-three AI powerhouse through NPU demonstration."

'Gwangju NPU Center' Under Budget Review

Prime Minister Kim Minseok is attending the plenary meeting of the Special Committee on Budget and Accounts held at the National Assembly on the 7th, responding to questions related to the 2026 budget. Photo by Yonhap News

The establishment of the Gwangju NPU Center is also being discussed by the National Assembly Special Committee on Budget and Accounts. Jeong Jinwook, a member of the Democratic Party of Korea (representing Gwangju Dongnam-gap), urged the government to provide active support, stating, "The AI-specialized pilot city must be pursued in Gwangju."

Assemblyman Jeong highlighted the inclusion in next year’s budget of ▲ 4 billion won for the design of the AI-specialized pilot city and ▲ 2 billion won for the basic planning of the AI Mobility National New City, emphasizing, "Gwangju, with the best-developed AI infrastructure and ecosystem, is the optimal location."

Jeong also proposed establishing the national NPU-dedicated computing center (1 trillion won) and the National AI Research Institute (600 billion won) in Gwangju, stating, "The Prime Minister must personally step forward to strengthen the competitiveness of domestic NPUs." In response, Prime Minister Kim Minseok replied, "I am well aware of the concerns of Gwangju citizens," and promised, "We will swiftly come up with alternatives."

An Do-geol, another Democratic Party lawmaker (representing Gwangju Dongnam-eul), called for concrete alternatives to the failed bid to attract the AI Computing Center, demanding that four tasks-NPU Center, National AI Research Institute, data center advancement, and smart mobility demonstration projects-be included in the main budget.

Seo Wangjin, a member of the Rebuilding Korea Party, remarked, "Gwangju and Jeonnam must jointly build a full-cycle AI ecosystem encompassing 'training, inference, and application,'" and pointed out, "The reversal of presidential campaign pledges has left local residents with a sense of loss."

In response, Minister of Science and ICT Bae Kyunghoon stated, "Gwangju and Jeonnam have ample potential to develop as AI pilot and demonstration cities," and added, "We will review measures to strengthen regional cooperation incentives and demonstration requirements."

Industry: "Domestic NPU Demonstration, Starting Point for Ecosystem Expansion"

The domestic AI semiconductor industry views Gwangju’s push for an NPU computing center as a stage for demonstrating Korean technology. The industry expects this project to serve as a turning point, expanding from research and development to real-world service environments.

An official from an AI semiconductor company commented, "A large-scale NPU data center is not just an infrastructure investment, but a strategic step to secure leadership in AI efficiency competition. The national NPU-dedicated computing center will become a symbolic milestone for Korea to lead a paradigm shift in AI infrastructure."

Shin Jungwoo, CEO of AX, said, "The essence of the GPU debate is closely linked to 'Sovereign AI,' and now countries and companies are moving to secure their own models and computational infrastructure." He added, "To build good models, infrastructure must be concentrated rather than scattered. Data centers and computational resources only become effective when they are gathered together. Regardless of politics or regionalism, infrastructure should be focused and centralized."

Gwangju City: "NPU and GPU in Parallel... A Hub for Demonstrating Domestic Semiconductors"

Gwangju City Hall exterior view.

The city of Gwangju is designing the NPU computing center to simultaneously serve as a demonstration platform for domestic AI semiconductors and as a service infrastructure. The plan is to build it using both NPUs and GPUs, including essential training functions such as lightweight training and preprocessing using GPUs.

Addressing concerns about a "training-incompatible NPU-only center," Gwangju explained, "By allocating 30% GPU, we can handle fine-tuning (lightweight training) and preprocessing tasks for large-scale models." The initial concept involves a structure with 70-80% NPU and 20-30% GPU in parallel.

Gwangju City emphasized, "NPUs are like long-range cruise engines, while GPUs are like cockpits responsible for complex takeoffs, landings, and flight planning. The harmony between the two engines is the core of AI services."

The city also stated, "While OpenAI and Google use GPUs for model development, they switch their infrastructure to inference-dedicated chips such as TPUs and ASICs for large-scale service stages. The Gwangju NPU Computing Center will become a national bridgehead for securing high-efficiency inference infrastructure based on domestic chips."

Gwangju pointed out that there has been a lack of large-scale demonstration environments for processing massive data in Korea, and views this center as the first platform capable of validating domestic NPUs. Choi Taejo, Director of the Artificial Intelligence Industry Office, said, "Our goal is to secure a high-efficiency inference infrastructure based on domestic chips, and as long as the government is committed, it is entirely feasible." He added, "Gwangju will not just be a city that develops AI, but will leap forward as a city that provides AI services efficiently."