South Korea has achieved a milestone in building independent artificial intelligence infrastructure: Daum, one of the country's largest web portals, now powers its AI search summaries using only domestically developed technology. The service, called AI Overview, runs on Upstage's Solar large language model (LLM), a type of AI trained on vast amounts of text to understand and generate human language, paired with FuriosaAI's second-generation neural processing unit (NPU), called Renegade. An NPU is a specialized computer chip designed specifically to run AI tasks efficiently on local devices rather than relying on distant cloud servers.

This partnership between Daum, Upstage, and FuriosaAI represents what industry leaders are calling Korea's first commercialized full-stack sovereign AI case, meaning the entire technology stack, from the semiconductor chip to the AI model to the user-facing service, relies on domestic innovation. The arrangement signals a broader shift in how countries are approaching AI independence, moving away from reliance on foreign technology giants.

Why Does Sovereign AI Matter for Search?

Real-time search summaries present a unique challenge for AI systems. Large language models are trained on historical data and can sometimes generate plausible-sounding but false information, a problem known as hallucination. For Daum's AI Overview, providing the latest information is critical. The team addressed this by using a technique called prompt engineering, which constrains the model to answer only based on current information provided to it, rather than relying on knowledge learned during training.

"For AI Overview, it is important to provide the latest information based on real-time data, but pre-trained knowledge could cause hallucinations. We applied harness engineering to the LLM so that it answers accurately only within the information I provide, instead of the knowledge it has learned so far, ensuring it provides only the latest information," explained Lee Gun-soo, CEO of AXZ, the company that operates Daum.

Lee Gun-soo, CEO of AXZ

The infrastructure supporting this service is substantial. FuriosaAI has deployed three computing nodes with 24 Renegade chips total, processing 500 million tokens per day, where a token is roughly a word or small piece of text. Despite this scale, the Renegade NPU achieves performance comparable to Nvidia's H200, a high-end graphics processing unit (GPU) commonly used for AI workloads, while cutting costs by at least 1.5 times.

How Are Korean Companies Building Competitive AI Infrastructure?

• Custom Chip Development: FuriosaAI designed the Renegade NPU in-house, including the compiler that maps AI models onto the accelerator hardware and the serving engine for optimal performance, eliminating dependence on foreign chip architectures.
• Domestic Language Models: Upstage's Solar LLM provides the intelligence layer, allowing Korean companies to control their own AI reasoning without licensing from international AI labs.
• Real-World Deployment: Rather than remaining theoretical, these technologies are already running production services at scale, processing hundreds of millions of tokens daily for millions of users.

The cost advantage of the Renegade NPU is particularly significant for companies considering AI transformation, or what industry professionals call AX work. By reducing infrastructure costs by 30 percent or more compared to GPU-based systems, the technology makes AI deployment more economically feasible for mid-market and enterprise organizations.

"To process AI summaries, we have deployed three nodes and 24 chips, handling 500 million tokens per day. We developed everything in-house, including the compiler that maps AI models well onto the accelerator and the engine for optimal serving, achieving performance on par with Nvidia's H200," stated June Paik, CEO of FuriosaAI.

June Paik, CEO of FuriosaAI

What's Next for Daum's AI Services?

Daum's AI Overview currently accounts for approximately 20 percent of the portal's total search query volume, and the company plans to expand this share over time. Beyond simple text summaries, the roadmap includes vertical-specific applications such as shopping and restaurant searches, where AI can provide more contextual and useful results. The company is also preparing to launch personalized AI agents, described as a "one agent per person" service, which would deliver customized information and recommendations to individual users.

"AI Overview currently accounts for about 20% of the total query volume, and we plan to gradually increase this. Beyond simple summaries, we are expanding into verticals such as shopping and restaurant searches, and we are also preparing to distribute a user-specialized 'one agent per person' through Daum," noted Lee Gun-soo.

Lee Gun-soo, CEO of AXZ

For Upstage and FuriosaAI, the Daum partnership serves as a proof point that Korean AI technology can compete with international alternatives. Both companies are actively encouraging other organizations to adopt their solutions during AI transformation initiatives. The collaboration demonstrates that building sovereign AI infrastructure is not merely a government policy goal but a commercially viable strategy that delivers real cost and performance benefits.