Lee Gyu-lee is a business writer at The Korea Times, focusing primarily on IT & telecommunications, the Ministry of Trade, Industry and Energy and KOTRA. Prior to this, she has covered a wide range of cultural news, from film, television and K-pop to lifestyle and fashion.
Upstage, FuriosaAI team up to power Daum's AI search with homegrown stack
From left, Upstage CEO Kim Sung-hoon, FuriosaAI CEO June Paik and AXZ CEO Lee Keon-su pose during a conference at AXZ's headquarters in Seongnam, Gyeonggi Province, Tuesday. Courtesy of Upstage
By Lee Gyu-lee
Upstage is partnering with its subsidiary AXZ, operator of internet portal Daum, and artificial intelligence (AI) chip startup FuriosaAI to run Daum’s new AI Summary search feature, which the companies describe as Korea’s first commercially deployed full-stack sovereign AI architecture.
The collaboration brings together Upstage’s Solar large language model (LLM) on FuriosaAI’s RNGD neural processing unit (NPU) to power Daum’s AI Summary search feature in a single, end‑to‑end stack.
The companies said the deployment marks Korea’s first commercial example of an AI stack built entirely with domestic technologies, spanning AI infrastructure, foundation models and end‑user services.
To mark the launch, Upstage CEO Kim Sung-hoon, AXZ CEO Lee Keon-su and FuriosaAI CEO June Paik met Tuesday at AXZ's headquarters in Seongnam, Gyeonggi Province to discuss the partnership and the future of sovereign AI.
Daum’s AI Summary feature, which launched its beta service earlier this month, analyzes web documents using an LLM to generate comprehensive answers to search queries.
"When users enter a keyword into Daum’s search engine, the service provides both a concise summary and the supporting sources. As content changes, the AI automatically updates the results so users always receive the latest information,” Lee said during the conference.
He added that the system automatically updates content when source pages change and uses a “hybrid search” approach combining keyword and vector‑based retrieval to feed Solar the most relevant, timely information while tightening hallucination controls.
“We strengthened our harness engineering so that the system delivers only the most up-to-date, accurate information to users,” he said.
For inference, Paik said the AI Summary service currently processes about 500 million tokens per day across three computing nodes.
"We designed both the hardware and software from the ground up, including a compiler that efficiently maps AI models onto the accelerator and an optimized serving engine, allowing us to achieve performance comparable to Nvidia's H200,” he said.
He added deploying FuriosaAI's NPUs reduces inference costs by more than 50 percent.
"Domestic NPUs can dramatically reduce token costs while offering better performance per watt and overall cost efficiency than graphics processing units (GPU),” he said, calling domestic accelerators a key enabler for large‑scale AI deployments.
“Given the current geopolitical environment, it is strategically important to combine sovereign AI semiconductors, models and services."
The AI Summary feature currently handles about 20 percent of Daum's total search queries. However, AXZ plans to gradually expand coverage into vertical services such as shopping and restaurant recommendations, while ultimately offering personalized AI agents for users.