Daum AI summary service powered by Renegade and Solar

CEOs of three firms hold talks at AXZ headquarters on sovereign AI expansion

From left: Kim Sung-hoon, CEO of Upstage; Baek Jun-ho, CEO of Furiosa AI; and Lee Geon-su, CEO of AXZ. [Provided by Upstage]
From left: Kim Sung-hoon, CEO of Upstage; Baek Jun-ho, CEO of Furiosa AI; and Lee Geon-su, CEO of AXZ. [Provided by Upstage]

Upstage is joining forces with AXZ, the operator of the Daum portal, and Furiosa AI to commercialize Korea's first full-stack sovereign AI — connecting domestic hardware to software and services in a single integrated stack. The three companies plan to use Daum's AI summary service as a springboard to expand sovereign AI deployments built entirely on homegrown technology.

Upstage announced Wednesday that it is building a full-stack sovereign AI commercialization case with AXZ and Furiosa AI, combining a domestic neural processing unit (NPU), AI model and live service.

Under the three-way partnership, Furiosa AI's NPU "RNGD (Renegade)" will run Upstage's large language model (LLM) "Solar" to power Daum's "AI Summary" service. Upstage said the collaboration marks the first sovereign AI case in Korea to link every layer of the AI stack — from infrastructure and model to live service — using domestically developed technology.

On Wednesday, Upstage CEO Kim Sung-hoon, AXZ CEO Lee Geon-su and Furiosa AI CEO Baek Jun-ho gathered at AXZ's headquarters in Pangyo, Seongnam, Gyeonggi Province, for a roundtable discussion on each company's plans and the three-way sovereign AI partnership.

Lee opened by explaining Daum's AI summary service. "Daum AI Summary is a feature where an LLM directly analyzes web documents and compiles comprehensive answers to search queries," he said. "When a user types a keyword into the Daum search bar, it provides a concise summary along with supporting sources, and if the content changes, the AI automatically reflects the update so users can check the latest information."

Lee also described the development process behind the AI summary feature. "We advanced our harness engineering to control LLM hallucinations and ensure only up-to-date information is delivered," he said, adding that the service uses hybrid search — combining keyword search and vector-based search — to feed Solar with highly relevant information.

Baek then explained how the infrastructure supporting the service operates. "To handle AI summary processing, we are processing 500 million tokens a day across three nodes," he said.

He added that the team carried out optimization work to map AI models efficiently onto the accelerator, developing a compiler and a serving engine as part of a ground-up hardware-software design. "We are delivering performance on par with the H200," he said. Introducing Furiosa AI's NPU cuts inference costs by more than 50 percent.

"Domestic NPUs can dramatically reduce token costs and offer better performance-per-watt and price-to-performance ratios than GPUs," Baek said. "Given the current international landscape, combining sovereign AI semiconductors, models and services also makes strategic sense."

Upstage, Daum and Furiosa AI are building a full-stack sovereign AI commercialization case through a three-way partnership. [Provided by Upstage]
Upstage, Daum and Furiosa AI are building a full-stack sovereign AI commercialization case through a three-way partnership. [Provided by Upstage]

The three companies also outlined their future plans. AXZ said it aims to expand Daum's AI services beyond simple summarization toward a "one agent per user" model. "The AI Summary feature, currently in beta, handles about 20 percent of total query volume, and we plan to gradually increase that share," Lee said. "We are looking to expand vertically into areas such as shopping and restaurant search, and we are also preparing to roll out a personalized 'one agent per user' experience through Daum." He added, "Please use the AI summary service — it supports Korea's AI ecosystem."

Furiosa AI said it plans to expand its product lineup, including Renegade. "We are expanding our product line to support advanced agentic AI services at competitive token prices," Baek said. "We will continue to improve performance and cost-effectiveness across chip generations to build a competitive edge over global semiconductors." He added that Renegade delivers performance comparable to GPUs at half the cost, and encouraged companies undertaking AI transformation to consider it seriously.

The supply outlook for Renegade was also discussed. When Kim said Upstage could need up to 10,000 RNGD units by year-end, Baek said with confidence, "Renegade is currently in mass production, and we can supply 10,000 units without any problem."

Upstage said it will continue to sharpen the performance of its domestic LLM to compete with models from global big-tech firms. "Solar's competitive strength lies in its global frontier-level performance achieved through our own compact model design, combined with specialized Korean-language capabilities," Kim said. "We ask for your attention to Upstage's next-generation model, which will be unveiled soon through our independent AI foundation model project and other initiatives."


chami@heraldcorp.com