If you come to Daum and use "AI Overview," you are using Upstage's large language model (LLM) "Solar," and that in turn uses our NPU. It would be great if people used it a lot so we can create this kind of virtuous cycle.
Kim Sung-hoon, CEO of Upstage, said this at a joint briefing by Upstage, AXZ and FuriosaAI on the 15th. The plan is to apply Solar to the portal Daum's AI search service "AI Overview" and run it on FuriosaAI's domestically made Neural Processing Unit (NPU) "RNGD (RNGD)" to link the domestic AI model, semiconductor and internet service. The three companies introduced this as "Korea's first full-stack sovereign AI commercialization case."
Currently, about 24 RNGDs are deployed across three server nodes for Daum AI Overview, processing about 500 million tokens per day. Based on Daum's search usage, the company aims to increase Solar usage and, in turn, expand demand for domestic NPUs.
However, it is uncertain whether this business can be a breakthrough to improve profitability for Upstage, which has continued to post losses. According to Upstage's 2025 audit report, operating revenue was 24.8 billion won, up 78% from 13.9 billion won the previous year, but operating expenses were 55.3 billion won, about 2.2 times revenue. It recorded an operating loss of 30.5 billion won and a net loss of 28.4 billion won.
If AI Overview usage increases, usage of Solar and RNGD also rises, but inference expenses for servers and electricity increase as well. The briefing did not disclose how much Daum's search volume and ad revenue have grown since adopting AI Overview, or whether revenue per user exceeds inference expenses. Although the operation scale of processing 500 million tokens per day was presented, there was no explanation of how this consolidates into Upstage's sales and profit.
◇ "Processing 500 million tokens a day with about 24 RNGDs"
Baek Jun-ho, CEO of FuriosaAI, said, "Currently, about 24 RNGDs are deployed across three server nodes to accelerate the Solar model used for AI Overview," adding, "We process about 500 million tokens per day." A token is the basic unit an AI model uses to recognize sentences and generate answers.
Lee Geon-su, CEO of AXZ, said, "AI Overview is currently exposed on about 20% of all search queries on Daum," adding, "We are preparing to expand it to more than half going forward."
AI Overview is a service in which, when a user enters a search term, related web documents are found and Solar shows the key content and evidence in answer form. Lee cited up-to-date information and hallucinations as key tasks in the development process.
Lee said, "Even if the model has pre-trained knowledge, we have advanced harness engineering so it answers only within the information provided through search," adding, "We are delivering highly relevant, up-to-date information with hybrid search that combines keyword and vector search."
However, specific indicators to objectively judge service quality—such as hallucination rate, source match rate and user satisfaction—were not disclosed. Lee said, "If satisfied, please press the upvote (recommend); if not, press the downvote (not recommend)," noting that user feedback will be used to improve quality.
◇ "On par with the H200"… comparison conditions undisclosed
FuriosaAI argued that RNGD delivers response performance equivalent to Nvidia's high-performance AI accelerator H200, while offering better power efficiency and price competitiveness.
Baek said, "We co-designed hardware and software, from the compiler that places AI models on the accelerator to the engine for optimal serving," adding, "We achieved performance on par with the H200 with better performance per watt and cost-effectiveness."
At the briefing, a demo ran Solar on the H200 and RNGD respectively. Kim assessed the response speeds of the two systems as "almost similar."
Baek said, "The cost efficiency—throughput divided by total cost of ownership (TCO), including chip purchase price and power consumption—is currently at least about 1.5 times," adding, "Our goal is to cut expenses to about half through additional optimization."
However, the conditions under which RNGD and the H200 were compared were not disclosed. It was not confirmed whether model precision and quantization method, input/output token length, concurrent users and batch size were the same. Detailed performance figures, such as time to first token, tokens per second and concurrent users handled per server, were also not provided.
It is also unclear whether the claim of about 1.5 times cost efficiency means actual inference expenses were reduced by that much. Since the current cost-savings effect and the post-optimization target are not distinguished, it is hard to judge economic viability in a large-scale commercial service environment based on the demo alone.
◇ "Able to supply 10,000 RNGD units within the year"
When Kim asked whether 10,000 RNGD units could be supplied by year-end, Baek replied, "We are currently in mass production and can easily supply 10,000 units." However, this is FuriosaAI's statement about supply capability and does not mean that Upstage or AXZ actually ordered 10,000 units.
Baek said, "Domestic NPUs have better performance per watt and cost-effectiveness than GPUs, allowing token expenses to be reduced," adding, "Given the international situation and supply chain diversification, a combination of sovereign AI Semiconductor, model and service is necessary."
AXZ plans to expand AI Overview into category searches such as shopping and restaurants, and to provide user-specific AI agents in the long term. Lee presented as an example a stock news agent that summarizes and delivers information every morning on Samsung Electronics, SK hynix and Micron in the United States. The specific launch timing was not disclosed.
This collaboration is meaningful in that it showed domestic AI models and semiconductors operating in an actual internet service. However, AXZ, which adopted Solar, is the operator of Daum that Upstage acquired. This is different in nature from a case where an independent external customer compared the performance and price of multiple AI models and accelerators and then chose Solar and RNGD.
An IT industry official said, "For Upstage to make Daum AI Overview a breakthrough for improving profitability, it needs to consolidate increased usage into ad or paid service revenue and secure external paying customers beyond Daum," adding, "Proving a case that delivers stable service at lower cost than Nvidia-based systems and generates actual revenue remains a task."