[Edaily Han Kwangbeom Reporter Shin Young-bin] The three consortia selected for “AI for All”—a free AI infrastructure project for the public—will simultaneously launch beta services this October and begin competing in the market. SKTelecom(017670), Kakao(035720), and KTCorporation(030200)have each presented telecommunications channel integration, integration of everyday-life data, and portal search and optimized routing as their key differentiators.

During the Q&A session at the “AI for Everyone” project kickoff briefing held on the 4th at the Korea Press Center in Seoul, the three-company consortium announced that it will launch a closed beta and basic chatbot services in October and begin full-scale service enhancement in preparation for the official launch in December.

The three consortia specifically revealed their respective core strategies for delivering services that are distinctly different from global AI offerings.

The SKTelecom consortium highlighted its strengths as the unique accessibility of a telecommunications provider combined with specialized partners. Kim Ji-hoon, Head of SKTelecom’s AI Business Division, explained, “Along with our strengths in data and services, we will leverage the differentiated scenarios offered by the ‘phone and text’ channels.” Kim Jin-woo, CEO of Liner and a participant in the SKT consortium, added, “Stable, general-purpose services serve as the ‘anvil,’ while service partners that delve deeply into specific areas, such as job preparation, act as the ‘hammer.’” The consortium is currently testing a mix of approximately three models, including K2.

The Kakao consortium has made a bold move by emphasizing data integration closely tied to daily life and “convenience.” Kim Se-woong, Vice President of Kakao AI, stated, “While it’s difficult to immediately surpass the performance of U.S. Big Tech AI right now, we will focus this year and next on creating AI that the public can use with true convenience.” He further emphasized, “Especially in areas under government control—such as finance, healthcare, and public services—if data and services are sufficiently interconnected, we can achieve what foreign AI cannot.” The consortium plans to operate its application models primarily around LG Corp. ExaOne and Kakao Kanana.

The KTCorporation consortium highlighted its data integration with the “Daum” portal and specialized vertical operators. Lee Jin-hyung, Head of KTCorporation’s AX Business Division, said, “We are pursuing the ‘Ieum Inside’ strategy with Daum to make search more accurate, and we are preparing to optimize startup models such as Upstage and to provide public services in collaboration with SALTLUX Inc.” Lee Geon-soo, CEO of Daum, who joined the consortium, said, “We will combine real-time data from Korean vertical service providers—such as shopping, restaurants, and real estate—with RAG (Reasoning, Augmented, and Generative) to provide accurate, real-time information.” KTCorporation plans to operate a total of seven models—including the core engine (Upstage), NPU offloading (KTCorporation Mideum), and public agents (SALTLUX Inc)—by automatically routing them according to specific tasks.

The government reaffirmed its policy of providing unprecedented budgetary support to back this service initiative. Kim Kyung-man, Director General of the AI Policy Bureau at the Ministry of Science and ICT, stated, “Of next year’s 250 billion won budget, approximately 170 billion won will be used to lease about 2,000 NVIDIA B200 GPUs, while 30 billion won will be directly allocated to each of the three consortia—10 billion won per consortium—to be used for system development and collaboration.” The remaining 50 billion won will be used to fund the expansion of external specialized agents.

In addition, the government plans to pursue separate policies to foster a domestic NPU (AI semiconductor) ecosystem, while holding further discussions on joint branding and the adoption of a common service name across the three companies’ services to improve user accessibility.

The following is the full transcript of the Q&A session from the “AI for All” project kickoff briefing.

- When will each consortium launch its beta service? Which services do you plan to introduce first during the beta phase?

△Kim Ji-hoon, Head of the AI Business Division at SKTelecom: We anticipate the beta phase to begin in October, after Chuseok. The first features we plan to roll out are the chatbot and search integration functions we discussed today.

△Lee Jin-hyung, Head of KTCorporation’s AX Business Division = We originally aimed to launch the basic features in late September to early October. However, a key feature of our service is that it must be tightly integrated with the ecosystem’s search functions for the public to truly benefit from it. We are currently discussing the beta schedule with the government. If a closed beta is necessary, the schedule may be slightly delayed. As for the general features themselves, we should be able to release them in late September or early October according to the original schedule.

△Kim Se-woong, Vice President at Kakao = We are targeting mid-October and plan to launch the beta service with a basic chatbot.

△Kim Kyung-man, Director General of the Artificial Intelligence Policy Bureau at the Ministry of Science and ICT = We anticipate that the beta service will initially be conducted in a closed format, focusing primarily on testing its completeness. According to the business plan, the official launch is scheduled for December. Since this is a new service, there are several areas requiring caution—such as personal information—so we are considering conducting thorough testing with a limited group of testers in the early stages. We will discuss the specific schedule with the three companies again and provide an update.

-Kakao has set a Monthly Active User (MAU) target. What are the user targets for SKT and KTCorporation? Also, why hasn’t the government disclosed the detailed evaluation scores for each consortium?

△Kim Ji-hoon, Head of SKTelecom’s AI Business Division = The SKTelecom consortium has set its MAU target for the end of this year at 5 million—the same as Kakao’s—and aims for 10 million next year.

△Lee Jin-hyung, Head of KTCorporation’s AX Business Division = Our consortium includes the Daum search service, and Daum itself has a very large MAU. Our MAU may vary depending on how many of those users migrate to our platform. We are projecting around 5 million for this year. We also plan to prepare our systems to accommodate more users next year.

△Kim Kyung-man, Director General of the Artificial Intelligence Policy Bureau at the Ministry of Science and ICT = It is true that the government prioritizes fairness and objectivity above all else. However, we also gave significant consideration to the wishes of the participating companies. Generally, the top-ranked company may wish to have its scores disclosed, but other companies often do not. In this instance as well, we sought the opinions of the participating companies, and that served as one of the criteria for not disclosing the scores. The scores of the three selected operators are nearly identical; the differences are minimal. In contrast, there was a noticeable gap between them and the remaining participating operators, and those companies also did not wish to have their scores disclosed. The evaluation criteria are detailed in the official announcement. The key evaluation factors include the stability of the business and how the ecosystem will be expanded—elements that were highlighted in the presentations made by the consortia today.

- How specifically do you plan to allocate next year’s 250 billion won budget for “AI for All”? How will follow-up support for each consortium be determined?

△Kim Kyung-man, Director General of the Artificial Intelligence Policy Bureau at the Ministry of Science and ICT = Next year’s budget is set at 250 billion won and can be broken down into three main categories. First, approximately 170 billion won covers the costs associated with GPU usage. Our calculation is based on the cost of leasing approximately 2,000 B200 units. Since this level sufficiently exceeds the GPU computational capacity currently required for domestic chatbot services, we expect it to cover next year’s projects to a certain extent. 30 billion won will be directly allocated to the three consortia, with 10 billion won going to each. Each consortium can use this funding to cover the costs of operating or building systems and collaborating with individual participating companies. We plan to use the remaining 50 billion won for agent expansion. While this project is based on chatbots, it holds greater significance for agent expansion, such as public agents or personalized services. We aim to create a structure where agents developed externally can be seamlessly integrated into and delivered via the three platforms. By around March of next year, we should be able to see which services are actually seeing the highest usage. Since GPU support is likely to be scaled up based on actual usage, consortia that secure a large number of subscribers and demonstrate high activity will be able to secure more resources. However, what the government wants to see is not merely competition, but a scenario where participants compete while also collaborating and growing steadily together.

-You mentioned reducing dependence on foreign AI, but there is a high reliance on NVIDIA for hardware such as GPUs. To what extent do you plan to utilize domestically produced AI accelerators in “AI for All”?

△Kim Kyung-man, Director General of the Artificial Intelligence Policy Bureau at the Ministry of Science and ICT = We are separately implementing policies to promote the use of domestically produced accelerators. If the public widely adopts “AI for All,” the volume of inference will increase significantly, and this could help open up a substantial market in terms of demand for accelerators. In addition to “AI for All,” we are separately implementing various policies to expand the use of domestically produced NPUs. We believe the ideal outcome is for this initiative to significantly increase the number of users, leading to widespread adoption of domestic models and accelerators in the process. We are securing a separate budget for accelerator-related policies to foster the ecosystem.

△ Baek Jun-ho, CEO of FuriosaAI = Domestic NPUs are rapidly advancing to fully support proprietary foundation models and agentic AI. We will strive to ensure that NPUs offer sufficient high performance and cost-effectiveness so that the public can use them without inconvenience. We believe that a significant portion of the tokens generated by “AI for All” should be generated primarily using domestic NPUs. We will work together with the consortium to make this a reality.

△Lee Jin-hyung, Head of KTCorporation’s AX Business Division = Many domestically produced NPUs are specialized for inference. A highly efficient configuration can be achieved by appropriately combining GPUs and NPUs. For example, this involves dividing tasks such as prefill and decode between the GPU and NPU to utilize them appropriately. We tested the Rebellion NPU with domestic models, and it delivered impressive performance. We plan to increase efficiency by appropriately combining GPUs and NPUs, and we anticipate that the proportion of NPU usage will gradually increase in the future, particularly in terms of operating costs.

- Deputy Prime Minister Bae Kyung-hoon stated, “If there’s no differentiation from foreign AI, ‘AI for All’ will fail.” What do the various consortia see as the key differentiators from foreign AI?

△ Kim Se-woong, Vice President of Kakao = What we are aiming for is convenient AI. Looking at the current landscape—including models, Agent AI, platforms, and the entire ecosystem—I don’t think it will be easy to surpass the AI of U.S. Big Tech companies based solely on performance right now. Of course, we must catch up. However, it’s not enough for just one technology to advance; the entire ecosystem must advance together. This year and next, we plan to focus on creating AI that the public can use with true convenience. In particular, in areas under government jurisdiction—such as finance, healthcare, and public services—I believe that if the data and services that “AI for Everyone” can utilize are sufficiently interconnected, we’ll be able to achieve what foreign AI cannot.

△Kim Ji-hoon, Head of the AI Business Division at SKTelecom = I think everyone here would agree that we can clearly establish a competitive edge in data and services. Furthermore, telecommunications companies are part of all three consortia. We, in particular, aim to leverage the unique scenarios offered by the “phone calls and text messages” channel.

△Kim Jin-woo, CEO of Liner = When operating B2C services, it’s essential not only to fully ensure basic functions like stability, speed, and convenience but also to prepare features that deeply address users’ needs at precisely the right moment. For example, for job seekers, we need to delve deeply into areas where the quality of the output has a significant impact—such as writing a strong cover letter. On the other hand, everyday conversations must be easy and quick for anyone to use. The SKT consortium includes many service partners capable of delving deeply into these specific areas. These partners act as the “hammer,” while stable, general-purpose services serve as the “anvil.” APIs and data made available by the public sector can also be key elements in creating these specialized services.

△ Lee Jin-hyung, Head of KTCorporation’s AX Business Division = Our biggest concern was how to differentiate ourselves from foreign AI products. General-purpose search or general-purpose conversation alone won’t be enough to win back users who are already familiar with ChatGPT. So, our first focus is “search.” Together with Daum, we are pursuing the “Ieum Inside” strategy to make search more accurate. Second is model competitiveness. We plan to utilize “Token Factory” to support the continued development of startup models with global competitiveness, such as Upstage’s proprietary foundation model. We also plan to actively leverage domestic technologies that can differentiate us from global services, such as NC AI’s 3D generation model. Finally, recognizing that public services are one of the areas where citizens are most comfortable using AI, we are preparing related services in collaboration with companies like SALTLUX Inc.

△Lee Geon-soo, CEO of Daum = Even global generative AI often fails to provide accurate answers to questions necessary for daily life, such as shopping, finding good restaurants, or searching for rental properties or apartments. In these areas, accurate databases and real-time data are crucial. We must not recommend non-existent apartment listings or invent products that don’t exist. What we prioritize is RAG (Retrieval-Augmented Generation) based on actual databases and real-time information. For example, if you ask a global AI system directly about today’s AI briefing, it may struggle to provide an accurate answer, but Daum Search displays real-time articles written by reporters. The AI can use this as a basis for its response. Furthermore, even Google does not possess the full extent of real-time data held by Korean vertical-market players like Danawa, Musinsa, and Zigbang. I believe connecting this data is the starting point that sets us apart from global AI systems.

-Kakao has disclosed the usage ratios for ExaOne and Kanana. What models do SKT and KTCorporation plan to use, and in what proportions?

△Kim Ji-hoon, Head of the AI Business Division at SKTelecom = The guidelines initially proposed by the government were in the form of a 50-30-20 ratio. The maximum share that any single model can account for is 50%. We are currently testing several models internally. Based on actual test results, we’ve found performance differences between models depending on the service type, so we haven’t set a rigid limit of 49% or 50% for any specific model. We are considering a plan to use a mix of about three models, including K2. Rather than setting exact ratios just yet, we’re mapping out which models are best suited for specific usage scenarios.

△Lee Jin-hyung, Head of KTCorporation’s AX Business Division = We utilize a total of seven models. Five models handle general responses, while the remaining two are a safeguard model and a model related to STT and TTS. Similar to SKT, we intend to comply with government guidelines but will not operate in a way that artificially sets ratios and negatively impacts performance. The Upstage model plays a major role in the core engine, and tasks requiring long-context processing are automatically routed to models that excel in that area. KTCorporation’s “Belief” model is optimized for domestically produced NPUs, so we utilize it via offloading; for public agents, we plan to use models from SALTLUX Inc., which has extensive experience developing related services—selecting the optimal model for each specific task.

- If three “AI for All” services are launched, will citizens have to sign up for each one separately? Does the government have any plans to create a common point of contact or an integrated usage method?

△Kim Kyung-man, Director General of the Artificial Intelligence Policy Bureau at the Ministry of Science and ICT = We discussed this issue with the participants of the three consortia even before the public briefing. We are currently reviewing concepts such as joint branding and joint marketing. Since each platform will operate its services by emphasizing its own unique characteristics and features, it is likely that sign-ups will be handled separately. However, we will hold further discussions with the three operators and their respective consortia on how to establish a joint brand on top of that. We are also discussing various ideas regarding service names and common approaches to improve accessibility.