Kakao is pairing artificial intelligence with KakaoTalk and has opted for a strategy focused on cutting costs. It plans to process about half of user requests on smartphones rather than on servers, aiming to reduce server inference costs by up to 90 percent.

Jeongsin-a (정신아), the designated CEO of KakaoAI, said at a press briefing on Aug. 21 related to the spinoff that KakaoAI's competitiveness lies in an on-device-first AI structure. On-device AI processes data on the device without sending it to external servers. Jeong explained, "About half of user requests are handled immediately on the device by our lightweight model, Kanana Nano, and only complex requests are passed to server AI models and specialised agents."

Kakao is also developing technology to decide where requests should be sent. The AI Conductor being built by Kakao is orchestration technology that assesses a request's complexity and difficulty and selects and connects the appropriate model. Jeong said combining on-device-first processing and intelligent model routing would keep performance while cutting server inference costs by up to 90 percent. She said the company aims to build a structure in which inference costs do not increase linearly even as users grow.

It is also moving to reduce the size of server models. Jeong said, "We are distilling the knowledge of large models and making them lightweight to implement comparable context-understanding performance at one-tenth the size of global large models." Distillation is a technique that transfers what a large model has learned to a smaller model. She said on-device models have already secured performance exceeding global models, and that Kakao is applying them to Kanana and KakaoTalk while reducing memory and battery use through its own tokenizer and lightweighting technology. Benchmark scores or comparison models supporting the performance claims were not presented on the day.

Kakao also plans to continue developing its own models. It said that while most large overseas models target business-to-business markets, Kakao needs models that fit its services, understand user context and reduce costs. It added that with external AI providers it will approach the expansion of the agent ecosystem rather than pursuing large deals such as equity partnerships. It explained it would bring existing Kakao ecosystem businesses in as partners for the Model Context Protocol (MCP) and agent-to-agent (A2A) communication.

◆Can AI DAU of 20 million be achieved?

KakaoAI's envisioned final form of KakaoTalk is a gateway where multiple AIs come and go. Jeong said it will shift to a method in which users convey only their intent in natural language, and agents connect to needed services, execute them and deliver results. The structure would place a general-purpose assistant as a default within KakaoTalk, with specialised agents by domain added as needed.

Jeong cited as KakaoTalk's competitiveness that after capturing user context, it completes service connection and execution within KakaoTalk as a single flow. KakaoAI will hand over its role in managing subsidiaries to KakaoX through the split and concentrate resources on AI, advertising and commerce.

This cost structure is a premise of KakaoAI's financial targets for 2030. KakaoAI presented goals of achieving revenue of more than 6 trillion won and EBITDA of more than 2 trillion won by 2030, with average annual revenue growth of about 20 percent. It set an operating profit margin of more than 30 percent and return on equity (ROE) of more than 25 percent.

It identified three revenue sources: agentic advertising, agentic commerce and subscriptions. It plans to launch new advertising products in which AI identifies user intent, suggests products and links through to purchases, and to enter search advertising with a target of securing more than 5 percent share of the search advertising market by 2030. In commerce, it will link external vertical services based on A2A and grow transaction fees into a new revenue source. It plans to raise AI revenue to a double-digit share of total revenue by 2028 and to exceed 1 trillion won in 2030.

It also disclosed the basis for its target of 20 million daily active users (DAU) for AI in 2030. Jeong said 30 million users currently engage in active daily conversation in group chat rooms. She said applying a 60 to 70 percent monthly-to-daily user ratio from Kanana Intalk, and even using a conservative 80 percent, would result in more than 20 million users. She also cited that the time spent by users who use AI is more than 50 percent longer than for those who do not.

Kakao also expects the time users spend on KakaoTalk to increase by more than 50 percent by 2030 from the current level. It calculates that as users accumulate experience interacting with agents, they will stay longer, expanding space to attach advertising and commerce revenue.

Investment funding will be covered by internal cash flow. Jeong said the outstanding amount of existing convertible bonds is minimal and that investment will be executed based on EBITDA and return on investment. She said KakaoAI's EBITDA is expected to significantly exceed planned capital expenditure (CAPEX).