[Edaily Reporter Yun Junghoon ]Kakao(035720)has revealed why it selected Amazon Web Services (AWS) as its core infrastructure partner for artificial intelligence (AI) model training. Kakao’s strategy is to flexibly secure graphics processing unit (GPU) resources and enhance the stability and efficiency of large-scale model training to support the government’s “AI for All” initiative and the development of its own large language model (LLM), “Kanana.”

Kim Se-woong, Vice President of AI Synergy at Kakao, unveiled Kakao’s AI infrastructure strategy and examples of technical collaboration with AWS at the “AWS AI Wave” event held on the 23rd at Parnassus Tower in Gangnam, Seoul.

Kakao plans to use KakaoTalk, the nation’s most widely used platform, as a gateway to expand AI agent services into the daily lives of all citizens. As part of the government’s “AI for All” initiative, Kakao is also working to build an agent platform that allows anyone to enjoy AI agents in their daily lives; the Kakao consortium includes major domestic partners such as LG Corp. and Lunit Inc.

At the same time, the company is accelerating the development of its proprietary AI model, Kanana. Kakao plans to develop a diverse product lineup—ranging from frontier-class models designed for top-tier performance to mid-sized models, small language models (SLMs), and lightweight models for on-device use—to allow for flexible selection based on specific service environments.

The challenge lies in securing the massive amount of GPU resources essential for model development in a stable and efficient manner. Training large-scale models requires thousands of computing resources, but GPU demand fluctuates constantly depending on the development stage and model size. As new GPUs are released every year, the hardware requirements for the development team also change constantly. Relying solely on in-house assets leads to limitations, such as resources lying idle after a project ends or an inability to supply them in a timely manner when needed.

To address these challenges, Kakao adopted a clear dual-infrastructure strategy: △ training on the AWS cloud and △ inference (service) on its own infrastructure and government-subsidized GPUs.

Kim Se-woong, Vice President of Kakao, explained, “While there is high internal demand for GPUs, there is significant waste once training is complete, and we face the challenge of securing the latest GPU models released each year in a timely manner.” He added, “We adopted a strategy where Kakao focuses solely on model development, while entrusting the operation of the training platform to AWS HyperPod so that developers can concentrate exclusively on research.”

Ensuring technical stability and building an uninterrupted training environment were also key reasons for choosing AWS. Kakao is actively utilizing “Amazon SageMaker HyperPod,” AWS’s distributed AI training infrastructure service. This is because even if some compute nodes fail during large-scale distributed training, HyperPod automatically detects and replaces them, immediately resuming training from the saved checkpoint, thereby dramatically reducing the overall training time.

Vice President Kim emphasized, “When we tested this internally, even if just one GPU failed, the entire node became unusable, resulting in significant operational downtime,” adding, “The biggest reason we chose AWS HyperPod was that shortening development time through ‘uninterrupted training’ was our top priority.”

In addition, Kakao established a management system that uses AWS Console tools to allocate limited resources efficiently and fairly among internal development teams. Vice President Kim added, “By clearly understanding resource utilization and efficiency through the AWS Console, we were able to reduce resource competition and conflicts between teams and enable rapid onboarding.”

The proprietary model “Kanana,” built in this way, has been registered on “Jumpstart,” AWS’s AI marketplace, and is rapidly gaining traction among global companies and developers. Going forward, Kakao plans to further strengthen its CSP partnerships while exploring the expansion of cloud-based inference capabilities.

Vice President Kim Se-woong stated, “Thanks to AWS’s proven development framework and support for next-generation training infrastructure, we were able to rapidly onboard the Kanana model and share our commercialization know-how with developers.” He added, “Going forward, we will expand our collaboration with various CSPs not only in model training but also in the inference domain to usher in an era where all citizens can enjoy the benefits of AI.”