[Edaily Reporter Kim Hyun-ah ] Kakao(035720)is changing its server operation model to accommodate the expansion of its artificial intelligence (AI) services. The approach involves maximizing the utilization rate of its own infrastructure under normal conditions and only supplementing it with external cloud resources when demand surges. Following efforts to reduce AI inference costs, the company’s strategy is to further improve server operational efficiency, creating a structure where costs do not rise sharply even as the user base grows.

Jeong Shin-ah, CEO of Kakao, stated at a press conference announcing the corporate spin-off, “We will maintain performance through on-device processing and intelligent model routing while reducing server inference costs by up to 90 percent,” adding, “We will build a service structure where inference costs do not increase linearly even as the user base grows.” The plan is to process requests that can be handled on the device using lightweight models, rather than routing all requests to large AI models, and to pass only complex requests to the server-based AI models.

This time, the company is focusing on cost efficiency not only in the inference methods of AI models but also in the infrastructure operations system that supports them. Rather than simply increasing the number of servers, the company plans to maximize the utilization of existing resources and utilize external computing resources only when additional resources are needed.

Baek Jeong-tae, a Kakao presenter, introduced the company’s direction for FinOps—a cloud cost management framework—at the “if(kakao)2026” developer conference on the 14th. Presenter Jeong Won-cheon explained the automatic resource scaling features of DKOS, the company’s internal Kubernetes platform, and case studies of its application in hybrid cloud environments.

Reducing Idle Servers and Calculating Costs Per Service

The issue Kakao identified is that there are many cases where servers are procured but not fully utilized. As AI features and services increase, so does the demand for computing resources; however, maintaining all necessary servers at all times leads to a significant cost burden. Conversely, indiscriminately reducing servers simply because of low utilization makes it difficult to respond to outages or sudden spikes in traffic.

Kakao decided to link resource usage and costs by service to distinguish between necessary resources and those that are over-provisioned. By integrating servers, databases (DBs), and shared platforms with the service catalog, the company identifies which services are using how much resources and generating what costs.

The company is also revising its cost management criteria. Instead of merely tracking the overall increase in server costs, it will measure “unit economics”—costs relative to service performance, such as the cost per 1,000 successfully processed requests or the cost of delivering 1 million messages. The goal is to assess the efficiency of infrastructure investments by analyzing both actual throughput and costs, even as AI capabilities expand.

Goal: 20–30% Reduction in Computing Resources

Actual resource reduction efforts are also underway. Presenter Jeong Won-cheon explained that, according to Kakao’s investigation, workloads were allocated approximately 30% more computing resources than their actual usage. Kakao is introducing a feature that automatically reduces resource requests to match actual usage, consolidates remaining tasks onto certain servers, and then reclaims idle servers.

The goal is to reduce computing resources by 20–30% through this approach. However, this is an expected figure rather than an actual result achieved across all services, and the scope of application is being gradually expanded.

In-house servers for regular use, public cloud during traffic spikes

The company is also changing its approach to securing servers. Kakao plans to pool spare resources from each service into a shared resource pool and, when demand from multiple services increases simultaneously and becomes difficult to handle, secure additional resources from the public cloud. This is a hybrid cloud strategy in which regular demand is handled by in-house data centers, while temporary demand is met with external resources.

To achieve this, the company connected public cloud nodes to its existing Kubernetes clusters. This allows developers to utilize external computing resources as needed—using the existing workflow—without having to deploy services separately to a distinct cloud environment.

Kakao validated this hybrid configuration on the KakaoTalk Gift Queue service, which frequently experiences traffic surges. During holidays or special occasions, traffic can increase up to 16 times the usual volume. To address this, the company utilizes predictions based on historical data, AI-driven detection of traffic spike patterns, and real-time threshold alerts.

Currently, the company has manually configured some features to verify their functionality and is proceeding with automation efforts. The target completion date for full automation is the end of November 2026.

As competition in the AI sector expands beyond model performance to include operational costs and profitability, the importance of infrastructure efficiency is also growing. The key challenge for Kakao is determining how much it can reduce the cost burden associated with expanding AI services by simultaneously lowering inference costs and server operating expenses.