[Edaily Kim Hyun-ah· Reporter Lee So-hyun] The cost structure of AI for businesses is changing. Providing AI services costing around 30,000 won per month per employee is just the beginning. As AI moves beyond simply reading documents, coding, and searching to performing tasks through independent reasoning, subscription fees are no longer the sole factor determining costs. The actual volume of data processed—that is, token usage—is the key factor.

While paid generative AI tools like ChatGPT and Claude have fixed monthly subscription fees, the number of tokens required by companies varies significantly depending on the task. This explains the growing cost disparity between simple queries, the analysis of documents spanning hundreds of pages, and agent tasks involving multiple rounds of reasoning.

“Token welfare”—which goes beyond simply providing an AI account to guaranteeing the usage quota needed to work—is emerging as a new type of work infrastructure. Just as PCs and the internet were once the basic work environment, in the AI era, AI access rights and sufficient usage quotas have become key factors determining productivity.

◇ AI Access Leads to Productivity Gaps… Companies Embrace “Token Welfare” Access to

AI is directly linked to actual labor productivity. A study by the National Bureau of Economic Research in the U.S. found that when customer support staff utilized AI, their hourly productivity increased by an average of 14%.

Companies have also begun to view AI as part of their work infrastructure. SOCAR(403550)provides AI tokens for work purposes to all employees—not just developers, but also those in planning and marketing—and offers additional support once the basic allowance is exhausted. KTCorporation(030200)is expanding the scope of AI usage by leveraging tools such as Copilot, ChatGPT, Gemini, and GitHub Copilot for work tasks and automating repetitive tasks through AI agents.

Samsung SDS (SAMSUNG SDS CO., LTD.(018260)) displays individual AI usage on a dashboard and manages usage statistics by service on a weekly basis. It also provides a model selection guide tailored to specific work levels and plans to introduce a “Smart Router” in October that automatically selects the most suitable model based on the query and the nature of the task.

SKTelecom(017670)Rather than simply reducing token usage, the company is focusing on improving efficiency while maintaining quality. It is applying techniques such as prompt caching—which summarizes the beginning of long conversations or reuses repetitive instructions—and Search-Augmented Generation (RAG) optimization to narrow the search scope. It is also implementing a system that allocates tokens differently based on the accuracy required for each task.

◇ As Token Allocations Increase,

So Do

Cost Pressures… Managed Through Routing and Caching

As AI usage expands, so does the

cost

burden. In particular, when agents repeatedly perform tasks such as coding, searching, and planning, the number of tokens required per task and the inference costs also increase.

Rather than unilaterally restricting usage, companies are responding by selecting appropriate models based on task difficulty and reducing unnecessary computations.

LG CNS (#LGCNES) applies model routing and prompt caching to AgenticWorks, routing simple tasks to cost-effective models and complex analyses to high-performance models. Frequently used instructions and internal policies are stored for reuse.

The company believes that as AI usage increases, investment efficiency can be maximized by managing actual throughput and costs together, rather than simply focusing on the number of users.

◇ The Success or Failure of “AI for All” Also Depends on the “Production Cost per

Token

”

To ensure the continued expansion of corporate token-based benefits and the nationwide popularization of AI, it is crucial to lower the production cost per token. For the “AI for All” initiative currently being promoted by the government, the challenge extends beyond simply making AI available to as many citizens as possible to establishing a sustainable, low-cost inference structure. SKTelecom, KTCorporation, and Kakao are scheduled to launch “AI for All” in December.

Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon recently appeared on KBS’s “Sunday Diagnosis Live,” stating, “Every citizen should be able to use AI just as they learn to read Hangul,” and unveiled the “AI for All” initiative, which aims to make advanced AI services available to everyone free of charge. He also highlighted the need to expand investment in computing infrastructure—such as AI data centers and domestically produced AI semiconductors (NPUs)—to support this initiative.

◇ Data Centers Evolving into “Token Factories”… Inference Infrastructure Is the Decisive Factor

The benchmarks for competition are also shifting in the AI infrastructure industry. Beyond simply how many high-priced GPUs a company has secured, the new competitive edge lies in how many tokens can be produced as cost-effectively as possible using the same computing resources.

Cho Kang-won, CEO of More, a company specializing in AI infrastructure software, stated at the “AI Infrastructure Summit 2026,” “The competitive edge of AI data centers has shifted from the race to secure GPUs to how many tokens can be produced cost-effectively using the same resources,” adding that data centers must evolve into “token factories” that produce tokens with ultra-high efficiency.

More unveiled distributed inference software that combines different accelerators—such as NVIDIA, AMD, and domestically produced NPUs—for parallel computation. Reducing token costs is leading to infrastructure competition that encompasses not only GPUs and NPUs but also low-power data centers, compilers, and distributed inference software.

As AI becomes the foundational operational infrastructure across all industries, the question of “how many people use AI” is being joined by “how efficiently AI is used.” For companies, ensuring sufficient AI access and managing costs have emerged as new challenges, while for nations, building an infrastructure ecosystem that lowers the per-token cost of inference has become a key priority.