'AI for everyone' gov't initiative faces feasibility questions

Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon, center, speaks during a policy briefing at Yeongbingwan guest house in Cheong Wa Dae in Seoul, Thursday. Joint Press Corps

The government plans to roll out a general artificial intelligence (AI) service allowing unlimited use for all Koreans without any fees, but the initiative is facing questions about its feasibility, particularly over how it would cover token costs and ensure the operator's financial sustainability.

During a policy briefing with President Lee Jae Myung Thursday, the Ministry of Science and ICT announced that it will introduce “a ChatGPT-like chatbot service” by using Korea’s proprietary AI model by the end of this year.

The government also plans to integrate public services into the platform, allowing users to search and apply for government services through the AI model.

Once the service gets on track, the government plans to upgrade it next year into an AI agent that will offer greater accessibility than a chatbot, with the goal of providing every Korean with a personal AI agent.

"In addition to a general-purpose AI chatbot service, we will continue to expand public AI agent services," Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon said. "Starting in 2027, we will advance the service into a one-agent-for-every-citizen model.”

The ministry has already begun accepting applications from service providers. It plans to select two or three operators in August, launch a beta service in September and roll out the full service in December.

ChatGPT, Claude and Gemini app icons are seen in this June 5 photo illustration. Reuters-Yonhap

To qualify, at least 50 percent of the AI models used for the service must be domestically developed foundation models that were trained independently. In addition, operators will be required to allocate at least 30 percent of the service's AI model usage to domestically developed models created by other companies.

The timeline is tight, with selected operators expected to launch a beta service just one month after being chosen. However, the government believes the schedule is achievable because leading domestic companies have been developing their own foundation models under the national foundation AI model project, a separate project also led by the government.

"The project had already been known within the industry before it was officially announced, and since companies have already been developing chatbot services based on their own models, the timeline for turning them into actual services appears feasible," an official at an AI service company said. "What remains unclear is how token costs will be covered and how the service's financial sustainability will be maintained."

Nvidia's Blackwell architecture / Courtesy of Nvidia

For the project, the government plans to allocate 256 of Nvidia's B200 graphics processing units (GPUs) to both operators. The operators will be required to estimate expected user numbers and query volumes on their own and provide matching investments accordingly.

The government's allocation of 256 B200s per operator is substantial in absolute terms, but it is difficult to conclude that the capacity will be enough for a general-purpose AI service geared toward the entire population.

According to Nvidia's benchmark data, a single B200 GPU can generate about 10,000 output tokens per second when running Llama 3.3, a medium-sized model with 70 billion parameters. Assuming each user receives responses at a rate of 50 output tokens per second, one GPU could simultaneously serve about 200 users, meaning 256 GPUs could theoretically handle responses for around 51,000 users simultaneously.

Assuming each user asks four questions a day, with each response averaging 500 tokens, and peak demand reaches six times the daily average, this translates into a theoretical capacity of about 18.4 million daily active users.

The government estimates that around 23 million people in Korea, or 44.5 percent of the population, use generative AI. Given that most already rely on overseas services such as ChatGPT, Gemini and Claude, the infrastructure could be sufficient in terms of initial user volume if the government-led service remains at supporting roles.

However, this assessment is based on the assumptions that operators use lightweight and highly optimized models, that the number of users does not surge immediately, and that queries remain short and simple.

Given that some Korean companies are developing models with more than 500 billion parameters, the number of users that each GPU can serve simultaneously could fall sharply.

gettyimagesbank

SK Telecom, for instance, has developed A.X K1, a model with 500 billion parameters, while also releasing a lighter model with 7 billion parameters. Although smaller models can respond faster and serve more users with fewer GPUs, they may struggle with complex queries or tasks such as integrating public services as wanted by the government. This can potentially reduce the overall quality of the service, forcing operators to face a dilemma between maintaining service quality and sustaining a free, unlimited service.

Another uncertainty is that the service is primarily aimed at people who are not familiar with AI. As conversations become longer, token usage and related costs can rise significantly. The promise of free and unlimited access could also invite existing paid AI users’ hefty workloads, and the government's plan to expand the service into AI agents may increase operating costs.

The government said it plans to begin providing state support for GPU costs and service operations from next year, but the size of the funding has not yet been determined.

The government believes operators will be able to recover costs by introducing their own revenue models, such as advertising. Industry officials, however, said the advertising model remains largely undefined and questioned whether it would be appropriate to include ads in a government-operated AI service.

"More details are expected to emerge as the project moves forward," the official said. "If the service targets the whole nation, companies may be willing to participate despite the cost burden given the nonfinancial advantages such as access to large-scale user data."