Korea is preparing to deploy a free, unlimited artificial intelligence service for its citizens, a plan underpinned by a significant investment of 256 Nvidia B200 graphics processing units allocated to service operators. The Ministry of Science and ICT intends to launch a ChatGPT-like chatbot by year’s end, with ambitions for a model by 2027.

However, the initiative faces questions regarding financial sustainability, as an official at an AI service company notes, with uncertainty surrounding how “token costs will be covered.” The government estimates 23 million Koreans, or 44.5 percent of the population, already use generative AI, raising the stakes for this ambitious expansion of access.

Korea’s 2027 “One Agent Per Citizen” AI Service Vision

This substantial allocation demonstrates a proactive approach to providing the necessary computational infrastructure for a nationwide artificial intelligence service, moving beyond pilot programs to a fully-fledged deployment. The B200, NVIDIA’s latest generation GPU, is designed for large-scale AI workloads and is central to the government’s plan to deliver a consistently available service, the company says. Beyond hardware, the Ministry of Science and ICT mandates that 30 percent of AI model usage within the service must be allocated to domestically developed models created by companies other than the primary operators.

This requirement is a deliberate strategy to foster competition and innovation within the Korean AI ecosystem. NVIDIA’s NVQLink, an open architecture for integrating quantum processors with GPU supercomputers, could play a role in supporting these diverse models and accelerating development, though the source does not detail specific integration plans.

Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon outlined the government’s long-term vision, stating, “In addition to a general-purpose AI chatbot service, we will continue to expand public AI agent services,” and further, “Starting in 2027, we will advance the service into a one-agent-for-every-citizen model.” The project had already been known within the industry before its official announcement, and since companies have already been developing chatbot services based on their own models, the timeline for turning them into actual services appears feasible. More details are expected to emerge as the project moves forward, but the government’s commitment to providing universal access to AI services is clear.

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.

Nvidia B200 GPU Allocation & Potential User Capacity

This hardware investment underpins the ambitious initiative, but raises questions about sustained performance under anticipated user loads. Nvidia’s B200, released earlier this year, delivers significant performance gains; however, the actual number of concurrent users the infrastructure can support remains dependent on model complexity and query volume, according to the company. Nvidia benchmarks indicate a single B200 GPU can generate approximately 10,000 output tokens per second using the Llama 3.3 model, a 70 billion parameter system.

Assuming a response rate of 50 tokens per second, one GPU could theoretically serve around 200 simultaneous users. Consequently, the allocated 256 GPUs possess a theoretical capacity to handle roughly 51,000 concurrent users. This translates to an estimated 18.4 million daily active users, assuming four queries per day averaging 500 tokens each, with peak demand six times the daily average.

Given that approximately 23 million Koreans already utilize generative AI, the initial infrastructure may prove sufficient if the government-led service initially supplements existing platforms. However, this calculation relies on several assumptions. The development of larger, more complex models, such as SK Telecom’s A.X K1 with 500 billion parameters, could significantly reduce the number of users each GPU can support simultaneously. This presents a trade-off between service quality and maintaining a free, unlimited offering.

The potential for increased token usage with longer conversations and the influx of existing paid AI users further complicates capacity planning. Nvidia’s partnerships also play a role in enabling this infrastructure, the firm reports. The 2023 integration of Quantum Machines OPX with Nvidia H100 GPUs, and the 2025 designation of QuEra Computing as an NVQLink launch partner, demonstrate the company’s commitment to hybrid quantum-classical workflows.

The government intends to begin providing financial support for GPU costs and service operations next year, but the specific funding amount remains undetermined. Industry officials question the viability of relying solely on advertising revenue to offset costs, especially for a government-operated service.

“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,” the official added. The government expects more details to emerge as the project progresses, but the long-term financial sustainability of the initiative remains a key concern.

What remains unclear is how token costs will be covered and how the service’s financial sustainability will be maintained.

Token Costs & Financial Sustainability of Unlimited Access

These allocations, combined with planned state support for GPU expenses beginning next year, aim to establish the infrastructure for a free, unlimited artificial intelligence service accessible to all citizens. However, the scale of sustaining such a service hinges on managing token costs, a concern voiced by industry officials. This requirement, while intended to support local innovation, introduces a further layer of complexity to cost management, as domestically produced models may have different resource demands than established, commercially available options.

The government anticipates operators will estimate user numbers and query volumes to justify matching investments, but the accuracy of these projections remains uncertain given the novelty of a truly unlimited, free service. The existing adoption of generative AI within Korea, with 23 million people already utilizing such tools, presents a unique dynamic.

This substantial existing user base suggests a considerable demand for the service, but also raises questions about the incremental impact of offering free access to those already paying for AI solutions. Longer conversations, a likely outcome with unlimited access, will drive up token usage and associated expenses, potentially exacerbating financial pressures. The government’s plan to expand the service into AI agents may further increase operating costs, demanding efficient resource management and potentially necessitating the exploration of revenue models beyond advertising.