South Korea’s Ministry of Science and ICT opened bidding on July 13 for “AI for All,” a national project that would offer residents a free, unlimited-use chatbot built on domestic AI models, along with a public AI agent for government information and administrative services.
The schedule is tight. Bids will be accepted until 5 p.m. on Aug. 11, operators are due to be selected by late August, a beta version is scheduled for September, and the full service is set to open nationwide in December.
Government support comes in GPUs, not direct cash
Rather than issuing a straight cash grant, the government will provide computing resources. The tender says 512 NVIDIA B200 GPUs will be allocated this year to the winning bidders.
If two companies are selected, each would receive 256 GPUs. If three are chosen, the top-ranked operator would get 256 GPUs and the second- and third-ranked operators would receive 128 each. Final allotments may still be adjusted based on each bidder’s needs and the evaluation outcome.
South Korean media, citing government benchmarks, said the subsidy is valued at 6.6 million won per GPU per month. At 256 GPUs, that comes to about 20.275 billion won a year, or roughly $13.5 million. At the full 512-GPU level, the total approaches $27 million.
The companies selected will not receive support entirely for free, but their required matching contribution is far below the government share. Large companies are expected to contribute about 10% of the government support amount, mid-sized firms 6%, and small and medium-sized companies 5%.
The support is not limited to a single year. Starting in 2027, operating costs will be covered directly through the government budget, with support running through Dec. 31, 2030.
At least 80% of output must come from Korean models
The tender requires that more than 50% of service usage come from domestic AI models that meet the standard for an “independent AI foundation model.” It also requires operators to connect models from other domestic companies for more than 30% of usage. The ratios are measured by output tokens.
Foreign models are allowed only for functions domestic systems genuinely cannot handle, including certain high-difficulty reasoning or multimodal capabilities. Those uses are excluded from government subsidies, and excessive reliance on foreign models can lower a bidder’s evaluation score.
In practice, that means at least 80% of output must come from South Korean models, and operators cannot rely on a single in-house model alone. Companies without their own model are still eligible to participate. The ministry said bidders may use other firms’ domestic foundation models through APIs, open-source versions, or licensing agreements.
An information session held on July 21 drew about 100 companies, including LG U+, SK Telecom, Naver Cloud, Kakao, and a number of AI startups. Officials expect roughly 10 companies to submit bids.
User experience will matter more than benchmark scores
The government said the evaluation will not revolve around technical benchmark performance alone. Instead, it will place greater weight on “the performance users actually feel.” After launch, the project will be reviewed once a year, and operators with weak user numbers may be replaced.
That makes user scale a core operating metric, a different framework from the one large South Korean companies have typically faced in government contracts.
The local-model requirement ties back to a separate national AI program
To understand what counts as a domestic model, the report points back to South Korea’s 213.6 billion won “independent AI foundation model” program, valued at about $142 million and described by local media as the country’s “national representative AI” initiative.
Under that program, five teams are to be selected in 2025: Naver Cloud, LG AI Research, SK Telecom, Upstage, and NC AI. Their target is to build homegrown models that reach 95% of the performance of the latest global models.
Questions remain over whether 512 B200 GPUs are enough
Opposition voices in South Korea have openly challenged the scale of the compute allocation. The criticism says 512 B200 GPUs account for only 4.4% of the country’s roughly 11,600 B200-equivalent compute resources already committed to independent model efforts, academic-industry research, and national programs.
An AI industry executive told South Korean media that once 512 GPUs are split among two or three operators, “there is a significant possibility” of insufficient compute capacity during nationwide peak usage.
The concern is not limited to chatbot traffic. The service is also expected to handle multimodal requests and connect with administrative systems, which would increase the computing cost of each query.
The ministry pushed back on that argument, saying this is not a base-model training project and does not require thousands of GPUs. Companies are also expected to add infrastructure on their own. In the early stage, the ministry said, the service will focus on basic chat functions and simple public services, and token consumption “will not be too excessive.”