The South Korean government is using 512 B200 GPUs to secure free AI for the public, aiming for domestically developed models to capture 80% of usage.

On July 13, South Korea’s Ministry of Science and ICT (과학기술정보통신부, hereinafter “the Science and ICT Ministry”) issued the “모두의 AI” (“Everyone’s AI”) tender to select 2 to 3 private-sector operators. Using domestic AI models, the program will create a universal chatbot that is free for everyone and not subject to usage limits, plus a public AI agent that will serve as a proxy for public services. A beta version will be launched in September, and the service will officially go live in December. Bids close at 5:00 p.m. on August 11. The government will provide 512 NVIDIA B200 cards as subsidies, with a maximum of 256 cards per single operator. Foreign models are only allowed to be used for functions that domestic models cannot accomplish.
(Background: South Korea’s side proposes an $880 billion AI national strategy—semiconductors, physical AI, and data centers, with three core priorities leading ahead)
(Additional context: The Taiwanese government is actively building “sovereign AI.” Wu Cheng-wen, chairperson of Taiwan’s National Science and Technology Council: forming a national special committee for an AI strategy)

Table of contents

Toggle

• The government will be the landlord for compute power, while operators pay only 10%
• Domestic models must account for 80%; Naver has already been screened out
• 512 cards to shoulder service for 51 million people

On July 13, the South Korea Science and ICT Ministry (과학기술정보통신부, hereinafter “the Science and ICT Ministry”) officially launched the “모두의 AI” tender, which can be directly translated into Chinese as “Universal AI.” The government will choose 2 to 3 private companies to build a set of general-purpose chatbots that everyone can use for free and without any usage cap using domestic AI models, and also a public AI agent that can help citizens query and apply for government services.

The timeline is extremely tight. Bid acceptance runs until 5:00 p.m. on August 11 this year. The operators will be selected at the end of August. A beta version will be released in September, and the service will be officially opened to the entire country in December.

This time, the Korean government is not giving out money—it is giving out graphics cards. The tender specifies that this year the government will provide 512 NVIDIA B200 GPUs to the winning bidders. If only 2 companies are selected, each side will receive 256 cards; if 3 companies are selected, the first-ranked operator will receive 256 cards, while the second- and third-ranked operators will each receive 128 cards. The final quantities will then be adjusted according to each operator’s needs and the results of the evaluation.

The government will be the compute-power landlord; operators pay only 10%

How is the subsidy calculated? A baseline cited by Korean media from the government is 6.6 million KRW per GPU per month. Using 256 cards, over a year this is about 20.275 billion KRW, which is roughly $13.5 million. If all 512 cards are fully deployed, it comes close to $27 million.

Winning bidders do not get it for free, but their contribution is not equivalent. The private-sector matching investment ratio is about 10% for large enterprises of the amount supported by the government, 6% for mid-sized enterprises, and 5% for small and medium-sized enterprises. The government pays 90% and operators pay 10%—this ratio basically already makes it clear who is the organizer and who is the contractor.

The support is not only for one year. Starting in 2027, operating expenses will be covered directly from the government budget. The support period will run continuously until December 31, 2030.

Domestic models must take 80%; Naver has already been screened out

The tender requires that at least 50% of the winning bidder’s service must use domestic AI models that meet the standard of “independent AI foundation models.” In addition, the service must also integrate other companies’ domestic AI models of at least 30%. The proportions are calculated based on output tokens. Foreign models are only permitted for functions with high-difficulty inference or multimodal capabilities—functions that domestic models truly cannot provide. They are not included in the government subsidy scope, and using too many will result in deductions from the evaluation scores.

In other words, 80% of the output must be produced by running Korea’s own models, and you cannot put all your eggs in just one in-house model—points will also be available for competitors. This rule resolves two issues at the same time: it hardens and increases the actual usage volume of domestic models, while also preventing a single major company from monopolizing the entire national-level entry point.

Companies that do not have their own models can still participate. The Science and ICT Ministry explained that operators can use other parties’ domestic foundation models through APIs, open-source versions, or licensing agreements. The explanation session scheduled for July 21 is expected to bring about 100 companies, including LG U+, SK Telecom, Naver Cloud, Kakao, and a number of AI startups. The official estimate is that in the end, roughly around 10 companies will actually submit bids.

The evaluation does not follow a run-speed or benchmark test model. The official said that “the performance that users actually feel” is more important than technical benchmarks. The evaluation will be conducted once every year, and operators whose performance is not good enough in users’ eyes will be replaced. Making the number of users the KPI is a completely different set of game rules compared with what Korean big firms that are used to government tenders are accustomed to.

To understand who qualifies as a “domestic model,” you need to look back at Korea’s 213.6 billion KRW (about $142 million) “independent AI foundation model” program—what Korean media refer to as “the national representative AI.” In 2025, 5 elite teams were selected: Naver Cloud, LG AI Research Institute (LG AI연구원), SK Telecom, Upstage, and NC AI. The goal is to develop self-built models whose performance reaches 95% of the latest global models.

512 cards to shoulder service for 51 million people

The opposition in South Korea raised direct skepticism, stating that 512 B200 cards account for only 4.4% of the roughly 11,600 B200-equivalent compute power that South Korea has currently invested in independent models, industry-academia research, and national programs. An AI industry executive told Korean media that once the 512 cards are split among 2 to 3 operating companies, during peak demand for the universal service, the likelihood of “insufficient compute power” is quite high.

Moreover, this service is not just for chat—it also needs to handle multimodal requests and connect with administrative systems. The compute consumption for each query is multiplied. The Science and ICT Ministry’s rebuttal is that this is not a project to train foundation models, so it does not need thousands of GPUs as a matter of course. Operators already have to supplement the basic infrastructure themselves, and in the initial stage it covers only basic chatting and simple public services, so “the token consumption volume will not be too excessive.”