South Korea's Ministry of Science and ICT opened bidding this week for "AI for All," a government-funded chatbot that will be free to every citizen with no usage caps. The project competition runs until August 11, with two or three private operators selected by the end of that month.
The GPU math doesn't work yet
Seoul is handing the winners a shared pool of 512 Nvidia B200 accelerators to get the thing running. That is the number worth staring at. South Korea has roughly 52 million people, and a technical analysis puts a 512-GPU cluster serving a 7-billion-parameter model somewhere in the range of 2,500 to 7,500 concurrent responses.
Call that a pilot. The government does not, but the arithmetic does.
The plan is that the rest of the country's compute picks up the slack later: the ministry's own fleet of around 13,000 GPUs, plus capacity from the operators of the National AI Computing Center. Whether any of that is contractually available to a free consumer chatbot in September is not something the announcement addresses. Budget support for running the service arrives only in 2027, which means the first winter is on the vendors.
The 50% rule
Here is the condition that actually shapes the product. At least half the system must run on domestic Korean foundation models meeting the ministry's independent-model standard. On top of that, at least 30% has to come from a Korean model built by someone other than the winning operator. So a company cannot just plug in its own LLM and collect the check.
That routes inference through the five consortia Seoul already bankrolled: Naver, LG AI Research, SK Telecom, Upstage, and NC AI. It is a closed loop, and it is deliberate. The sovereign model program selected those teams last year with the stated goal of hitting 95% of frontier performance.
Ninety-five percent of a moving target, a year later, is a strange thing to promise citizens who can already download ChatGPT. Ministry Minister Bae Kyung-hoon called the foundation-model program "the launchpad for 'Everyone's AI,'" which was true in the sense that the launchpad now exists and the rocket is a procurement notice.
What sovereignty means here
Critics have pointed out the obvious: the chips are Nvidia's, the software stack is CUDA, and the training infrastructure is American. Domestic Korean AI models sit on top of hardware Seoul cannot manufacture and does not control. The project answers this by drawing the sovereignty line at the model layer instead of the silicon, which is either a pragmatic compromise or a rebrand, depending on how generous you feel.
The operators are also expected to fund themselves eventually, and the ministry has said the quiet part: companies should build revenue models using the prompt data they collect from users. Free service, monetized queries. Not a new bargain, just a state-sanctioned one.
Applications close August 11. Beta opens in late September.
