finops.work

Deep dive · 7 parts

Free AI for a nation

An admirable ambition, and a complex test of public value. What the service delivers, and how it earns trust, protects privacy and serves sovereignty, will decide what it is worth.

TL;DR

  • South Korea’s AI for All is in a closed test, with a public launch planned for December 2026. Three consortia run it on 512 state-owned Nvidia B200 GPUs; the proposed 2027 budget of 250 billion won ($186 million), mostly to lease about 2,000 more, has yet to pass.
  • The government’s stated reasons are a gap in AI skills it does not want to become a social gap, and dependence on foreign services. Whether a free domestic chatbot answers either is my question, not its claim.
  • This series is a deep dive into what that hardware might deliver, built from benchmark scenarios and what is known of the programme today, not from measurements of the service. The technology will keep improving; how fast those gains reach this service is uncertain.

A closed test of AI for All (모두의 AI) was due to begin on 16 October 2026, for staff of the consortium companies and public bodies; a public beta is under review, and the official launch is planned for December 2026.[3] The promise is a chatbot for every citizen with no charge and no usage limit, and a public agent that finds the public services you are due and applies for them.[27] Three consortia won: SK Telecom, Kakao with the LG companies, and KT. They share 512 Nvidia B200 GPUs the government owns; the 2027 proposal adds 250 billion won ($186 million), most of it to lease about 2,000 more.[11]

Why the government is doing it

It began as a campaign pledge in April 2025: a free “Korean ChatGPT” for every citizen, part of an “AI basic society”.[30] The ministry’s tender call of July 2026 gave two reasons. Using AI has become a basic skill, like reading or arithmetic, and the gap in who uses it must not become a social and economic gap; about a third of citizens do not use AI, and most who do are on free foreign services.[27] And those free services limit use and leave citizens exposed to foreign companies’ decisions: a price rise, or a service withdrawn.[27] Hence “our” AI for everyone, built mostly on domestic models (chapter 5).

The longer plan is agents: from 2027 the chatbot is to grow into agents that act for people, until every citizen has one.[27] The minister calls the service the calculator and computer of the AI age.[27] At selection and kick-off two more aims were named: a chatbot that surpasses foreign AI, and a new domestic AI market for companies.[12, 18] Four goals, then: access and skills for everyone, independence from foreign providers, public services by agents, and a market for domestic models and companies.

What I make of it

This is my reading, and it is mostly questions; the chapters work through them.

Access. The surveys the ministry cites say that for most non-users the barrier is skill, not price, and free foreign chatbots already exist (chapter 6). Will a free domestic one reach the third who don’t use AI, or mostly the people who already do?

Capacity. “No usage limit” is a promise per person on a fixed number of GPUs. Can the fleet keep it at the busiest hour, and at what speed (chapters 2 and 3)?

Sovereignty. The ministry’s second reason is dependence on foreign companies, and the rules answer it with model shares: at least 80% of the answers written by Korean models (chapter 5). Would that limit foreign influence over the country’s data and the insights drawn from it? Where a model’s weights come from is one part of that. The others are where prompts are processed and kept, who can read them and under which country’s law, what the operators may do with them, and which contracts and hardware the service depends on (chapters 4 and 5). A Korean model on a Korean platform keeps the weights at home; it doesn’t by itself keep the prompts private, and the chips are American.

Public value. The subsidy runs through the three largest platforms. Does that build a market for domestic AI or narrow one, and what will the state count as success (chapter 6)?

Each number says what kind it is when it first appears: reported, measured, a scenario, an assumption, or my view. The small numbers after a claim point to the sources in the appendix.

The parts

  1. Budget per citizen is not cost per active user Spread over every resident, the proposed 2027 budget for free AI is a few thousand won a year. That is the fiscal burden of making it available to everyone, not what the service costs for each person who uses it.
  2. What 512 GPUs can deliver, and what they draw The promise is unlimited use; the supply is a fixed number of GPUs. What they can write is a scenario built from benchmarks of stand-in models, a daily average says nothing about the busiest hour, and electricity is the smaller bill.
  3. Agents are a different workload, and waiting is the only price One measured coding-agent session wrote about 95 times the output of an assumed chat user's day, which is why agents need their own forecast. The rules forbid charging and usage caps, so if peak demand exceeds usable capacity, slower answers and queues do the rationing.
  4. Who pays for what, and what the deal rewards The state leases the GPUs and grants each consortium 10 billion won; the operators carry the service and must find revenue against rivals that are already free. What that rewards is a hypothesis about incentives, not a measured outcome.
  5. Sovereign models on leased American chips The rule that makes the service sovereign counts the output tokens each model writes, and says nothing about the chips. Of the eight listed models, two were trained from scratch by their makers' account, three grew from their makers' earlier models, and for three the start isn't published.
  6. My view: what the money should buy, and what it could crowd out Free access is not useful adoption, and the subsidy runs through the platforms that already have the users. I would judge the programme on quality-checked tasks, their full cost and what they add, and ask the winners for open terms in return for the GPUs.
  7. Appendix: evidence, arithmetic and sources Every number in the series, what kind it is, how it was computed and where it comes from. The chapters cite the numbered sources at the end of this page.