South Korea's AI Gamble

Author | Wan Lianshan

Recently, South Korea fired the first shot in the global "AI public utility" initiative.

According to the "AI for All" plan announced by South Korea's Ministry of Science and ICT, the goal is to provide free, unlimited generative AI services to all 51 million citizens, explicitly promising "no subscription fees and no token usage limits."

Where does the money come from? It is directly tied to registration, public housing applications, and tax filing systems, with operating costs covered directly by the fiscal budget.

Progress has already moved from slogans to selecting participants: on Aug. 28, three consortia led by SK Telecom, Kakao, and KT were selected; on Sept. 4, the government held a project launch meeting to promote nationwide use by the end of the year. 

On the surface, this looks like a government-led money giveaway.

But a closer look shows it is far more than just "welfare."

AI is a tool, and it will also become a means of production.

Spending money and guiding policies to get everyone using AI is equivalent to distributing the future means of production into everyone's hands.

Regardless of success or failure, this is undoubtedly a grand social experiment.

01 First, look at the real problem.

South Korea has indeed enjoyed hardware dividends in the AI era, but who has heard of any South Korean domestic large model?

According to Wiseapp·Retail data: in April 2026, South Korea's ChatGPT monthly active users were about 23.45 million, Gemini about 8.45 million, and Claude about 2.41 million.

What about domestic models? Almost no one uses them.

AI competition has obvious scale effects. Without users, products are hard to improve, and revenue cannot support R&D; when products lag behind, users are even less willing to come.

After one cycle, the most proficient skill of startup teams may be reduced to revising financing plans.

Therefore, the "AI for All" plan stipulates that at least 50% of services must use South Korean models that meet the domestic foundation model standards, and at least 30% must use models developed by other South Korean companies; foreign models may be used to a limited extent for necessary functions, but those parts do not receive government support. 

At the same time, public procurement, industrial support, and technological sovereignty are tied together.

Domestic companies gain computing power support and application scenarios, citizens gain free tools, and the government hopes to retain domestic R&D, operations, and service integration capabilities.

The government pays and locks 80% of computing power for domestic models, effectively helping the domestic industry cross the most difficult initial stage.

How much exactly?

On Sept. 4, MoneyToday reported, based on project documents obtained from the National Assembly office, that the plan for 2027-2030 is 250 billion won per year.

170 billion won for GPU rental, 30 billion won for the three consortia, and 50 billion won for expanding the AI agent ecosystem.

Spread across South Korea's population of about 51.61 million in 2026, that is about 4,800 won per person per year.

This amount is clearly insufficient, even if doubled.

So how can this budget work?

It relies on mixing different usage intensities and allocating computing power across different tasks. 

In the plan announced by SK on Aug. 19, it explicitly mentioned that simple questions are handled by lightweight models, while complex tasks use high-performance models.

Model routing, combined with inference optimizations such as caching, batch processing, and quantization, can reduce per-task costs.

The service is also prepared to retain commercial revenue.

According to a report by Financial News on Sept. 7, the government is considering allowing moderate advertising and discussing charging for premium services after the second half of 2027.

So the fiscal path should be: first invest in initial computing power, then support services with the annual 250 billion won plan, and finally rely on broader AI research and infrastructure investment.

Of course, the final total cost will also need to include corporate investment and expansion from actual usage, which cannot be estimated at present.

However, once this path is successfully established, South Korean companies will undoubtedly gain a first-mover advantage in domestic government and life services.

To turn this convenience into economic growth, South Korea still faces a challenge.

02 According to research data released by the Bank of Korea in June 2026, average AI use reduces working hours by 3.8%, about 1.5 hours per week, corresponding to a potential productivity increase of about 1.0%.

But the study did not find that the saved time generally translates into actual output growth.

The effect is more pronounced among self-employed individuals, professionals, and high-intensity AI users.

This reflects the productivity J-curve in the diffusion of general-purpose technologies: tools enter first, followed by training, process restructuring, and organizational adjustment, with benefits gradually emerging.

Free access for all can lower the technical threshold, but to truly unleash output, companies must be willing to adjust work methods, and employees must have the motivation to save time.

Otherwise, while the average "AI usage" increases, the average overtime hours may also increase.

Deeper changes will occur in the distribution of benefits.

The government pays for users, and the first to secure definite demand are computing power suppliers, model companies, and platform operators.

For small businesses, public platforms may reduce development and customer acquisition costs; for consumers, they may reduce transaction costs.

But when chat interfaces further control reservations, shopping, and service recommendations, platforms will also gain new distribution control.

Who gets recommended, who pays commissions, and who can access will all affect the competitive landscape.

This is precisely why fiscal subsidies must consider usage effectiveness.

According to a report by Financial News on Sept. 7, GPU support will be allocated based on differences in users and usage performance, with the first evaluation possibly as early as March 2027.

In specific implementation, task completion rates, error rates, and actual time saved all need to be included in the evaluation.
 

The labor market will also adjust accordingly.

AI assistance can enable less experienced people to complete more tasks, but it may also compress positions originally reserved for newcomers to practice.

Older people can more easily access services, while whether young people can continuously accumulate skills and earn income will also determine how far this reform can go.

After tools become widespread, training, job adjustments, and income distribution must keep up.

The current assessment is that South Korea has the conditions to make "AI for All" a practical public service and expand the market for domestic models.

Existing user channels, fiscal support, and local service connections provide the foundation for its implementation.

Other countries may also see a possibility: through public procurement, transform AI from a personal subscription product into a social service available to all.

But to achieve widespread productivity dividends, longer-term organizational and institutional changes are needed.

This is a more profound change.

03 The greatest aspect of this experiment is the attempt to deliver the most important production tool of the future to every citizen in the form of a public service.

South Korea's Ministry of Science and ICT explicitly stated in the project announcement that the direction of continuous upgrading is "one AI agent per person," allowing citizens to participate in economic and social activities through their own agents and share the benefits brought by AI.

Giving money increases current purchasing power; giving production tools attempts to enhance the ability to generate income in the future.

The role of fiscal expenditure extends from subsidizing consumption to supporting ordinary people's participation in production.

For example, a young person wants to start a small business and needs to do promotion, organize customer needs, analyze orders, and occasionally modify a website.

Even if AI tools can be used, the token consumption is a considerable expense.

If public AI can assist with part of this, the minimum investment to start a business may decrease.

This is not just lowering the threshold; it expands the production resources an individual can mobilize.

A person's experience, judgment, customer relationships, combined with affordable and effective tools, have the opportunity to support businesses that were previously impossible.

This change may also affect the bargaining power of workers.

If a person can only work relying on software, equipment, and organizational systems provided by the company, the cost of leaving is relatively high.

If they can independently take on some business with the help of public AI, or more easily learn new skills and switch tracks, then when facing employers, they have more external options.

Having the ability to choose another path gives you stronger confidence to negotiate better terms with more leverage.

This also pushes the issue of social distribution a significant step forward.

When discussing income distribution, we often think about how the government adjusts through taxes and welfare after wealth is created.

What "AI for All" attempts is to intervene at an earlier stage: before wealth is generated, help more people obtain the tools to participate in wealth creation.

And in what capacity ordinary people participate in growth: merely as consumers purchasing products, or as individuals who can use new tools, provide services, and run businesses.

By covering part of the tool costs, the government reduces the restrictions on acquiring production capacity caused by family background, enterprise size, and personal income.

When ordinary people have an extra production tool in hand, they gain an additional livelihood path, more confidence in career choices, and ensure they benefit from the gains of technological progress.

Everything will get better.