Korea's Free AI for Everyone Could Repeat the Public Delivery App Disaster

Seoul’s gamble on sovereign AI

South Korea has selected SK Telecom, Kakao, and KT to build and run “Everyone’s AI” — a government-funded platform promising free, unlimited access to AI services for all citizens. The announcement came August 28 from the Ministry of Science and ICT, which chose three bidders out of six applicants. This year, each will receive 512 GPUs. Starting in 2027, the government will cover the full cost of nationwide service.

On paper, this looks like a bold move toward digital sovereignty. On paper, many things look bold.

The real question is whether a service built by committee and launched by mandate can compete with ChatGPT, which already commands 23.67 million monthly active users in South Korea — more than double second-place Gemini’s 9.47 million. Korean users have already calibrated their expectations to frontier-level foreign AI. They will judge Everyone’s AI within hours of beta launch.

The GPU allocation alone tells part of the story. Five hundred and twelve accelerators per company sounds substantial, but OpenAI reportedly trains its models on tens of thousands of GPUs. Even Google’s mid-tier deployments run well past that threshold. The math suggests Korea is buying a service that will be competitive in narrow domains — customer service automation, document summarization, basic coding assistance — but fundamentally outmatched on the reasoning benchmarks that drive daily user loyalty.

The public delivery app ghost

IT industry observers are not hiding their skepticism. The most cited cautionary tale is the public delivery app, launched under similar pretenses: lower fees for small merchants, free access for consumers, government backing. It lasted barely long enough to become a punchline. Delivery Hero’s Baemin and Coupang Eats were untouchable. Users do not switch apps for patriotism.

The public delivery app was supposed to compete with private platforms on price and convenience while offering transparency that operators like Baemin deliberately obscured. Instead, it launched late, functioned poorly, and collapsed within two years after absorbing roughly 30 billion won in taxpayer funds. Merchants who had been promised preferential treatment found themselves on another platform with similar commission structures. Consumers who had been promised lower prices found limited selection and clunky interfaces.

Everyone’s AI faces the same structural problem. A government contract does not manufacture product-market fit. And unlike the delivery app, where the alternative was just paying commission fees, the alternative here is ChatGPT — already fluent, already fast, already free at the tier most users need.

Why the consortium model matters

The three selected operators are not startups. They are telecom and platform giants with entrenched ecosystems. SK Telecom and KT control Korea’s mobile infrastructure. Kakao owns one of the country’s dominant messaging platforms. Their selection signals a preference for players who can integrate AI into existing services — messaging, mobile plans, enterprise offerings — rather than building something truly new.

That is a strategic choice with real consequences. If Everyone’s AI becomes another channel for these companies to lock users into their walled gardens, it will reinforce the very dependencies Korea claims to want to reduce. The government may be buying sovereignty in name while consolidating private control in practice.

SK Telecom has already been testing AI-integrated phone plans and home broadband bundles. KT has been pushing AI into its B2B portfolio and smart-city projects. Kakao operates an AI assistant embedded in its messaging app with millions of daily interactions. All three have every incentive to treat Everyone’s AI as infrastructure for their own businesses rather than as open public utility.

Alternatively, a properly open architecture could let independent developers build on top of the infrastructure, creating a domestic AI ecosystem that competes on features rather than subsidies. Nothing in the Ministry’s announcement suggests that path is being prioritized. The stated plan emphasizes chatbots, public-service agents, and startup collaborations — all useful, all incremental.

The performance trap

Government AI projects elsewhere have stumbled on the same rock. When a service is positioned as free and nationally backed, users expect parity with paid alternatives. They get neither the incentive structure nor the talent pipeline that drives frontier model development. Most domestic AI efforts in Korea still fine-tune or wrap foreign models rather than train independently. The 512 GPUs per company are a start, but they are not comparable to the thousands of chips deployed by OpenAI, Google, or Anthropic.

The talent gap is equally significant. Korea’s top AI researchers increasingly work for American labs or remote positions. The domestic talent pool that remains tends to specialize in applied engineering rather than foundational model research. Without a credible path to independent frontier development, Everyone’s AI will likely run on wrapped or lightly customized versions of models built elsewhere — potentially including the very models it is supposed to replace.

The Ministry claims the selected bidders presented plans that “surpass foreign AI” in specific use cases. That claim is impossible to verify before launch and easy to disprove once users type a prompt. The moment someone asks a simple math question or requests a nuanced analysis, the comparison becomes unavoidable.

Second-order effects

Beyond the immediate product competition, there are structural risks. If Everyone’s AI succeeds in attracting even modest adoption, it will divert developer attention and user attention away from independent Korean AI startups. Small companies will either build on top of the subsidized infrastructure — becoming dependent on government terms — or compete against a free alternative backed by national budgets. Neither outcome strengthens the domestic ecosystem the policy claims to protect.

There is also the question of data. A national AI platform will process enormous volumes of citizen interactions. Who controls that data? What happens if the government uses usage analytics to inform policy decisions or if data leaks occur? The Ministry has not published a detailed data governance framework, which makes this a live concern rather than speculation.

The subsidy model also creates perverse incentives for the selected operators. Revenue is decoupled from user satisfaction when the government pays regardless of adoption metrics. There is no market discipline forcing continuous improvement. Projects like this tend to degrade slowly rather than collapse dramatically — the Korean delivery app followed exactly that pattern, fading into irrelevance over years rather than failing overnight.

What happens next

Beta testing is expected soon, with full service targeting year-end. Within days of launch, usage data will reveal whether Koreans adopt the platform or treat it as a curiosity. If adoption stays flat, the project becomes another line item critics will cite when debating tech spending. If it gains traction, it will likely do so because one of the three operators successfully integrated it into a product people already use — which means Everyone’s AI becomes Everyone’s AI inside KakaoTalk or SKT’s app, not a standalone public utility.

The political dynamics around this project will shape its longevity far more than product quality. Once the government commits budget through 2027 and beyond, exit becomes politically expensive. Politicians who championed the initiative cannot easily admit failure. Contractors have already invested in hiring and infrastructure. The pattern in Korean public IT is not cancellation but extension with modified promises — a failing project gets reframed as a phased rollout rather than a mistake.

International competitors will continue advancing regardless of Korea’s timeline. By the time Everyone’s AI reaches full deployment, the models it is competing against will be years ahead. This is not a static comparison. A service that seems adequate on launch day will look stale within months.

The consortium model chosen today will determine whether Korea’s AI infrastructure opens up to independent builders or closes into another telco-controlled loop. The GPUs are real. The budget is real. The track record of government-led consumer tech in Korea is not.