South Korea’s Ministry of Science and ICT opened bidding on July 13, 2026 for a free, unlimited AI chatbot for all 52 million residents, with a beta service due in September and full rollout by December. Firms that win must route at least 50 per cent of user queries through domestic Korean foundation models, with an additional 30 per cent from other domestic models.
The project’s pilot runs on just 512 Nvidia B200 GPUs — an American silicon stack that no domestic-model mandate can bypass. That hardware leash exposes a vulnerability the equity headlines skip.
The tender rules are explicit: any consortium building South Korea’s free AI assistant must use a Korean foundation model meeting ministry standards for at least half of all inference. An additional 30 per cent must draw from other domestic models. Foreign models, if permitted at all, receive no government subsidy.
The same 81-page call for proposals says nothing about the chips underneath. Every GPU provisioned for the pilot is an Nvidia B200. The software stack is CUDA, a proprietary American layer that no South Korean laboratory can replace. The hardware chain runs through Santa Clara, not Seoul.
Bae Kyung-hoon, the deputy prime minister and science minister, calls the project “the calculator and computer of the AI era.” The government’s framing positions universal AI access as a basic right, akin to education. President Lee Jae-myung has promised daily support so all citizens “can enjoy the benefits of AI technology development.”
The question the tender cannot answer is whether a handful of smaller domestic models — none above 32 billion parameters — running on a pilot-scale GPU fleet can deliver parity with the free foreign services most Koreans already use. The equity promise stands or falls on a supply chain it barely controls.
The mandate that hardware cannot enforce
The five eligible consortia — Naver Cloud, LG AI Research, SK Telecom, Upstage, and NC AI — offer models in the 7‑to‑32‑billion parameter range. Frontier models from OpenAI and Google run above 100 billion. The government targets 95 per cent of frontier performance; independent benchmarks on general tasks remain limited.
HyperCLOVA X, trained on 6,500 times more Korean data than GPT‑4, outperforms it on Korean-language benchmarks. SK Telecom’s AX 3.1 Lite reaches roughly 96 per cent of the top score on the KMMLU2 reasoning test. Those numbers are narrow and language-specific. They do not settle the broader capability question.
The pilot’s 512 Nvidia B200 GPUs can process an estimated 2,500 to 7,500 simultaneous chatbot responses. For a population of 52 million, that is a rationing threshold, not a universal-access capacity. A larger national GPU fleet — 13,000 ministry-owned chips now, a target of 50,000 by 2030, plus over 7,000 B200 units already allocated to cloud operators — exists on paper. The tender makes no commitment to expand the pilot’s compute mid‑contract.
“AI for Everyone is more than just a service; it is the calculator and computer of the AI era,” Bae Kyung-hoon said. Ministry officials told bidders that access and capability gaps “could lead to new social and economic inequalities, so domestic AI services must expand.” A government survey reveals the gap is already wide: 31.9 per cent of digitally vulnerable groups had used AI, against 59.4 per cent of the general public.
The gap the tender is designed to close is not theoretical.
| Country | Current rule | New rule | Effective date |
|---|---|---|---|
| South Korea | Fragmented across 19 AI bills | AI Basic Act consolidates regulation, creates National AI Committee and AI Safety Research Institute, imposes risk-based duties on high-impact AI systems | January 22, 2026 |
| South Korea | No mandated domestic AI model usage | “AI for Everyone” tender requires ≥50% of inference via domestic foundation models plus ≥30% via other domestic models | Tender opened July 13, 2026 |
| South Korea | AI access not guaranteed | Government commits to free, unlimited AI chatbot for all residents through 2028 | Full service by December 2026 |
| South Korea | No proactive public-service AI agent | Public-service agent to identify individual benefit eligibility and assist applications | 2027 |
Kakao Corp. representatives, leaning on the KakaoTalk platform and the Kanana model, say they aim to deliver “AI services closely integrated with daily life.” The public-service agent, due in 2027, is supposed to find benefits individuals are missing and help them apply — a concrete promise that will test whether a domestic model can handle the messy administrative data of a modern welfare state.
The pilot hardware sits on a larger foundation. The ministry’s own GPU fleet counts 13,000 units, with a target of 50,000 by 2030. A sizeable national deployment — 260,000 GPUs announced at the October 2025 APEC Summit — is backed by Samsung, SK Group, Hyundai, and Naver. Those chips are earmarked for the ecosystem, not the free chatbot. The tender allocates no budget line for scaling the chatbot’s compute once user demand arrives.
The GPU leash no policy can remove
No AI Basic Act clause addresses the CUDA dependency. Every B200 GPU in the pilot requires Nvidia’s proprietary software to operate. In 2023, US export restrictions reminded Seoul that chip supply can be turned into a policy instrument overnight. A national AI access program built on foreign hardware inherits a geopolitical fragility that no domestic-model mandate can nullify.
The closest Western parallel is the European Union’s emerging digital public infrastructure push, but no EU member state yet offers a single nationwide free AI assistant for all residents. Western efforts remain fragmented and market-driven, letting cloud providers and US tech giants control most everyday AI access. South Korea’s scheme hardwires domestic models and GPU support into a state-backed platform.
The beta launch in late September will deliver the first verdict. Strong latency and uptime numbers on 512 GPUs would prove the domestic stack can scale. Weak performance would leave the ministry with a choice: ration access, or quietly accept more foreign models and dilute the sovereignty mandate. Traffic-volume metrics alone cannot reveal which platforms are being served, leaving the true picture harder to verify independently.
Bae calls the project the calculator of the AI era. Calculators run on batteries anyone can buy. This calculator runs on Nvidia chips. The supply chain, not the policy text, will determine how universal “universal” really becomes.
Beyond the headline
The Bigger Picture
South Korea’s universal AI assistant is not a consumer product. It is a structural bet that the state can prevent the exclusion of previous technology waves — broadband, smartphones — from repeating at AI scale. By repositioning access as public infrastructure rather than a private subscription, Seoul challenges the US-led, platform-centric model directly.
The Power Behind It
Real control over the rollout does not sit in Seoul. Nvidia’s hardware roadmap and global export policies will shape capacity far more than any ministry tender. The domestic-model quota creates a market for Korean labs, but the underlying compute layer remains American. That imbalance gives foreign suppliers quiet leverage over both social policy ambitions and industrial strategy.
What Isn’t Being Said
Official documents focus on access and innovation. They sidestep how prompt logs and interaction data from tens of millions of citizens will be governed — whether domestic AI firms can monetise those datasets, how long they are retained, and how cross-border data flows intersect with US-based hardware and software. Including those questions changes the project from a pure equity story to one about long-term data power.
The bets that will be tested within weeks
With bids due August 11 and a beta service weeks away, the project moves from policy to procurement. Four groups face immediate decisions.
- Western AI Policy Professional
South Korea’s universal-access model, backed by a dedicated AI Act and direct GPU subsidies, is a live policy experiment. Compare it to the EU’s digital public infrastructure framework; no member state has replicated it yet. The outcome will either strengthen or weaken the case for state-backed free AI access in Western capitals.
- US-based Semiconductor Procurement Manager
Every B200 GPU in this project comes from Nvidia, and the CUDA stack is non-negotiable. Track Nvidia’s investor and regulatory filings on export controls and data-centre deployments. A tightening of chip export rules would hit this program faster than any Korean policy change.
- Western Investor with APAC Tech Exposure
The five consortia — Naver Cloud, LG AI Research, SK Telecom, Upstage, NC AI — now have a state-guaranteed market for their models. Assess their ability to reach the 95% frontier performance target; the gap between their benchmark scores and real-world user expectations will determine commercial viability beyond the tender period.
- European Digital Public Services Strategist
South Korea is testing a full-stack public AI assistant, from payment to benefit navigation. Monitor the September beta’s uptime and the 2027 agent launch. If it works at scale, it will become the reference case for government-driven, domestic-model AI services that Europe currently lacks.
FAQ
Who can use South Korea’s free AI assistant?
The service is intended for all residents, but you will likely need a Korean mobile number and real-name verification to sign up. The tender calls for nationwide business-to-consumer capability, and access will probably run through major platforms like KakaoTalk or telecom apps. Details on foreign resident access and language support are still undefined in the current tender materials.
Will the government use citizen prompt data for commercial purposes?
Tender guidance indicates companies may develop business models using user prompt data, which could mean commercial exploitation of logs from the free chatbot and public-service agent. Privacy and retention rules will fall under Korea’s Personal Information Protection Act and the AI Basic Act’s risk frameworks, but specific limits on training domestic models with citizen data are not yet spelled out. Future regulatory clarifications are expected.
What happens to the free service after 2028?
The government has committed to free, unlimited use through 2028, with direct budget support for operating costs starting in 2027. After that, documents suggest private co-investment and firms developing their own revenue models, which could introduce premium tiers or specialised paid services. The baseline free assistant is likely to remain, but its scope could shift depending on fiscal decisions and market appetite for enhanced features.
Explainer
- AI Basic Act
- South Korea’s foundational artificial intelligence law, effective January 22, 2026. It consolidated 19 previous AI-related bills, established a presidential National AI Committee and an AI Safety Research Institute, and introduced risk-based obligations for high-impact AI used in healthcare, employment, education, finance, and public safety. The Act frames universal AI access as part of a national “AI basic society” — a concept that repositions the technology as public infrastructure akin to roads or schools.
- HyperCLOVA X
- Naver Cloud’s large language model, trained on 6,500 times more Korean-language data than GPT-4. It outperforms GPT-4 on Korean benchmarks and is one of the five domestic models eligible under the government’s “AI for Everyone” tender. Unlike general-purpose English-centric models, HyperCLOVA X is optimised for Korean-language tasks, which makes it a central piece of the country’s sovereign AI strategy.
- Nvidia CUDA
- A proprietary parallel-computing platform and programming model developed by Nvidia. CUDA is the software layer that allows GPUs to perform general-purpose processing, and it is deeply embedded in virtually every AI training and inference workflow. Because CUDA is closed-source and controlled by a US company, any AI service that relies on Nvidia GPUs — including South Korea’s universal chatbot — inherits a dependency that cannot be replaced by domestic policy.
- AI digital divide
- The measurable gap between groups in the ability to access and use artificial intelligence tools. In South Korea, a government survey showed a 27.5-percentage-point gap: only 31.9% of digitally vulnerable citizens had used AI, compared to 59.4% of the general public. The divide is particularly pronounced across age, income, and regional lines, compounding earlier digital exclusions from the broadband and smartphone eras.
- National AI Committee
- A top-level government body created by South Korea’s AI Basic Act, chaired by the president. It sets national AI strategy, oversees the AI Safety Research Institute, and coordinates the government’s AI infrastructure investments, including the “AI for Everyone” project. The committee replaces what was previously a fragmented policy landscape, centralising decision-making for the country’s AI push.
RELATED STORIES
