Two Tracks, One Finish Line: Korea Confirms Frontier AI Push While Saving Its Sovereign Model Project

For most of late September, South Korea’s AI policy looked like it was about to eat itself. Headlines in the local press warned that the government was “weighing halting” its national representative AI foundation-model project — the flagship “Sovereign AI Foundation Model” initiative that had been narrowed to a three-team race just weeks earlier. Industry observers openly questioned whether the whole enterprise would be folded into something new, or quietly cancelled.

On September 30, Science and ICT Minister Bae Kyung-hoon put an end to the speculation. The sovereign foundation-model project will continue, he confirmed — but it will no longer carry the country’s frontier ambitions alone. Starting next year, the government will launch a separate, substantially larger frontier AI initiative, and the two programs will operate as what officials are now calling twin pillars of Korean AI strategy.

What was announced

The confirmation, reported by Seoul Economic Daily and MLex on September 30, resolves a week of contradictory signals. Speaking to reporters, Bae said Korea’s sovereign AI foundation-model project — known as the K-AI project — “will continue,” while the country simultaneously pursues a dedicated frontier AI program designed to produce models comparable to the world’s leading systems.

The word “separate” is doing a lot of work here. Earlier reporting from Chosun Biz on September 28 and Maeil Business on September 29 had suggested the existing foundation-model project might be halted mid-stream, with its technology and personnel absorbed into the new frontier effort. Fn News reported a plan to combine roughly KRW 5 trillion in next-year frontier AI funding with the assets of the existing project. Industry voices then pushed back hard: in a September 30 Chosun Ilbo piece, AI industry representatives argued that both programs matter — the frontier track “will break through South Korea’s technological limits,” while the homegrown foundation model serves as a second pillar anchoring domestic AI adoption.

The minister’s answer effectively splits the difference. Nothing gets cancelled; instead, Korea doubles down on both tracks at once.

The numbers behind the pivot

The frontier initiative is not a modest line item. Yonhap reported on September 8 that the Ministry of Science and ICT plans to invest about 4.7 trillion won (roughly US$3.5 billion) in a project to develop a homegrown frontier AI model — a figure that had already grown from the 3.5 trillion won outlined in July, when Bae first sketched the two-track strategy at his first-anniversary press briefing. Reports since then have floated figures as high as 5.59 trillion won for the frontier budget starting next year, though the final allocation will be settled in the annual budget process.

Compute is the other big number. The July briefing put the hardware need at roughly 10,000 GPUs — Bae estimated around 10,000 Nvidia Vera Rubin GPUs costing approximately KRW 3.5 trillion on their own. That is a serious bet for a country whose earlier sovereign AI compute earmarks (about 1.5 trillion won for 13,000 GPUs, per ministry planning documents) were widely judged insufficient for frontier-scale training runs.

Why the foundation model project was under pressure

To understand why the project nearly got shelved, it helps to look at its design. Launched in August 2025, the Sovereign AI Foundation Model project put five consortia — led by Naver, LG AI Research, SK Telecom, NC AI, and Upstage — on government funding, GPUs, and datasets, with teams to be evaluated and eliminated through 2026 and into mid-2027. By August 2026 the field had narrowed to three teams. The government’s original target was telling: a Korean foundation model reaching “about 95% of the performance of frontier-class AI models.”

That 95% framing was always the project’s quiet vulnerability. Sovereign AI, in the procurement sense — Korean models serving Korean public institutions and businesses, with usage quotas requiring that at least half of government AI usage flow to qualifying domestic models — does not strictly require beating GPT or Gemini. But the global race moved faster than the project’s elimination schedule. By mid-2026, with U.S. labs shipping near-Astra intelligence at a fifth of flagship prices and Chinese open-weight models closing the gap from the other side, a “95% as good” domestic model looked less like sovereignty and more like a consolation prize. Chosun’s September 16 analysis was blunt: opaque review processes and tight funding had split the government’s AI portfolio, with the new frontier project (~4.7 trillion won) threatening to overshadow the foundation-model track.

The strategic logic: two pillars, not one bet

Minister Bae has been consistent about the underlying rationale. “We cannot afford to slow down the pace of AI development,” he said in mid-September, arguing that the gap between countries that have secured AI capabilities and those that have not will harden into permanent economic and security hierarchies. He has also framed a sovereign frontier model as a matter of avoiding technological subordination — preventing Korea from becoming dependent on foreign frontier models for its industrial base.

Keeping both tracks alive is a hedge against two different failure modes. Kill the foundation-model project, and Korea loses the ecosystem scaffolding — domestic model developers, Korean-language datasets, public-sector adoption pipelines — that the past year of investment built. Kill the frontier ambition, and Korea consigns itself to being a fast follower in the layer of the stack where the actual capabilities are being defined. The two-pillar structure lets the government pursue frontier-scale training runs without dismantling the adoption-focused foundation work, and without forcing a single consortium to serve two very different missions.

It also mirrors a broader pattern among middle-power AI nations. Sovereign AI rhetoric has shifted from “our own ChatGPT” toward a portfolio view: sovereign access (compute, foreign frontier APIs on acceptable terms), sovereign capability (domestic frontier training), and sovereign diffusion (models embedded across the economy). Korea’s announcement is one of the clearest articulations yet of that three-way split becoming explicit policy.

What to watch

Three things will determine whether this is strategy or sprawl. First, the final 2027 budget: if the frontier line item lands near the 4.7–5.6 trillion won figures being floated, the project is real; if it shrinks, the two-pillar language was cover for a quiet wind-down of ambition. Second, GPU procurement — the 10,000-GPU estimate assumed Vera Rubin availability on schedule, and global accelerator demand remains brutal. Third, what happens to the three surviving foundation-model consortia: if their funding and compute access grow alongside the frontier program, the ecosystem reading holds; if they wither on the vine while the new project hoovers up talent, the industry’s warning about “two pillars” becoming “one pillar plus a mascot” will have been prescient.

For now, Korea’s sovereign AI experiment lives — bigger, more expensive, and more ambitious than originally designed. The 95% model is no longer the goal; it is the floor.