Roche has its AI factory. Eli Lilly has its. Now South Korea wants one too — and it is using public money to build it. In the global race for AI drug discovery, the competition has quietly shifted. The real battleground is no longer which company holds the most promising drug candidate pipeline. Instead, it is who controls the computational infrastructure — the racks of high-end GPUs that train the models doing the searching. For Korean biotech firms, that gap has been painfully wide. However, the government just made its most concrete move yet to close it. 52 Projects, 3,000 GPUs — What the Government Actually Decided On July 1, the Ministry of Science and ICT (MSIT) convened the sixth Science and Technology Cabinet Meeting — a cross-ministerial forum where Korea's top research and industrial policy decisions are ratified — and formally approved the National AI Project. Out of 121 applications submitted by 28 ministries, 52 projects from 25 ministries were selected. Each winning project receives an allocation of advanced GPUs drawn from a national pool of 10,000 cards the government has already secured. In total, approximately 3,000 GPUs will flow to these projects. Distribution begins in April, staggered according to each project's start date. Furthermore, MSIT plans monthly usage reviews; projects that underperform or misuse resources will have their GPU allocations clawed back and redistributed. The stakes are real. Global Market Insights projects the AI drug discovery market will grow from $3.1 billion in 2025 to $43.9 billion by 2035 — a compound annual growth rate of 30.5%. Korea is betting it can claim a slice of that market by building the infrastructure now. Where the GPUs Are Going: A Biomedical Focus The biomedical sector dominates the list. MSIT itself leads with the single largest allocation: 208 GPUs for its Bio-Medical AI Technology Development programme. In addition, it is funding an "AI-Native Advanced Bio Autonomous Laboratory" (48 GPUs) — essentially a self-driving lab where robotic systems and AI models run experiments with minimal human input. The Ministry of Health and Welfare receives allocations across three projects. Its flagship "AI Basic Healthcare" initiative (80 GPUs) targets voice-based clinical note summarisation and AI-assisted medical imaging. Meanwhile, a separate track funds an AI biomedical foundation model for drug development (16 GPUs) and a medical AI data utilisation platform (64 GPUs). Other agencies join the push. The Ministry of Food and Drug Safety gets 32 GPUs to build an AI-powered pharmaceutical review and industry support system. The Korea Disease Control and Prevention Agency (KDCA) runs two projects: an AI-based national infectious disease response system (8 GPUs) and a cohort and biobank research data platform (16 GPUs). Outside biomedicine, the Ministry of Trade, Industry and Energy lands 304 GPUs for its software-defined vehicle and autonomous driving model project — the single largest non-bio allocation on the list. Why Computational Power Is Now the Drug Pipeline Traditional drug development takes 10 to 15 years and costs over $1 billion per approved molecule. AI, however, is compressing that timeline by two to three years on average. More striking is the clinical success rate: molecules identified through AI-driven screening reach Phase 1 clinical trials with an 80–90% success rate, far above the historical industry average. Korean firms have felt the computational gap acutely. By contrast, global Big Pharma — armed with thousands of in-house GPUs — can iterate on protein folding models and generative chemistry tools at a pace domestic players simply cannot match without external support. That is precisely the gap the national GPU pool is designed to address. Domestically, results are already emerging. KAIST's "K-Fold" protein structure prediction model achieves accuracy close to Google DeepMind's AlphaFold 3, but processes predictions up to 30 times faster — typically within one minute. Lunit, a Korean medical AI company listed on KOSDAQ, has deployed its clinical decision support system (CDSS) in hospital settings with a 94% diagnostic concordance rate. Both are products of earlier government-backed compute support. "AI is evolving into an independent research agent with autonomy across the entire drug development lifecycle," said Pyo Joon-hee, director of the AI Drug Research Institute at the Korea Pharmaceutical and Bio-Pharma Manufacturers Association. "The next five years will be a structural design competition. Overcoming data barriers and building self-driving labs will be the core differentiators." The Foundation Model Strategy — and Its Limits Korea's approach differs from simply buying access to existing Western AI platforms. The government is funding the development of Korean-specialised foundation models — large AI systems trained on domestic biomedical data, tuned for Korean patient populations and regulatory frameworks. These models are planned for open-source release, making them available to Korea's approximately 850 pharmaceutical and biotech companies. Choi Dong-won, MSIT's director of AI infrastructure policy, framed the ambition clearly: "Even within five months, we developed bio-medical specialised AI models capable of competing in the global market. We will continue supporting commercialisation in high-value sectors like diagnostics and drug development." Nevertheless, risks remain. Foundation model quality depends on data volume and diversity. Korea's biomedical datasets, though growing, are smaller than those available to US or European counterparts. In addition, GPU allocations — while meaningful — are modest by Big Pharma standards. Roche and Eli Lilly each operate internal GPU clusters that dwarf the 3,000 cards Korea is distributing across 52 projects. The monthly reallocation mechanism is therefore critical: idle compute must move quickly to projects that can use it. For investors, the structure of this programme matters as much as the headline numbers. GPU access, once a barrier to entry, is becoming a government-provided resource — similar to public research grants. As a result, smaller Korean biotech firms that previously could not afford large-scale AI model training may now compete on more equal footing. Watch the open-source model releases: if K-Fold and its successors gain international adoption, the licensing and service revenue potential is substantial. Looking Ahead: Self-Driving Labs and the Long Game The near-term deliverables are clear: GPU distribution from April, monthly performance reviews, and reallocation of underused resources. However, the longer strategic horizon points toward autonomous laboratories — facilities where AI models design experiments, robotic systems execute them, and results feed directly back into model training. Korea's inclusion of an "AI-Native Autonomous Bio Lab" in this round signals that this is not a distant ambition. The global AI drug discovery market will not wait. Therefore, the quality of Korea's medical data infrastructure and the continuity of its compute investment will ultimately determine whether this GPU injection translates into globally competitive drug pipelines — or remains a well-funded domestic exercise. Minister Bae Kyung-hoon, who doubles as Deputy Prime Minister for Science, described the initiative as "an AI highway built on government seed investment, on which national AI innovation will begin to genuinely take root." The metaphor is apt. Highways are only as useful as the vehicles that run on them. Korea now has the road. The question is how fast its biotech sector can drive.