South Korea is preparing to launch a roughly $3.5 billion program to develop a frontier AI model as early as March 2027, opening a competition to select participants1. The announcement lands alongside Doosan Robotics' win of two national R&D projects to build collaborative robots that think on domestic AI chips without cloud tethers3,4,5, and a national budget picture that channels roughly ten times more funding toward frontier language models than toward physical AI6,7. Together, the three developments crystallize a structural tension in Korea's AI strategy: the country is betting heavily on software-layer scale while its hardware-intensive manufacturing base gets a fraction of the investment.
Why it matters
Physical AI, the application of artificial intelligence to robotics, manufacturing, and logistics in the real world, is where national competitiveness in hardware-intensive industries will be decided. Korea's manufacturing base, from shipbuilding to nuclear energy, depends on skilled welders and assembly workers whose ranks are thinning. Doosan Robotics CEO Park In-won said "labor shortages and the aging of skilled workers in manufacturing are structural challenges that cannot be addressed simply by increasing the number of automated systems"2. Doosan's projects aim to replace that dependency with on-device intelligence. But the budget arithmetic raises a question: can a handful of flagship projects substitute for a systemic commitment?
The big picture
The frontier AI program, reported by The Korea Herald, would represent the single largest line item in Korea's AI budget and is set to open a competitive selection process. Separately, Doosan Robotics announced on October 6 that two proposals, one under the K On-Device AI Semiconductor Technology Development Program and another under the Robot Industry Technology Development Program, were selected as national R&D projects. Both are overseen by the Ministry of Trade, Industry and Energy and the Korea Evaluation Institute of Industrial Technology (KEIT). The combined R&D budget is approximately ₩98.9 billion, with about ₩68.1 billion in government funding.
The first project develops the core capability: a Korean-made AI system-on-chip integrating a neural processing unit for AI computation with real-time communication and control functions, embedded directly in collaborative robots. The goal is perception, judgment, and control without an external PC or cloud connection. Doosan plans to develop three collaborative robot models with payload capacities of 5 kilograms, 10 kilograms, and 20 kilograms, as well as an industrial humanoid capable of dual-arm operation and omnidirectional movement. The company aims to commercialize the robots as standard products beginning in 2031. Research partners include DEEPX, SAIGE, and Changwon National University's industry-academic cooperation foundation, with Doosan Enerbility participating as an end-user.
The second project applies that technology to intelligent welding, where AI will learn from skilled welders' movements to generate welding paths and weaving motions. Doosan aims to achieve more than 95% alignment between robot-generated welding paths and those of skilled workers, with path deviation limited to within 3 millimeters. The technology will initially target nuclear power equipment production before expanding into shipbuilding, industrial plants, and defense manufacturing.
Pull back to the national ledger, and the picture shifts. South Korea's government AI budget for next year will reach ₩9.4211 trillion, approximately $7 billion. Physical AI receives ₩462.2 billion, approximately $344 million, or 4.9% of the total. Frontier-grade AI model development is slated to receive roughly ₩4.7 trillion, approximately $3.5 billion. Physical AI funding has surged 10.9-fold in two years, yet the absolute gap between the total AI budget and the physical AI budget will widen to ₩8.9589 trillion next year.
ANALYSIS Doosan's projects are structured to prove that Korean-made silicon can run inference at the robot's edge, a sovereignty play that reduces dependence on cloud infrastructure and, implicitly, on foreign chip architectures. The on-device framing is deliberate: if a cobot in a shipyard or defense facility must phone home to a data center for every weld decision, latency and security both suffer.
Democratic Party lawmaker Hwang Jung-ah has urged the government to design cross-program linkages connecting frontier AI outcomes to manufacturing, robotics, and logistics sectors. Hwang's proposed Physical AI Special Act has been referred to the National Assembly's Science and ICT Committee legislation subcommittee.
The legislative push reflects a recognition that raw budget growth is insufficient when the structural tilt favors language models over embodied applications. Hwang's framing points to a coordination failure: frontier AI research and physical AI deployment sit in separate funding silos with no mandated handoff.
Japan offers a contrast. Through NEDO, Japan has selected Noetra and the National Institute of Advanced Industrial Science and Technology to operate an AI Robot and Physical AI Multimodal Foundation Model Development Project. Noetra plans to build AI computing infrastructure equipped with approximately 27,500 GPUs, using NVIDIA's Rubin products, with the system scheduled to go live in June 2028. Total support over five years could reach ¥1 trillion, approximately $6.3 billion. Japan's integrated roadmap, language models feeding into physical-world AI under a single institutional umbrella, is precisely the linkage Hwang is demanding Korea build.
What's next
Korea's frontier AI model competition is set to open as early as March 2027. Hwang's Physical AI Special Act faces committee review in the National Assembly. Doosan's first project will run for 45 months under KEIT oversight. Noetra's GPU infrastructure is scheduled to go live in June 2028.