Korea makes the world's memory and little else in chips
Two firms give Korea a commanding share of global memory production while the country holds a low single-digit share of the far larger system semiconductor market, and two decades of diversification policy have not moved that ratio much.
Korea’s position in semiconductors is two entirely different positions held simultaneously. In memory, Samsung Electronics and SK hynix together account for something in the region of seventy percent of global DRAM revenue and roughly half of NAND flash, a concentration with few parallels in any manufactured product. In system semiconductors — the logic, analogue, sensor and processor devices that make up around two thirds of the world semiconductor market by value — Korea’s share has been estimated in the low single digits, commonly cited at around three percent, with the fabless design segment alone closer to one percent. The country dominates the smaller category and is marginal in the larger one.
The lopsidedness is not an accident of neglect. Memory rewarded exactly what the Korean industrial system was built to supply: enormous, repeated, counter-cyclical capital investment in identical products sold on price, decided quickly by owners who could absorb losses through a downturn to take share afterwards. System semiconductors reward the opposite. They are heterogeneous, low-volume relative to memory, defined by customer-specific design, and the value sits in architecture, software tooling and long design-win relationships rather than in yield on a standard part. A structure optimized for the first is not converted into the second by capital.
Foundry was supposed to be the bridge, and it is where the ambition has been most tested. Samsung Foundry runs advanced logic processes and competes at leading nodes, but its global share has hovered around a tenth against TSMC’s dominant position, and the shortfall is not primarily about lithography. It is about the ecosystem a foundry customer buys into: verified intellectual property blocks, process design kits, packaging options, and a bench of design service houses that can take a customer’s specification to tape-out. Building that is slower than building a fab, and a foundry that also owns a chip design business must additionally persuade fabless customers that their designs are safe inside it — a structural conflict TSMC does not have.
The domestic fabless sector is where the diversification argument runs into its hardest numbers. Korea has several hundred fabless companies, and a small number hold durable positions in specific niches — display driver ICs, automotive vision processors, embedded controllers. Below that tier the picture is of firms that struggle to cross from a first product to a portfolio. Access to advanced-node shuttle runs is expensive and rationed; licensing costs for processor cores and interface IP are paid to foreign vendors before any revenue exists; design automation tools come from three American suppliers; and a Korean startup is competing for a domestic customer who can buy a proven part from an established vendor with a decade of software support behind it.
Talent allocation compounds it. Two enormous employers offering memory and foundry careers at the top of the domestic pay scale absorb much of each graduating cohort of chip designers, and a fabless startup cannot match the compensation, equipment or stability. That constraint is downstream of the structure rather than the cause of it. Successive programmes since the 2019 system semiconductor strategy have funded design centres, shuttle access, IP subsidies and graduate tracks, and aggregate market share has moved little — the interventions look correctly aimed and too small against the gravitational pull.
The strategic case for persisting is straightforward. Memory is cyclical, capital-hungry and increasingly contested by state-backed Chinese entrants at the mature end, and an industry whose earnings swing with one commodity’s price cycle is exposed in a way a diversified one is not. The case against forcing it is equally straightforward: comparative advantage is real, and money spent manufacturing a fabless sector against the structure of the economy may buy less than the same money spent extending the memory lead into high-bandwidth and packaging, where the AI build-out has made memory unusually strategic. Twenty years of policy has assumed the first answer. The market share statistics have not yet agreed.