Every AI accelerator shipped today needs large amounts of memory, and most of that memory is made by two South Korean companies. In the second quarter of 2026, Counterpoint Research estimated that Samsung held about 39 percent of global DRAM revenue and SK hynix about 26 percent, with US-based Micron close behind at 25 percent. Together the two Korean firms accounted for roughly two thirds of the market. [1][2]

The industry is also central to the Korean economy. According to Korean government figures reported by the Korea Herald, the country's total exports reached a record 709.7 billion US dollars in 2025, and semiconductor exports alone reached a record 173.4 billion dollars, up 22.2 percent from 2024. The trade ministry attributed the growth to demand from AI data centres and higher memory prices. [3]

High-bandwidth memory, or HBM, is where the link to AI compute is most direct. HBM stacks several DRAM dies vertically and places them next to the GPU or accelerator, giving far more bandwidth than conventional memory modules. The supply of HBM has repeatedly been one of the limiting factors on how many high-end GPUs can be built, which is why the Korean memory makers attract so much attention from cloud operators and investors. [4][5]

Samsung's route into chips began in 1974, when Lee Kun-hee bought half of the struggling Hankook Semiconductor with his own money; the company was fully acquired in 1977 and renamed Samsung Semiconductor in 1978. After Samsung founder Lee Byung-chul toured the United States in 1982 and saw American firms losing ground to Japanese memory makers, Samsung publicly committed in early 1983 to large-scale investment in memory chips. [6]

Samsung licensed 64K DRAM technology from Micron and SRAM technology from Sharp, hired engineers who had worked at US chip companies, and began operating its Giheung fab in 1984. Samsung's own timeline records the development of a 64Kb DRAM in 1983, the industry's first 64Mb DRAM and the top DRAM market share in 1992, the world's largest memory market share in 1993 and the largest flash memory share in 2003. In 2013 it began what it describes as the first mass production of 3D vertical NAND. [6][7]

SK hynix has a parallel history. It was founded in 1983 as Hyundai Electronics, which began building its plant in Icheon that year and started mass production of 256K DRAM in 1985. It merged with LG Semiconductor in 1999, became Hynix Semiconductor in 2001, and took the name SK hynix in 2012 after SK Telecom acquired a controlling stake. In 2018 it joined a consortium that bought Toshiba's memory business, now Kioxia, and in 2020 it began buying Intel's NAND and SSD business. [8]

The Intel deal was completed in two phases. SK hynix paid 6.61 billion dollars at the end of 2021 for Intel's SSD business and its NAND fab in Dalian, China, which became the subsidiary Solidigm, and a further 1.9 billion dollars in March 2025 for the remaining intellectual property, research assets and staff, for a total of about 8.85 billion dollars. [9]

The result is a strong position in NAND flash as well as DRAM. TrendForce estimated that in the second quarter of 2026 Samsung was the largest NAND supplier with 23.06 billion dollars of revenue, about 29.3 percent of the market, and that SK hynix Group including Solidigm was second with 14.27 billion dollars, ahead of Micron, Kioxia and SanDisk. [10]

SK hynix was an early mover in HBM and has used that head start to become the main supplier for AI accelerators. Its corporate history lists the world's first HBM3 in 2021, mass production of HBM3E in 2022, the first 12-high HBM3E stack in mass production in 2024, and completion of HBM4 development in 2025. In September 2025 the company said its HBM4 doubled the number of I/O terminals to 2,048, ran above 10 Gbps compared with the 8 Gbps JEDEC standard, and improved power efficiency by more than 40 percent. [8][4]

Samsung announced in February 2026 that it had begun mass production and commercial shipments of HBM4, built on its sixth-generation 10nm-class DRAM process with a 4nm logic base die. Samsung stated a consistent transfer speed of 11.7 Gbps and bandwidth of up to 3.3 terabytes per second per stack, and said it expected HBM sales to more than triple in 2026 compared with 2025. [5]

Market share has moved quickly. Counterpoint Research data reported by Korea JoongAng Daily put SK hynix at 50 percent of HBM revenue in the second quarter of 2026, down from 58 percent in the first quarter, while Samsung rose from 21 percent to 33 percent and Micron slipped from 21 percent to 18 percent. Most HBM revenue in that quarter still came from HBM3E, with HBM4 shipments expected to become more visible in the second half of the year. [11]

The financial effect of the AI memory cycle is large. SK hynix reported second-quarter 2026 revenue of about 79.3 trillion won and operating profit of about 60.5 trillion won, a 76 percent operating margin, and said mass shipments of HBM4 had begun in the quarter with a production ramp planned for the second half. [12]

SK hynix is headquartered in Icheon in Gyeonggi Province, where its first plant was built in 1983. Its newer M15X fab in Cheongju, North Chungcheong Province, is intended as its main HBM production hub; the company began installing equipment there in late 2025, and in its second-quarter 2026 results it listed accelerating M15X mass production among its priorities, together with a new P and T7 advanced packaging plant and an M17 NAND base. [8][13][12]

The largest new project is the Yongin semiconductor cluster in Gyeonggi Province. SK hynix plans four advanced fabs on a site of about 1.97 million square meters, as part of a long-term plan the company has valued at around 600 trillion won. In February 2026 its board approved an additional 21.6 trillion won for the first fab, bringing the committed total for that fab to about 31 trillion won, and moved the opening of its first cleanroom forward from May 2027 to February 2027. [14]

TrendForce reported in July 2026 that SK hynix had started ordering equipment for the first Yongin fab, which is expected to begin with roughly 20,000 wafers per month of sixth-generation 10nm-class DRAM, with products including HBM4E. Samsung is separately planning fabs in the government-backed Yongin National Industrial Complex, with the first reportedly targeted for 2029. [15]

Samsung's main memory expansion today is at Pyeongtaek, south of Seoul. According to TrendForce, its P4 building now produces HBM and high-end DRAM, while the P5 building, on which construction started in 2022, is targeted for completion around 2030 and would raise capacity for HBM, DRAM, V-NAND and foundry chips. Samsung is also reported to be breaking ground in September 2026 on a 6 trillion won HBM packaging plant at its Onyang campus. [16]

Korean support for the chip industry has relied mainly on tax credits, policy loans and public infrastructure rather than large direct grants of the kind used in the United States. In May 2024 the Ministry of Economy and Finance announced a 26 trillion won support package, including a 17 trillion won programme of loans at preferential rates, a larger ecosystem fund for materials, parts and equipment suppliers, and more than 5 trillion won for research and training over three years. [17]

The legal basis for the main tax incentive is an amendment to Korea's Restriction of Special Taxation Act passed in March 2023 and commonly called the K-Chips Act. It raised the base tax credit for investment in facilities for national strategic technologies, including semiconductors, from 8 to 15 percent for large companies and from 16 to 25 percent for small and medium-sized firms, with an additional credit for investment above a company's recent average. [18]

The 2024 package extended the period of that national strategic technology tax credit and committed the government to help with roads, water supply and power for the Yongin cluster, with the aim of halving the time needed to start construction. Separately, a regulatory change that allowed the floor area ratio of industrial complexes to be relaxed by up to 1.4 times enabled SK hynix to fit more cleanroom space into its first Yongin fab. [17][14]

In our view, this model of tax relief, cheap credit and state-provided infrastructure has helped Korean firms keep investing through downturns, but it also ties public finances and regional planning closely to one cyclical industry. Power and water supply for the new clusters are likely to be as important as subsidies in determining how fast capacity arrives, and both are worth watching for anyone assessing the long-term reliability of memory supply.

The memory market is highly cyclical and competition is intensifying. Counterpoint noted that SK hynix's DRAM share fell in the second quarter of 2026 partly because it had signed long-term supply agreements at fixed prices below later market levels, while Micron was gaining ground; the firm also highlighted rapid growth at China's CXMT. [1][2]

Geography is a second consideration. Much of the world's advanced DRAM and HBM is made within a relatively small area of South Korea. For buyers of AI infrastructure this does not usually show up as a direct risk, but it helps explain why governments in the United States, Japan and Europe have tried to attract memory and packaging investment of their own, and why any disruption in Korea would be felt across the whole AI hardware supply chain.

For people renting GPUs, Korean memory output is one of the less visible factors behind prices and availability. When HBM is scarce, fewer top-end accelerators ship and new generations reach clouds more slowly. When new Korean capacity at Cheongju, Pyeongtaek and Yongin comes online, the supply of HBM-heavy GPUs such as current and next-generation data centre parts should become less constrained, although demand has so far kept pace with every expansion. On Kovara you can compare current GPU cloud prices across providers, explore which regions offer the latest HBM-equipped GPUs, or ask Kova how memory supply is affecting the prices you see.