South Korea’s largest technology manufacturers have announced a combined $550 billion investment programme to expand production of high-bandwidth memory (HBM) chips, directly addressing a supply constraint that has throttled artificial intelligence infrastructure deployment globally. Samsung Electronics and SK Hynix disclosed the commitment on Monday, with initial production capacity increases scheduled for 2027.
The investment represents the semiconductor industry’s most substantial response yet to what engineers have termed “RAMageddon” — a shortage of specialised memory chips required for training and running large language models and other AI systems. Current HBM supply cannot meet demand from hyperscalers including Microsoft, Google, and Amazon, which have delayed data centre expansions due to component unavailability.
HBM chips stack multiple memory dies vertically to achieve bandwidth exceeding 1 terabyte per second, a specification essential for AI accelerators such as Nvidia’s H100 and upcoming Blackwell processors. Each advanced AI server requires between 8 and 16 HBM modules, but global production capacity remains concentrated among three manufacturers: SK Hynix, Samsung, and Micron Technology.
According to industry analysts at TrendForce, HBM accounted for just 3% of total DRAM revenue in 2023 but is projected to reach 18% by 2026 as AI infrastructure spending accelerates. The South Korean investment aims to triple domestic HBM production capacity within four years, establishing new fabrication facilities in Pyeongtaek and expanding existing plants in Icheon.
The announcement carries immediate implications for the AI infrastructure market. Cloud service providers have reported quarter-long lead times for AI-capable servers, with memory availability cited as the primary constraint. Nvidia has publicly acknowledged that HBM supply limits its ability to fulfil orders for AI accelerators, despite having secured long-term agreements with all three major producers.
Samsung’s participation marks a strategic shift after the company lost early market share to SK Hynix in HBM3 production. Industry sources indicate Samsung experienced yield issues with its fifth-generation HBM that delayed qualification for Nvidia’s platforms. The investment programme includes dedicated research facilities for next-generation HBM4 specifications, expected to double bandwidth to 2 TB/s.
For enterprise AI deployment, the capacity expansion offers a timeline for when infrastructure constraints may ease. Organisations that have deferred private AI implementations due to hardware unavailability can anticipate improved supply conditions in late 2027, though pricing will depend on whether demand growth outpaces the new capacity.
The geopolitical dimension cannot be ignored. South Korea’s investment consolidates its position in a technology layer critical to AI competitiveness, whilst the United States and European Union have announced separate semiconductor programmes focused primarily on logic chips rather than memory. China’s domestic HBM development remains constrained by equipment export restrictions that limit access to advanced packaging technology.
Market analysts at Bernstein estimate the investment will generate approximately 15,000 specialised manufacturing jobs and require significant expansion of South Korea’s chemical supply chain for advanced materials including through-silicon vias and micro-bump interconnects. Equipment manufacturers Applied Materials, ASML, and Tokyo Electron stand to benefit from fabrication tool orders.
The timeline presents risks. If AI model architectures evolve to reduce memory bandwidth requirements — through techniques such as quantisation or sparse attention mechanisms — demand projections underpinning the investment may prove optimistic. Conversely, continued scaling of model parameters could render even the expanded capacity insufficient.
Industry observers will monitor several indicators: Samsung’s success in qualifying HBM3E for Nvidia platforms, which would validate its manufacturing improvements; pricing trends for HBM modules, which currently command 3-5x premiums over conventional DRAM; and whether alternative memory technologies such as GDDR7 gain adoption in AI inference workloads where extreme bandwidth proves unnecessary.
The $550 billion commitment establishes South Korea’s semiconductor sector as the primary responder to AI infrastructure constraints, with implications extending beyond immediate supply relief to longer-term competitive positioning in the technology stack underlying artificial intelligence deployment at scale.
