SK hynix has approved 54 trillion won (approximately $38 billion) in new manufacturing investments, authorizing the construction of two semiconductor fabrication plants aimed at expanding production of the memory chips that power artificial intelligence systems. The South Korean memory giant's decision, announced Friday, underscores its conviction that demand for AI infrastructure will remain robust for years—and that memory, rather than processors, has become the industry's primary bottleneck.

The investment follows record quarterly results that highlight the central role of high-bandwidth memory (HBM) in the current AI buildout. As previously reported by Cryptopolitan, SK hynix posted revenue of 79.3 trillion won in the second quarter, with net income reaching 60.5 trillion won, driven by intensifying demand for HBM technology. The company has already begun high-volume deliveries of HBM4 chips as it works to maintain its leadership position in the AI memory market.

The capital commitment is significant for the broader AI ecosystem because advanced accelerator devices depend on multiple HBM stacks integrated alongside processor dies using sophisticated packaging technologies such as TSMC's CoWoS. Expanding GPU production alone would be insufficient if HBM supply cannot keep pace. TSMC's own CoWoS capacity has been a widely tracked constraint in the AI supply chain, and packaging bottlenecks have at times limited accelerator shipments even when sufficient HBM dies were produced.

According to a report by Omdia, memory is projected to account for more than 50% of global semiconductor market revenue by 2026, fueled by AI-driven demand for DRAM and NAND. TrendForce separately reports that DRAM and HBM supply will remain tight at least through 2027, as advanced packaging capacity stays constrained despite industry-wide capacity expansion efforts.

SK hynix supplies HBM to Nvidia and other AI accelerator manufacturers, making its production roadmap a critical barometer for the pace of AI infrastructure growth. The investment also follows Nvidia CEO Jensen Huang's announcement of a $500 billion partnership with SK Group, further highlighting the strategic importance of memory manufacturing.

Allocation of the 54 Trillion Won

Of the total commitment, 35.2 trillion won will fund construction of the Y2 DRAM fab in Yongin, while 19.1 trillion won will support the M17 NAND flash fab at SK hynix's Cheongju campus. According to the official announcement, these projects form part of a broader development strategy in which the company has already dedicated 600 trillion won to the Yongin Semiconductor Cluster and an additional 100 trillion won to expand Cheongju. Construction of the neighboring Y1 fab is already underway.

The Yongin Semiconductor Cluster represents one of the largest concentrated chipmaking buildouts in South Korea's industrial history and is central to the country's efforts to maintain its competitive edge in memory manufacturing against escalating investment from the United States, Japan, and the European Union, all of which have launched multibillion-dollar subsidy programs to expand domestic semiconductor production.

SK hynix captured 58% of global HBM revenue in the first quarter of 2026, according to Counterpoint Research's Memory Tracker, far ahead of Samsung and Micron, which each held 21%. Both competitors are pursuing their own HBM capacity expansions—Samsung has been qualifying its HBM3E products with major accelerator customers, while Micron has begun ramping HBM3E volume production—though neither has yet closed the gap in market share.

HBM has become one of the scarcest components in the AI supply chain. Unlike conventional DRAM, it requires stacking technology, through-silicon vias (TSV), and complex multi-chip packaging, making fabrication exceptionally challenging.

Omdia projects 94.1% growth in the semiconductor industry for 2026, driven by AI-related demand for NAND and DRAM.

Supply Constraints Already Reshaping Product Roadmaps

Tight memory supply is already influencing product development plans across the industry. Nvidia is reportedly evaluating 8-Hi HBM and HBM4 configurations for its next-generation Rubin Ultra accelerator, rather than the originally planned 12-Hi HBM4e, amid concerns that reduced DRAM availability in 2027 could constrain HBM module production.

This dynamic means memory suppliers increasingly dictate how quickly AI hardware reaches the market. Expanded memory capacity unlocks higher accelerator shipment volumes, while shortages force manufacturers to redesign products around available memory.

Construction Timeline

Analysts expect supply conditions to remain tight because the new fabs will not come online for several years, meaning the capacity unlocked by this investment will address demand in the 2029–2030 timeframe rather than the current crunch.

The Y2 fab will produce both HBM and next-generation DRAM products. Construction is scheduled to begin in July 2027, with the cleanroom expected to be ready by June 2029. Groundbreaking for the M17 fab is planned for February 2027, with its cleanroom projected for completion by December 2028.

SK hynix characterized the investment as foundational groundwork for a long-term industry transition rather than a response to a short-lived trend. Citing Omdia forecasts, the company noted an expected 19% compound annual growth rate in global DRAM and NAND demand through 2030, reinforcing its conviction that large-scale memory manufacturing capacity is essential for competitive positioning in the AI era.