SK hynix approved $38.1 billion in capital expenditure Thursday for two new memory fabrication facilities in South Korea, the largest single-company semiconductor commitment since TSMC's Arizona plan in 2020. The board authorized construction of a high-bandwidth memory and DRAM line in Yongin and a NAND flash complex in Cheongju. Neither fab delivers meaningful output before late 2027, with volume production targeted for 2029.

The approval arrives as HBM3E supply already runs six quarters behind datacenter demand. Nvidia, Broadcom, and AWS have pre-committed 80% of SK hynix's 2025 HBM wafer starts under forward contracts disclosed in February earnings. Spot HBM3 pricing now sits at $2,400 per unit, triple the 2022 baseline, with no correction visible in forward curves through Q1 2026. SK hynix shipped 38% of global HBM volume in 2024, second only to Samsung's 41%, but the gap between capital deployment and tape-out means the duopoly pricing power extends through the decade.

The Yongin facility earmarks $22 billion for HBM and next-generation DRAM, with advanced packaging infrastructure embedded in the same clean room to reduce interposer transit time. The Cheongju NAND plant takes the remaining $16.1 billion and focuses on 300-layer-plus V-NAND for enterprise SSDs, a segment where SK hynix trails Kioxia and Micron in density but leads in yield consistency. Construction permits cleared in March; groundbreaking occurs in Q3 2025. The timeline assumes no permitting delays and stable won-dollar exchange rates, neither guaranteed.

This capex cycle matters because it cements a four-year window where hyperscalers cannot diversify memory supply without accepting performance downgrades or inventory risk. Micron's Idaho HBM line adds capacity in 2026, but at 15% lower bandwidth than SK hynix's latest HBM3E specs. Samsung's Taylor, Texas DRAM fab remains 18 months behind schedule per March filings. Chinese CXMT and YMTC face sub-90nm node restrictions under October 2023 export controls, effectively removing them from the frontier memory conversation. SK hynix's move is less about growth ambition and more about holding the choke point.

Operators should track three forward inflection points. First, SK hynix's July earnings call will detail HBM4 architecture and whether 2029 volume production aligns with Nvidia's Rubin Ultra platform launch, likely Q4 2029. Second, watch Samsung's April capex guidance—if they match or exceed $32 billion, the HBM duopoly calcifies; if they undershoot, SK hynix's pricing power extends another two years. Third, monitor Micron's Q2 2025 conference call in late June for any Idaho fab acceleration, which would be the only near-term supply relief vector.

The $38.1 billion is not deployed to solve the bottleneck. It is deployed to own it.