SK Hynix Aims to Double Wafer Capacity

Memory, not logic, is the binding constraint on AI infrastructure right now, and HBM capacity decisions made in 2026 determine accelerator availability in 2029-2030. A commitment from the world's leading HBM supplier to double wafer capacity is the strongest supply-side signal of this cycle - but the same executive's forecast that the shortage persists until 2030 tells practitioners the relief arrives at the end of the planning horizon, not the beginning.

SK Group chairman Chey Tae-won told reporters at Computex in Taipei that SK hynix will double its memory wafer capacity over the next five years; Reuters carried the direct quote: "We are going to double the whole capacity over the next five years." Chey repeated his earlier forecast that the AI-driven memory shortage could run until 2030. He declined to put a cost on the buildout but said 2026 capital spending would climb well above the 30.2 trillion won spent in 2025, per Tom's Hardware, which also reports SK hynix has filed to list American depositary receipts in New York this year. The announcement marks a reversal from March, when Chey said at Nvidia's GTC that adding capacity was not planned. In a piece of trade-show theater that doubled as a demand signal, Nvidia CEO Jensen Huang stopped at SK hynix's booth and wrote "Please Make More" on a displayed HBM4E wafer the same afternoon, per Tom's Hardware.

HBM is the immediate driver of wafer demand: stacked high-bandwidth memory consumes several times more wafer area per bit than commodity DRAM, so wafer-equivalent demand rises even when bit growth is moderate. HBM's high margins have pulled capacity allocation toward advanced packaging and HBM lines across the industry. Counterpoint Research figures cited by Reuters put SK hynix at about 58% of the global HBM market in the first quarter - a concentration that makes single-supplier decisions like this one systemically important.

Two timing realities cut against near-term relief. Building out from an existing site takes at least three years and a greenfield fab more than five, placing even an immediate construction start near the tail of the shortage window Chey himself describes. And doubling wafer starts does not double HBM output if advanced-packaging capacity lags behind. For teams budgeting large training clusters, this argues for locking in memory-heavy capacity commitments early rather than waiting on announced supply.

• •Quarterly HBM shipment and wafer-start data from industry trackers such as Counterpoint Research and TrendForce
• •SK hynix capex disclosures, fab construction milestones, and the ADR listing timeline
• •Samsung and Micron capacity responses, which determine whether HBM concentration eases or hardens