SK Hynix Plans to Double Production Capacity in Five Years
Alina Collins
SK Group Chairman Chey Tae-won said SK Hynix will double its memory production capacity within five years and aims to become the primary HBM supplier for NVIDIA's Vera Rubin system — the most explicit top-level commitment yet to AI-driven memory expansion, signaling a shift from reactive supply to preemptive capacity build-out.
What is SK Hynix betting on with a five-year doubling?
Chey's logic chain is straightforward: AI infrastructure → massive "AI factories" → core consumable is HBM (high-bandwidth memory — chips that stack multiple memory layers vertically for ultra-fast data throughput). Doubling capacity is an active bet on that demand curve.
This means → SK Hynix is no longer waiting for orders to arrive before expanding. This reflects a management conviction that AI compute demand is not a short-term pulse but a multi-year structural build.
Memory expansion cycles typically run two to three years. A five-year window means new capacity comes online between 2027 and 2030 — overlapping squarely with the expected AI compute surge.
Why name-drop NVIDIA's Vera Rubin?
Vera Rubin is NVIDIA's next-generation flagship AI training architecture after Blackwell, expected to use HBM4 memory. By calling out the intent to be its primary supplier, Chey is locking in a seat for the next arms race.
This means → SK Hynix aims not just to defend its lead in HBM3E, but to push the competitive front line to HBM4 — a product not yet in mass production. In plain terms = win this round and claim the next one early.
As the dominant force in the global HBM market, the statement doubles as a signal to NVIDIA: our capacity roadmap can sync with your chip cadence.
Who controls the HBM4E timeline?
Chey was notably more cautious on HBM4E: the product roadmap will "depend on customer demand." This means → SK Hynix will not push technology transitions unilaterally but will stay tightly coordinated with lead customers — primarily NVIDIA and Google.
This reflects a defining trait of the HBM market: HBM chips must be deeply co-engineered with the packaging architecture of specific AI accelerators. The supplier's R&D cadence is effectively set by a handful of large buyers.
HBM4 has entered production ramp this year. HBM4E is positioned as the 2026–2027 technology node. In plain terms = the final shape of that roadmap directly determines how high a moat SK Hynix can build — and how much pricing power it holds — in this cycle.
Content is for reference only, not financial advice.