SK Hynix Chairman: Memory Chip Shortage Most Severe in 2027, Capacity Expansion Must Be Tied to Long-Term Orders
Nashnova编辑部
SK Group Chairman Chey Tae-won warned that 2027 will bring the most severe memory chip shortage yet. SK hynix has drawn up a $72 billion expansion plan, but every new fab must be backed by long-term customer orders — price hikes are already fueling chip inflation that consumers ultimately bear.
Why does the shortage peak in 2027?
AI demand for memory chips is surging, yet supply-side expansion cycles are long and yields constrained. The gap between capacity and demand hits its worst point in 2027.
Chey added that shortages could persist through 2030 — until humanity reaches AGI (artificial general intelligence), supply is unlikely to catch up.
This means → this is not a one-quarter mismatch but a structural shortage that could span three to five years.
How will $72 billion be spent — and why must orders come first?
SK hynix plans to double capacity within five years at a total cost of $72 billion, with one hard prerequisite: every fab investment must be tied to a long-term customer order. No demand backing, no factory.
In plain terms = the memory industry has been burned by blind expansion before — demand receded, overcapacity cratered prices, and the whole sector bled. This time SK hynix insists on "order first, build second," shifting risk onto buyers.
Site selection hinges on three conditions: stable power and water, sufficient industrial land, and a complete chemicals-and-equipment supply chain. Indiana is on the shortlist, but no timeline or scale has been locked in.
What do the two risk-sharing models mean?
SK hynix is rolling out two new structures: joint-venture fabs with customers, where both sides invest and share risk; and "Memory as a Service" — customers lease memory capacity on demand instead of buying chips outright.
This means → the first model locks in big customers' capacity commitments; the second lowers the capex barrier for smaller buyers. Both paths point to one goal: making buyers share the expansion risk.
Where is the HBM production bottleneck?
HBM — high-bandwidth memory, a chip designed for AI that stacks multiple layers for ultra-high data throughput — consumes 4× the wafer area of standard DRAM per chip, and advanced stacking yields remain constrained.
In plain terms = on the same production line, the wafer area that once produced four standard memory chips now yields just one HBM. The lines haven't shrunk — AI chips' physical footprint has "eaten" the output.
This reflects a structural problem: the current shortage is not just about building fabs too slowly but about per-line output being drastically diluted by HBM.
How important is Nvidia — is the customer base too concentrated?
Chey stated plainly that Nvidia is SK hynix's most important customer today. Nvidia's GPUs depend on SK hynix HBM, and continued AI compute growth keeps pulling in memory orders.
He called Nvidia CEO Jensen Huang's model of coordinating the entire supply chain a key driver of industry progress, adding that "without Nvidia, a complete AI infrastructure ecosystem would not exist."
SK hynix also maintains long-term supply relationships with Google, Microsoft, and Meta. This means → Nvidia is the largest customer but not the only one; concentration risk is manageable for now.
Who bears the price hikes — and what should we watch for in 2027?
Chey apologized to the market for rapid memory price increases, acknowledging that the rises are fueling chip inflation across the industry and are ultimately borne by consumers.
He called it an outcome SK hynix does not want, yet cannot resolve quickly while supply trails demand.
Whether effective supply materializes on schedule by 2027 is the key checkpoint for this memory up-cycle — whether expansion plans land and HBM yields break through will determine if pricing pressure eases or intensifies.
Content is for reference only, not financial advice.