Chey Tae-won: HBM Capacity Gap May Persist, Memory Shortage Most Severe Next Year
Nashnova编辑部
SK Hynix chairman Chey Tae-won expects next year to be the worst year for memory shortages, because new fab capacity takes four to five years to build while AI-driven demand has already exploded — a gap that shifts pricing power firmly toward upstream suppliers.
Why does he say next year will be the tightest?
Building a new memory fab from groundbreaking to volume production takes four to five years; AI-driven memory demand has exploded within the past year alone.
This means → supply physically cannot catch up with the demand curve, and next year is when the gap peaks.
Chey compared today's AI to "a four-year-old child." In plain terms = AI is still in its infancy — every year it matures, it consumes more memory, not less.
He projects global memory capacity may need to reach nearly five times current levels over the next decade, with AI agent usage growing tenfold in five years.
How are Nvidia, TSMC, and SK Hynix locked together?
Chey stated plainly that Nvidia is SK Hynix's most important customer, because Nvidia is building not just GPUs but an entire AI ecosystem.
The supply-chain dependency runs both ways: without TSMC's advanced manufacturing, GPUs cannot scale; without enough HBM — high-bandwidth memory, a type of stacked memory designed specifically for AI chips — TSMC's capacity cannot become a complete AI system.
This reflects a mutually locked chain among the three companies — a bottleneck at any link stalls the entire chain.
Chey revealed he met TSMC chairman C.C. Wei roughly a month ago; both sides reached consensus on capacity expansion.
What is "chip inflation" — and why should consumers care?
Chey used the term "chip inflation" to describe rapidly rising memory prices, noting the pressure is spreading to PCs, smartphones, automobiles, and consumer electronics.
He pointed out that Apple has already had to raise flagship product prices, and there is no short-term fix.
In plain terms = AI chips are consuming a huge share of wafer capacity, leaving less for everyday products — so consumers pay more for phones and laptops.
A key number: HBM consumes more than four times the wafer area of standard DRAM. This means → every additional HBM chip produced crowds out four times as much conventional memory capacity.
How does SK Hynix plan to expand without repeating past overcapacity mistakes?
SK Hynix has announced plans to double capacity over the next five years, but Chey acknowledged most customers think that is still not enough.
To guard against overinvestment, the company re-verifies real market demand before every expansion and requires customers to sign long-term supply agreements.
It is exploring two risk-sharing models: fab joint ventures and "memory as a service." In plain terms = customers co-fund the fabs or pay by usage, rather than SK Hynix bearing the entire capital risk alone.
How realistic is a U.S. memory fab?
Chey did not commit to building a full memory fab in the United States.
He stressed that site selection is not just about land, water, and power — it requires the ability to bring along six to seven hundred materials, chemicals, and service suppliers to form a full industrial ecosystem.
This means → memory manufacturing is not a single-factory decision but an entire supply-chain migration; the U.S. does not yet have that ecosystem in place.
What is the key verification point for markets?
The time gap between the supply expansion cycle and the demand explosion is the core pillar of the current HBM shortage thesis.
Whether that gap begins to narrow after next year will be the critical checkpoint for markets to reassess valuations across the entire memory chain.
In plain terms = if next year truly is the tightest as Chey predicts, the pricing-driven bull case for memory stocks holds; if the gap closes sooner, valuations need recalculating.
Content is for reference only, not financial advice.