A Chairman On Jeju Argues For Lower Prices
SK Group chairman Chey Tae-won stood in front of reporters at the Shilla Jeju hotel and argued that the price of the product his company sells should come down. The setting was the press briefing at the 49th Korea Chamber of Commerce and Industry Jeju Forum, reported on 19 July. "Prices have to normalize. Otherwise, the market shrinks and competitors flood in."
Read that sentence as a supplier's statement about its own constraint. A company whose problem is margin does not campaign for lower prices in public. A company whose problem is that it physically cannot make enough does, because a shortage that persists long enough turns into an invitation for every rival with a fab roadmap to walk through the door.
For a European operator planning 2027, that is the useful signal in the whole briefing. The scarcity being described is capacity scarcity. Price is the symptom the seller would like to reduce, and the seller cannot reduce it, because the only cure is output that does not exist yet.
The Order Book For 2027 Is Already Written
The demand for 2027 has already been placed, and it was placed by buyers who moved before you did. Chey said SK hynix customers have requested 60 to 100 percent more AI memory for 2027 than for 2026, and that he anticipates overall semiconductor demand growth of at least 50 to 60 percent.
Those are requests booked against a supplier who is already the dominant seller. In the first quarter of 2026, SK hynix held 58 percent of HBM revenue, with Samsung at 21 percent and Micron at 21 percent. The largest seller in the market is telling you that his customers have asked for roughly double, and that he is not saying yes to all of it.
The practical consequence is unglamorous. A 2027 procurement plan that assumes memory can be sourced in 2027 is not a plan, it is a queue position that has not been requested yet. The volume being fought over now is the volume that ships in the year you are budgeting for.
February 2027 Is Late, And It Is The Fast Version
The capacity answer lands inside the demand year, not ahead of it. The Yongin clean room opening was pulled forward to February 2027 from May 2027. Cheongju M15X is being converted into a dedicated DRAM base for HBM. In March 2026, SK announced an additional 21.6 trillion won of investment, or 14.52 billion dollars, roughly 13 billion euros and around 11 billion pounds, including a 3.87 billion dollar HBM advanced-packaging plant in Indiana in the United States.
Pulling a clean room forward by three months is an aggressive move, and it is worth noticing what the aggressive version still produces. Wafers do not become qualified HBM stacks on opening day. Packaging capacity in Indiana has to be built, staffed and yielded. The output that eases 2027 arrives during 2027 at the earliest, and much of it later.
So the shortage is not a market that failed to notice a problem. It is a market that noticed, spent 13 billion euros, and still cannot deliver the relief before the demand arrives.
The Queue Becomes A Political Object
Chey's second remark moves the whole problem out of procurement and into policy. "Right now, companies absorb the pressure. Governments will start pressuring other governments soon." That is a chairman saying the allocation queue will stop being a commercial arrangement between a supplier and its customers.
For a European operator this changes what a supply agreement is worth. When allocation becomes a matter that ministries raise with each other, a buyer's position in the queue stops being purely a function of contract terms and volume history. It becomes exposure to whichever bilateral conversation is going well that quarter, and Europe is not the party with the largest share of that conversation.
The defensible response is to be inside the contract before the politics start, with volume committed, dates fixed and penalties that make displacement expensive for the supplier. A buyer holding a signed multi-year allocation is a buyer whose position is harder to reassign quietly.
Size The Plan To Secured Memory
The 2027 inference footprint should be sized to memory already secured, not to GPUs that can be bought. That inversion is the entire owner action, and it runs against the way most capacity plans are still written, which starts from accelerator counts and treats memory as an accompanying line item.
In practice that means contracting multi-year DRAM and HBM volume now, at whatever price normalization the chairman is promising, and accepting that the contract is the asset rather than the unit price. It also means the finance case for 2027 should be built twice: once on the memory under contract, and once on the memory an operator hopes to add opportunistically, with the second number treated as upside rather than baseline.
Owners who do this convert a supply risk into a known cost. Owners who do not will discover in early 2027 that their accelerators are installed, their power is contracted, and the memory to fill them was sold eighteen months ago.
Read next: The 518 Billion Dollar Memory Fix Lands in 2028 | Three Firms Now Decide If Your Phone Stays Cheap
