SK hynix said October 28 that DRAM and high-bandwidth memory volumes are effectively sold out through 2029, combining signed long-term supply agreements with customer demand under active negotiation. Signed LTAs already cover more than half of 2027–2028 production capacity. The claim lands on a quarter when revenue topped 100 trillion won for the first time — and when Seoul needs to know whether those bookings are binding demand or forward-looking inventory risk.
Seoul, Oct. 28 — SK hynix said its DRAM and high-bandwidth memory output is effectively sold out through 2029, a characterization it offered during its third-quarter earnings briefing as investors asked how far long-term supply agreements (LTAs) can stretch demand visibility. Yonhap and English-language outlets including the Korea Herald reported the statement on the same timeline as a record quarter that sent the Kospi sharply higher on memory euphoria.
The company framed the outlook as signed LTAs plus high-probability volume still under negotiation with hyperscaler and accelerator customers — not merely aspirational forecasts. Volumes already secured under LTAs account for more than half of total production capacity for 2027 and 2028, SK hynix said, and it intends to keep mid- to long-term LTA coverage at a similar share while leaving the remainder flexible for next-generation products and spot conditions.
What “Sold Out” Actually Means
Memory executives have used scarcity language before. The Culled distinction here is contractual, not rhetorical.
Signed LTAs bind capacity and pricing mechanics for defined periods — the same instrument Micron and peers have cited when AI customers prepay for HBM mix. Pipeline demand under discussion is not yet inked; SK hynix described it as tied to concrete customer purchase plans with high realization probability. Uncontracted capacity remains available for HBM4 transitions, yield learning, and market shocks.
Kim Ki-tae, head of HBM sales and marketing, said customers are requesting more than twice the volume SK hynix can supply, according to the Korea Herald — a demand statement that supports scarcity pricing but does not by itself prove every gigabit ships. DRAM marketing head Park Joon-deok pushed back against AI bubble narratives, citing paid AI subscriptions and capex spreading beyond U.S. megacaps; management also pointed to shortages lasting through 2030 given fab lead times.
What the briefing establishes: multi-year visibility on memory output at the world’s leading HBM vendor after revenue crossed 100 trillion won in a single quarter and operating margin approached 80 percent. What it does not establish: that every “sold out” wafer is non-cancellable if model training economics or export rules shift.
From GPU Queues to Wafer Reservations
The AI economy’s binding constraint is migrating down-stack. A year ago the public story was GPU allocation. After SK hynix’s print — and Seoul’s October memory repricing — the operative question is guaranteed access to physical components: HBM stacks, DRAM die, and the capex to expand cleanroom footprint without repeating governance detours that delay U.S. NAND funding.
HBM4 shipment momentum and improving yields were central to management’s defense of sustained pricing. Samsung’s following earnings print becomes the cross-check: if two Korean memory leaders both describe multi-year LTAs at majority capacity, platform power in AI infrastructure sits with fabs, not model labs — at least until alternative memory architectures or packaging paths scale.
Treat “sold out through 2029” as evidence that buyers are pre-purchasing optionality on silicon, not as proof the cycle cannot break. The dangerous version of this story is the same as every memory supercycle: customers double-order, fabs over-build, and LTAs reprice when spot collapses.
What Would Change the Verdict
Disclosure of cancellation terms, take-or-pay clauses, and volume flex in major LTAs would clarify how much of the 2029 picture is hard backlog versus negotiated intent. Samsung’s guidance on HBM4 pricing and contract coverage is the nearest peer datapoint. U.S. hyperscaler capex cuts or export-control shocks that idle accelerator lines would test whether pipeline demand converts to shipments.
Until then, SK hynix’s statement is the clearest corporate admission yet that AI infrastructure competition has moved from compute auctions to multi-year memory reservations — with more than half of mid-decade capacity already spoken for on paper.
Continue reading
Sources
Yonhap (Oct 28, 2026); Korea Herald; SBS; Seoul Economic Daily; SK hynix Q3 2026 earnings call; Culled coverage of Micron margins and Seoul memory rout.
