SK Group chairman: HBM/memory is the AI buildout's binding constraint, with the shortage forecast to run to ~2030
At Computex on 2 June 2026, SK Group chairman Chey Tae-won (not a separately-quoted SK Hynix CEO) said the AI-driven memory/HBM shortage will likely persist through about 2030, with industry wafer supply running more than 20% below demand and a greenfield fab taking five-plus years to build; SK Hynix aims to roughly double wafer capacity over five years and has effectively sold out 2026 HBM. This is a forward-looking forecast from an interested supplier, not a settled fact, and it is the memory maker who benefits from a shortage narrative. The widely-cited figure that AI data centers will absorb ~70% of (high-end) memory in 2026 is a separate analyst/industry estimate, not part of Chey's remarks.
Why it matters (spine). Compute is not only logic; it is the physical memory the logic runs on, and the buildout's binding constraint may be the wafers, not the models. An interested supplier now puts the AI memory/HBM shortage on a multi-year clock — running to roughly 2030 — because industry wafer supply sits more than 20% below demand and new capacity takes years to stand up. Treat this as signal, not settled fact: it is a forward-looking forecast from the memory maker who profits from a scarcity narrative, and 2026 HBM being "sold out" is itself a sales claim. But even discounting the incentive, the shape holds — while reasoning trends toward free, the atoms it runs on stay scarce for years, which is exactly the kind of durable input a moat is built on.
What would prove it wrong (falsifier). A demand collapse or a step-change in HBM yield closes the supply gap well before 2030.
Live-status check + triangulation. Quoted from the SK Hynix newsroom's Computex 2026 review, attributing the remarks to SK Group Chairman Chey Tae-won (Computex, 2 June 2026): "Building a new fab takes at least three years; a greenfield development starting from land preparation takes five or more." Corroborated by TrendForce (SK chair sees memory shortage through 2030, eyes capacity doubling), TheElec, and TechRadar (data centers to take ~70% of high-end memory in 2026 — a separate industry/analyst estimate). Fetched from the SK Hynix newsroom page in-session.
Correction carried forward (discipline). (1) The draft claimed both SK Group's chairman AND SK Hynix's CEO put the shortage at 2030 — every primary and secondary source attributes the remarks to Chairman Chey Tae-won alone; no SK Hynix CEO is quoted separately (the co-attribution was fabricated). (2) The draft folded "~70% of memory output to AI data centers" into Chey's Computex remarks — the ~70% figure is a separate industry/analyst estimate (data centers to take ~70% of HIGH-END memory in 2026, per TrendForce/Counterpoint-associated reporting), not stated by Chey, and it is not "memory output" overall. (3) The draft said flatly "a new fab takes 5+ years" — Chey's own words (SK newsroom): a new fab takes at least three years, and a greenfield build from land preparation takes five or more; the 5+ figure applies specifically to greenfield-from-scratch. (4) The draft tier "verified" was downgraded to "signal": the assertion is a forward-looking forecast from an interested supplier, and the original capsule carried a fabricated co-attribution plus a mis-attributed statistic.
Card it as the load-bearing constraint on the Moat Clock: while reasoning trends toward free, the atoms it runs on stay scarce for years.
domains: machine-layer · hbm · memory · supply-constraint · moat · compute