The memory supercycle in three acts: SK Hynix's $12.1B special dividend, Goldman cutting KOSPI by 25% while calling it the world's cheapest, and 2028's turning point

The memory supercycle's capital drama is running in three acts: SK Hynix reportedly plans to pay out about 16 trillion won ($12.1 billion) in a special dividend — funded by the proceeds of its full exit from Kioxia; Goldman slashed its KOSPI 12-month target from 12,000 to 9,000 (-25%) while simultaneously warning "the market is pricing in an excessively pessimistic fundamental outlook"; and SemiAnalysis founder Dylan Patel's read is that substrate and memory capacity constraints won't ease soon, with industry memory gross margins reportedly at 88-90%^[1][2][3].

SK Hynix's $12.1 billion red envelope: the Kioxia exit's curtain call

Per Korea's Dong-A Ilbo, citing investment-banking sources, SK Hynix plans to return the bulk of its Kioxia sale proceeds to shareholders as a special dividend, with the formal announcement expected at the Q3 earnings call on October 28. The numbers' provenance: when Bain Capital led the 2018 acquisition of Toshiba Memory (later Kioxia), SK Hynix bought Kioxia shares and convertible bonds through two SPVs; as Bain sold the final Kioxia tranche this year, Hynix completed its exit — booking about 20 trillion won ($15.1 billion) in net gains after costs, of which $12.1 billion now returns as a special dividend^[1]. Cashing out a legacy investment at the peak of the AI memory boom, then returning it as cash — the timing is itself a statement about where the memory cycle stands.

Goldman's contortion: a 25% cut while calling it the world's cheapest

Capital markets are pricing along a different track. Goldman's strategists cut the KOSPI's 12-month target from 12,000 to 9,000 in their October 25 note — against a backdrop of the index falling nearly 40% from June's peak; yet Goldman stresses the revised target still implies over 60% upside from current levels, and "the market is pricing an excessively pessimistic fundamental outlook"^[2]. The data are extreme: the KOSPI's 12-month forward P/E has fallen to about 4x — three standard deviations below its historical mean and an extreme unseen among 18 trillion-dollar-plus markets since 2006 — roughly half Indonesia's Jakarta Composite at 8x^[2]. Money has begun responding: Korean retail investors are reportedly piling back into Samsung and Hynix leveraged ETFs — domestic dip-buying in the memory duo intensifying even as prices slump^[2].

The SemiAnalysis roundtable: bottlenecks, 88% margins and 2028's pivot

Industry research supplies the cycle's technical base map. SemiAnalysis founder Dylan Patel told the CASPA annual conference that AI compute demand remains strong, but capacity constraints in substrates, memory and other key links are capping infrastructure expansion; financing costs have surged (the 10-year Treasury above 5%, some neoclouds' unsecured bonds near 15%), making capital harder to obtain for emerging clouds without long-term customer contracts — whether AI investment proceeds depends on financing capacity as much as demand^[3]. Memory's supply-demand picture splits in two: HBM demand is brisk with industry gross margins reportedly at 88-90% today; but if new capacity concentrates into production after 2028 while demand growth slows, memory prices and margins could come under pressure — 2028 is this memory bull market's first preset inflection^[3].

What to watch

First, SK Hynix's October 28 earnings — the dividend's official confirmation and HBM order guidance. Second, Samsung's capex language at its own report the same week. Third, fund flows at a 4x P/E KOSPI — whether retail dip-buying spreads to institutions. Fourth, the 2028 capacity timeline — it decides whether this memory cycle is a "long bull" or a "bubble"^[1][2][3].

The three acts’ shared protagonist: HBM’s pricing power

Laid side by side, the three materials share one protagonist: HBM (high-bandwidth memory). Hynix’s dividend confidence comes from the Kioxia legacy stake, but its current earnings engine is HBM — locked supply to Nvidia’s AI chips gives it the best seat among the memory trio; Goldman’s “excessively pessimistic” read on the KOSPI rests precisely on the market pricing memory stocks like the old commodity cycle while ignoring HBM’s contract-locked nature; and SemiAnalysis’s reported 88-90% gross margins quantify the richness of this lane — against a long-run memory margin midpoint of 30-40%. Understanding HBM’s pricing power explains why this memory boom differs from prior cycles: the last was a supply-demand mismatch cycle; this one is a mandatory component inside the AI-compute architecture — as long as AI training and inference keep expanding, HBM’s demand curve does not obey the traditional memory cycle function. The risk is equally clear: if 2028’s new capacity releases into slowing AI growth, HBM’s high margins will also be the steepest part of the drawdown.