SK hynix: HBM turns “cyclical memory” into a structural winner

Erik@YWR on SK hynix Inc. (000660-KRX | skhynixinc00)

8/8/2025

Summary

SK hynix (000660 KS) is leading the most important shift in memory in decades: the rise of High Bandwidth Memory (HBM) to relieve the “memory wall” throttling AI accelerators. Unlike past DRAM cycles, HBM’s economics are different—capacity is built against orders, packaging yields are a moat, and wafer allocation tightens legacy supply—so profits look more durable than the market assumes. In your analysis, SK hynix has >60% HBM share, enjoys a measurable yield/process advantage, and is effectively the preferred partner to the leading GPU vendor—positioning it at the center of AI capex with “Nvidia-like” growth but a far lower multiple. In Q2 revenues were up ~35% with operating profit +68% and net profit +70% . There is a potential price target of KRW 56,000 EPS by 2027 and a rerating to ~10x P/E (~KRW 560,000 per share, +~110%) if investors recognize the structural change.

Thesis

1) The AI memory wall made HBM mission-critical—and SK hynix leads As GPUs sprint ahead, system performance is increasingly constrained by memory bandwidth and capacity. That bottleneck thrust HBM from a niche curiosity into the default companion for state-of-the-art accelerators. SK hynix pioneered stacked-DRAM HBM (initially eight dies) as far back as 2015—years before the mainstream AI use case was obvious—and kept iterating. Today it holds over 60% HBM market share and sits “at the centre of AI development.” That leadership isn’t just about first-mover advantage; it is process-driven. HBM is as much a packaging challenge as it is a memory challenge: stacking ultra-thin dies, managing thermals and power delivery, and—crucially—bonding. Samsung pursued TC-NCF while SK hynix (with Namics) advanced MR-MUF, and the yield gap has been large enough to cost Samsung meaningful share and contracts. Yield is the moat; it compounds volume leadership, lowers cost, and makes you the safer bet for hyperscalers who can’t afford schedule risk. Why this matters for the P&L: HBM isn’t a commodity DRAM stick—the integration, validation, and thermal/mechanical constraints create switching costs. As long as hyperscalers and leading GPU vendors prioritize time-to-deploy for AI clusters, the supplier who can ship predictable volumes at high yield commands better pricing and visibility. That’s SK hynix right now. 2) The business model changed: from “build-and-sell” to “order-make-deliver” Classic memory cycles were brutal because producers built capacity ahead of demand, chasing economies of scale. The revenue line went parabolic, then ASPs collapsed as inventories swelled. With HBM, the flow is different: SK hynix is booking orders first, then building capacity and shipping against committed demand. That “order-make-deliver” discipline softens the amplitude of the cycle and supports margin durability. Just as important, wafer allocation has gone in reverse: HBM die sizes are increasing (fewer chips per wafer), and SK hynix (and peers) are redirecting wafer starts away from legacy DRAM/NAND into higher-margin HBM. That tightens supply in legacy categories and props up pricing there, providing a second-order earnings boost while HBM ramps. It’s the mirror image of the old dynamic that perpetually flooded the market. 3) HBM is packaging-heavy—technical moats are real, not theoretical Memory investors have been trained to assume that any advantage evaporates within 12–24 months as everyone copies the next node. HBM breaks that pattern because packaging yields (stacking, bonding, underfill) and thermals dominate cost and reliability. SK hynix’s MR-MUF bonding process has delivered better yields than Samsung’s TC-NCF—so much so that it contributed to SK hynix winning the lead HBM slots with the top GPU provider. The company is already researching hybrid bonding next, which raises the bar again. This is hard tech, not just lithography. When the #1 memory company by history struggles to get an HBM generation out with competitive yields, you know you’re not trading a pure commodity. That’s the structural, non-consensus angle: hardware process edges are moats—and in AI build-outs, moats convert into P&L. 4) Financials reflect the new structure—and could still be conservative Revenues are growing +35%, with operating profit +68%, and net profit +70%. The core driver being HBM mix, yield leverage, and tight legacy supply. Looking ahead, we assume “only” ~20% growth in 2026–2027, which takes EPS to KRW 56,000 by 2027. Put a 10x multiple on that and you get KRW 560,000—about +110% upside. It’s “Nvidia like growth on ~7x earnings,” given gross margins approaching 50%—a striking spread versus AI beneficiaries that trade at software-like multiples. Is 10x fair? If HBM remains order-backed and capacity additions are rational, cyclical downside should compress versus history. That merits some multiple expansion. Meanwhile, if hyperscalers overbuild AI memory pools (common in early platform phases), SK hynix gets a volume/mix tailwind; if spend normalizes, the wafer-allocation dynamic can still keep legacy pricing supported. 5) Why the market is mispricing it • Anchoring to the old DRAM playbook. Investors remember every painful down-cycle and reflexively fade memory spikes. But HBM’s process/packaging stack and order discipline are new; assuming the old pattern repeats ignores the changed physics and incentives. • Underestimating yield moats. Yield differences aren’t just a talking point—they gate supply, determine who wins sockets, and move margins for years, especially when qualification cycles are long and risk tolerance is low among hyperscalers. • Ignoring second-order benefits. Shifting wafers to HBM shrinks supply in commodity DRAM/NAND; even a temporary reprieve on pricing expands EBIT dollars far beyond HBM alone. • “It’s still memory.” Yes, but the most strategic kind. In AI datacenters, idle GPU minutes are a five-figure opex line item—operators will overprovision HBM before they risk starving compute. The value of reliable HBM supply is not symmetrical with commodity DIMMs. ( “hardware is the new software” theme.) 6) Catalysts • HBM ramps/attach rates. Watch announced HBM content per accelerator generation and SK hynix’s volume milestones. Sustained leadership here validates the yield/process moat. • Legacy pricing resilience. Evidence that DRAM pricing stays tighter due to wafer reallocation will reinforce the “reverse capacity” narrative. • Process transitions. Updates on hybrid bonding or further MR-MUF improvements would extend the moat. • Customer wins/visibility. Continued alignment with the leading GPU vendor and initial inroads with alternative accelerators (or incremental hyperscalers) de-risk concentration.

Risks

1. Samsung catch-up / share loss. The key bear case is that Samsung closes the yield gap quickly and re-wins leading GPU sockets, eroding SK hynix’s HBM pricing power and growth visibility. 2. Reversion to old memory behavior. If “order-make-deliver” slips back into build-ahead, inventories rise and ASPs crack. This is the primary structural risk because it collapses the very differentiation that’s supporting margins. 3. Customer concentration / pricing pressure. Being deeply tied to the leading GPU vendor is a strength and a dependency. If that buyer tilts allocations, dilutes content, or renegotiates aggressively, it can sting. 4. Technology transition risk. Moving from MR-MUF to hybrid bonding (or to next-gen HBM4/5) could reset yields temporarily. If a competitor nails the transition first, the share dynamic can flip. 5. Supply-chain & process shocks. HBM requires specialized materials, precise thermomechanical controls, and advanced substrates. Any disruption (equipment bottlenecks, underfill/adhesive issues, substrate tightness) can cap shipments. 6. Geopolitics / export controls. As a Korea-based supplier into global AI stacks, SK hynix can be exposed to shifting export regimes. (General industry risk; not specific to your note.) SK Hynix is building a $3.8 billion chip packaging plant in Indiana, but its existing fabs are in South Korea and China.

📈 Price Targets

Tags