Samsung's Quiet HBM4 Comeback Is Rewiring the Memory Chip Order

The Number That Should Make English-Language Desk Heads Pause

LS Securities just cut SK Hynix’s target price by 27 percent — from 3.3 million won to 2.4 million won — while simultaneously raising Samsung Electronics’s target from 400,000 won to 450,000 won. Two days ago, no one writing about the AI chip boom in English was paying attention. This matters.

The adjustment isn’t about weaker demand. It’s about who gets to supply it.

HBM4 Is the New Battleground

High-bandwidth memory is the bottleneck product of the AI era. Every major data-center GPU — NVIDIA’s H100, Blackwell, and beyond — depends on it. SK Hynix held that throne for years. Samsung, its fiercest rival, had been lagging behind, especially on the latest generation.

Then HBM4 arrived.

According to LS Securities, Samsung’s HBM4 share of total HBM shipments surged from roughly 5 percent in the first quarter to about 35 percent in the second quarter. That is not a gradual improvement. That is a inflection. Meanwhile, Samsung’s HBM4 mixed yield improved by more than 5 percentage points quarter over quarter — a pace that LS called notably smoother than Samsung’s messy debut with HBM3E.

The implication: Samsung is learning faster than anyone expected, and SK Hynix’s decade-long lead in HBM yield and reliability is narrowing.

Second-Order Effects No One Is Tracking

The yield shift has consequences that extend far beyond Samsung and SK Hynix’s balance sheets. NVIDIA, which has staked its roadmap on HBM availability, now faces a supplier landscape that is less concentrated than it was six months ago. When LS raised its Samsung target and cut SK Hynix’s by such a wide margin, it was implicitly acknowledging that NVIDIA’s supply risk is migrating — from a single-source dependency toward a credible dual-source architecture.

That matters for pricing, inventory planning, and the broader AI infrastructure buildout. GPU makers that were quietly hedging their bets with SK Hynix can now diversify their sourcing without sacrificing performance. The margin compression LS projected for SK Hynix’s HBM business — from 80 percent down to roughly 60 percent — is a structural change, not a cyclical bump. At the scale of AI data-center deployments, a 20-point compression translates into tens of billions of won in annual profit per competitor.

There are also downstream effects on the broader DRAM market. Samsung’s HBM4 success is being funded by its conventional memory division, which LS noted is enjoying a rare dual lift in both earnings estimates and valuation multiples. That pattern — a high-margin product subsidizing a core business upgrade — rarely sustains across memory cycles, but LS sees it as plausible here given the volume trajectory.

Who Gains, Who Loses

Samsung gains first and most visibly. LS raised its HBM shipment and profitability estimates for the company, noting that each additional percentage point of HBM4 mix is flowing into earnings faster than previously assumed. Analyst Jung Woo-sung warned that once HBM4 starts contributing meaningfully to Samsung’s memory division profits, it could reshape the earnings profile of Samsung’s entire conventional DRAM business.

SK Hynix loses in a subtler but more economically consequential way. LS did not downgrade its rating — it remains a buy. But it slashed its 2027 HBM operating margin forecast from an eye-popping 80 percent down to roughly 60 percent. That gap matters enormously at scale. SK Hynix’s HBM business is already generating more profit per chip than any other memory segment on the planet. A 20-point compression is a multibillion-won revision.

The Premium Is Normalizing

The core of LS’s thesis is that the market has been pricing SK Hynix as a scarcity play. Its early HBM4 dominance gave it pricing power — what LS calls the “supplier premium.” Samsung’s rapid HBM4 ramp means that scarcity is evaporating. Major customers, particularly NVIDIA, have a credible second source. The industry call is not that demand is slowing; it’s that the supply side is becoming less concentrated.

Jung was explicit: the adjustment reflects a competitive landscape normalizing, not a demand collapse.

For global investors, this distinction is everything. The popular narrative in English-language coverage has been that SK Hynix is the AI memory winner and Samsung is the laggard catching up. LS’s report turns that story inside out — Samsung may be catching up faster than the narrative assumes, and the upside of that catch-up is being captured by Samsung, not SK Hynix.

What Happens Next

The near-term risk for SK Hynix is clear: every additional wafer Samsung ships to NVIDIA or other AI GPU makers reduces the monopoly premium SK Hynix has enjoyed. Jung noted that further yield improvements could temporarily restore margin pressure, but the directional trend favors Samsung.

Samsung’s upside scenario is equally concrete. LS projects that if HBM4 volume continues to expand as shipments confirm, the company could see simultaneous upgrades to both earnings estimates and valuation multiples — a rare dual lift that rarely happens in memory cycles.

SK Hynix still has a path. Jung cited a potential 2028 price-to-book ratio of 1.6 times if new AI demand sources open up and yields continue to improve. But that ceiling is lower than the one LS originally imagined.

Why English-Language Coverage Has Missed This

The dominant framing in global financial media has been straightforward: SK Hynix is the HBM champion, Samsung is the underdog. That framing was true in 2024 and early 2025. But Samsung’s Q2 HBM4 data — a sevenfold increase in shipment share in a single quarter — is the kind of signal that should trigger a rewrite of that story. The fact that it hasn’t yet is a coverage gap, not a market signal.

LS Securities is a Korean broker speaking to a domestic audience that already knows the competitive dynamics intimately. An English-language audience, reading headlines about SK Hynix’s stock run and NVIDIA’s supply chain dependencies, is likely to walk away thinking nothing has changed. That would be wrong.

The HBM4 arms race is no longer a one-horse story. And the horse that was supposed to be trailing is now pulling ahead.