SK Hynix Target Raised as Memory Shortage Signals AI Cost Shock

The Call That Underlines a Bigger Problem

Meritz Securities lifted its target price for SK Hynix to 3.2 million won from 3.1 million won. The 3.2% adjustment sounds routine. The reasoning behind it is not.

Research analyst Kim Sun-woo flagged what he called a clear memory supply shortage extending into 2027. He also noted that HBM4 supply has now reached a scale where its profit contribution is accelerating, and that mobile customers who once held leverage are signing contracts on terms increasingly favorable to SK Hynix. The implication stretches far beyond a single Korean chipmaker.

This revision matters because it is one of the few recent analyst moves that explicitly ties semiconductor pricing dynamics to the broader AI infrastructure buildout. It is not simply a stock-specific upgrade. It is a signal that the memory market — long treated as a cyclical backwater compared to logic chips — is now at the center of the AI cost equation.

What Changed in Q3

Meritz expects SK Hynix’s third-quarter operating profit to reach 80.65 trillion won — a 32% jump over the previous quarter and roughly 3% above its own prior estimate. The revision is driven almost entirely by DRAM. Shipments are projected to grow 13% quarter-over-quarter while average selling prices climb 22%. NAND is also climbing, with volume up 4% and pricing up 15%, adding an estimated 17 trillion won in operating profit to the already heavy 63 trillion won from DRAM.

For years, SK Hynix held back on aggressive price increases to preserve long-term customer relationships. That restraint is fading. Kim Sun-woo said the effects of past price hikes are now finally flowing through to earnings, and the company’s highest HBM sales share — previously a drag on average pricing power — is becoming a profit engine as HBM4 ramps.

The sequencing here is significant. The quarter-over-quarter volume growth alone would have been strong under normal market conditions. But when combined with the pricing acceleration, it produces a compounding effect on operating margins. This is the kind ofearnings profile that typically emerges not from operational improvement alone, but from structural supply-demand imbalance — and the analyst’s language makes clear that this is exactly how he reads the data.

HBM4 Is the Pivot Point

High-bandwidth memory has been the bottleneck in every major AI training and inference cluster. SK Hynix already supplies the bulk of HBM used in NVIDIA’s top-tier GPUs. The shift to HBM4 is not a generational upgrade in name alone — it is a structural change in how much memory capacity each AI accelerator can access per second. That shift is what is locking in SK Hynix’s margin expansion.

Meritz highlighted that SK Hynix’s HBM sales mix is the highest in the industry. For a while, that dominance meant the company absorbed pricing pressure because buyers could play competitors against each other. Now the leverage is tilting. With HBM4 entering full-scale supply, each unit carries more value, and the number of qualified buyers remains concentrated among the same handful of hyperscalers and chip designers. That concentration is what turns SK Hynix’s HBM leadership from a vulnerability into a pricing advantage.

The second-order effect of this dynamic is worth noting. As each generation of HBM commands higher per-unit pricing and more demanding qualification timelines, smaller memory producers face increasing barriers to entry. SK Hynix’s ability to qualify HBM4 with NVIDIA’s GPU architectures while Samsung struggles with yield challenges reinforces a widening gap that is unlikely to narrow in the near term.

Who Actually Loses When Memory Tightens

The easy reading is that SK Hynix wins. The harder reading is that someone else pays for it. Data-center builders — the cloud providers, AI platform operators, and infrastructure funds deploying GPU clusters — are the ones buying memory in volume. When the supply curve steepens, they absorb the cost. Their margins compress whether they acknowledge it in earnings calls or not.

This is not a short-term blip. Kim Sun-woo’s forecast that the supply shortage is clear through 2027 means the cost pressure extends well beyond the current AI capital cycle. Companies that assumed memory would remain abundant relative to GPU supply are now revising that assumption upward.

The consequence is a shifting cost structure across the AI value chain. Memory was long considered a relatively stable line item — inexpensive enough that budget discussions rarely dwelled on it. That era is ending. When DRAM prices climb 22% in a single quarter and HBM commands a premium that continues to expand, the total cost of a GPU server node rises meaningfully. Deploying thousands of those nodes changes the economics of every AI project in progress.

The Broader Signal for AI Hardware Economics

The memory market is not isolated. It sits at the intersection of three forces: AI workload growth, fab capacity constraints, and customer concentration. Every new generative model pushes peak memory requirements higher. Every new fab takes years to come online. Every major buyer negotiates from a position of fewer alternatives.

Meritz noted that AI investment competition among big tech companies will continue to underpin medium- to long-term demand. That is the core driver. But the deeper story is what happens when that demand meets a supply side that cannot respond fast enough. Memory becomes a capacity constraint on the entire AI stack — not just a component cost.

One implication that often goes unexamined is the risk of deployment deferral. If memory costs rise faster than budgets anticipated, data-center operators may slow or rephase their GPU purchases. This would not necessarily reduce long-term demand for memory, but it could produce lumpy purchasing patterns that make capacity planning more difficult for everyone in the supply chain. The alternative — absorbing higher costs — simply shifts financial pressure elsewhere in the economy.

SK Hynix’s Share Buyback Adds Another Layer

SK Hynix is wrapping up a 40 trillion won share repurchase program. Meritz expects an immediate new capital return policy to follow. That sequencing matters. It tells you the company believes the current earnings strength is durable enough to sustain shareholder payouts without jeopardizing the capex required to expand HBM and DRAM capacity.

If the supply shortage persists into 2027, that capex pressure will be real. Memory fabrication requires enormous investment, and scaling HBM4 production means competing for the same advanced packaging capacity that TSMC and others control. SK Hynix’s financial flexibility right now is an asset. It may not stay that way for long if demand keeps outpacing supply.

The share buyback also sends a market signal that the company views its current valuation as reasonable relative to the durability of its earnings environment. Management teams do not typically launch capital return programs during periods of uncertainty. This one appears designed to communicate confidence precisely when the broader market may be questioning how long the memory boom can last.

What Comes Next

The target price revision is a small number on a stock quote. The thesis underneath it is larger. Memory supply is tightening because AI demand is growing faster than fab capacity can expand. SK Hynix is positioned to capture disproportionate benefit because of its HBM leadership. And data-center builders are the ones who will feel the squeeze first.

For investors, the question is whether SK Hynix’s margin expansion is sustainable or whether it triggers a chain reaction — higher GPU costs, slower deployment cycles, or a shift in bargaining power among the majors. For the rest of the market, the signal is clearer: the AI hardware cost structure is changing, and memory is the leading indicator. The question now is not whether this adjustment has already begun, but how far the reallocation of cost and margin will travel before the next cycle turns.