Korean Memory Stocks Are Pricing in Nothing—And That Could Be Their Biggest Mistake

The Inventory Number That Changes Everything

Samsung Electronics and SK Hynix—the two companies that together control roughly half the world’s memory chip output—have less than ten days of inventory sitting in warehouses right now.

Ten days. Not the comfortable buffer firms typically hold to absorb demand shocks or production hiccups. This is the kind of number that appears during acute supply crunches, not during a routine recovery cycle. And according to KB Securities, this is only the opening move.

The broker, long regarded as one of Korea’s sharper research outfits on semiconductors, says next year will bring the tightest supply environment in memory history. Not a mild deficit. Not a modest gap between supply and demand. A structural squeeze where the volume of memory chips actually available for sale could run out entirely.

What makes this moment distinct from previous upcycles is the speed with which the inventory buffer has evaporated. In earlier cycles, memory manufacturers typically maintained a several-week cushion, replenishing stockpiles as demand recovered. The current posture—operating with barely enough product to meet immediate orders—suggests the industry has crossed a threshold where the normal oscillation between surplus and scarcity may no longer apply.

Why AI Spending Is Rewiring Memory Demand

The mechanism behind this isn’t subtle. Cloud giants are dramatically accelerating capital expenditure on AI infrastructure, and KB estimates that spend will climb 60 percent year over year to $1.3 trillion next year.

That figure already includes the kind of aggressive reinvestment cycles that followed the generative AI breakout—cloud AI services, token-based billing, agentic AI deployment, model hosting. These are not speculative line items. They are revenue-generating categories that require physical hardware attached to them.

Here’s what makes the demand shift qualitatively different from prior memory upcycles: memory’s share of that AI infrastructure spend is compounding rapidly. KB projects memory will account for 14 percent of AI infrastructure investment in 2025, 40 percent this year, and 57 percent next year. TrendForce, the market research firm, puts next year’s figure even higher—at 68 percent.

That’s roughly a fourfold increase in two years. Memory is no longer a peripheral component of AI servers. It is the core input.

The implication extends beyond raw volume. Each terabyte of AI training capacity requires exponentially more memory bandwidth than traditional compute workloads. A single NVIDIA H100-based training cluster can consume 288 gigabytes of HBM per GPU, and next-generation designs are pushing that number even higher. The memory requirements aren’t just growing—they’re scaling at a different order of magnitude than anything the industry has calibrated to.

The Shortage Won’t Stop at HBM

Most market commentary focuses exclusively on high-bandwidth memory—the HBM stacks that sit directly alongside GPU packages in AI training and inference servers. HBM is undeniably critical, and shortages there have been well documented. But KB’s analysis flags something more alarming for the broader semiconductor ecosystem: the shortage is already spreading beyond HBM into general-purpose DRAM and NAND.

AI server deployments are pulling demand simultaneously for HBM, server-grade DDR5, and enterprise SSDs. Each of those categories competes for the same underlying wafer fabrication capacity. The result is a cascading constraint.

The math is blunt. KB expects DRAM and NAND bit-demand growth next year to outpace bit-supply growth by more than 10 percentage points. That is an exceptionally wide gap for a market that typically balances within a few points over a cycle. Sustaining that gap requires either significant price escalation or severe allocation cuts—or both.

This is not a localized problem confined to data centers. Consumer electronics manufacturers—smartphone makers, PC OEMs, automotive electronics producers—are now competing for the same base memory that once served their categories almost exclusively. The ripple effects are already visible in lead times. Orders for standard DDR5 modules that previously shipped within weeks now face multi-month queues. Enterprise SSD availability has tightened considerably, pushing some cloud operators to reconsider storage architectures altogether.

HBM4 Is Eating Its Own Supply Chain

There is a specific technological bottleneck amplifying the problem. HBM4, Samsung and SK Hynix’s next-generation high-bandwidth memory, requires approximately three times the wafer production capacity per unit compared to standard DRAM. Every wafer routed to HBM4 is three wafers subtracted from the pool available for commodity memory products.

As both companies accelerate HBM4 ramp—which they must do to meet NVIDIA, AMD, and custom AI accelerator demand—the general-purpose DRAM supply picture deteriorates further. The very product delivering the industry’s highest-margin growth is simultaneously tightening supply for the rest of the market.

This is not a hypothetical. It is a direct reallocation of finite fab capacity, and it means that even customers who do not need HBM—the smartphone makers, the PC manufacturers, the data-center operators buying standard DDR5—will feel the squeeze.

The dynamic creates a self-reinforcing loop. As HBM demand intensifies, both Samsung and SK Hynix are incentivized to divert even more capacity toward it, which further constrains standard DRAM supply, which drives prices higher, which validates the capital allocation toward HBM, which pulls still more capacity. The cycle accelerates with each pass.

Second-Order Effects Across the Supply Chain

The consequences extend well beyond the memory manufacturers themselves. Foundry partners are feeling the pressure—TSMC and Samsung Foundry are seeing elevated demand for advanced packaging capacity needed for HBM integration, creating bottlenecks that constrain all customers, not just memory players. Equipment suppliers are experiencing a secondary demand surge as both Korean memory giants invest in new fab capacity and retool existing lines for HBM4 production.

On the demand side, smaller cloud providers and AI startups are being priced out. When enterprise-tier memory allocations are prioritized for hyperscalers, independent AI companies face either cost penalties or delays that could slow their product roadmaps. Some are exploring alternative architectures—computational memory, near-memory processing—in ways that could reshape the competitive landscape over the next several years.

Geopolitical dimensions add another layer. South Korea’s memory industry is concentrated in a narrow geographic footprint vulnerable to seismic and weather risk. Any disruption to the Korean fab base would instantly amplify the shortage globally, and both Samsung and SK Hynix have been quietly diversifying some production exposure, though the scale of that diversification remains limited relative to total output.

The Valuation Disconnect

Despite all of this, the stock prices tell a different story. Samsung Electronics and SK Hynix shares have dropped roughly 38 percent from their recent peaks over the past three months. At current levels, both trade at approximately 3 times forward earnings. That is not merely low by historical standards. It is deep into what KB characterizes as extreme undervaluation territory.

The broker expects both companies to post the largest earnings resets in their respective histories over the next three years, supported by persistent pricing power and sustained capacity constraints. Samsung and SK Hynix are also expected to maintain aggressive shareholder return programs—dividends and buybacks—at scales unseen in recent memory.

For investors focused on multiples rather than megatrends, the valuation math is seductive. For investors tracking actual inventory numbers and AI capex trajectories, it looks like a pricing error.

The disconnect is particularly striking because the market has priced these stocks as if the memory cycle is approaching its peak rather than its inflection point. Historical cycle patterns suggest that when inventory falls below two weeks of supply and demand growth materially outpaces supply growth, earnings revisions tend to move sharply upward—not downward. The stock market, however, has been moving in the opposite direction.

What Comes Next

The critical question is timing. Memory cycles are notoriously difficult to call precisely, and “historic shortage” language from research reports can arrive later than market expectations. But the leading indicators—inventory below 10 days, HBM4 consuming disproportionate wafer capacity, cloud capex forecasts climbing 60 percent—are all pointing in the same direction simultaneously.

For the global supply chain, the implication is clear: prices for consumer electronics built around these chips are likely to remain elevated longer than most analysts expect. For Korean equities specifically, the gap between operational reality and market pricing represents one of the most asymmetric setups in the current semiconductor complex.

The memory shortage is not a future scenario. It is a present condition, and it is intensifying. Samsung and SK Hynix are running their businesses on a knife’s edge—minimal inventory, maxed-out capacity, rising prices, and earnings that are about to reset materially higher. The market hasn’t caught up yet. When it does, the repricing will likely be violent in both directions.

The shortage is already here. The question is whether the market decides to notice before it’s too late.