For decades, investing in memory semiconductors was largely an exercise in understanding cycles.
Demand accelerated. Memory prices rose. Producers increased investment. Supply eventually caught up, inventories accumulated, prices fell, capital expenditure was cut, and the cycle started again.
This extreme earnings volatility also shaped the way memory companies were valued.
At the bottom of the cycle, profits could fall so sharply that conventional price-to-earnings ratios became misleading or even meaningless. Investors therefore often turned to price-to-book ratios as an anchor for valuing enormously capital-intensive semiconductor businesses through the cycle.
The AI hardware revolution does not make that framework obsolete.
Nor does it mean the memory cycle is dead.
But it may be changing the architecture of the cycle itself.
That distinction is the starting point for understanding both South Korea's growing importance in the AI supply chain and the emerging valuation debate surrounding Samsung Electronics and SK hynix.
AI Is Turning Memory Into a System-Level Constraint
The first phase of the generative-AI boom was largely understood as a GPU story.
Increasingly, it is becoming clear that it is also a memory story.
Modern AI accelerators can perform enormous numbers of calculations, but those processors continuously need access to equally enormous quantities of data. As models grow and AI workloads become more demanding, the speed at which data can move between processors and memory becomes a critical constraint on system performance.
This is why High Bandwidth Memory, or HBM, has moved from a specialized memory product to a strategic component of AI infrastructure.
HBM stacks multiple DRAM dies vertically and connects them through high-density interconnects, dramatically increasing the amount of data that can be transferred between memory and an accelerator.
The significance goes beyond faster DRAM.
The architecture increasingly requires advanced DRAM processes, sophisticated stacking and packaging, thermal management, power efficiency and close integration with the logic chips underneath the memory stack.
Samsung's HBM4 illustrates that convergence. The product combines sixth-generation 10-nanometer-class DRAM with a 4-nanometer logic base die. Samsung began mass production and commercial shipments in February 2026 and says HBM4 can operate at up to 13Gbps, delivering maximum bandwidth of roughly 3.3 terabytes per second. news.samsung.com/global/samsung-ships-industry-first-commercial-hbm4-with-ultimate-performance-for-ai-computing
The implication is important.
At the technological frontier, memory is becoming more deeply integrated into the performance of the computing system itself.
HBM Changes More Than Demand
It would be easy to describe HBM simply as a fast-growing new category of DRAM.
That misses the more important economic change.
HBM affects not only how much memory is demanded, but how memory is developed, manufactured and sold.
Advanced HBM consumes sophisticated DRAM manufacturing capacity. It requires additional stacking and packaging processes. Qualification is closely connected to specific accelerator platforms and customer roadmaps. And successive generations are increasingly combining memory-process technology with customized logic.
The relationship between NVIDIA and SK hynix illustrates this shift.
In June 2026, the two companies announced a multiyear technology partnership to co-develop next-generation memory aligned with NVIDIA's AI infrastructure roadmap. The collaboration covers memory for Vera Rubin AI supercomputers and Vera CPUs and extends into personal and physical AI platforms. The agreement addresses the extended development cycles, advanced fabrication requirements and capital investment necessary to sustain future AI-memory supply. nvidianews.nvidia.com/news/sk-hynix-ai-factory
That begins to look different from the traditional model of interchangeable commodity memory sold primarily into a spot- and contract-price cycle.
It does not mean HBM is immune from competition or oversupply.
It means that a differentiated layer is emerging inside a fundamentally cyclical industry. That distinction will become central to valuation.
The AI Effect Is Spreading Beyond HBM
The structural change does not stop with HBM.
One of the most important characteristics of the current cycle is the way AI demand is transmitting into the broader memory market.
HBM competes for manufacturing resources that could otherwise be allocated to conventional DRAM. At the same time, AI data centers require far more than accelerators: high-capacity server DRAM, enterprise SSDs and other parts of the memory hierarchy expand alongside AI compute.
The transmission mechanism increasingly looks like this:
AI infrastructure investment → HBM demand → capacity reallocation → tighter conventional DRAM supply
while, on the demand side:
AI infrastructure investment → more servers and storage → greater conventional memory demand
The two effects can reinforce each other.
The consequences became particularly striking in early 2026. TrendForce estimated that sharp conventional DRAM price increases pushed per-wafer revenue for DDR5 64GB RDIMM above HBM in the first quarter of 2026, with estimated profitability also surpassing HBM. TrendForce also estimates that HBM wafer input among the three major suppliers could rise from roughly 18% of total DRAM wafer input at the end of 2025 to 22% in 2026 and 30% in 2027. trendforce.com/presscenter/news/20260602-13074.html
This does not mean conventional DRAM has suddenly become technologically superior to HBM. It shows how strongly the AI boom can propagate through the economics of the broader memory supply chain.AI is not merely creating a premium memory product.
It is altering the allocation — and opportunity cost — of memory manufacturing capacity.
Why South Korea Matters
This brings us to South Korea. The global AI hardware supply chain is geographically fragmented.
The United States occupies a dominant position in accelerator design and hyperscale computing. Taiwan is indispensable to advanced semiconductor manufacturing and packaging.
South Korea occupies another critical layer: memory.
Samsung Electronics and SK Hynix are two of the world's dominant DRAM producers, and both are increasingly directing technology and manufacturing resources toward AI-oriented memory.
The macroeconomic consequences are already visible.
South Korea's semiconductor exports reached $192.4 billion in the first half of 2026, an increase of 162.6% from a year earlier. Remarkably, that six-month figure alone exceeded the country's previous full-year semiconductor export record of $173.4 billion, set in 2025. Korea's Ministry of Trade, Industry and Resources said the first-half surge was supported by higher memory contract prices and expanding AI-server investment. english.motir.go.kr/eng/article/EATCLdfa319ada/2677/view
The monthly data show how powerful the acceleration became. In June alone, semiconductor exports reached $44.82 billion, up 199.5% year over year, crossing $40 billion in a single month for the first time.english.motir.go.kr/eng/article/EATCLdfa319ada/2677/view
These numbers are extraordinary. But their significance goes beyond export growth. They demonstrate how directly the global AI infrastructure buildout is now transmitting into the Korean economy.
Korea does not dominate the GPU. It does not dominate leading-edge foundry manufacturing. But it occupies a disproportionately important position in one of the resources without which advanced AI processors cannot operate efficiently at scale.
That gives Korean memory producers strategic importance far beyond what their role in the traditional PC and smartphone memory cycles would suggest.
Samsung and SK Hynix Are Not the Same AI Story
Yet investors should resist treating Samsung Electronics and SK Hynix as interchangeable expressions of the same theme.
SK Hynix entered the AI-memory transition with a particularly strong HBM position and increasingly deep integration with NVIDIA's AI ecosystem.
The 2026 multiyear partnership with NVIDIA is important not simply because it involves additional memory supply. It connects future memory development more directly with NVIDIA's infrastructure roadmap. nvidianews.nvidia.com/news/sk-hynix-ai-factory news.skhynix.com/multi-year-tech-partnership-with-nvidia/
The company's technology roadmap is also advancing rapidly. In June 2026, SK hynix began shipping samples of its 12-layer HBM4E to major customers. The product reaches a maximum 16Gbps per pin, while SK hynix says its Advanced MR-MUF technology reduces thermal resistance by 17% compared with HBM4. news.skhynix.com/12-layer-hbm4e-sample/
Samsung's position is different.
It combines one of the world's largest DRAM manufacturing bases with foundry, logic and advanced-packaging capabilities. That broader semiconductor footprint could become increasingly relevant as HBM architectures bring memory and logic closer together.
Samsung began mass production and commercial shipments of HBM4 in February 2026 using its 1c DRAM process and 4nm logic base die. At the time, the company said it expected 2026 HBM sales to more than triple from 2025 and that it was proactively expanding HBM4 production capacity. news.samsung.com/global/samsung-ships-industry-first-commercial-hbm4-with-ultimate-performance-for-ai-computing
Then the roadmap accelerated. On May 29, Samsung announced that it had begun shipping 12-layer HBM4E samples to major global customers. The product offers a stable 14Gbps pin speed scalable to 16Gbps, with maximum bandwidth of 3.6TB/s per stack.
This makes the competitive landscape more dynamic than a simple snapshot of current HBM market share would suggest.
Each generation — HBM3E, HBM4, HBM4E and eventually custom HBM — creates another opportunity for technology leadership, customer qualification and market share to change.
That matters for equity valuation.
If technological execution produces persistent differences in margins, returns on capital and earnings durability, Samsung and SK Hynix should not necessarily trade according to the same valuation logic even though both remain exposed to the same underlying memory cycle.
Today's Shortage Creates Tomorrow's Supply
This is where the argument against a permanent AI-memory “supercycle” becomes important.
High prices create high profits → High profits attract capital → And capital eventually creates supply.
That mechanism has not disappeared. Micron provides a useful global comparison.
In fiscal Q3 2026, Micron reported revenue of $41.46 billion, compared with $9.30 billion in the same quarter a year earlier. GAAP gross margin reached 84.6%. Yet perhaps the more important number for understanding the next cycle is capital expenditure: Micron invested $7.1 billion in a single quarter. https://investors.micron.com/financials/quarterly-results/default.aspx
Micron's SEC-filed Form 8-K provides an additional primary-source cross-check of the results.
At the same time, Micron said its multiyear Strategic Customer Agreements should enhance the durability and predictability of its financial performance. Those developments capture the tension at the heart of the new memory cycle.
On one side, customer relationships may be becoming longer-term, technology barriers higher and AI-related demand more structural. On the other, extraordinary profitability is generating extraordinary investment.
The first force can weaken some characteristics of the old commodity cycle. The second recreates them. AI demand may be structural. Semiconductor supply still responds to returns on capital.
The question is not whether the cycle disappears. It is what the next cycle looks like when these two forces collide.
A Differentiated Layer Inside a Cyclical Industry
This is the central idea.
AI is not transforming memory from a commodity industry into a non-cyclical technology industry. It is creating a more differentiated layer inside an industry that remains fundamentally cyclical.
HBM customer qualification matters.
Packaging capability matters.
Power and thermal performance matter.
Time-to-market matters.
Increasingly, the ability to combine advanced memory with logic technology matters.
But capacity still matters.
Inventory still matters.
Pricing still matters.
Capital expenditure still matters.
The interaction between them — rather than the disappearance of either one — is what investors need to understand.
Why This Changes the Valuation Question
This brings us back to P/E and P/B.
Historically, P/B worked particularly well as a valuation anchor for memory companies because earnings were extremely cyclical.
At the bottom of the cycle, earnings could disappear while the productive assets, technology and balance sheet of the company remained. P/E could therefore send exactly the wrong signal: a memory stock could appear most expensive precisely when earnings were closest to the bottom.
Book value provided an alternative anchor. But what happens if AI changes the earnings structure without eliminating cyclicality?
Suppose the next generation of memory producers experiences higher normalized margins, stronger through-cycle ROE, longer customer commitments and greater product differentiation — while still remaining exposed to capacity and pricing cycles.
Then P/B should not disappear.
But P/E may deserve greater weight.
This is the possibility at the center of the GeoMarketSignal Memory Cycle Series:
The AI hardware revolution may be shifting Korean memory stocks toward a new valuation equilibrium between P/B and P/E.
P/B continues to reflect the industry's capital intensity and cyclical downside.
P/E increasingly reflects the possibility of more durable structural earnings.
The investment problem is determining how much of today's extraordinary profitability represents the top of another memory cycle — and how much represents a genuine upward shift in normalized earnings power.
South Korea is where that question matters most. Samsung Electronics and SK Hynix now sit at the intersection of two identities. They remain cyclical memory manufacturers. But they are also becoming strategic suppliers to the global AI infrastructure buildout.
The valuation regime that emerges from that tension may look very different from the one investors used in previous memory cycles.
And that is where we turn next.
The AI Memory Cycle & Korean Semiconductors
This article is Part I of the GeoMarketSignal AI Memory Cycle series.
← Previous: AI Is Reshaping the Memory Cycle: How Should Investors Value Korean Semiconductor Stocks?
Next → Part II: The Memory Cycle Is Not Dead: Why P/E and P/B Send Opposite Signals