For decades, investors in Korean memory stocks learned a counterintuitive lesson: sometimes, the time to become interested was when the price-to-earnings ratio looked expensive — and the time to become cautious was when it looked cheap.
There was a reason.
Memory semiconductors were among the world's most cyclical industries. At the bottom of the cycle, collapsing memory prices crushed earnings. P/E ratios could rise dramatically even as share prices approached a bottom.
Near the top of the cycle, the opposite happened. Memory prices and profits surged, earnings expanded rapidly, and P/E ratios fell. Stocks could suddenly look remarkably cheap precisely when the earnings cycle was becoming increasingly mature.
For this reason, investors in companies such as Samsung Electronics and SK hynix could never rely on P/E alone. P/B, inventory levels, memory pricing, capital expenditure, capacity utilization and expectations for the next supply cycle all mattered.
Then came artificial intelligence.
And the old valuation framework became considerably more complicated.
AI Is Changing the Economics of Memory
The AI revolution is often described as a software revolution. But underneath large language models, AI agents and increasingly sophisticated applications lies an enormous physical infrastructure.
GPUs need memory. AI servers need memory. Data centers need memory. And as computing power increases, moving enormous quantities of data quickly enough to feed those processors has become one of the central technological challenges of AI infrastructure.
That has elevated high-bandwidth memory, or HBM, from a specialized memory product into a critical component of advanced AI computing.
The impact is spreading beyond HBM itself. Memory manufacturers are allocating valuable production capacity toward HBM and server products, constraining conventional DRAM supply at the same time that AI infrastructure is creating additional memory demand.
TrendForce currently expects AI inference, expanding server deployment and HBM's increasing consumption of DRAM production capacity to keep the DRAM market structurally tight into 2027. Yet it also sees a different picture emerging in NAND, where additional capacity and weaker consumer demand could create looser supply conditions. trendforce
That divergence itself is important.
There may no longer be one simple "memory cycle."
Why South Korea Is at the Center of This Debate
Few economies are more exposed to this transformation than South Korea.
Korean semiconductor exports reached an all-time high of $173.4 billion in 2025, up 22.2% from the previous year. Korea's Ministry of Trade, Industry and Resources attributed the strength partly to AI data-center demand and higher memory-chip prices. Ministry of Trade
Meanwhile, the technological competition in HBM continues to accelerate. Samsung Electronics began mass production and commercial shipments of HBM4 in February 2026 and followed with HBM4E samples in May. https://news.samsung.com/global/samsung-ships-industry-first-commercial-hbm4-with-ultimate-performance-for-ai-computing
This places Korea at the intersection of several forces simultaneously:
AI infrastructure investment. Memory supply and demand. Semiconductor capital expenditure. Technological competition. And increasingly, geopolitics.
For investors in Korean equities, however, this creates a difficult problem.
How should these companies now be valued?
Two Opposing Views of the Memory Cycle
The debate can broadly be divided into two camps.
The first argues that AI has structurally changed the memory industry.
HBM has higher technological barriers than conventional commodity memory. AI infrastructure spending is creating new sources of demand. Advanced memory is becoming increasingly important to overall computing performance, while the production requirements of HBM can also constrain conventional DRAM capacity.
From this perspective, historical memory valuation ranges may underestimate the industry's new earnings power.
The second camp sees a much more familiar story.
Memory remains a manufacturing business. High profitability eventually encourages capital expenditure. New fabs eventually produce new supply. Competitors expand capacity. Customers accumulate inventory. Demand growth eventually slows.
And when supply begins to exceed demand, prices fall.
From this perspective, declaring the end of the memory cycle may simply be another version of the dangerous phrase investors hear near the top of many cycles:
"This time is different."
Both arguments contain something important.
And that is precisely the problem.
The P/E Vs P/B Problem
If the traditional memory cycle remains largely intact, investors should remain cautious about interpreting a low P/E ratio as evidence that a memory stock is cheap.
Peak earnings can produce deceptively low multiples.
P/B can therefore remain valuable because book value is considerably less volatile than earnings across the cycle. Historically, investors could compare valuations against the company's asset base while watching memory prices, inventories, capex and the supply-demand balance for clues about where they stood in the cycle.
But AI introduces a new question.
What if normalized profitability itself is changing?
If HBM, AI servers and increasingly memory-intensive computing structurally raise margins, return on equity and the durability of earnings, then simply applying the P/B ranges of previous memory cycles may also become misleading.
This is where I believe the most interesting valuation debate begins.
The question may no longer be:
Should investors use P/E or P/B?
The better question may be:
How much of today's earnings should be treated as cyclical, and how much deserves a structural AI premium?
That distinction matters enormously.
Treat too much of the current profitability as permanent, and investors risk paying a structural-growth valuation near the peak of a semiconductor cycle.
Treat all of it as temporary, and investors may value a transformed industry using a framework designed for an economic structure that no longer fully exists.
The market may therefore be searching for a new equilibrium between P/E and P/B — between the old memory cycle and the new economics of AI hardware.
Finding that equilibrium will be one of the central questions of this series.
We Are Entering Uncharted Territory
There is another reason I am reluctant to declare either side of the debate correct.
We have never experienced an AI hardware investment cycle of this scale before.
The current market is not simply repeating the PC cycle, the smartphone cycle or even the cloud-computing cycle.
AI training is increasingly being joined by inference and agentic AI, potentially expanding memory demand across HBM, server DRAM and other parts of the memory hierarchy. TrendForce, for example, has sharply increased its forecasts for the global memory market as it expects agentic AI and inference workloads to drive substantially greater memory requirements. https://www.trendforce.com/presscenter/news/20260529-13068.html
At the same time, today's shortage contains the seeds of tomorrow's supply response.
Samsung, SK hynix and Micron are expanding capacity and advancing manufacturing processes. New facilities will eventually come online. And the economics of HBM versus conventional DRAM can themselves alter how manufacturers allocate production. TrendForce has already noted that conventional DDR5 server memory became more profitable per wafer than HBM in early 2026, illustrating how quickly production incentives can change. https://www.trendforce.com/presscenter/news/20260602-13074.html
This is exactly why simple narratives are dangerous.
AI is structural. Semiconductor supply is cyclical. Both can be true at the same time.
And Then There Is Geopolitics
The valuation problem becomes even more complicated when geopolitics enters the equation.
Samsung Electronics and SK hynix are Korean companies operating across a technology system increasingly divided by U.S.–China strategic competition.
U.S. export controls affect access to semiconductor manufacturing technology in China. In 2025, the U.S. Bureau of Industry and Security removed Samsung China Semiconductor and SK hynix Semiconductor China from its Validated End-User authorization list, changing the regulatory framework surrounding their Chinese manufacturing operations.
Meanwhile, Korea is attempting to strengthen its domestic semiconductor ecosystem. In August 2026, the Korean government announced plans for a KRW 5 trillion semiconductor fund targeting materials, components, equipment and fabless companies, alongside another KRW 5 trillion in trade finance for suppliers. https://www.reuters.com/world/asia-pacific/south-korea-establish-fund-semiconductor-materials-parts-equipment-official-says-2026-08-10/
China's own memory ambitions add another variable.
The future valuation of Korean semiconductor companies therefore cannot be understood through earnings and book value alone.
It increasingly sits at the intersection of:
P/E + P/B + ROE + Memory Pricing + Capacity + AI Capex + Technology + Geopolitics.
That is the framework this series will explore.
What This Series Will Examine
This opening article does not attempt to answer the valuation question.
Instead, it sets out the question we will try to answer.
The series will proceed in three parts.
Part I — AI Is Reshaping the Memory Cycle: Why South Korea Matters
We will begin with the industry itself.
Why has HBM become so important to AI computing? How is HBM changing the economics of DRAM production? Why does Korea occupy such an important position in this market?
And importantly, we will examine why Samsung Electronics and SK hynix should not necessarily be viewed as identical AI-memory stories.
Part II — The Memory Cycle Is Not Dead: Rethinking P/E and P/B for Korean Semiconductor Stocks
This will be the core valuation study.
We will examine how P/E and P/B behaved through previous memory cycles and why memory stocks have historically produced the strange phenomenon of looking expensive near the bottom and cheap near the top.
Then we will test the harder question:
Has AI changed normalized earnings, margins and ROE enough to justify a new valuation range?
Rather than assuming the answer, we will look at the data.
Part III — AI, China and the New Korean Semiconductor Order
Finally, we will broaden the analysis beyond Samsung and SK hynix.
China's memory ambitions, U.S. technology controls, Korea's industrial policy and the restructuring of semiconductor supply chains could affect not only the two memory giants but also Korea's wider ecosystem of materials, equipment, components, packaging and testing companies.
This is where semiconductor analysis becomes geopolitical analysis — and where geopolitics ultimately returns to the stock market.
The Question We Are Trying to Answer
AI has clearly changed the memory industry.
But change is not the same thing as the elimination of cyclicality.
The mistake may be to frame the debate as a binary choice between a permanently transformed AI growth industry and an ordinary semiconductor cycle approaching another peak.
Reality may lie somewhere between those extremes.
The challenge for investors is to determine where.
How much of today's extraordinary profitability is cyclical?
How much represents a structural change in the economics of memory?
What valuation premium, if any, should investors assign to that structural change?
And ultimately:
Where is the new equilibrium between P/E and P/B for Korean semiconductor stocks in the age of AI?
That is the question this series will attempt to answer.
The AI Memory Cycle & Korean Semiconductors
This article is the introduction to the GeoMarketSignal AI Memory Cycle series.
Next → Part I: AI Is Reshaping the Memory Cycle: Why South Korea Matters