As shares of Samsung Electronics and SK hynix keep hitting record highs day after day, the industry's attention has already turned to the next question: when and how will the boom end? This reflects a kind of learned wariness the memory chip sector has acquired over decades of cyclical ups and downs. The memory semiconductor industry has ridden a roller coaster of booms and busts on a regular basis.
The perception that the memory industry is a highly cyclical business swinging between extremes was cemented by the "chicken game" of the 1990s. After a brief boom driven by the PC adoption craze, the global DRAM market saw prices collapse as more than 10 corporations across the United States, Japan, Europe, Korea and Taiwan began building fabs in a race. By the early 2010s, German and Japanese corporations went bankrupt or were sold off one by one. Since then, the DRAM market has shifted to a tight oligopoly in which Samsung Electronics, SK hynix and Micron control most of the market.
Market researchers and investment banks differ on how long the current supercycle, triggered by artificial intelligence (AI) infrastructure investment, will last, but the consensus is that a sharp downturn is unlikely to return. Memory semiconductors remain a cyclical industry, but the nature of the cycle has changed. In particular, analysts say that diversification of demand stemming from AI in PCs, mobile devices and servers, along with changes in suppliers' capital expenditure strategies, will allow a certain level of profitability defense even in downturns.
◇ Three-firm oligopoly and node conversions aim to block a "supply bomb"
According to Counterpoint Research on the 16th, as of the first quarter of 2026, the three companies—Samsung Electronics (38%), SK hynix (29%) and Micron (22%)—share 90% of the global DRAM market. The three corporations are survivors that have already weathered multiple downturn cycles in the past. Since 2020, when signs of oversupply appear in the DRAM and NAND flash markets, the three have implicitly kept supply prices within a certain range through aggressive output cuts, production adjustments or process conversions.
Capital expenditure approaches have also changed. All three are pursuing conservative capacity expansion, converting commodity DRAM to HBM (high bandwidth memory) and DDR4 to DDR5 at existing fabs instead of building new plants. While a new fab can become a supply bomb two to three years after ground is broken, node conversions limit the increase in supply while boosting productivity.
In particular, the emergence of high bandwidth memory (HBM), which is crucial for AI infrastructure, has strengthened the fundamentals of the memory market. In particular, Samsung Electronics and SK hynix are allocating about 30% to 40% of their total DRAM output to HBM. As a result, supply of commodity DRAM used in PCs, mobile devices and servers has declined, sustaining price increases. In a recent report, Kiwoom Securities described this as a "supplier-led shortage cycle," projecting that if Samsung Electronics and SK hynix maintain their current investment stance, the tight supply-demand balance will continue. In fact, market researcher TrendForce projected that memory shortages will persist through 2027, citing AI data center demand and conservative capital spending.
◇ Reliance on big tech is a risk… "From a supply cliff to a demand cliff"
There are, of course, new risk factors. Dependence on big tech for sales and operating profit has grown excessively. It now takes only one purchasing decision by a few big tech firms to sway global memory supply and demand. In the case of SK hynix, its sales dependence on Nvidia is estimated to be close to 30% of the total. The purchasing volumes and supply prices of major customers such as Nvidia and Broadcom could have an even greater impact on earnings going forward.
A semiconductor industry official said, "While diversifying demand sources has built the resilience to lessen the blow during downturns, it has also created the limitation of deepening reliance on big tech," adding, "When the infrastructure investment boost from Generative AI or agentic AI ends, the sales and operating profit of Samsung Electronics and SK hynix are highly likely to be swayed by the investment needs of U.S. big tech."
That said, the prevailing view in the market is that a cliff-like downturn like in the past is unlikely to recur. Starting in the second half of this year, commodity memory prices will face gradual downward pressure, but a "diverged cycle" is expected in which AI-focused, high-value products such as HBM remain relatively strong. Morgan Stanley said, "Memory has become the decisive bottleneck resource for AI infrastructure, and long-term agreements (LTA) are turning a traditionally cyclical business into one with stable profitability and long-term cash flows," adding, "There is no immediate fix for the memory shortage, and this situation will persist for the time being."