As the age of artificial intelligence (AI) arrives, the long-standing formula for the memory semiconductor industry's cycle is being shaken. Analysts say the traditional memory cycle—marked by repeated gluts, price plunges, and production cuts—is entering a new phase on the back of expanding AI demand and long-term supply contracts.
U.S. memory semiconductor corporations Micron Technology said at its fiscal year 2026 third-quarter earnings release on the 24th (local time) that it expects the tight supply-demand balance driven by AI-led demand across every institutional sector and structural supply constraints to continue even after 2027. With the spread of Generative AI and inference AI, demand for data center DRAM and NAND flash is surging, and the company said the supply shortage could become prolonged.
This powerful demand is fundamentally transforming the structure of the memory market. In the past, the memory industry was like a rain-fed field that swung with consumer cycles for front-end IT devices such as PCs and smartphones, but now infrastructure investment by big tech corporations is steering market conditions. In fact, Micron said that for the first time this year, the total addressable market (TAM) for data center-bound memory on a bit basis surpassed half of the entire memory market. A single server used for AI training and inference needs far more memory than a conventional server. With the spread of agentic AI and replacement demand for existing servers added on, observers say the very foundation of memory demand has expanded significantly compared with the past.
These solid fundamentals were proven by record results. Micron posted third-quarter revenue of $41.46 billion and earnings per share (EPS) of $25.11, delivering an earnings surprise that far beat market expectations. Thanks to rising prices and increased sales of high-value-added products, the gross margin reached a record high of 84.9%. Guidance for the next quarter also topped expectations, with revenue of $50 billion and EPS of $31. In after-hours trading on the New York stock exchange, the stock jumped more than 12%. Wall Street expects the strength in the memory cycle to continue as AI-driven memory demand outpaces supply growth for the time being.
Experts especially point to the "change in product characteristics" as the key driver of the new memory cycle. If DRAM in the past was a "commodity" with identical specifications whose prices repeatedly crashed and spiked depending on supply, then high-bandwidth memory (HBM) in the AI era is closer to a "customized strategic asset" designed to fit a specific customer's chip. HBM has a more complex manufacturing process than general-purpose DRAM and requires advanced packaging technology, making it difficult to ramp up capacity in a short time. The industry says bottlenecks are appearing not only in wafer capacity but across advanced packaging processes. As a result, major memory makers have already presold a significant portion of their HBM volumes for several years.
With supply so severely constrained, the sales model has also fully shifted to a "build-to-order" approach. Rather than producing first and selling to match market prices as in the past, companies are entering production after signing long-term contracts with customers seeking to secure volumes in advance, thereby controlling earnings volatility.
Micron said it signed a total of 16 Strategic Customer Agreements (SCA) this quarter. SCAs are multi-year contracts stronger than existing long-term agreements (LTA), including customer volume commitments and price floors, which increase revenue visibility. These are long-term supply contracts covering about 20% of the company's total DRAM volume and about one-third of its NAND volume. The company expects this to reduce earnings volatility from the past "chicken game" era and increase management predictability.
Korean companies are showing a similar trend. SK hynix has secured a significant portion of HBM supply volumes through advance contracts with major customers including Nvidia. Samsung Electronics is also focusing on securing key AI customers as it expands its HBM business. The industry expects the share of long-term supply contracts in the HBM market to grow further.
The investment race to meet growing AI demand is also intensifying. Micron guided net capital expenditures (CapEx) of about $27 billion (about 41.67 trillion won) for this fiscal year, and is pursuing new fab construction in Idaho and New York, as well as expansion of advanced packaging capacity in Singapore. Samsung Electronics and SK hynix are also accelerating investments in HBM capacity expansion and advanced packaging.
Some in the industry are cautious, noting that rising depreciation expenses from large-scale capital investments could weigh on profitability over the long term. They also say that if profitability concerns over AI services persist or rivals expand capacity faster than expected, a supply glut phase could reemerge.
Even so, as memory in the AI era establishes itself not just as a commodity but as a strategic asset, there is growing assessment that the memory industry's cycle could take on a long-term, stable pattern different from the short-term business cycles of the past.