According to industry specialist TrendForce, the global memory market could reach over $1,280bn by 2027, up from a previous forecast of $842.7bn. For 2026, the estimate has also been raised to $889.3bn, from $551.6bn previously and $224.7bn in 2025, representing a projected y-o-y increase of approximately 300%.

This shift in scale is explained by the evolution of AI use cases. Following an initial phase dominated by large model training, demand is gradually shifting towards large-scale inference, where AI agents multiply queries, reasoning cycles and temporary memory requirements. The management of "KV cache," essential for maintaining conversational context and avoiding data recalculation at every step, has thus become a central performance factor for LLM models. This dynamic directly supports demand for DRAM and HBM, fueling the stockmarket surge of players such as Micron, Samsung, SanDisk, SK hynix, and Kioxia in recent months.

The pressure is further intensified as HBM production consumes more wafers - the thin circular slices of semiconductor material - thereby limiting available capacity for traditional DRAM used in PCs and strengthening manufacturers' pricing power. TrendForce now estimates that the DRAM market could reach $618.7bn in 2026, a 303% annual growth rate, followed by $903.3bn in 2027.

"NAND Flash," primarily used in SSDs, is also benefiting from the growing need for high-speed storage in data centers. TrendForce has raised its 2026 forecast to $270.6bn, with expected growth of 280.7%, reaching nearly $379.4bn in 2027. In this context, the alliance between Japan's Kioxia and America's SanDisk is expected to increase capital expenditure by approximately 41% y-o-y to $4.5bn, according to reports cited by TrendForce.

This acceleration is accompanied by a restructuring of the value chain, with Korean groups Samsung and SK hynix concentrating a larger portion of their resources on HBM, while Kioxia and SanDisk seek to capitalize on opportunities in NAND. For investors, memory is no longer merely a bet on the PC or smartphone cycle, but has become one of the potential bottlenecks of AI infrastructure.