Memory Is the Bottleneck, but Not Forever

Memory has transitioned from being the least glamorous semiconductor business to a highly profitable industry driven by artificial intelligence (AI) acceleration. Global memory revenue is projected to increase by 250 percent in 2026, reaching over $800 billion. Micron, a leading memory producer, reported a fiscal third-quarter revenue of $41.5 billion, more than quadrupling from the previous year, and guided to an 86 percent gross margin for the next quarter. Similarly, SanDisk's gross margin improved from 22.5 percent to 78.4 percent within a year.

The surge in memory demand is primarily attributable to AI accelerators, which require high-bandwidth memory (HBM) for efficient data processing. Inference servers also rely on ordinary DRAM, while vector databases, model checkpoints, and retrieval systems utilize flash storage. As a result, hard drives have become essential for storing vast amounts of data generated by these systems.

However, this demand is not evenly distributed across the storage market, as different products have varying economics and value propositions.

HBM, in particular, stands out as a key bottleneck in the memory market. It is a stacked DRAM package designed to work in tandem with AI accelerators, enabling data movement without straining the GPU or custom AI chip. HBM requires advanced packaging, thermal management, and a dedicated logic base die, making it a capital-intensive product. The scarcity of HBM is real, as supply cannot be easily adjusted to meet demand. As a result, HBM is expected to command high multiples and reward execution by suppliers.

Conventional server DRAM, while less exotic, benefits indirectly from HBM due to the increased memory requirements in AI servers. These servers demand more memory than traditional machines, and AI creates direct demand through model weights, embeddings, retrieval indexes, training checkpoints, and key-value caches. Enterprise products fetch premiums for endurance, firmware, latency consistency, and power efficiency.

However, the underlying NAND bits remain relatively interchangeable, and supply becomes relatively elastic once factories are running.

On the storage side, nearline hard drives serve as an archive for data that accelerators cannot handle due to their slow processing speed. Seagate and Western Digital now operate in a two-vendor market for high-capacity cloud drives, where higher areal density leads to lower cost and power consumption per terabyte. Despite the significant capital expenditures required for HBM development, an 86 percent gross margin is an attractive prospect for investors.

The bear case for the memory market does not rely on a collapse in AI demand. Instead, it hinges on the faster growth of supply relative to customers' urgent requirements or a pause in accelerator architecture development by AI manufacturers. In a fixed-cost industry like memory, even marginal price differences can significantly impact profits.

Thus, ordinary DRAM has emerged as a misunderstood winner in the current memory cycle, benefiting from price increases and limited supply across various applications such as phones, PCs, automobiles, and industrial products.