Many people buy AI memory, only focusing on HBM.

But the real shortage is much bigger than that.
Goldman Sachs and JPMorgan's assessments are consistent: this memory shortage has already exceeded normal cycles, and from 2024 to 2027, the supply and demand gap for DRAM, NAND, and HBM will continue to widen.
This is not just a matter of one or two quarters; it’s a multi-year price increase cycle.
The reason is simple: HBM is placed next to GPUs and is the most direct bottleneck for AI training, which everyone knows, but now AI inference workloads are also growing rapidly.
Inference requires server DRAM and enterprise SSDs, not HBM.
In other words, the entire memory industry chain from top to bottom is under simultaneous pressure.
The five companies in this sector each cover different levels:
Micron is the only U.S. company capable of producing HBM, DRAM, NAND, and SSDs simultaneously, and cloud service providers will increasingly rely on it for supply chain diversification.
Expected revenue growth of 188% in the next fiscal year, with a P/E ratio of 11.1 times, still not extreme relative to cycle peaks.
SK Hynix is the clearest leader in HBM, a core supplier for NVIDIA, with new wafer capacity and logic node upgrades widening its moat.
Expected revenue growth of 212%, with a valuation of 6.3 times.
SanDisk has the most direct relationship with NAND; cloud providers are migrating active data from slow storage to dense flash arrays, making SanDisk the most immediate beneficiary of this trend.
Revenue growth of 209%, but its valuation has already reached 12.6 times, relatively more expensive than others.
Samsung has the broadest coverage, producing HBM, DRAM, NAND, and SSDs, tracking the memory cycle in the most diversified way, but its yield for HBM has always been a weak point.
Kioxia focuses on NAND, with an expected revenue growth of 311%, the highest among this group, but its valuation at 9.7 times is also the most expensive.
Looking at these five companies together, there is a pattern: the higher the revenue growth rate, the less necessarily the higher the valuation.
The market’s pricing for this sector has not yet fully reflected the ongoing nature of the supply and demand gap.
If AI inference demand expands as expected, DRAM, NAND, and HBM will become the most persistent bottlenecks in the entire AI technology stack.
This logic will be repeatedly validated in the future.