Storage is shifting from “picking up money” to “making money.”

One character difference, but the difficulty is ten times higher.
Last year’s market was an explosion from “zero to one.”
HBM has gone from a niche product to the heart of AI infrastructure—demand slope is nearly vertical, supply wasn’t prepared at all, and the three leading manufacturers sold out their planned HBM capacity for 2026 early.
Buy it and it goes up—pullbacks are shallow. As long as you hold still, you can earn several times.
This easy-money phase is getting narrower.
Samsung plans to expand HBM capacity by 50% in 2026. SK hynix aims to double wafer capacity within five years. The first plant for the Lung-in cluster is scheduled to start production next year in Q1. Today it just raised $26.5 billion on Nasdaq, with most of it going toward capacity expansion.
Micron is building plants in Idaho, New York, and Virginia at the same time, with long-term investment exceeding $200 billion.
A scenario where all three expand production together has never been a good sign in the history of the storage industry.
There’s another data point most people haven’t noticed: TrendForce statistics show that in 2026 Q1, HBM’s single-wafer revenue has already been overtaken by DDR5 64GB RDIMM.
HBM’s margin advantage is narrowing. Manufacturers are starting to reallocate capacity between HBM and traditional DRAM, and the pricing game will be far more complicated than in the past.
Demand from AI for storage is still accelerating. For Nvidia Rubin Ultra, single-card HBM capacity is pushed to 384GB; HBM needs to account for 30% of DRAM total wafer investment by 2027.
The ceiling of demand can’t be seen.
But when all three expand capacity at the same time, margins flatten from their peak, and HBM’s annual pricing mechanism is renegotiated, these variables stacking together will make volatility much larger than before.
Every dollar earned later will test judgment and patience more than what came before.