The Memory Cycle Looks Different This Time, and That Is Exactly the Trap

Memory has always been the most brutal corner of semis. Boom, glut, crash, repeat. So when people tell me "this time is different", my ears go up, in both directions.

A quick history lesson

For thirty years the memory business followed the same script. Prices spike, everyone builds fab capacity, supply floods in about eighteen months later, and the whole group craters. Investors who bought the top of a memory cycle have been carried out more times than anyone wants to admit.

What is genuinely new

The AI angle is not marketing this time. High bandwidth memory is a real bottleneck for AI accelerators, the products are sold on long contracts rather than spot pricing, and capacity for 2026 is reportedly already committed. That changes the rhythm. Long contracts smooth out the violent spot price swings that used to define the cycle. SK Hynix reportedly eyeing a US listing to raise as much as $14B for more capacity tells you the operators themselves think the demand is durable.

Where it could break

Here is the catch. Every prior cycle also felt structurally different at the top. The thing that ends a memory boom is always the same, too much capacity arriving at once. If three giant suppliers all expand into the same demand curve, 2027 could look a lot less friendly than 2026. Contracts protect you until they reset.

My takeaway

I am long the theme but I refuse to pretend the cycle is dead. I trim into vertical moves and I keep a list of what would tell me supply is catching up. Respect the chart, but never forget what business this is.