Originally posted January 25, 2026
Memory. 메모리. 内存. MEMORY!
Are you tired of hearing about it yet?
"NAND this, HBM that."
"Demand is astronomical, we're still early!"
"Its gone way too parabolic, it's all cyclical - short this now!"
What if it is cyclical? What if the entire market is cyclical?
I think that argument once held weight, but not anymore.
We are in an era of human history that is progressing faster -- with more demanded from us and of us, than any period before.
The Memory "Cycle"
Markets tend to model not only this memory super cycle, but nearly everything attached to AI, on the precipice of the "AI Datacenter Boom."
That is already outdated.
Demand for datacenter buildout has not yet been satisfied.
The infrastructure simply needs to be built and the capital deployed.
The next step-function in demand isn't just larger models and perfect inference solutions. It's models that act (agentic), robots that move (physical AI), and infrastructure that leaves Earth (space compute).
The demand cycle is clear.
Supply is where the picture blurs.
The Supply "Relief"
New fabs will add wafer capacity.
Advanced packaging will expand.
HBM lines will scale.
But most of this supply arrives 2027+.
And more importantly:
Memory supply ≠ deployable supply.
If demand accelerates in 2026-2027 due to the rise of agentic and physical deployments, the "new supply" shows up as already breaking.
Yes supply is coming, but the timeline matters.
The Storage "Glut"
It's easy to focus on HBM because training AI models is what dominates the conversation.
This is where confusion starts -- when SanDisk (SNDK) and Seagate (STX) move alongside HBM companies like Micron (MU).
"NAND is commodity, it will flood"
People underestimate deployment demand because current consumer models still struggle with persistent memory.
But Physical AI and Agentic systems do not just think.
Longer context
More caching
"Always-On" workloads
They store.
They log.
They retrieve and re-access massive amounts of data.
That shifts the demand further toward enterprise NAND and high-end SSDs.
That is infrastructure storage demand --
and it scales with deployment, not just training.
Robots aren't just inference endpoints like the today's models.
They are sensor-to-action loops which means memory demand spreads into:
Factory servers
Warehouses
Autonomous vehicles
This turns memory and storage into an industrial input and not just a cloud input.
Now place this in space compute where you cannot easily add or replace memory. Will we ever outpace demand?
Offline compute requirements
Radiation constraints
Intermittent connectivity
Space doesn't need a lot of volume today -- but it increases the number of environments where memory must exist and must be reliable.