If I Had to Full-Port Into One AI Theme…
I would long memory.
GPUs do the computing, but memory stores the data and feeds it to the chips. As AI moves from text to images, video, reasoning, and agents, memory requirements increase dramatically.
HBM is a big reason why. It’s the high-speed memory stacked next to GPUs, allowing them to process massive amounts of data faster. The problem is that HBM uses roughly 3x more wafer capacity than traditional DRAM, so every wafer shifted toward HBM tightens supply across the entire memory market.
-> Demand is accelerating as AI clusters get larger
-> New memory fabs take years and tens of billions to build
-> HBM consumes significantly more production capacity than standard memory
-> Tight supply creates pricing power for manufacturers
WSTS expects memory revenue to reach $800B in 2026, up 250% YoY, followed by $1.1T in 2027, up another 32%.
Despite that growth, Samsung, SK Hynix, Micron, and SanDisk all trade around 7x 2027 earnings or lower.
If the group rerates to just 15x earnings, that’s 100%+ upside before factoring in any additional earnings growth.
SK Hynix is the HBM leader, Micron is the cleanest U.S. memory play, Samsung provides scale across the industry, and $DRAM offers exposure to the entire memory cycle in a single ticker.
I don’t see a realistic scenario where memory demand slows meaningfully over the next few years.
Huge demand + constrained supply = pricing power = rocket.
