The next memory trade is still the memory trade (receipts from one year ago included)

For context, I was pitching SK hynix one year ago Jun 2025.

I was pitching SK, Micron, SanDisk (and Qualcomm) Jan 2026 again

https://www.reddit.com/r/stocks/s/APCJwmjU20
https://www.reddit.com/r/ValueInvesting/s/Luvl4kY8Hy https://www.reddit.com/r/ValueInvesting/s/QcraWYZtJx
https://www.reddit.com/r/stocks/s/fILYpK6Ivo

(I was extremely lucky with fate and Providence, rather than any skill on my part)

I would still pitch the 3 big memory players - SK Hynix ($SKHY) Samsung and Micron ($MU) today. The core reason being forward P/Es are single digit (around 7-8x), unheard of in this space.

The reason is because the market is STILL pricing these businesses as a cyclical commodity. With dramatically elevated capex plans, supply is estimated to increase by approximately two times in the next 5 years. However, demand, particularly for agentic inference and higher context, is expected to go up several fold over the next 5 years. So even if memory chips remain a commodity, demand is going to far outstrip supply \*\*even with the massive capacity build out\*\*.

Secondly, I urge folks to read up on custom HBM and after that, memory-on-logic configurations. Just type this into your chatbot of choice and ask them to explain it to you. From Nvidia Feynman chip (late 2027) onwards and for many of the custom Asics in the near future (prime example is the custom inference chip that is being made by Qualcomm), the DRAM stack seats directly on top of the compute chip the GPU. The only way this can happen is by co-design between the chip vendor Nvidia or Google or Qualcomm and the memory maker. At that point, the co-design and customization explicitly means that the high end memory chips is no longer a commodity. You can't just jam a Chinese memory chip onto your GPU at that point, you are committed to one supplier with whom you codesigned the product.

Tl;Dr: 1. memory chip demand outstrips supply, even with capacity expansion over the next 3-5 years. 2. Custom HBM and memory-on-logic tips from 2028 onwards means high-end memory is no longer a commodity. Market has still not priced in either.

When it does, a fair multiple is the PE multiple given to Nvidia or ASML, about 20-30x. So that's a 4x on \**multiple expansion alone\*\* from current levels, and that is not accounting for increased earnings from higher chip sales volumes going into 2027 and 2028.