Simplifying the memory trade... 🫡 The fastest AI chips in the world are useless if you can't feed them data fast enough, and there are several different types of memory doing that work. Each one is a different business with different economics and different winners. The squeeze starts at the top. Near memory is HBM, the high bandwidth stuff sitting right next to the GPU doing 1-20 terabytes per second. This is the actual bottleneck on every $NVDA system shipping today and $MU is the cleanest American pure play on it. Sk Hynix and Samsung control most of the market globally, but if you trade US listed names, Micron is the trade. Then comes main memory, which is the DDR5 doing long context inference at 100-500 GB per second. Micron again. Same playbook. Expansion memory is where it gets interesting and where most retail isn't paying attention. This is the CXL layer that lets servers pool memory across the rack, and it's a real growth market right now. $MRVL and $MCHP build the controllers that make it work, and $ALAB is the one that nobody had on their radar a year ago and now everyone wishes they did. $RMBS sits in the IP layer collecting royalties. Context memory is the SSD and NAND layer doing session memory at 20-200 GB per second. Micron plays here too... $WDC and $SNDK are the public American pure plays, and this is the layer that benefits most from agentic AI specifically because agents need persistent session memory at scale. And at the bottom you have the data lakes, which is petabyte scale storage for RAG, documents, and training files. $DELL, $NTAP, $HPE, and $IBM are the names selling the actual infrastructure here, with Western Digital underneath them on the drives. I also shared an important position update with indicator suite users today too. Now zoom out... like I noted the other day, Jensen said agentic AI needs 1000X more compute. You can't 1000X the compute and leave the memory stack alone... I think all dips are for buying.