$Sandisk(SNDK.US) A simple analogy: Imagine a chef working on a tiny cutting board (HBM). Every time a customer adds a request—"No onions... Wait, add onions... Now make it vegetarian... And add a side dish"—these sticky notes (context/KV cache) pile up on the board. Eventually, the board is covered in notes, and the chef can't chop; this expensive chef just stands idle. The so-called "architectural refactoring" is simply establishing a reasonable kitchen workflow: Keep the most urgent notes on the cutting board (HBM), move "important but not immediate" notes to the prep table next to it (DRAM), and store the rest in a nearby filing cabinet/pantry (enterprise SSD). Then, you add a runner and an organizer (DPU + network) to fetch and place the right notes at the right time, allowing the chef to cook at full speed—this means higher throughput, lower per-token cost, and less wasted GPU time.$Micron Tech(MU.US)