Memory stocks took a heavy hit this wk but DDR5 server RDIMM spot just moved +27.9% in 30 days to $1,375. that's the memory market pricing in AI server demand before the supply side can respond.

on CXL: CXL could ease the capacity/utilization problem, increasingly in GPUs not just CPUs DRAM is bolted to each individual server. If one box has spare memory and the one next to it is starved, there's no way to share that idle capacity is "stranded" (Microsoft has said ~25% of the DRAM in its Azure fleet just sits there doing nothing). Since memory is roughly half the cost of a server, that's a lot of money idling. CXL unlocks three things: expansion (add capacity beyond the physical DIMM slots), pooling (a shared pool multiple servers draw from on demand), and sharing (multiple hosts hitting the same memory). The payoff is utilization. you stop provisioning every box for its worst case, so you buy less total memory for the same work. The catch is latency: CXL-attached memory sits ~200-500ns away vs ~100ns for local DRAM, so it's a tier below your main memory. good for colder data, KV-cache offload, big in-memory databases. I think of CXL as the shared storage closet down the hall cheaper per GB, much bigger, but you have to walk there It's not a DDR5 replacement, and importantly it does nothing for HBM bandwidth, which is a separate tier entirely.