Your GPU re-prefills the same 15,000-token prompt ten thousand times a day. KV-cache offloading to DRAM and NVMe turns that recompute into a cheap fetch — 10x users.