Huawei gives AI clusters 64 PB of KV cache

Huawei introduced OceanStor M900, a context-memory system that claims 64 PB of shared KV-cache capacity for a single AI SuperPod cluster, according to Huawei’s September 17 announcement. The system extends KV-cache storage across on-chip memory, DRAM, and SSDs; its capacity and performance claims are vendor-reported, not an independent benchmark. Long-running agent memory is becoming a hardware procurement variable, so builders should compare cache reuse against the retention costs in Temporal’s agent-history analysis before treating more memory as automatic savings.