At the VentureBeat AI Impact Series, WEKA field CTO Alan McSeveney told analyst Rob Strechay that GPUs running LLM inference can spend up to 60% of compute recomputing already processed context. KV-cache eviction can push utilization as low as 30%, while optimization can bring it above 90%; he argued NVMe flash is a better fit for KV-cache storage than expanding DRAM amid budget, power and data center space constraints.