Purpose This study investigates how micro-level GenAI infrastructure optimisation – specifically CPU thread tuning on NPU-accelerated inference – affects enterprise knowledge management, organisational resilience, and digital transformation outcomes. Design/methodology/approach We propose the Infrastructure-to-Knowledge Outcomes (I2KO) pathway as an infrastructure-level operationalisation linking service performance distributions to SECI knowledge flows, Kolb's learning cycle, and dynamic capabilities. Using Qwen2.5–3B on KT ATOM + NPUs, we benchmarked 70 workloads across eight categories, including n = 15 multiturn scenarios for socialisation-phase analysis. We introduce the Resilience Degradation Index (RDI = P95/P50) to capture tail-risk exposure invisible to average-centric metrics. Findings Optimal thread configurations improved average throughput (+8.8%) and mean latency (−1.6%) but increased P95 latency (+12.1%) and context scaling sensitivity (+30.4%). The multiturn analysis suggests that tail-latency degradation increases with conversational turn depth across the n = 15 workload set, with optimisation benefits concentrating in single-turn tasks while tail-risk accumulates in conversational and large-context workloads; This directional pattern (anchored by n = 13, five-turn) requires replication. Organisational implications are theoretically inferred and await field validation. Practical implications We propose a four-layer governance stack (Policy, Control, Monitoring, Review) and deployable design patterns – Lite Tier, Analytical Tier, Memory Broker, Context Pipeline – with illustrative SLO thresholds (e.g. P95 30s for Socialisation; scaling factor 9.0 for Externalisation) derived from HCI response-time research and enterprise SLA precedents. Originality/value This study provides an initial empirical operationalisation of performance-distribution effects on enterprise knowledge capabilities, extending IT business value and dynamic capabilities theory by disaggregating infrastructure performance into efficiency-oriented (P50) and predictability-oriented (P95) dimensions.
Yong-Jae Lee (Fri,) studied this question.