As AI applications move from training into production, inference workloads are creating distinct infrastructure requirements around hardware selection, rack density, latency, and geographic placement. Uptime Institute's 2026 survey found more operators reporting peak rack densities of 30 kW or higher, while Nvidia's GB200 NVL72 rack consumes approximately 120 kW at full load. Air cooling remains suitable for lower-density CPU-based configurations, but higher-density GPU deployments require direct-to-chip liquid cooling, and Schneider Electric emphasizes coordinating rack, power, and cooling designs together. The Akamai-Anthropic agreement, in which Anthropic committed $11.6 billion over seven years for CPU-based cloud capacity with a potential $9 billion expansion, reflects how inference demand is already shaping colocation purchasing.
Inference workloads are arriving at a different scale and distribution pattern than training clusters, meaning facilities built around large GPU training deployments may not be economically suited for the next wave of AI demand. Operators who plan capacity using average rack densities or training-centric assumptions risk misallocating capital as inference becomes the dominant production workload.
The shift from training to inference as the dominant AI workload is a significant structural story affecting data center architecture, power density, and hardware procurement across the industry. Named topic (AI inference, data center design) is distinct from the already-published list.