AI Drives Rapid Evolution of Data Center Infrastructure

The article explains that AI workloads are reshaping data center infrastructure.

• Main findings and technical impacts:AI workloads cause power oscillations that can fluctuate by more than 50% in mere microseconds, leading to microburst-driven UPS battery cycling that can reduce battery lifespan by 50% or more and shorten replacement cycles from three-to-five years to ~18 months. Cooling thresholds change at ~40 kW per rack (air-cooling limit) while industry roadmaps (e.g., NVIDIA) target 600 kW–1 MW per rack, forcing direct liquid-to-chip cooling, much faster response times (from 20–30 minutes down to ≤2 minutes), and requirements for backup cooling power and thermal storage.
• Background, responses, and infrastructure actions:UPS vendors are exploring alternatives like capacitors; operators face high retrofit costs for chilled-water and coolant distribution infrastructure, leading to the rise of purpose-built AI facilities and modular construction that scale incrementally. The article notes an efficiency paradox: AI-optimized GPUs deliver substantially more computation per kilowatt, incentivizing optimized power and cooling architectures despite higher infrastructure costs.