AI infrastructure is evolving faster than the data centers built to support it. While a data center may be designed for a 20–25 year lifecycle, GPU and AI technologies can evolve dramatically within just a few years, driving rapid increases in rack power density, cooling requirements and overall facility demand.
This presentation explores how data centers can be designed today to remain adaptable to the next generations of AI. It will examine the journey from GPU to rack, from rack to data hall, and ultimately from data center to grid, highlighting the interaction between IT workloads, electrical infrastructure, liquid cooling, heat rejection, controls, energy systems and the utility grid.
A key focus will be future-ready design and commissioning, moving beyond simply demonstrating that a facility works at today's design conditions, towards establishing and validating the operating envelope within which the infrastructure can safely accommodate future increases in AI density and changing technologies.
The presentation will explore how design flexibility, modularity, capacity margins, system interfaces, simulation and performance-based commissioning can help data-center owners manage uncertainty without attempting to predict exactly what the next GPU generation will look like.
The central question is simple: How do we build a 25-year data center for an AI technology that may change every 3–5 years?