Digital Realty CTO warns against GPU-only AI planning

The author argues that designing successful AI infrastructure requires evaluating location, power, cooling, and connectivity as a single unified system rather than focusing solely on acquiring GPUs. While training requires massive centralized power and cooling, inference and agentic workflows introduce a placement problem where latency and data proximity dictate success. Consequently, organizations must strategically balance central campuses and distributed metro facilities to align infrastructure investment with workload value. Open article →