Issue Existing data centres were built for air-cooled racks that AI hardware does not fit
Summary
Deloitte argues that the current state of enterprise infrastructure creates its own barrier to AI. Existing data centres were built with raised floors, standard cooling, orchestration based on private cloud virtualisation and conventional workload management, all designed for rack-mounted servers cooled by air. AI infrastructure has quite different specifications, from networking between GPUs to advanced interconnects such as InfiniBand, which call for architectures that traditional enterprise environments do not have. The report warns that this physical mismatch could become a primary bottleneck as AI adoption expands. Dell's John Roese adds that most enterprise infrastructure reflects decisions made before AI existed, so retrofitting it is hard. The chapter points toward purpose-built AI data centres with more GPUs per CPU, optical networking and new cooling, including direct liquid cooling that can be at least twice as energy-efficient as free air cooling.
Classification
Evidence 1
- Tech Trends 2026: As technology innovation and adoption accelerate, five trends reveal how successful organizations are moving from experimentation to impact Deloitte Insights (Deloitte Development LLC, US Office of the CTO) page=36;section=The AI infrastructure reckoning: Optimizing compute strategy in the age of inference economics / The infrastructure mismatch 2025-12 accessed 2026-07-26
Constituent trends 1
Directly linked signals 1
Relation types: constitutes · direct_urgent
Public id: fm-949d73636b9b
