AI has turned power from a facilities question into a board-level infrastructure constraint. Traditional enterprise data centres were already large electricity users, but AI clusters raise the density and volatility of demand. Training runs, fine-tuning jobs and inference surges can create fast-changing load profiles, which means grid connection, storage, backup generation and cooling design now shape where AI capacity can realistically be deployed.
The International Energy Agency has warned that AI-focused data centres can behave more like power-intensive industrial facilities than conventional IT rooms. The issue is not only total annual consumption; it is timing, concentration and flexibility. A 100 MW site that ramps quickly can stress local networks if it is not coordinated with utilities. That is why demand response, batteries, thermal storage, flexible workload scheduling and on-site power management are becoming part of the AI data-centre playbook.
For Singapore and the wider region, the lesson is clear: AI infrastructure policy cannot be separated from energy policy. Enterprises buying AI services should ask where workloads are hosted, how renewable claims are backed, what cooling technology is used and whether providers can shift non-urgent compute away from peak grid periods. The cheapest compute may not be the most resilient compute once power, carbon and regulatory risk are priced in.
