Data center operators squeezed by power constraints have a set of near-term tools to keep AI growth on track, including efficiency improvements and more flexible scheduling of workloads. But according to a new Infosys analysis of AI energy demand through 2030, those measures will buy time rather than eliminate the need for new electricity infrastructure.
The report, discussed by Infosys executive vice president Mit Majumdar in an interview with Data Center Knowledge, argues that demand-side measures such as AI-driven energy management, advanced cooling, and peak-load optimization can create meaningful headroom faster than building new generation, transmission, or grid upgrades. Majumdar did not provide a numerical estimate of how much additional AI capacity these measures could unlock.
Not All Workloads Can Flex
A central finding is that flexibility depends heavily on workload type. Model training, retraining, and batch processing can often be scheduled around grid conditions, but real-time inference and customer-facing applications generally cannot. That means a facility running both kinds of workloads cannot treat its entire AI load as available for curtailment.
Majumdar said this distinction needs to show up in commercial agreements, not just technical planning. He pointed to a need for workload-aware service-level agreements, demand-response incentives, and pricing structures that reward customers for shifting non-urgent computing to times when the grid has more capacity.
Building in Stages Rather Than Locking In Assumptions
The report also highlights a planning challenge: AI adoption is driving computing demand upward even as new models, chips, and cooling techniques change how much power that compute actually requires. The analysis recommends matching models to tasks and applying techniques like quantization, pruning, and knowledge distillation to cut computational load, along with measuring energy use and carbon emissions at the workload level.
Majumdar advised developers to build in phases rather than committing an entire project to today’s assumptions, allowing capacity to scale as demand becomes clearer and technology improves. The aim, he said, is to avoid both overbuilding and underestimating long-term demand.
Utilities Need Proof, Not Projections
For utilities weighing whether demand-side savings can support new data center connections, the report argues that projections alone are insufficient. Operators should be able to show historical performance, provide real-time monitoring, and make contractual commitments to reduce load when the grid needs it.
Majumdar also called for shared accountability if promised flexibility does not materialize in practice, with operators backing performance claims through guarantees and utilities setting clear standards for qualification and verification.
For hosting providers and data center operators planning AI capacity, the takeaway is that efficiency and flexible scheduling remain useful tools, but they work best paired with realistic, staged infrastructure investment rather than as a substitute for it.