Researchers say software can ease AI data-center power pressure

Energy-aware inference, training and scheduling can cut use without sacrificing much performance, offering operators more headroom before new infrastructure is needed.

Researchers and infrastructure operators are making the case that software changes, not just new chips and more power, can ease AI data-center energy pressure. Work from the University of Michigan’s ML.Energy group found that FP8 inference on Alibaba’s Qwen 3 235B A22B Thinking model used about one-third less energy than bfloat16, and Chung’s Perseus optimizer cut training energy by up to 30% without lowering throughput. Nvidia is also using workload-aware Blackwell power profiles, which it says can save up to 15% of energy while preserving 97% or more of performance and lift throughput by as much as 13%. Separate research from ETH Zurich points to shifting flexible jobs across time and geography to follow grid conditions, though data-sovereignty rules and moving large datasets can limit that approach. The overall message is that software is not a replacement for better hardware or new infrastructure, but it can squeeze more useful work out of existing AI clusters.