Software could be the easiest fix for hyperscalers' AI power squeeze, researchers say

Researchers from ML.Energy and ETH Zurich argue that software optimization can cut data center energy use without replacing hardware. The IEA expects AI to consume about 945 TWh by 2030, roughly Japan's current electricity use. Tests showed up to 30% savings in model training and up to 15% with Nvidia Blackwell power profiles.

• IEA: AI workloads will need about 945 TWh of electricity by 2030
• Running Qwen 3 235B inference in FP8 used a third less energy than bfloat16
• The Perseus optimizer cut training energy by 30% without losing throughput
• Nvidia Blackwell power profiles save up to 15% energy while keeping 97% performance