As AI data centers face soaring electricity demand, researchers are pushing back against a hardware-only mindset. The Uptime Institute's 2025 survey shows average power usage effectiveness has barely changed in six years, and servers account for roughly 60% of a modern data center's electricity draw. That makes software a promising lever, since chips are slow and costly to replace while systems software, algorithms, and applications can be adjusted more quickly.
Evidence from the ML.Energy initiative supports this view. Tests of Alibaba's Qwen 3 235B A22B Thinking model found that running inference in FP8 consumed a third less energy than bfloat16 on problem-solving tasks. Separately, the Perseus training optimizer slows underutilized parts of a large-model training job so they finish alongside busier parts, cutting training energy by up to 30% without reducing throughput or changing hardware.
Hardware vendors are also pursuing software-level gains. Nvidia's Blackwell power profiles tune GPU compute, memory frequencies, power limits, NVLink states, and cache settings to match workloads, saving up to 15% energy while retaining 97% or more of performance. Other practical fixes include retiring legacy equipment running old code, caching repeated AI prompts, routing routine requests to smaller models, and shifting non-urgent batch jobs to times or regions with lower-carbon electricity.
The source notes that software is not a silver bullet: data sovereignty rules and the difficulty of moving large datasets can limit workload shifting, and Jevons' paradox means efficiency gains may be consumed by generating even more tokens. Still, researchers see software as a control layer that makes the best use of every watt, especially when power is the core bottleneck in AI data centers. The source does not present competing views, so the article reflects a single perspective.```json {