The evidence is concrete. Tests by the ML.Energy initiative found that running inference on Alibaba's Qwen 3 235B A22B Thinking model in FP8, a lower-precision number format, used a third less energy than bfloat16 on problem-solving tasks. Jae-Won Chung, a University of Michigan PhD candidate behind the initiative, also built Perseus, a training optimiser that slows underworked parts of a large-model training job so they finish alongside busier ones, cutting training energy by up to 30% without reducing throughput or touching the hardware.

Industry is already moving this way. Nvidia's Blackwell power profiles adjust GPU frequencies, power limits and cache settings to match the workload, which the company says can save up to 15% of energy while keeping 97% or more of performance, and let power-constrained sites fit in more GPUs for up to 13% higher throughput. Simpler wins include right-sizing cloud instances, caching repeated AI prompts, routing routine requests to small models, and capping output lengths.

Software can also shift when and where electricity is drawn. Sophie Hall, a doctoral student at ETH Zurich who has studied workload shifting across Google's data-centre fleet, argues the question is not just how much power is used but when and where, and how that interacts with the grid. Delaying non-urgent batch jobs until local demand falls, or routing them to regions with spare capacity, can help. There are limits: data sovereignty rules can stop workloads crossing borders, and physically moving large datasets between sites is far from trivial.

None of this makes software a substitute for better chips and new power infrastructure. And efficiency has a catch known as Jevons' paradox: if every watt saved simply generates more tokens, total electricity use may not fall. As Chung put it, "Power is the core bottleneck in AI data centers. We really want to make the best use of every watt we consume."