A study notes that AI workloads could drive data‑center electricity use up to 945 TWh by 2030 – roughly Japan’s current consumption. The researchers argue that instead of only building more efficient GPUs, operators should look at the software running on servers. Since servers use about 60 % of a data‑center’s power, cutting unnecessary compute could save a lot of energy. They point out that average power usage effectiveness has barely moved in six years, suggesting hardware gains alone aren’t enough.
Why it matters
If hyperscale operators improve software efficiency, they could lower electricity bills and reduce strain on already tight power grids.
