AI task energy use falls ~10x per year, yet data center power still jumped 50%: IEA

import_jmr · x · 2026-08-14

Per the IEA's latest report, energy use per AI task is falling roughly 10x a year — faster than almost anything in energy history. The gains come from real architecture research: better embeddings, small domain-adapted models, smarter attention and more efficient chips; Gemini Encoder-Heavy and Embedding run the same task orders of magnitude more efficiently at equal or better quality.

Yet the efficiency dividend doesn't shrink the bill: electricity use from AI data centers still rose 50% last year. A simple text query uses about 0.05 Wh, while an agentic task that reasons and calls tools can use 50. Every efficiency gain gets spent on heavier workloads — reasoning, agents, video — with task complexity outpacing efficiency gains. Breaking the loop, the author argues, takes a real shift in how AI software and hardware are built, from new model architectures to neuromorphic chips and quantum computing.

Original post →

More from Infra

Infra channel →