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.
More from Infra
- Three CoreWeave co-founders have sold $3.3B in shares since IPO amid credit worries — SumitGup · 2026-10-16
- Cherry pi on a Raspberry Pi: the tiny-model rumor developers are watching — sull · 2026-10-16
- CoreWeave's debt costs falling: DDTL spreads tightened from S+9.62% to S+2.25% — BenBajarin · 2026-10-16
- SoftBank lines up US IPOs for SB Energy and AI robotics spinoff Roze, valuations top $50B — nordicinst · 2026-10-16
- TRL reproduces Modal's mixed-precision async RL: plain GRPO collapses, SC+MIS holds 0.72 — QGallouedec · 2026-10-16
- Hugging Face scales Indic pretraining with 2.8T tokens and an open data pipeline — Hugging Face Blog · 2026-10-16