The growing energy demand of AI

Sustainability · 2026-03-29 · 9 min read

By Ravi Patel, Sustainability & agriculture

Training gets the headlines. Inference, running billions of times a day, is where the electricity actually goes.

A training run is a large, one-off cost that is easy to point at. Serving a popular model is a small cost repeated continuously, and at scale the second comfortably exceeds the first over a model’s lifetime.

That shifts the engineering priority toward efficiency per request: smaller models where they suffice, caching identical answers, batching requests, quantisation, and simply not invoking a large model for tasks a lookup would settle.

The physical constraints are local. Data centres need grid connections that take years to obtain, cooling that in some designs consumes significant water, and neighbours who notice both. Siting has become the hardest part of expansion.

There are genuine offsetting effects — better grid forecasting, materials research, logistics optimisation — but they should be argued for specifically rather than used as a blanket justification for whatever gets built.

The reporting standard worth pushing for is boring and comparable: energy and water per thousand requests, disclosed by region, alongside whether the electricity is genuinely additional clean generation or a certificate purchased elsewhere.

Tags: ai, energy, infrastructure

Ravi Patel — Ravi reports on agriculture, food systems and sustainability for ESPYCRUX, with a habit of following the money before the technology. He grew up around a family farm and it shows in the questions he asks.