How do efficiency gains and new AI capabilities balance each other?

The IEA frames the growth of AI's electricity demand as the net result of three trends moving in different directions at once: "improvements in efficiency, surging uptake, and changing model capabilities, which can unlock new and, in many instances, far more energy-intensive use cases." On the efficiency side, the report says software and hardware advances have resulted in "the energy use per AI task dropping by at least an order of magnitude annually in recent years," to the point that "simple text queries now typically consume less electricity than running a television over the same period of time" — and that if every conventional internet search were replaced by a simple AI text query, total consumption would stay "less than 4 terawatt-hours (TWh) ... annually, equivalent to less than 1% of total data centre consumption today." Set against that gain, the report says "new energy-intensive AI applications are increasingly being launched and used, such as those for video generation, reasoning and agentic tasks," and that "these kinds of tasks can consume hundreds or thousands of times more energy per query than simple text generation."

The IEA's point is that efficiency and intensity move independently: a per-task efficiency curve falling "at a rate unprecedented in energy history" does not by itself predict total electricity demand, because the task mix is shifting toward categories — video generation, reasoning, agentic workflows — whose per-query energy cost is orders of magnitude above the simple text query the efficiency figure describes. The report applies a parallel logic to economic growth: an AI-driven boost to GDP "does not translate one-to-one into higher energy demand" because the growth is "concentrated in knowledge-intensive services and high-income countries, where the elasticity between energy demand and economic activity is lower," putting the energy effect of an AI growth dividend at "between 1-4% in 2035" above the no-AI-boost trend.

This note is a companion to How fast will data center electricity demand actually grow?, which carries the resulting aggregate demand numbers; this note carries the IEA's explanation for why those numbers keep climbing despite fast efficiency gains. The "order of magnitude annually" efficiency claim sits alongside two other fast-moving but differently-measured rates among the nearest notes: Epoch AI's finding that How fast is AI inference actually becoming more efficient? isolates a slower, cost-focused efficiency rate once competition and hardware price declines are stripped out, and How fast is AI capability becoming cheaper to produce? makes a similar "faster than history" claim but for the price of a fixed performance level rather than energy per task — the three are not measuring the same quantity and should not be read as corroborating a single number.

The excerpt does not quantify how the three trends net out going forward, or give a timeline for when energy-intensive task categories (video, reasoning, agentic) might become a majority of AI workload rather than a high-cost minority; it also does not break the "hundreds or thousands of times" multiplier down by task type, so the claim establishes direction and rough scale, not a number that could be used to forecast total demand from a changing usage mix.