Does making AI models smarter each year actually mean they use less energy to get the same job done?

This explores the difference between two ways of measuring AI progress: how much cheaper the model's computation gets each year (algorithmic efficiency), and how much energy it takes to get a real piece of work done (energy per task). They sound like the same thing, but the corpus suggests they come apart in ways that matter.

Start with the cleanest number. Researchers who stripped out falling hardware prices and price wars by studying open-weight models estimate that better algorithms alone make inference about 3x more efficient each year. That figure roughly matches separate measurements of energy efficiency, so it holds up across methods How fast is AI inference actually becoming more efficient?. On that view, the two measures line up: a smarter algorithm needs less compute, so it uses less energy for the same output. The catch is the phrase "the same output." The International Energy Agency (IEA) treats energy per task as falling by orders of magnitude, but it says total electricity demand is the net of three forces: those efficiency gains, surging adoption, and a shift toward heavier tasks like video generation and reasoning How do efficiency gains and new AI capabilities balance each other?. A "task" isn't fixed. When people stop asking for a short answer and start asking for long chains of reasoning or a video, energy per task can go up even while every token gets cheaper.

The second gap is physical. Algorithmic efficiency is usually counted in operations, but on real hardware, moving data can cost more than computing with it. MobileLLM shows that on phones, running the same transformer block twice uses less latency than loading a second set of weights from memory Does recomputing weights cost less than moving them on mobile?. An approach that looks wasteful on paper (doing the math twice) turns out cheaper in practice. So energy per task depends on where the bottleneck sits on a given device, not just on how many operations an algorithm needs.

The third gap is about what counts as one task. For multi-step agent work, cost and latency pile up over the whole episode, which is why a compact 35B model trained on execution data can sit at the cost-performance sweet spot against much larger models Does model efficiency matter more than peak capability for real work?. MAKER goes further: break a problem into tiny enough steps, vote at each one, and small non-reasoning models can complete million-step tasks without errors Can extreme task decomposition enable reliable execution at million-step scale?. Cheaper tokens don't always give you cheaper solved problems, though. Non-reasoning models can't catch up with reasoning models even with unlimited inference budget, because the training is what makes extra tokens useful Can non-reasoning models catch up with more compute?. Comparisons of search strategies point the same way: the total compute spent and the quality of the scoring signal matter more than which framework you pick Does the choice of reasoning framework actually matter for test-time performance?. Energy per task rewards spending compute where it does useful work, not just spending less of it.

The less obvious point: "per task" can include the human too. One study finds AI doesn't cut total task time. It moves time away from doing the work and toward writing prompts and checking outputs Does AI really save time, or just change how we spend it?. If you counted the whole cost of finishing a job, a model that gets 3x more efficient each year could still leave the job itself costing about the same. The corpus has good material on the model side and the system side, but nothing yet that measures energy per task end-to-end with humans included. That's a real gap in the collection.