AI + EARTH · AN EDITORIAL JOURNAL
ARCHIVED ESSAY · THE EFFICIENCY QUESTION
If AI gets more efficient, why does its energy demand keep growing?
THE EARTH REPORT
Last week, we followed a prompt into the physical world. This week, we change the scale of the question: from the resources behind one answer to the resources behind all of them.
A smaller footprint per task matters. To understand the whole picture, we also need to know how much we are asking the system to do.
TWO FINDINGS TO HOLD TOGETHER
This week’s question, grounded in research published in 2025 and 2026.
01 / PER TASK
33×
Google reports a 33-fold reduction for the median Gemini Apps text prompt between May 2024 and May 2025.[1]
Company-reported. Not independently verified. Specific to this product and period.
02 / ACROSS THE SYSTEM
+17%
The IEA estimates that global data-center electricity demand grew 17% in 2025.[2]
Global annual estimate. Includes AI and other data-center workloads.
FOUR HONEST ANGLES
The Win
Google attributes its reported improvement to work across models, hardware and serving systems. Reducing the resources needed to deliver an answer is meaningful engineering progress.[1]
The Tradeoff
The IEA points to growing use and more demanding applications, including video, reasoning and agents. Efficiency is one part of a changing system.[2]
What We Learned
A claim about one task and a claim about the whole system answer different questions. Before comparing them, check the task, time period and measurement boundary.
What We Can Do
When evaluating an AI workflow, ask for resources per useful result and total resources over time. Define the result first. A smaller number means more when we know what it describes.
UNDER THE SURFACE · AN INTERACTIVE ILLUSTRATION
Start with half the energy per task and three times as many tasks. Move either slider and watch the total change.
ILLUSTRATIVE ENERGY UNITS · SAME SCALE
+50% total energy
50% of the original energy per task × 3 times the activity = 150% of the original total.
An illustration, not a footprint calculator. Tasks are assumed identical; only energy per task and task count change. No training, hardware manufacture or separate facility overhead is modeled. The values are not Google or IEA measurements.
THE QUESTION TO CARRY FORWARD
Keep the progress in view.
Keep the whole system in view, too.
EVIDENCE, METHOD & CREDIT
Reviewed September 16, 2026. Publication dates and measurement periods are kept separate.
Google Cloud · Company-reported measurement
How much energy does Google’s AI use? We did the mathPublished August 21, 2025
Median Gemini Apps text prompt, May 2024 to May 2025. Google states that its data and claims have not been independently verified. This is not a universal AI footprint or a full lifecycle assessment.
International Energy Agency · Intergovernmental analysis
Key Questions on Energy and AI — Executive summaryPublished April 2026
Global data-center electricity demand in 2025 includes workloads beyond AI. The report distinguishes efficiency, uptake and changing applications; the two figures on this page are not a matched dataset.
The interactive example is original explanatory arithmetic. The practical questions are Groundwork’s editorial guidance. Neither is a measured claim about a particular product.
Atmospheric photography: Luca Bravo (forest image) and David Marcu (landscape image), via Unsplash. Landscapes are not evidence of AI’s environmental performance.