Gemini's 0.24 Wh: what Google measured, and what it left out

What was measured

Google's paper "Measuring the environmental impact of delivering AI at Google scale" reports that the median Gemini Apps text prompt in May 2025 consumed 0.24 Wh of energy, emitted 0.03 gCO₂e and consumed 0.26 mL of water. The method is what makes it unusual: the measurement is taken across the production fleet and includes the accelerators doing the work, the host CPUs and memory, idle machines held in reserve for traffic spikes, and the data centre's own overhead for cooling and power conversion. Google calls it the comprehensive method and contrasts it with the active-chip method used by most benchmarks, which gives 0.10 Wh for the same prompt.

The improvement

Over the twelve months to May 2025, energy per median prompt fell 33× and carbon 44×, with no change to the buildings. The drivers were new TPU generations, serving software, batching and smaller models for simpler requests. It is the single best piece of evidence that per-prompt costs can fall fast, and it comes with the obvious corollary that Google's total electricity rose 37% in the same period.

What was left out

The distribution. A median says nothing about the long tail. Epoch's independent work shows long inputs costing 10 to 100 times a typical prompt; Google has the data to publish percentiles and has not.

Training. The paper covers inference. No training-emissions figure exists for any Gemini model; the index estimates about 9,000 tonnes for a frontier run on TPUs at Google's grid intensity, labelled as an estimate.

The water boundary. 0.26 mL counts water evaporated at Google's own data centres, derived from a fleet WUE of 1.15 L/kWh. The water consumed generating the electricity is excluded, as The Verge and others pointed out. Mistral's full-scope figure for a comparable response is 45 mL.

Verification. The study is self-reported. No third party has reviewed the fleet data behind it, and the index's disclosure checklist marks that item unmet.

How the index scores it

Efficiency 90 with an A grade, the highest in the index: a disclosed, measured figure with a published method. Transparency 82 rather than higher because of the four gaps above. The carbon per prompt, 0.03 g, is market-based and reflects Google's contracted clean energy; the location-based figure would be higher.

What it set as the standard

Any company can now be asked: comprehensive or active-chip? Median or mean? On-site water or full scope? Google's paper made those questions answerable, which is why the index uses its definitions as the baseline for every estimate on the site.