Google recently published a research report on the energy consumption of its AI model Gemini.

Blog: https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference

Technical report: https://services.google.com/fh/files/misc/measuring_the_environmental_impact_of_delivering_ai_at_google_scale.pdf

The report says that processing a median Gemini text prompt consumes only about 0.26 milliliters of water, roughly five drops; 0.24 watt-hours of electricity, equivalent to watching television for less than nine seconds; and produces 0.03 grams of carbon dioxide emissions.

Note: The median is a statistical measure used to describe the central tendency of a dataset. It is the value in the middle after a set of data is sorted in order. Here, it means researchers measured the resource consumption of multiple Gemini text prompts and then ranked all consumption data, including water use, electricity use, and carbon emissions, separately.

Google said it reduced the energy consumption of a single text prompt by 33 times and cut its carbon footprint by 44 times between May 2024 and May 2025.

Google also stressed that its measurement method is more comprehensive than traditional approaches. It includes not only the energy consumed by AI accelerators while active, but also server idle power, auxiliary hardware consumption, and total overhead from data center cooling and power distribution.

Google attributed the low figures to its “full-stack” approach to efficiency optimization, spanning model architecture, algorithms, hardware, and other layers. Specifically:

At the model architecture level, Gemini is based on the Transformer architecture and is 10 to 100 times more efficient than the previous state-of-the-art language modeling architecture. Google has also adopted other optimization methods, including MoE and hybrid reasoning modes.

At the algorithm level, Google uses methods such as AQT, or Accurate Quantized Training, to maximize efficiency and reduce energy consumption during inference without lowering response quality.

Google also uses techniques such as speculative decoding, in which smaller models make predictions first and larger models quickly verify them. Distillation uses more capable large models as teacher models to create lighter and more efficient serving models, such as Gemini Flash and Flash-Lite.

At the hardware level, Google has spent more than a decade designing TPUs from the ground up with the goal of maximizing performance per watt. Google also co-designs AI models and TPUs to ensure software can fully exploit hardware advantages, while enabling the hardware to run efficiently alongside future AI software. Google’s latest-generation TPU, Ironwood, is 30 times more energy-efficient than the earliest publicly disclosed TPU and is far more energy-efficient than general-purpose CPUs on inference tasks.

At the software level, the XLA machine learning compiler, Pallas kernels, and the Pathways system enable model computations written in high-level languages such as JAX to run efficiently on TPU inference hardware.

Highly efficient data centers: Google’s data centers are among the most efficient in the industry, with a fleet-wide average PUE of 1.09.

Google has also continued to optimize its cooling systems, balancing energy use, water use, and carbon emissions locally, guiding cooling method choices, and limiting water use in water-stressed regions.

Experts Are Skeptical

Although Google highlighted its efforts to improve per-unit efficiency, several experts questioned the completeness of the study’s methodology and conclusions.

Indirect Water Consumption Is Ignored

Shaolei Ren, an associate professor at the University of California, Riverside, and Alex de Vries-Gao, founder of Digiconomist, said one of the report’s biggest omissions is that it does not account for “indirect water use.”

Google’s estimate of 0.26 milliliters covers only the direct water used by data centers to cool servers. But the power plants that supply electricity to data centers, whether natural gas or nuclear, also consume large amounts of water for cooling and turbine operation.

De Vries-Gao described Google’s disclosed figures as “basically just the tip of the iceberg,” because water use in power generation often far exceeds the direct consumption of data centers.

Incomplete Carbon Accounting

Experts also said Google shared only a “market-based” measurement method for carbon emissions, which allows companies to offset emissions by purchasing renewable energy certificates. That does not fully reflect the company’s actual impact on local power grids.

Experts said Google should follow the internationally recognized Greenhouse Gas Protocol and also include “location-based” carbon emissions data. That metric more accurately reflects the real emissions produced by the mix of clean energy and fossil fuels on the grids where data centers operate. Professor Ren said, “That’s the ground truth.”

Misleading Data Comparisons

In the report, Google compared its extremely low water consumption figure with data as high as 50 milliliters from Professor Ren’s earlier research, saying its result was “orders of magnitude” lower. Ren said the comparison was “apples to oranges,” because his earlier study included total direct and indirect water use, while Google counted only the former.

Google also used the “median” to prevent outliers from skewing the results, while previous studies have generally used the “average.” Experts criticized Google for not providing the underlying details used to calculate the median, such as the number of words or tokens in the prompts, making it difficult for outsiders to verify how representative the results are.

The Jevons Paradox in the Bigger Picture

Although the efficiency of individual AI prompts continues to improve, experts warned that this could lead to a phenomenon known as the “Jevons paradox”: greater efficiency may spur more usage, causing overall resource consumption and pollution to rise rather than fall.

That concern is reflected in Google’s own sustainability report. The data shows that as the company has aggressively expanded AI, its “ambition-based carbon emissions” have risen 51% since 2019 and increased 11% last year alone.

De Vries-Gao put it bluntly: “If you look at the total numbers Google has disclosed, the situation is actually very bad.” He said the Gemini efficiency report released this time “does not tell the full story.”