How much energy does Google’s AI use?

 

In an effort to promote transparency and sustainability in AI, Google published an in-depth blog post on August 21, 2025, detailing how it measures the environmental impact of AI inference—the process where trained models generate outputs like text, images, or audio. This often-overlooked stage consumes energy and water and contributes to carbon emissions, but Google’s analysis offers one of the most comprehensive assessments to date.

According to Google, in May 2025 a single median text prompt processed by Gemini Apps consumed approximately 0.24 watt-hours (Wh) of electricity, emitted about 0.03 grams of CO₂ equivalent, and used roughly 0.26 milliliters of water—about the volume of five drops. For perspective, this is equivalent to the energy consumed while watching TV for less than nine seconds. Compared to the same period in 2024, these figures represent dramatic efficiency improvements: energy use per prompt dropped 33 times, and carbon emissions per prompt decreased 44 times, even as model quality and complexity improved.

What sets Google’s analysis apart is its full-system methodology. Rather than measuring only the active power consumed by AI accelerators like GPUs or TPUs, Google factors in total dynamic power—including CPU and RAM usage—as well as the energy used by idle components that are necessary for stability. It also includes overhead from data center operations, using the Power Usage Effectiveness (PUE) metric, and accounts for water usage tied to cooling systems. This yields a more realistic estimate of AI’s true environmental cost. By contrast, narrower estimates that ignore these variables tend to understate the footprint, placing it closer to 0.10 Wh per prompt.

The report highlights the layered innovations behind Google's efficiency gains. These include more energy-aware model architectures such as Mixture-of-Experts and Gemini Flash; algorithmic advancements like quantization; and high-performance custom hardware, particularly the new Ironwood TPU, which is 30 times more efficient than earlier generations. Efficient serving techniques like speculative decoding and capacity-aware idling also contribute to operational savings. All of this takes place in highly optimized data centers, with an average PUE of 1.09 and ambitious targets for 24/7 carbon-free energy use and water replenishment.

By sharing these metrics and methodologies, Google aims to foster industry-wide consistency in environmental reporting and help guide collective efforts toward more sustainable AI. Their work provides a benchmark for others, setting a new standard in how the tech industry measures and mitigates the impact of advanced machine learning systems.

Image: Measuring the Environmental Impact of AI Inference

Source: Measuring the Environmental Impact of AI Inference