The Carbon Cost of AI: How Google Measures Its Energy Use

Google’s Sustainability team recently released a technical paper explaining how it measures the environmental footprint of AI inference — the stage where trained AI models generate text, images, or predictions. While AI has the potential to drive massive economic and scientific progress, its growing energy demands raise important questions about sustainability.

How Much Energy Does an AI Prompt Use?

According to Google’s analysis, a typical Gemini text prompt consumes about 0.24 watt-hours of energy, produces 0.03 grams of Carbon emissions, and uses 0.26 milliliters of water — roughly equivalent to watching TV for less than nine seconds.

Interestingly, over just 12 months, the energy and carbon footprint of a Gemini text prompt have dropped by 33x and 44x respectively, It happens because due to the advancements in model efficiency, hardware, and data center performance.

Why Measuring AI’s Footprint Is Complex

Google emphasizes that measuring AI’s environmental cost isn’t as simple as looking at active chip usage. A full picture needs to include:

• Idle capacity kept ready for reliability and traffic spikes.

• CPU and RAM usage, which support model execution.

• Data center overhead, such as cooling and power distribution.

• Water consumption for cooling systems.

Many estimates in the public domain focus only on GPU/TPU consumption, which Google says can underestimate the true footprint by more than half.

Efficiency Through a Full-Stack Approach

Google attributes its efficiency gains to a full-stack strategy — improving AI across every layer, from hardware to algorithms to data centers. Key contributors include:

• Smarter model architectures like Mixture-of-Experts, which only activate parts of a model needed for a task.

• Algorithmic improvements such as quantization, reducing energy use without sacrificing accuracy.

• Optimized serving methods like speculative decoding and distilled models (e.g., Gemini Flash) that handle queries with fewer computations.

• Custom TPUs designed for maximum performance per watt, now 30x more efficient than the first generation.

• Ultra-efficient data centers running at a fleet-wide average PUE (Power Usage Effectiveness) of 1.09, powered increasingly by carbon-free energy.

Google 2025 Environment Report

Based on Google's 2025 Environmental Report, the carbon emitted from Google's data centres in the last year, 2024, was approximately 2.5 million metric tons of carbon dioxide equivalent (tCO2e).

For broader context, Google's total operational emissions (which include all Scope 1 and Scope 2 market-based emissions for all Google and Alphabet Inc. operations, including offices, not just data centres) for 2024 were 3,132,200 tCO2e

To know more about the environment report. - https://www.gstatic.com/gumdrop/sustainability/google-2025-environmental-report.pdf

Looking Ahead

Google acknowledges that the job isn’t done. It continues to push for further efficiency improvements while pursuing its broader goals of 24/7 carbon-free operations and water replenishment. By publishing its methodology, Google hopes to set a standard for how the industry measures the true environmental footprint of AI.

As we build more powerful AI, we must remember the planet isn’t ours alone.

Reference Links

1. Research Paper - https://arxiv.org/abs/2508.15734

2. Distilling a knowledge in Neural Network - https://arxiv.org/abs/1503.02531

3. 2025 Environment Report - https://www.gstatic.com/gumdrop/sustainability/google-2025-environmental-report.pdf