Google measures and shares the environmental cost of AI prompts

Artificial intelligence has become a part of daily life, powering everything from search engines to chatbots. But what is the hidden environmental price of every AI answer we receive? A new study from Google offers the most detailed picture to date. By tracking the real-world energy, carbon, and water footprint of serving billions of AI prompts, the company shows that the average query is surprisingly efficient, but warns that scale makes the impact impossible to ignore.

“Until now, most figures about AI’s environmental cost were based on rough estimates,” the authors write. “Our study is the first to measure these metrics directly in production at a global scale.”

Measuring AI in the wild

The research focused on Google’s Gemini AI assistant, which serves millions of users worldwide. Previous studies often relied on model assumptions, hardware specifications, or lab benchmarks. Google instead used in-situ telemetry across its data centers to capture the full picture: not only the energy of AI accelerators, but also the host CPUs, idle backup capacity, and overhead like cooling systems.

The results are striking. In May 2025, the median Gemini Apps text prompt consumed just 0.24 watt-hours (Wh) of energy, less than watching nine seconds of television. In terms of water, the equivalent query used about 0.26 milliliters, roughly five drops. The carbon footprint came out at 0.03 grams of CO2 equivalent.

“These numbers are substantially lower than many public estimates,” the report notes. Some earlier analyses suggested anywhere between 1.8 and 7 Wh per prompt, while others put water use at 45–50 milliliters.

No escape: AI is everywhere - and what that means for the planet

Even if you want to avoid using AI, you probably can’t. The rising energy costs are worrying, an MIT Technology Review roundtable showed.

Read More

The efficiency challenge

How did Google manage to drive down the numbers? According to the report, a mix of software and hardware innovations made a dramatic difference. Between May 2024 and May 2025, the company achieved a 33-fold reduction in per-prompt energy use and a 44-fold drop in carbon emissions.

Key factors included:

• Smarter model architectures, such as mixture-of-experts systems, that only activate the parts of a model needed for a given query.
• Speculative decoding, where a smaller model drafts an answer that the main model only verifies, reducing unnecessary computation.
• Custom-built TPUs, with the latest generation 30 times more energy-efficient than the first.
• Optimized batching and serving, which keep accelerators busy instead of idling.
• Cleaner energy sourcing, with Google buying carbon-free power wherever possible.

The improvements highlight the importance of looking beyond AI model design. “Comprehensive environmental metrics are critical to properly incentivize efficiency opportunities across a large-scale, globally distributed production fleet,” the study stresses.

Watt Matters in AI

Watt Matters in AI is a conference that aims to explore the potential of AI with significantly improved energy efficiency. In the run-up to the conference, IO+ publishes a series of articles that describe the current situation and potential solutions. Tickets to the conference can be found at wattmattersinai.eu.

View Watt Matters in AI

Why boundaries matter

A key contribution of the paper is its definition of the measurement boundary: the scope of activities included in calculations. Narrow studies that only count GPU energy tend to underestimate the true cost. By including idle capacity and cooling overhead, Google finds that a realistic AI prompt consumes 2.4 times more energy than what narrower approaches suggest.

This matters because inconsistent methodologies have led to wildly different claims about AI’s footprint. For example, recent disclosures from OpenAI, Mistral, and academic groups each reported numbers that differed by orders of magnitude. Without a common standard, policymakers and users cannot make meaningful comparisons.

“Environmental metrics need to be actionable and comparable,” the authors argue. “Otherwise reported figures can vary by orders of magnitude for similar tasks, hindering transparency and accountability.”

Drops in the ocean, or a growing wave?

So, should we stop worrying about AI’s environmental impact if each prompt only costs a fraction of a second of TV time and a few drops of water? Not quite.

While the per-prompt figures look small, AI queries now number in the billions per day. At such a scale, even tiny efficiencies add up to massive savings - or waste. The report acknowledges this tension: “While the impact of a single prompt is low compared to many daily activities, the immense scale of user adoption globally means that continued focus on reducing the environmental cost of AI is imperative.”

Google says it is committed to further reducing water use in high-stress regions and pursuing its ambition of operating on 24/7 carbon-free energy. But the broader lesson is that transparency itself is a form of efficiency: knowing what to measure, and how, is the first step toward cutting waste.

Setting a standard for the industry

The study is more than just a progress report on Google’s Gemini. It is also a call for the AI industry to adopt consistent, comprehensive metrics. “We hope this work contributes to ongoing efforts to develop efficient AI at this critical time,” the authors conclude. With this statement, they are somewhat doing a disservice to the creators of the French Mistral: that company had already provided a transparent overview of its own energy consumption.

For users, the message is twofold. On one hand, your chat with Gemini or another AI assistant is far less resource-intensive than you might fear. On the other hand, the sheer volume of AI use worldwide means efficiency gains, no matter how technical or invisible, have planetary consequences.

In other words, every drop counts.