Google’s AI Transparency: What We Learned from Its New Gemini Energy Report

Artificial intelligence is increasingly becoming part of our daily lives, but behind every prompt and response lies a real environmental cost. Until now, the scale of that impact has often been hidden. In a significant step forward, Google has released a detailed technical report on the energy, water, and carbon costs of its Gemini applications. As reported by MIT Technology Review, this is the most transparent disclosure yet from a major technology company.

The headline finding is striking: the median Gemini text query consumes just 0.24 watt-hours of electricity. To put that in perspective, that’s about the same as running a microwave for a single second. The report also calculates associated impacts—approximately 0.03 grams of carbon dioxide emissions and 0.26 milliliters of water, or five drops, per query. While these figures may appear minimal, their significance lies in the transparency and methodology behind them. Here is a link to their technical report

Breaking Down the Numbers

Importantly, Google didn’t only count the energy drawn by its custom AI chips (TPUs). According to MIT Technology Review’s summary of the report, the company also included supporting infrastructure. The TPUs themselves accounted for 58% of total energy demand, while CPUs and memory contributed 25%. Backup systems absorbed another 10%, and the final 8% was attributed to data centre overheads such as cooling and power conversion.

This level of disclosure matters because, until now, researchers have struggled to measure AI’s true energy footprint without full access to company operations. Academics like Mosharaf Chowdhury and Jae-Won Chung from the University of Michigan praised Google’s report as the most comprehensive analysis to date, and potentially a “keystone” for future AI energy research.

Efficiency Gains and Limitations

Another noteworthy detail is the dramatic fall in energy use per prompt. Google reports that a Gemini text query in May 2025 consumed 33 times less energy than it did in May 2024. This improvement is credited to model refinements and software optimisation.

However, as Google’s own scientists emphasise, not all prompts are created equal. A request to summarise several books or to run reasoning-intensive tasks will naturally consume far more energy than a short factual query. And this report is limited to text-based prompts—image and video generation are known to be significantly more resource-intensive.

Broader Implications

The report also sheds light on Google’s approach to emissions accounting. Rather than using grid averages, the company applies a “market-based” estimate reflecting its renewable energy purchases. Since 2010, Google has signed agreements for more than 22 gigawatts of clean power, which lowers its reported carbon intensity.

Still, as MIT Technology Review highlights, transparency only goes so far. Google has not disclosed the total volume of Gemini queries processed daily—data that would allow researchers to calculate the platform’s overall footprint. Climate researchers such as Sasha Luccioni have welcomed this publication but continue to call for a standardised AI energy score, similar to the Energy Star label for household appliances. For more information into AI energy usage check out MIT's other research

A Step Forward

Google’s report does not close the debate about AI’s environmental costs, but it does open a valuable window. By sharing detailed, verifiable numbers, the company has set a new benchmark for openness in the AI industry. The challenge now is to ensure that this transparency becomes a norm, not an exception, and that industry, researchers, and policymakers work together to establish clear standards for measuring AI’s true impact. Here at tax the ai, we are very supportive of this transparency.