Asking a simple question to artificial intelligence might seem harmless, but even a “thank you” has an environmental cost. Each interaction involves electricity consumption, water use for cooling, and carbon emissions. For the first time, Google Cloud has released official data on Gemini's impact.

The report shows not only the average consumption per prompt but also direct comparisons with ChatGPT, YouTube, and even Google Search itself, revealing numbers that surprise due to the efficiency achieved.

Why measuring AI's environmental impact matters now

Artificial intelligence has moved from a laboratory experiment to becoming a part of everyday life, from quick searches to digital assistants and content generation. This rapid growth is accompanied by an ever-increasing energy demand, which is not neutral for the planet. Measuring the environmental impact of AI matters because it helps us understand the real cost of technologies we use almost invisibly.

Without clear metrics, it becomes difficult to scale the aggregate effect of billions of daily interactions. Each question seems irrelevant on its own, but on a global scale it represents electricity consumption, water for cooling data centers, and carbon emissions. By bringing these numbers to light, companies like Google clear the way for governments, society, and even users to discuss efficiency, regulation, and better usage practices.

More than a technical issue, it is a strategic topic: in a world seeking to reduce emissions and optimize resources, knowing exactly how much AI consumes today is the first step toward making it more sustainable tomorrow.

Gemini's numbers: energy, water, and CO₂ per prompt

Google Cloud's report brought, for the first time, official data on Gemini's environmental consumption in real use—that is, at the moment the user asks a question and receives an answer. These numbers help transform the impact of AI from something abstract into concrete information, comparable to activities in our daily digital lives.

Energy per prompt

Each query to Gemini consumes an average of 0.24 Wh, equivalent to watching TV for less than 9 seconds. This value represents an approximately 33-fold reduction compared to what was consumed a year ago, thanks to advancements in more efficient software and hardware.

Water use

The data centers' cooling process was also measured. Each interaction requires 0.26 milliliters of water, roughly five drops. Although small on an individual scale, the impact becomes relevant when multiplied by the global volume of daily interactions.

Carbon emissions

In terms of carbon footprint, the progress is also remarkable. Before the improvements, each Gemini prompt emitted an average of 1.32 g of CO₂. Today, emission has dropped to 0.03 g, a 44-fold reduction, achieved primarily through the use of renewable energy in Google's data centers.

Overall, these numbers show that artificial intelligence can be optimized to consume fewer resources and emit fewer pollutants. Even so, when multiplied by millions of interactions, the impacts remain significant and reinforce the importance of energy efficiency as a central part of AI's evolution.

Gemini vs. ChatGPT vs. Google Search: which consumes more?

When speaking of efficiency, comparison is inevitable: how does Gemini position itself against other digital services? According to the released data, a query to Gemini consumes an average of 0.24 Wh, less than the 0.34 Wh estimated for ChatGPT and even below the 0.3 Wh of a search on Google Search itself.

For reference, watching one minute of video on YouTube consumes about 0.2 Wh, meaning the difference between online activities can be smaller than we imagine, but every detail makes a difference on a global scale.

The most surprising result of the report is precisely the fact that Gemini has become more efficient than the traditional search engine, a symbol of everyday internet. This reversal shows not only technological progress but also how the pressure to reduce environmental impacts is already beginning to influence the design and operation of AI systems. The challenge, from here on, is to maintain this efficiency curve even as more complex and multimodal models become popular.

How Google calculates consumption: from the chip to the data center

To reach the numbers presented in the report, Google adopted a broader methodology than simply measuring the energy spent by the chip that processes the response. The calculation includes the entire chain involved in delivering AI: the active consumption of accelerators (such as TPUs), the energy used by support servers (CPU and memory), the idle capacity that must be maintained to ensure stability, and, finally, the overhead of the data center itself, which involves cooling, redundancy, and power conversion.

This approach, called a “comprehensive methodology,” contrasts with simplified analyses focusing only on the main hardware. The result is a picture closer to reality, though more complex to communicate. It is worth noting that under narrower methodologies, the expenditure per prompt can drop by half, but in return, it fails to reflect the systemic cost of maintaining AI on a global scale.

By opening up these behind-the-scenes details, Google not only demonstrates transparency but also establishes a methodological benchmark for the industry. This helps create fairer comparisons and pressures other companies to adopt similar criteria, preventing the environmental impact of AI from being underestimated.

Surprise in the report: Gemini more economical than search

The most unexpected finding from the report is that, in terms of efficiency, Gemini surpassed even Google Search. Traditionally seen as one of the lightest and most optimized tools on the internet, the search engine consumes an average of 0.3 Wh per query, while Gemini, despite being a much more complex generative AI model, recorded 0.24 Wh per interaction.

This result shows that recent advancements in hardware, algorithms, and energy management have not only reduced the environmental burden of AI but have also challenged the logic that more sophisticated systems must consume more resources. This is a symbolic turning point: the technology that seemed most costly becomes, in practice, more efficient than the product that has symbolized online speed and simplicity for two decades.

This reversal could serve as a milestone for the sector. If Gemini managed to reach this level of efficiency in a short time, it sets a precedent for other AI tools to also adopt increasingly sustainable standards without sacrificing performance or accessibility.

What the numbers don't tell: criticisms and limitations

Although impressive, the numbers in Google's report should be interpreted with caution. The study focuses only on the serving of Gemini, meaning the delivery of responses to the user, and does not consider model training, a stage that consumes significantly more energy and resources.

Furthermore, the methodology adopted is “market-based,” accounting for clean energy purchased by data centers but not necessarily reflecting the actual local impact of each facility.

Other critical points include the indirect water linked to electricity production in the energy mix and the decision to present the median consumption per prompt, which represents the typical user but ignores longer or multimodal queries that require much more computing power.

Experts point out that while the numbers provide transparency, they do not tell the whole story of AI's environmental impact.

Acknowledging these limitations is essential for strategic decisions to be made based on realistic data. Google's report is an important step, but it shows there is still a lot of room to improve metric standardization and reach a complete understanding of artificial intelligence's environmental cost.

Discover why artificial intelligence consumes so much energy in this other exclusive article from CodeBlog.

Sustainable AI: next steps in efficiency and transparency

Gemini's advancement shows that artificial intelligence can be optimized to reduce energy consumption, water use, and carbon emissions. But the challenge does not end here: for AI to be truly sustainable, it is necessary to combine efficient technology, clean energy, and transparency in disclosing environmental metrics.

Companies in the sector are beginning to adopt more rigorous measurement standards and detailed reporting, allowing for consistent comparisons between different models and services. Regulators and users also gain a strategic role, as conscious usage decisions, such as opting for shorter interactions or avoiding excess unnecessary prompts, can reduce daily impacts.

In the future, the balance between innovation and sustainability will depend on two factors: the continuous efficiency of hardware and algorithms, and a culture of transparency, where every digital interaction has its environmental cost made clear. Understanding these metrics is the first step toward responsible AI growth, offering benefits without compromising natural resources or contributing unnecessarily to the global carbon footprint.

AI has already become a reality, and its use spans from work to leisure. Searches, questions, image generation, coding, research, and writing through it are becoming more common and accurate every day. As consumers, we must keep ourselves updated regarding the impacts of this technology on our daily lives and on the world. Continue following CodeBlog to stay informed about the advancements in artificial intelligence.