Understanding Energy Usage in AI: A Look at Google’s Gemini
As artificial intelligence continues to evolve, understanding its energy demands becomes increasingly important. A recent exploration into Google’s Gemini AI sheds light on how varying prompts can require significantly different amounts of energy. One striking example provided by Dean is the process of feeding a multitude of books into Gemini and requesting a detailed synopsis of their content. This task, he notes, consumes much more energy compared to the average prompt due to the extensive processing required.
The Energy Demand of Complex Tasks
When we delve into the inner workings of AI, especially for more intricate tasks, we find that certain actions inherently demand more energy. Reasoning models, which rely on multiple cognitive steps to arrive at a conclusion, also demonstrate a higher energy load. The longer and more complex the task, the more resources are needed, emphasizing the intricacies of AI operations. This is not just about speed; it involves substantial processing power and time that can strain energy resources.
Limitation to Text Prompts
An important aspect to note is that the report focused solely on text prompts, not accounting for energy used in generating images or videos. Such tasks typically require far greater energy, as highlighted in previous analyses, such as one from MIT Technology Review’s Power Hungry series. The difference in energy demand between text and multimedia tasks underscores the diverse challenges that AI presents in various formats.
Declining Energy Consumption Over Time
Interestingly, Google reports a substantial decrease in the energy used for issuing prompts to Gemini. The company observed that the median energy consumption for a prompt in May 2024 was 33 times higher than in May 2025. This dramatic reduction is attributed to advances in AI model efficiencies and software optimizations, showcasing the tech giant’s commitment to refining its operations and reducing its environmental impact.
Greenhouse Gas Emissions and AI
When discussing energy, it’s crucial to consider its environmental implications. Google estimates that the median prompt generates approximately 0.03 grams of carbon dioxide. To arrive at this figure, the company multiplies the energy consumed per prompt by the average emissions per unit of electricity. However, Google adopts a unique approach, using a market-based emissions estimate. This accounts for their investments in clean energy projects, rather than relying on averages from traditional grids.
Commitment to Clean Energy
Since 2010, Google has been proactive in signing agreements to purchase over 22 gigawatts of power from renewable sources like solar, wind, geothermal, and advanced nuclear projects. This impressive commitment significantly lowers their emissions per unit of electricity, making their operations more sustainable. On paper, their emissions are about one-third of the average grid emissions, demonstrating their strategic efforts to promote an eco-friendly energy transition.
Water Usage in AI Operations
Beyond energy, AI data centers also necessitate water for cooling purposes. Google’s analysis reveals that each prompt consumes approximately 0.26 milliliters of water—roughly equivalent to five drops. This consumption further emphasizes the broader resource implications of AI technology, highlighting the intricate balance between technological advancement and environmental sustainability.
Transparency in AI Interaction
Ultimately, the goal of this report is to shed light on the energy consumption tied to user interactions with AI. Dean emphasizes the importance of understanding what goes into each AI response—not just for transparency but for fostering a more conscious engagement with technology. As users, knowing the energy and resource demands of our queries can inform more sustainable practices and encourage a mindset focused on efficiency and responsibility in the digital age.