0.3 watt-hours. That's roughly what a single Google search costs in energy. A ChatGPT query? About 2.9 watt-hours by older estimates, though GPT-4o has brought that down to around 0.3 Wh as well. A new analysis by researcher Quentin Adam pulls together the latest data on what your AI prompts actually cost in energy, water, and carbon - and the numbers are sobering even as efficiency improves.
The Per-Query Math
Text queries have gotten dramatically cheaper to run. Google says Gemini's energy cost dropped 33x between May 2024 and May 2025, landing at a median of 0.24 Wh per query. GPT-4o runs at roughly 0.3 Wh. At the individual query level, the difference between an AI prompt and a regular search has nearly vanished.
Image generation is a different story. Creating a single AI image takes 5 to 50 times more energy than a text query, according to Hugging Face research. A Stable Diffusion image runs about 0.7 Wh, and more complex models push well beyond that. Every "generate me a logo" request costs meaningfully more than a conversation.
Water is the hidden cost most people don't think about. Data centers need water for cooling, and a single Gemini query uses about 0.26 mL of water. That sounds tiny until you scale it: UC Riverside researchers found that 10 to 50 GPT-3 responses consume roughly one standard 500 mL water bottle. U.S. data centers are projected to use 150 to 280 billion liters of water per year by 2028. Microsoft's water consumption alone jumped 34% year-over-year in 2023.
The Macro Picture Is Less Reassuring
While per-query costs are falling, total consumption is rising fast because usage is exploding. Global data centers consumed about 536 TWh of electricity in 2025, roughly 2% of the world's total. In the U.S., data centers already account for 4% of electricity consumption, and that's expected to more than double by 2030. The IEA projects global data center energy use will exceed 1,000 TWh by 2030.
The geographic concentration makes this especially visible in certain regions. Data centers consumed 26% of Virginia's total electricity in 2023. That's not a typo - a quarter of an entire state's power grid feeding server rooms.
Tech companies' carbon emissions tell the same story. Combined emissions surged 150% in three years. Amazon's rose 182% since 2020, Microsoft's 155%, Meta's 145%, Google's 138%. These companies purchased 68.4 million carbon credits in 2025 (up 181%), and their combined AI infrastructure spending hit $320 billion in 2025. Without intervention, Accenture projects emissions could surge 11-fold.
What This Means for Daily Users
The practical takeaway isn't that you should stop using AI. The per-query cost genuinely has dropped, and for most people, your AI usage is a rounding error compared to driving a car or heating your home. But the numbers argue for being intentional rather than wasteful. Regenerating an image ten times because the hands look weird has a real cost. Running the same prompt through four different models to compare outputs multiplies that cost.
The bigger concern is structural. Efficiency gains per query are being outpaced by growth in total queries. Training GPT-4 alone consumed roughly 50 GWh of energy. U.S. coal generation rose nearly 20% recently, partly driven by data center demand. The AI industry's "we'll solve it with renewable energy" narrative hasn't kept pace with its actual power consumption growth.
These aren't reasons to avoid AI tools. They're reasons to pay attention to which providers are investing seriously in efficiency and renewables, and which are just buying carbon credits and calling it progress.