AI’s environmental footprint is shrinking, thanks to significant software and renewable energy gains over the past year.
AI’s Energy Footprint Shrinks, But Volume Remains a Concern
The environmental impact of a single text request from AI is surprisingly small. Estimates suggest a median Gemini Apps text prompt uses 0.24 watt-hours of energy, emits 0.03 grams of carbon dioxide equivalent, and consumes about five drops of water. This energy use is comparable to watching roughly nine seconds of television.
However, the sheer volume of these requests is undeniably high. Integrating AI operations into every single search query represents a massive increase in compute demand that simply didn’t exist a couple of years ago. This means that while the individual impact is minor, the cumulative cost is likely substantial.
The good news? Just a year ago, the impact would have been significantly worse. This improvement stems from both circumstantial advantages and deliberate optimizations. The surge in solar power, particularly in the U.S., has made it easier to source renewable energy, resulting in a 1.4x reduction in carbon emissions per unit of energy consumed over the past year.
The most significant gains, however, have come from software advancements. Different approaches have led to a remarkable 33x reduction in energy consumed per request.
Credit:
Elsworth, et. al.
To arrive at these figures, a team tracked requests and the hardware serving them over a 24-hour period, also noting idle hardware time. This provided an energy-per-request estimate that varies by the AI model used. For each day, the median prompt was identified to calculate the environmental impact.
Key optimizations include the “Mixture-of-Experts” approach, which activates only the necessary parts of an AI model for a given request, potentially reducing computational needs tenfold to a hundredfold. Developing more compact AI model versions also lowers the computational load. Efficient data center management, ensuring active hardware is fully utilized and idle hardware is in a low-power state, further contributes to these savings.
