The Compunding Carbon Cost of AI – EcoTech Insights
Sustainability analyst Bonnie Schneider unpacks a new study from Greenly revealing how AI’s carbon footprint grows with every query—not just during model development. From training to daily use, discover what’s driving emissions and how to reduce them.
Transcript
AI's carbon cost doesn't stop at training. Every prompt, every task fuels a growing footprint, and it's accelerating fast. That's the conclusion of a new study from Carbon accounting Firm Greenly, which shows that emissions accumulate from development to deployments.
Building a model like GPT-3 can emit as much carbon dioxide as 112 gas powered cars do in a year, and that's just the beginning. Envision endless queries used billions of times, it all adds up. Greenleaf's research also shows that not all AI tasks are created equal.
05 kilograms a staggering 60 fold difference. The gap widens further with added complexity. Supporting additional languages, for instance, typically requires separate rounds of training, compounding the carbon footprint with every expansion.
Fortunately though, the study does point to tangible ways to reduce AI's impact, such as training less often, reusing models, improving hardware and data centers, plus tapping into renewable energy. Combine that with efficient code and emissions could fall sharply. Reducing AI's carbon footprint isn't about choosing between training and usage.
It's about optimizing both.
