Artificial Intelligence Sustainability: A Practical Guide 2026

AI's carbon cost is real and growing — but the bigger risk is companies defaulting to maximum compute for every task. Training one model can emit 284 tonnes of CO2. Goldman Sachs projects data center demand up 160% by 2030. The fix isn't less AI: it's proportionality — right model, right task, right scale. Practical steps for SMEs on cloud selection, workflow efficiency, and what to measure before you have perfect metrics.

ELECTE is an AI-powered data analytics platform for European SMEs — turning raw data into clear, verifiable, actionable insight. Learn more at electe.net

The AI analysis 100,000+ readers trust. Join them:  

- Subscribe to the ELECTE newsletter

- Official merch

New episodes regularly. Subscribe wherever you listen.
Written and hosted by Fabio Lauria.

Transcript

SPEAKER_00 0:00

This is the Electi Review. Today, artificial intelligence has a carbon problem, but the real issue isn't whether to use AI. It's whether companies are using the right amount of it. Training a single large AI model can generate over 284 metric tons of CO2, equivalent to the lifetime emissions of five cars. Developing GPT-3 required roughly 1,287 megawatt hours of electricity, comparable to the annual consumption of 120 to 130 average American households. Goldman Sachs research projects that data center electricity demand could rise by 160%, adding approximately 200 terawatt hours per year between 2023 and 2030. By 2028, AI-related consumption alone could account for 19% of total data center energy needs. Here is the uncomfortable part. Making AI more efficient doesn't automatically reduce total consumption. When computing gets cheaper, companies use more of it. That's the efficiency paradox, and it's the core tension this article confronts. The article draws a clear line between two distinct problems. First, the sustainability of AI, the environmental footprint of the systems themselves. Second, AI for sustainability, using AI to cut waste, optimize energy in industrial facilities, and automate ESG reporting. That last category has nearly tripled in adoption over the past year. The most common mistake companies make is not technical. It's a decision-making failure. They default to the largest, most powerful model for every task, even when a lighter model or a traditional process would deliver the same result with a fraction of the compute. The principle the article defends is proportionality. Use the right model for the right task at the right level of capability. For SMEs, the practical guidance is concrete. Evaluate cloud providers by their regional energy mix, not just price. Eliminate redundant workflows where the same data is exported, copied, and reformatted multiple times, and measure what you can, even imperfectly. Fewer API calls per completed task, less data transferred per session, smaller average model size per operation. These are the signals that matter before you ever touch a carbon accounting system. The conclusion is direct computing is not a free resource. Every organization should treat it the way it treats budget or staff time, allocated only where it creates genuine value. That's the review.