The Trade-Offs Are Real
Every generative AI query consumes measurable resources: 1 kilowatt-hour of energy, 4 litres of water, and emits 466 grams of carbon. These aren't theoretical numbers - they're production realities that enterprises need to account for.
The impact is material. 42% of executives are re-examining previously set climate goals specifically due to generative AI's footprint. Amazon's 2024 emissions surged 6% compared to 2023, driven by AI embedding and datacenter expansion. Cornell researchers project AI could generate 24-44 million metric tons of CO₂ annually by 2030 in the US alone.
The Efficiency Gains Matter Too
But the story shifted in 2025. Google's median Gemini text prompt now consumes 0.24 Wh of energy and 0.26 mL of water - a 33× reduction in energy and 44× reduction in carbon compared to 2024. That's the difference between production-scale optimisation and early deployment.
Context matters: a 15-mile car commute produces about 6 kg of CO₂ - equivalent to tens of millions of AI prompts at 2025 efficiency levels. Datacenters power far more than generative AI; every search, email, and video stream uses the same shared infrastructure.
What CTOs Should Watch
Three things matter:
First, measurement tools are maturing. CodeCarbon, Green Algorithms, and emerging carbon APIs now provide production-grade tracking for ML pipelines. The shift from speculation to real data enables meaningful comparison.
Second, the net impact depends on deployment choices. BCG estimates AI could mitigate 5-10% of global greenhouse gas emissions by 2030 if used to optimise supply chains, reduce physical waste, and power energy management systems. Microsoft's research shows AI completing knowledge work tasks faster and more sustainably than humans - when implemented thoughtfully.
Third, transparency remains inconsistent across providers. This will likely follow cloud computing's path, where environmental disclosure eventually became standard practice. The enterprises asking for carbon metrics now will shape what becomes available.
The question isn't whether AI has an environmental cost - it does. The question is whether the efficiency gains and optimisation benefits outweigh it. History suggests we won't get that answer from vendor announcements. We'll get it from implementation data.