Planet & resources · Research field
AI’s resource footprint
Catalogued:
What it means
The material demands of developing and operating AI, including electricity, water, hardware, infrastructure and supply chains. Training and inference have different profiles. A universal “cost per question” obscures variation in models, hardware, workload, location and how an estimate is calculated.
The strongest case
Measuring resource use supports better infrastructure decisions and makes environmental tradeoffs visible. Useful applications may save resources elsewhere, but those savings require evidence rather than being assumed.
The difficult part
Efficiency per task can improve while total consumption rises. Local constraints matter, and neither one alarming analogy nor a company-wide renewable claim describes every deployment. Compare real alternatives using explicit system boundaries.
Sources & further reading
Sources support the factual context. The jokes are our responsibility.