What is the environmental impact of AI?

It's a real and rapidly growing energy demand, frequently exaggerated at the level of an individual query and frequently understated at the level of the infrastructure build-out. Both distortions are common, and the honest picture sits in between. **At the individual query level**: a single AI chat response uses meaningfully more energy than a web search, but the absolute number is small — comparable to running a lightbulb briefly. The viral comparisons ('one query equals a bottle of water') generally come from aggregate figures divided in questionable ways, or from studies about training rather than use. Personal query volume is not where your carbon footprint is decided; a single flight dwarfs a year of heavy chatbot use. **At the infrastructure level**, the concern is legitimate: - Data centres consumed roughly 1-2% of global electricity before the current build-out, and AI workloads are driving substantial growth on top of that. - Several major technology companies have reported rising emissions in recent years, reversing prior progress, and have explicitly attributed this to AI infrastructure. - Local grid impact is genuinely contentious — new data centres are being built faster than clean generation can be added in some regions, extending the life of fossil plants and raising local electricity prices. - **Water** is used for cooling, and in water-stressed regions this is a serious local issue regardless of the global average. Community opposition to data centre siting is often about water rather than carbon, and it's not unreasonable. **Training versus inference**: training a large model is a large one-time cost, widely cited. But inference — everyday use — now dominates total energy because it happens billions of times. The framing that focuses on training numbers is outdated. **What genuinely mitigates it**: efficiency gains have been substantial (energy per unit of capability has fallen dramatically), smaller models handle many tasks that don't need the largest one, and siting data centres where clean power is abundant makes a large difference. Efficiency alone doesn't solve it, though — cheaper inference has historically increased total use rather than reducing it. **Where to direct attention if you care about this**: not individual query guilt, which is negligible and mostly a distraction. The pressure points are transparency (most providers don't publish per-model energy figures, which makes accountability nearly impossible), siting and grid decisions, and whether the build-out is matched by genuinely additional clean generation rather than accounting arrangements. Those are policy and procurement questions, and they're where the actual environmental outcome gets decided.