Society has normalized the use of AI in everyday life, so much so that it’s easy to forget the cost that comes with each response. Electricity, water, and land footprints contribute to AI’s environmental costs, and we have a very narrow window to make AI usage feasible within environmental constraints.
Consider this: by 2030, global data centers powering AI will consume 945 terawatt-hours of electricity annually. That’s nearly triple the combined electricity use of Pakistan, Bangladesh, and Nigeria—three countries home to over 650 million people. And that’s just one piece of the whole picture. The associated water footprint will equal the basic annual domestic water needs of 1.3 billion people in Sub-Saharan Africa, while the land footprint will exceed 14,500 square kilometers, roughly twice the Jakarta metropolitan area.
The real problem? Most discussions about AI’s environmental impact focus mostly on carbon emissions. We are judging with incomplete knowledge. Every kilowatt-hour of electricity used carries a water footprint from cooling and power generation, and a land footprint from energy infrastructure and supply chains. But cutting carbon emissions by 70 percent may increase water use more than thirtyfold. So when companies say their data centers will be powered by renewables, everyone should know that the environmental burden might shift to regions already facing severe water or land stress.
The elephant in the room: Most discussions obsess over training massive models like GPT-4. But once deployed, inference—the continuous running of models to answer everyday prompts—accounts for 80 to 90 percent of total AI energy use. ChatGPT alone processes roughly 2.5 billion prompts per day, consuming approximately 383 GWh of electricity annually—equivalent to powering 500,000 people’s basic water needs and spanning over 800 football fields in land impact. This cost scales with every query, every image generation, every video you create.
Geography makes this worse. Over 90 percent of AI-specialized cloud computing is concentrated in two countries, while more than 150 nations lack sovereign AI infrastructure. Meanwhile, communities in places like Mexico and Uruguay bear the environmental burden of data centers sited in water-stressed regions while reaping none of the economic benefits. By 2030, AI infrastructure could generate up to 2.5 million tonnes of electronic waste annually, much of it processed in low-income economies with limited safeguards.
So what to do about it? Good news! This isn’t unsolvable. An environmentally-friendly report proposes a “responsible AI ecosystem” built on transparency, efficiency by design, equity and environmental justice, lifecycle responsibility, global cooperation, and sustainable use. Practically, this means that governments should integrate AI infrastructure into energy and water planning; companies should treat model selection and product defaults as footprint decisions; users should adopt “fit-for-purpose” approaches, selecting the lightest model that actually solves the problem, not the most powerful one available. Without resource budgets, token limits, and low-resolution defaults, efficiency improvements will simply be absorbed by volume growth—the so-called rebound effect.
The window is narrow, but it’s still open. We have the tools to limit AI’s environmental footprint. We just need to have the will to use them.